Deep learning decision-making system for short-wave and ultra-short-wave networking protocol optimization

Abstract

The invention relates to the technical field of wireless communication, and discloses a deep learning decision-making system for short-wave and ultra-short-wave networking protocol optimization, and the system comprises a physical layer feature obtaining module which is used for converting a baseband signal into a time-frequency atlas matrix; the link state prediction module is used for processing the atlas matrix based on a deep learning model so as to predict a residual effective time window of a link, and outputting an intervention instruction in combination with a service attribute and a risk level; the intervention event generation module is used for extracting interference texture Hash and distributing a preemptive intervention event log; the global strategy generation module is used for aggregating the log to identify the recurrent interference mode and generating a network global strategy containing a routing adjustment rule; and the strategy synchronization and execution module injects virtual cost into the protocol stack or executes radio frequency control according to the instruction or the strategy. According to the method, preemptive perception of link faults is achieved through texture prediction, virtual metric injection is used for decoupling physical connectivity and logic routing, and it is ensured that smooth switching of service flow is completed before physical interruption.

Description

Technical Field

[0001] This invention relates to the field of wireless communication technology, specifically to a deep learning decision system for optimizing shortwave and ultra-shortwave networking protocols. Background Technology

[0002] Shortwave and VHF self-organizing network communication systems are widely used in emergency rescue and tactical coordination scenarios, and their operating environment is highly dynamic and complex. Wireless signals face multipath effects, terrain obstruction, and variable electromagnetic interference during propagation. The movement of nodes causes real-time changes in the network topology. The stability of the communication link is directly constrained by the time-varying characteristics of the physical channel.

[0003] Existing mobile ad hoc network routing protocols, such as optimized link-state routing protocols or on-demand distance-vector routing protocols, primarily rely on network layer or link-layer statistical metrics to maintain network topology. These protocols typically use periodic handshake messages to probe the connectivity of neighboring nodes. Some improved solutions attempt to introduce cross-layer mechanisms, utilizing the physical layer's signal-to-noise ratio or received signal strength as auxiliary metric parameters. Routing algorithms calculate path costs based on these parameters, typically selecting the path with the fewest hops or the highest current signal strength for data forwarding.

[0004] However, this maintenance mechanism based on statistical indicators is inherently lagging. Routing protocols often only determine link failure and initiate route recalculation after multiple handshake messages are lost consecutively or the data retransmission rate increases. By this time, the physical link is usually already interrupted, and the service data flow immediately faces transmission gaps caused by the topology convergence process. A single signal-to-noise ratio (SNR) value only reflects the instantaneous energy state and is difficult to identify the microscopic textural evolution of radio frequency signals in the time-frequency domain. The system struggles to distinguish between instantaneous shadow fading and continuously approaching interference sources, lacks the ability to predict the remaining link lifetime, and the risk perception at the physical layer cannot be effectively translated into logical avoidance actions at the network layer. As long as the physical connection has not been completely broken, the routing protocol will still send data to high-risk nodes. Therefore, this invention provides a deep learning decision system for optimizing shortwave and ultra-shortwave networking protocols to address the shortcomings of existing technologies. Summary of the Invention

[0005] To address the shortcomings of existing technologies, this invention provides a deep learning decision system for optimizing shortwave and UHF networking protocols. This system solves the problems of existing shortwave and UHF networking routing protocols relying on statistical indicators to detect faults, which has a lag and lacks the ability to predict the evolution trend of channel interference. As a result, it is difficult to proactively avoid risks and achieve lossless service switching before physical link interruption.

[0006] To achieve the above objectives, the present invention provides the following technical solution: a deep learning decision-making system for optimizing shortwave and ultra-shortwave networking protocols, comprising the following steps: The physical layer feature acquisition module is used to capture raw signal data from the baseband of the wireless communication device and generate a time-frequency spectrum matrix representing the current physical state of the channel through time-frequency analysis. The link status prediction module is used to process the time-frequency spectrum matrix based on a deep learning model to predict the remaining effective time window of the link, and to arbitrate the service attributes of the data packets to be transmitted with the risk level parameters set by the strategy synchronization and execution module to output intervention instructions. An intervention event generation module is used to respond to the intervention command, extract the texture features that lead to the intervention, and encapsulate them into a preemptive intervention event log for network distribution; The global policy generation module is used to aggregate the preemptive intervention event logs to identify reproducible link problems and generate a global network policy that includes routing cost adjustment rules. The policy synchronization and execution module is used to inject virtual costs into the local network protocol stack or execute radio frequency control according to the intervention instructions or the network global policy in order to adjust routing behavior.

[0007] Preferably, the physical layer feature acquisition module includes: The windowing and framing unit is used to receive complex baseband signal streams and perform windowing processing on the signals according to the preset frame length and number of overlap points. The spectrum transformation unit is used to construct a parallel pipeline structure, perform a fast Fourier transform on the windowed signal, and convert the time-domain complex sequence into a frequency-domain complex sequence. The feature map generation unit is used to perform logarithmic mapping and dynamic range quantization on the frequency domain complex sequence, and to downsample the frequency domain dimension through a max pooling strategy, thereby splicing to generate the time-frequency map matrix of a fixed size.

[0008] Preferably, the link state prediction module includes: The deep feature extraction unit adopts a hybrid architecture of convolutional neural network and recurrent neural network to extract spatiotemporal features reflecting channel evolution from the time-frequency spectrum matrix and capture the morphology and dynamic change patterns of interference texture. The remaining lifetime regression unit maps the extracted spatiotemporal features to a scalar output to obtain the remaining effective time window, which represents the duration for which the link bit error rate remains below the limit under the current texture evolution trend. A multi-dimensional arbitration decision unit is used to calculate the dynamic safety threshold and generate the intervention command when the remaining effective time window is less than the dynamic safety threshold.

[0009] Preferably, when calculating the dynamic security threshold, the multi-dimensional arbitration decision unit: Parse the service priority parameters of the data packets to be transmitted, and read the currently effective risk level parameters from the policy synchronization and execution module; Based on the sum of the internal processing latency of the decision-making system and the handshake latency of the routing protocol, and superimposed with the weighted influence of the service priority parameter and the risk level parameter, a dynamic security threshold is calculated to ensure that the switching is triggered in advance in high-priority services or high-risk environments.

[0010] Preferably, the intervention event generation module includes: The context capture unit is used to record spatiotemporal metadata when the intervention occurs, including timestamps, center frequencies, and predicted duration of the blockade. The texture fingerprint extraction unit is used to extract feature vectors from the intermediate layer of the deep learning model of the link state prediction module and convert them into binary texture hash values ​​for digital classification of physical interference patterns. The event encapsulation and distribution unit is used to combine the spatiotemporal metadata with the texture hash value to construct the preemptive intervention event log, and to piggyback the transmission using the extended area of ​​the routing control message.

[0011] Preferably, when generating the texture hash value, the texture fingerprint extraction unit: Extract the output of the global average pooling layer of the convolutional neural network as a feature vector; The mean of the feature vector elements is calculated as the adaptive quantization threshold; The numerical values ​​of each dimension of the feature vector are compared with the adaptive quantization threshold to generate a binary sequence, which serves as the core identification identifier in the preemptive intervention event log.

[0012] Preferably, the global policy generation module includes: The log aggregation and cleaning unit is used to maintain a database based on a sliding time window, and to deduplicate and merge the preemptive intervention event logs reported by multiple nodes based on the Hamming distance of the texture hash values. The temporal correlation mining unit is used to cluster the aggregated logs and calculate the global hazard confidence of each interference cluster. The global hazard confidence is determined by the spatial coverage breadth and temporal recurrence frequency of the interference. The policy formulation and distribution unit is used to map interference clusters whose global hazard confidence exceeds a threshold to specific routing penalty costs, generate the global network policy, and distribute it to the entire network.

[0013] Preferably, the strategy synchronization and execution module includes: The policy incremental synchronization unit is used to listen to and parse the update packets of the network global policy, and write the newly added texture hash and penalty cost rules into the local policy library. The virtual metric injection unit is used to match the texture hash detected in real time with the local policy library during the routing calculation cycle. When the match is successful, the corresponding penalty cost is forcibly added on the basis of physical cost to generate a corrected link cost value and report it to the routing engine to decouple physical connectivity from logical routing priority.

[0014] Preferably, the strategy synchronization and execution module further includes a passive tracking and switchback unit for use during RF transmission channel shutdown: Keep the receiving channel open, continuously calculate the texture hash of the environmental signal, and update the interference dissipation index using an exponentially weighted moving average algorithm; When the interference dissipation index drops to the recovery threshold and the duration exceeds the safety protection time, it is determined that the interference source has left, and the penalty value in the virtual metric injection unit is gradually reduced until the link overhead returns to the physical true value.

[0015] Preferably, the physical layer feature acquisition module is configured as a hardware-accelerated IP core in a field-programmable gate array (FPGA), and the link state prediction module runs in the neural processing unit (NPU) of the embedded system-on-a-chip. The two share the time-frequency spectrum matrix data through direct memory access (DMA).

[0016] This invention provides a deep learning decision system for optimizing shortwave and ultra-shortwave networking protocols. It has the following beneficial effects: 1. This invention, through the cooperation of the physical layer feature acquisition module and the link state prediction module, transforms the radio frequency baseband signal into a time-frequency spectrum matrix. It uses a deep learning model to capture the texture evolution features of the interference signal, thereby predicting the remaining effective time window of the link. This enables communication nodes to trigger the route recalculation process in advance before the physical link bit error rate deteriorates to the interruption threshold. This overcomes the lag caused by the reliance on packet loss rate statistics in traditional networking protocols and ensures the continuous transmission of service data in strong interference environments.

[0017] 2. This invention utilizes virtual metric injection technology in the policy synchronization and execution module. When the system detects high-risk interference textures, even if the current signal-to-noise ratio still meets communication requirements, it will forcibly increase the overhead value of that link in the routing protocol. This approach forces the routing algorithm to automatically calculate and switch to a detour path based on the shortest path principle while keeping the physical layer connection active, thus achieving smooth migration based on risk prediction.

[0018] 3. This invention, through the collaboration of the intervention event generation module and the global policy generation module, extracts the texture hash fingerprint of the interference signal and aggregates observation logs from multiple nodes to calculate the global hazard confidence, thereby identifying reproducible interference source patterns. The resulting global network policy guides other nodes in the network to identify similar interference characteristics and uniformly executes route avoidance actions when an interference source approaches, effectively eliminating false alarms caused by multipath effects in single nodes and improving the overall resilience of the network. Attached Figure Description

[0019] Figure 1 This is a system architecture diagram of the present invention; Figure 2 This is a flowchart of the method steps of the present invention; Figure 3 This is a schematic diagram of the network topology and interference scenario according to an embodiment of the present invention; Figure 4 This is a schematic diagram comparing the time-frequency spectra of the physical layer signals according to an embodiment of the present invention; Figure 5 This is a comparison chart of end-to-end throughput during interference crossing in an embodiment of the present invention.

[0020] Among them, 10 is the physical layer feature acquisition module; 20 is the link state prediction module; 30 is the intervention event generation module; 40 is the global policy generation module; and 50 is the policy synchronization and execution module. Detailed Implementation

[0021] The technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0022] See attached document Figure 1 , Figure 1 This is a system architecture diagram according to an embodiment of the present invention. The present invention provides a deep learning decision system for optimizing shortwave and ultra-shortwave networking protocols. The system may include: a physical layer feature acquisition module 10, a link state prediction module 20, an intervention event generation module 30, a global policy generation module 40, and a policy synchronization and execution module 50.

[0023] The physical layer feature acquisition module 10 is used to capture raw signal data from the front-end baseband processing unit of the wireless communication device and generate a time-frequency spectrum matrix that can represent the current physical state of the channel through time-frequency analysis methods.

[0024] The link state prediction module 20 is used to receive the time-frequency spectrum matrix and predict and make decisions on the link state: based on a preset deep learning model, it processes the received time-frequency spectrum matrix to predict the remaining effective time window of the current communication link; when making intervention decisions, the link state prediction module 20 parses the service attributes from the header of the data packet to be transmitted and obtains the risk level parameters set by the current network global policy from the policy synchronization and execution module 50; based on the predicted remaining effective time window, the obtained service attributes, and the set risk level parameters, it performs arbitration to determine whether to execute a preemptive intervention action and output the intervention command.

[0025] Intervention event generation module 30 is triggered when the link state prediction module 20 outputs an intervention command. It is used to encapsulate the relevant parameters of the intervention behavior into a structured preemptive intervention event log and distribute it through the network.

[0026] The global policy generation module 40 is used to collect preemptive intervention event logs distributed by the intervention event generation module 30 from the network. By performing long-term pattern analysis on the collected log data, it identifies reproducible link problems in the network and generates a global network policy that includes routing cost adjustment rules and risk level settings.

[0027] The policy synchronization and execution module 50 is deployed on each node in the network to ensure that the network-wide policies generated by the global policy generation module 40 can be consistently obtained and applied by each node. After obtaining a new policy, a node applies the rules in the new policy to its local network protocol stack to complete a macro-level adjustment of protocol behavior.

[0028] See attached document Figure 2 , Figure 2 This is a flowchart of method steps according to an embodiment of the present invention. The present invention provides a deep learning decision-making method for optimizing shortwave and ultra-shortwave networking protocols, which may include the following steps: S100, the physical layer feature acquisition module 10 continuously captures raw signal data from the front-end baseband processing unit of the wireless communication device, and generates a time-frequency spectrum matrix that can represent the current channel physical state through short-time Fourier transform; S200, the link state prediction module 20 receives the time-frequency spectrum matrix and predicts the remaining effective time window of the current communication link based on the preset deep learning model; at the same time, it parses the service attributes from the header of the data packet to be transmitted and obtains the risk level parameters set by the network global policy from the policy synchronization and execution module 50; it arbitrates based on the remaining effective time window, service attributes and risk level parameters to determine whether to perform preemptive intervention. S300, if the link state prediction module 20 determines to perform a preemptive intervention action, it outputs an intervention command to perform a preset preemptive blocking operation, such as injecting a fake link interruption event into the network layer routing daemon process, or temporarily suspending the data transmission queue of the link in the network interface driver layer. S400, after the link state prediction module 20 outputs the intervention command, the intervention event generation module 30 is triggered, which encapsulates the relevant parameters of the intervention behavior into a structured preemptive intervention event log and distributes it through the network; S500, the global policy generation module 40 collects preemptive intervention event logs distributed by the intervention event generation module 30 from the network. By performing long-term pattern analysis on the collected log data, it identifies reproducible link problems in the network and generates a global network policy that includes routing cost adjustment rules and risk level settings. S600, the policy synchronization and execution module 50 obtains the global network policy generated by the global policy generation module 40, and applies the rules in the new policy to the local network protocol stack to complete the macro-adjustment of protocol behavior.

[0029] The physical layer feature acquisition module 10 is configured as a hardware-accelerated IP core on the logic side of the field-programmable gate array (FPGA) of the software-defined radio (SDR) platform. This module utilizes the pipelined parallel processing architecture of the FPGA to directly interface with the baseband IQ signal stream output from the RF front-end analog-to-digital converter (ADC), converting the high-sampling-rate time-domain signal into a low-dimensional time-frequency spectrum matrix. The physical principle behind this conversion process is that different types of RF interference (such as sweep interference and comb interference) and channel fading effects exhibit significantly different texture characteristics in the time-frequency domain. By converting a one-dimensional signal into a two-dimensional image, these microscopic physical features can be captured using computer vision techniques.

[0030] The physical layer feature acquisition module 10 specifically includes a windowing and framing unit, a spectrum transformation unit, a logarithmic mapping unit, and a feature map generation unit.

[0031] The windowed framing unit is used to receive continuous complex baseband signal streams. Since directly performing a Fourier transform on the truncated signal causes spectral leakage, the side lobes of strong interference signals can overwhelm weak signals, affecting texture clarity; therefore, windowing processing is necessary. The windowed framing unit has a fixed frame length and number of overlap points. For the input discrete-time signal Signal after windowing It is calculated using the following formula: ; In the formula, The number of points used in the Fast Fourier Transform (FFT) operation is configured based on the current operating mode, such as in shortwave narrowband mode. The value is 1024, in the ultra-shortwave broadband mode. The value is 4096; This is the index of the sampling point within the current frame, with a value ranging from 0 to... Number of overlapping points Typically set to That is, a 50% overlap rate, to ensure the continuity of the signal's time-domain information. and This represents the fixed weighting coefficients of the Hamming window function.

[0032] The spectrum transformation unit employs a parallel pipelined structure built upon the FPGA's internal DSPSlice resources to perform a radix-2 Fast Fourier Transform (FFT). This spectrum transformation unit receives a windowed time-domain complex sequence. Convert it into a frequency domain complex sequence The specific hardware implementation structure of FFT operations can be achieved by those skilled in the art using cascaded butterfly operation units and rotation factor lookup tables, which are well-known techniques in the field of digital signal processing and will not be elaborated upon here.

[0033] The logarithmic mapping unit is connected to the output of the spectrum transformation unit and is used to convert linear-scale spectral energy into logarithmic-scale dBm values ​​that conform to human visual perception and neural network input characteristics. This logarithmic mapping unit first calculates the frequency domain complex sequence. The square of the modulus, i.e., the power spectral density, is then subjected to a logarithmic transformation. The specific mapping calculation follows the formula: ; In the formula, Indicates the first The logarithmic power values ​​of each subcarrier, in dBm; and Representing frequency domain complex sequences respectively The real and imaginary parts; The normalized load impedance of the system is set to 50 ohms in this embodiment; This is a calibration constant for the RF link. This constant is obtained through factory calibration and stored in the FPGA's configuration register. It is used to compensate for the gain and feeder loss of the front-end low-noise amplifier (LNA), enabling... It can accurately reflect the received signal strength at the antenna aperture.

[0034] The feature map generation unit is used to convert a continuous power spectrum frame sequence into a fixed-size time-frequency map matrix to adapt to the fixed input dimension of the subsequent convolutional neural network. The feature map generation unit performs two sub-processes: dynamic range quantization and dimension normalization.

[0035] During dynamic range quantization, the feature map generation unit reads a preset sensitivity noise floor threshold. (e.g., -120dBm) and saturation level threshold (For example, -20dBm). These two thresholds are determined based on the receiver's hardware dynamic range specifications. The feature map generation unit will... Linear mapping to 8-bit unsigned integers This refers to the grayscale pixel value. The quantization formula is as follows: ; In the formula, This indicates a floor operation. If the result is less than 0, it is truncated to 0; if it is greater than 255, it is truncated to 255. Through this step, the energy fluctuations of the radio frequency signal are converted into changes in the brightness of image pixels.

[0036] During dimension normalization, the number of FFT points varies depending on the bandwidth of different frequency bands. It may be larger than the target map width. In the case of 64 pixels (e.g.), the feature map generation unit downsamples the frequency domain dimension using a max pooling strategy. The downsampling step size... Calculated as For the first output map Pixel column ( Its value is the maximum value of the input sequence within the corresponding interval: ; In the formula, To reduce the sampling step size; For the input sequence index; Input sequence values; It is a function for maximizing the value; Represents the time-frequency spectrum matrix at time... , No. The pixel values ​​in the column are processed using max pooling instead of average pooling to preserve the peak features of narrowband interference signals during dimensionality reduction, preventing weak interference textures from being buried by background noise during averaging. Simultaneously, in the time dimension, this unit employs a first-in-first-out (FIFO) queue to cache the most recently viewed data. The spectral data of frames (e.g., 64 frames) are finally stitched together to generate a resolution of The single-channel grayscale time-frequency spectrum matrix.

[0037] After the time-frequency spectrum matrix is ​​generated, the physical layer feature acquisition module 10 triggers the direct memory access (DMA) controller to write the spectrum matrix data into the shared memory region of the SoC system in burst transfer mode, and sends a completion interrupt signal to the link state prediction module 20. Through the above hardware implementation, the physical layer feature acquisition module 10 can continuously output visual texture data characterizing the physical state of the channel without consuming CPU computing power.

[0038] The link state prediction module 20 runs in the neural processing unit (NPU) or digital signal processor (DSP) of the embedded system-on-a-chip in the communication terminal. The working principle of this module is based on the spatiotemporal continuity of channel state evolution: the changes in interference sources (such as sweep frequency interference and comb interference) or fading characteristics in the physical channel are not discrete jumps on the time axis, but rather exhibit a gradual, textured characteristic with specific directionality or diffusion in the time-frequency domain. The link state prediction module 20 captures this textured evolution pattern through a deep learning model, thereby predicting the time of link failure before the signal-to-noise ratio deteriorates to the interruption threshold.

[0039] The link state prediction module 20 is logically divided into a deep feature extraction unit, a remaining lifetime regression unit, and a multi-dimensional arbitration decision unit.

[0040] The deep feature extraction unit is responsible for extracting spatiotemporal features reflecting channel evolution from the input time-frequency spectrum sequence. This deep feature extraction unit adopts a hybrid architecture based on lightweight convolutional neural networks (CNNs) and recurrent neural networks (RNNs). The input data is organized into tensor form. ,in For batch size, The time step (i.e., the number of consecutive historical frames input, such as 8 or 16 frames). and These represent the height and width of the graph, respectively.

[0041] In the spatial dimension, this deep feature extraction unit uses a depth-separable convolutional layer to construct a feature extractor. By utilizing a combination of channel-wise convolution and point-wise convolution, it significantly reduces floating-point operations while maintaining feature extraction capabilities, in order to extract edge and morphological features of textures such as sweep interference and comb interference.

[0042] In the time dimension, the deep feature extraction unit flattens the feature map output by the convolutional layer and inputs it into the LSTM layer of the Long Short-Term Memory network or the GRU layer to establish a dynamic model of texture features changing over time and capture the moving speed or energy diffusion rate of interference frequency points.

[0043] The Remaining Lifetime Regression (L-VRW) unit is connected to the output of the deep feature extraction unit. It maps the extracted high-dimensional features into a single scalar output, the Link Remaining Effective Time Window (L-VRW), through a fully connected layer. Here, L-VRW is physically defined as follows: under the current channel texture evolution trend, the bit error rate (BER) of the physical link remains within the limit of the system's forward error correction (FEC) capability (e.g., 10^-5). -3 The remaining lifetime regression unit (RBR) is the estimated duration of the remaining connection. Unlike traditional classification models that output discrete interference type labels, the RBR outputs continuous time predictions, quantifying the urgency of the impending link failure. The training of this prediction model employs supervised learning, with the training dataset derived from historical channel recordings. Its ground truth label is calculated by offline analysis of the time difference between the moment the BER actually exceeds the threshold and the current frame.

[0044] The multi-dimensional arbitration decision-making unit is used to comprehensively determine whether to trigger preemptive intervention based on L-VRW prediction values, business attributes, and global strategy parameters. This multi-dimensional arbitration decision-making unit first establishes a parameter acquisition mechanism: On the one hand, by detecting the DPI (Deep Packet Inspection) or reading the Type of Service (ToS) field in the IP header, the service priority parameter of the data packet to be transmitted can be parsed. Different types of services are mapped to values ​​between 0 and 1 (e.g., voice services are mapped to 0.8, and file transfer is mapped to 0.2). On the other hand, the risk level parameters in the currently effective network global policy are read from the shared memory area of ​​the policy synchronization and execution module 50. . It is a normalized floating-point number that represents the tolerance of the current network environment to link interruption.

[0045] To implement differentiated intervention strategies, the multidimensional arbitration decision-making unit calculates dynamic safety thresholds. This threshold represents the minimum time margin that the system must react in advance to ensure a seamless business switchover. The calculation follows the formula: ; In the formula, This indicates the internal processing latency of the system, including the time spent on model inference and interface calls, and is usually a fixed value (e.g., 10ms to 50ms). This represents the standard handshake delay required for the routing protocol to complete neighbor discovery and topology convergence. This value depends on the configuration of the routing protocol being run (e.g., it may be 2 to 5 seconds in OSPF). This represents the business priority weight, with a value ranging from 0 to 1. This is a risk level parameter, with a value ranging from 0 to 1; and The preset adjustment weight coefficient, typically set between 0.1 and 1.0, is used to adjust the impact of business importance and environmental risk on the decision threshold.

[0046] After completing the threshold calculation, the multi-dimensional arbitration decision unit executes comparison logic: when the real-time predicted L-VRW is less than the calculated L-VRW... When it is determined that the physical link is about to fail to meet the transmission requirements of the current service, and the remaining time is insufficient to support a conventional reactive routing switch, the unit immediately generates an intervention command and outputs it to the intervention event generation module 30. This intervention command includes the target network interface identifier and the suggested blocking duration. By introducing a weighted calculation of service attributes and global risk levels, this embodiment achieves a shift from single-physical-layer prediction to cross-layer service awareness-based intelligent decision-making. This ensures that when facing high-priority services or high-risk environments, the system can adopt a more aggressive preemptive switching strategy, thereby completing the routing migration before the physical link is completely interrupted.

[0047] When the link state prediction module 20 outputs an intervention command, the intervention event generation module 30 is activated, performing environmental context capture, texture fingerprinting, and event encapsulation and distribution operations. This intervention event generation module 30 converts the specific physical cause of this preemptive blocking—that is, the specific interference texture pattern in the i.e., the frequency domain—into binary structured data, enabling different nodes to identify and classify the same type of interference source. The intervention event generation module 30 is internally divided into a context capture unit, a texture fingerprint extraction unit, and an event encapsulation and distribution unit.

[0048] The context capture unit is responsible for recording the basic spatiotemporal metadata at the time of the intervention. This unit reads the system clock to obtain a UTC timestamp with microsecond precision, accesses the baseband processor's status register to read the current center frequency and bandwidth configuration parameters, and parses the type of action executed (e.g., "setting routing cost to the maximum" or "silencing physical interface transmission") and the model-predicted duration of the blocking from the intervention command. This basic metadata constitutes the header information of the preemptive intervention event PIE log.

[0049] The texture fingerprint extraction unit utilizes perceptual hashing technology to transform high-dimensional time-frequency spectral features into compact binary fingerprints, i.e., "texture hashing." The underlying principle is that the intermediate layer feature map of a deep learning model is essentially an abstract representation of the energy distribution topology of the input signal in the time-frequency domain, possessing translation invariance and scaling invariance. Therefore, extracting intermediate layer features and performing binarization compression can generate "digital fingerprints" with noise resistance against various interference sources.

[0050] The texture fingerprint extraction unit directly extracts intermediate layer data from the deep learning model of the link state prediction module 20. Specifically, the extraction location is the output of the global average pooling layer after the last convolutional layer and before the fully connected layer in the CNN model. The extracted feature vector is denoted as... This vector typically contains a set of 32-bit floating-point numbers.

[0051] To facilitate network transmission and subsequent efficient matching, the texture fingerprint extraction unit processes the feature vector. Perform binary hash encoding. Let the feature vector be... The dimension is To adapt to the hash encoding bit length (In this embodiment) (Set to 64), if Not equal to Then, a preset linear projection matrix will be used to project the data. Mapped to A dimensional vector. Let the mapped vector be of dimension 1. The value of each element is The texture fingerprint extraction unit first calculates the mean of the vector elements. This mean is used as the adaptive quantization threshold. Texture hash value The generation logic follows the following formula: ; ; In the formula, For the first The binary hash value of a bit represents the relative magnitude of the feature strength of that dimension with respect to the overall average strength; These are the element values ​​of the feature vector; For adaptive quantization threshold; For texture hash value; Indicates the first The weight of each binary bit in the integer representation; The total number of bits in the hash value is fixed at 64. Through the above calculation, the time-frequency image features are compressed into a 64-bit long integer. This hash value has Hamming distance comparability, meaning that for similar interference patterns (e.g., the same type of radar pulse with only a slight frequency shift), the generated hash values ​​differ by only a few bits, thus achieving digital classification of physical interference patterns.

[0052] The event encapsulation and distribution unit is connected to the texture fingerprint extraction unit and is used to construct and send the preemptive intervention event PIE log. This event encapsulation and distribution unit defines a compact binary log format, which includes, in sequence: event ID, timestamp, center frequency, and 64-bit texture hash value. Intervention action codes and reserved fields. At the network distribution level, the event encapsulation and distribution unit adopts a restricted flooding mechanism. This unit encapsulates PIE logs into UDP packets and sets the Time-to-Live (TTL) field in the IP header to a small integer (e.g., 3). This value is determined based on the average neighbor density of the network, aiming to limit packets to propagate only within a local neighborhood of 2 to 3 hops, preventing a network-wide broadcast storm.

[0053] To ensure reliable transmission of log data, the event encapsulation and distribution unit employs a "piggybacking" strategy. This unit maintains a queue of logs to be sent and monitors the transmission status of routing control messages at the network layer in real time. When it detects that a node is about to send a regular routing control message (such as a Hello packet or TC topology control packet from the OLSR protocol), it constructs a custom-type TLV (Type-Length-Value) data block. The Type field is identified as "Interference Event Log," and the Value field is filled with PIE log content. This TLV data block is then appended to the extension area of ​​the routing message and sent together. This mechanism leverages the periodic broadcast characteristics and link keep-alive mechanism of the routing protocol itself, achieving reliable propagation of interference events among neighboring nodes without introducing additional channel contention.

[0054] The global policy generation module 40 is deployed in cluster head nodes, gateway nodes, or ground command and control center servers within a hierarchical self-organizing network architecture. This module operates based on the principle of "multi-point collaborative sensing": due to the time-varying characteristics of wireless channels, the detection results of a single node often contain false alarms caused by shadow fading or sudden noise. The global policy generation module 40 aggregates observation data from nodes in different geographical locations, utilizing the continuity of spatial distribution and the temporal reproducibility of interference sources to eliminate isolated, sporadic events and extract common network environment features. The global policy generation module 40 includes a log aggregation and cleaning unit, a time-series correlation mining unit, and a policy formulation and deployment unit.

[0055] The log aggregation and cleaning unit is responsible for maintaining a global log database based on a sliding time window. This unit continuously receives preemptive intervention event (PIE) logs reported by the intervention event generation module 30 from various nodes via a network interface. To adapt to the dynamically changing network environment, the log aggregation and cleaning unit is configured with a sliding window duration. (For example, set to 3600 seconds, this value is usually set to 2 to 3 times the activity period of the interference source). Only log data within this time window is retained, and expired data will be automatically deleted or archived to local non-volatile storage media. At the same time, in the case where the same interference event may be reported by multiple neighboring nodes almost simultaneously, a deduplication operation is performed: if the Hamming distance of the texture hash values ​​of multiple logs is less than a preset threshold (e.g., 2 bits) and the timestamp difference is less than the synchronization error tolerance (e.g., 500ms, this value depends on the time synchronization accuracy of the entire network), they are merged into an aggregate record, and the list of reporting nodes is recorded.

[0056] The temporal correlation mining unit is used to identify statistically significant interference patterns from aggregated log data. Because physical signals exhibit multipath effects as they propagate along different paths, and because FPGAs have least-squares errors during spectral quantization, even for the same interference source, the texture hash values ​​extracted from different nodes may show subtle bit flips. Therefore, this temporal correlation mining unit first clusters the texture hashes based on Hamming distance. For two texture hash values... and The similarity judgment logic is as follows: ; In the formula, Indicates Hamming distance; and Representing hash values ​​respectively and In the The value (0 or 1) on each binary bit. This represents the XOR operation; The similarity clustering threshold is set to 4 in this embodiment. This threshold is chosen to ensure that signal sources of the same type can still be grouped together while tolerating no more than 6.25% (4 / 64) of feature bit differences, preventing feature splitting caused by channel distortion. Hash values ​​that meet the condition are grouped into the same interference cluster. .

[0057] The temporal correlation mining unit calculates each interference cluster Global hazard confidence This metric reflects the probability and extent of impact of a specific texture feature causing link failure. The calculation formula is as follows: ; In the formula, Indicates the current time window Internally reported as belonging to cluster The number of independent nodes in the intervention event, which characterizes the spatial coverage of the interference; This represents the total number of active nodes in the current network. This is the average frequency (times / minute) of the interference cluster appearing across the entire network, and this term characterizes the temporal recurrence frequency of the interference. This is the frequency sensitivity coefficient, ranging from 0.1 to 0.5. This coefficient is used to adjust the confidence level's sensitivity curve to frequency, preventing high-frequency noise from causing confidence level saturation too quickly. The formula indicates that the more nodes that detect an interfering texture and the higher its frequency of occurrence, the closer its hazard confidence level will be to 1.

[0058] The policy formulation and dissemination unit generates a network-wide policy (NGP) based on the calculated hazard confidence level. This unit maintains a mapping table that maps representative texture hash values ​​of interference clusters to additional costs for routing protocols. Regarding confidence level For interference patterns exceeding a preset trigger threshold (e.g., 0.3), this unit generates a penalty rule. The penalty cost is... The calculation follows a linear mapping relationship: ; In the formula, The unit link baseline cost defined for the routing protocol (e.g., typically defined as 1 in the OLSR protocol); and These are the minimum trigger threshold (e.g., 0.3) and saturation threshold (e.g., 0.8) for the confidence level, respectively. This is the maximum penalty multiplier factor, typically set between 5 and 20. The principle for determining this factor is that it must be greater than the average diameter of the current network (i.e., the maximum number of hops) to ensure that when the maximum penalty is applied, the path cost calculated by the routing algorithm through the interference area is significantly greater than the cost of the detour path, thus mathematically guaranteeing the effectiveness of route avoidance. This indicates rounding down to the nearest integer.

[0059] The generated Network Global Policy (NGP) is encapsulated as a configuration file containing a version number, an effective timestamp, and multiple "texture hash-penalty cost" rule pairs. The policy formulation and publishing unit distributes the updated NGP to the entire network via multicast, enabling all nodes to preventively avoid similar interference in their local routing calculations based on the policy.

[0060] The policy synchronization and execution module 50 is used to convert received global policies or local intervention commands into specific routing metric parameter adjustments and radio frequency transceiver control actions. This module decouples the actual connectivity of physical links from the routing priority of logical routes: even if the physical link can still communicate at the signal-to-noise ratio level, if high-risk interference is predicted, the system still forcibly increases the logical overhead of the link, automatically avoiding risky areas using the shortest path algorithm of the routing protocol, and achieving service migration before physical interruption. Specifically, the policy synchronization and execution module 50 includes a policy incremental synchronization unit, a virtual metric injection unit, and a passive tracking and backswitching unit.

[0061] The policy incremental synchronization unit is responsible for maintaining consistency between the local policy database and the latest state of the entire network. It receives network global policy (NGP) update packets from the global policy generation module 40 by listening to a specific multicast address. To reduce the bandwidth consumption of control signaling, this policy incremental synchronization unit employs a version comparison mechanism based on a logical clock. When the received NGP version number... Greater than the local version number At that time, the incremental synchronization unit of the strategy parses the difference part in the update package, that is, the newly added or modified "texture hash-penalty cost" rule, and uses atomic operations to write it into the strategy lookup table in the local shared memory mapping area to prevent data inconsistency caused by multi-threaded read and write.

[0062] The virtual metric injection unit is configured to implement "soft blocking" intervention at the routing layer. In standard link-state routing protocols (such as OLSR or OSPF), link cost is typically determined solely by packet loss rate or bandwidth physical parameters. In this embodiment, the virtual metric injection unit takes over the assignment of link cost by calling the API interface of the routing daemon or directly modifying the link structure data before the link-state database (LSDB) is generated. During each routing table recalculation cycle or Hello packet assembly cycle, the policy incremental synchronization unit reads the real-time texture hash output by the physical layer feature acquisition module 10. And match it with the local policy library.

[0063] If the currently detected texture hash matches the first one in the policy library According to this rule, the virtual metric injection unit will force a modification of the link cost value reported to the routing engine. The revised cost value is calculated as follows: ; In the formula, The original physical overhead is calculated based on the current physical layer signal-to-noise ratio (SNR) and bit error rate (BER). This is the set of all interference strategies matched at the current moment. For the first The additional penalty value stipulated in this strategy; This is an indicator function that outputs 1 when the condition is met, and 0 otherwise. The current system time; For the first The timestamp of the last time the policy was confirmed to be in effect; This is the effective lifetime of the policy (e.g., 60 seconds). This formula ensures that only active interfering policies within their validity period will affect routing costs, preventing outdated policies from causing suboptimal persistence of routing paths.

[0064] The passive tracking and back-off unit is responsible for monitoring and recovery during physical layer blocking. When the predicted interference risk is extremely high, causing the system to decide to temporarily stop signal transmission to avoid radar detection or strong countermeasures, traditional link detection mechanisms will fail. At this time, the passive tracking and back-off unit enters "passive texture tracking" mode. In this mode, the RF transmission channel is turned off, but the reception channel remains open, continuously sampling and hashing environmental signals.

[0065] To achieve oscillation-free smooth recovery, the passive tracking and back-off unit executes "negative dissipation decision" logic. Channel availability is inferred by statistically analyzing the frequency of target interference patterns in the received signal, and the passive tracking and back-off unit calculates the interference dissipation index. Its update formula uses the Exponentially Weighted Moving Average (EWMA) algorithm: ; In the formula, This is the current sampling time; For a moment Received environmental signal texture hash; The target interference texture hash that triggered this blocking; For the matching function, if the Hamming distance between two hash values ​​is less than the preset similarity threshold (e.g., a value of 4), then output 1 (indicating the presence of interference); otherwise output 0 (indicating the absence of interference). express Disturbance dissipation index at any given time; The smoothing factor is set to a range of 0.05 to 0.2. This range is chosen to smooth out instantaneous hash jumps caused by environmental noise while maintaining a second-level response capability to events where interference sources leave.

[0066] when Decrease to the preset recovery threshold (For example, 0.1 indicates an extremely low probability of interference) and the duration exceeds the safety protection time. (For example, 2 seconds, to prevent the ping-pong effect caused by intermittent transmissions from the interference source) When the passive tracking and back-cutting unit determines that the interference source has left or stopped working, the unit performs back-cutting in stages: first, it removes the transmission silence, allowing the transmission of routing Hello packets to probe physical connectivity; after receiving confirmation replies from neighboring nodes, it gradually and linearly reduces the additional cost in the virtual metric injection unit. Until the link overhead fully returns to normal. .

[0067] To better understand the technical solution of this invention, the following description is based on a specific application scenario.

[0068] Scene setting: such as Figure 3 As shown, assume a simple diamond-shaped self-organizing network system consisting of 4 nodes. Node A is the video reconnaissance source node, and node D is the command and control center receiving node. There are two possible paths in the network: Path 1 (Main Path): A to B to D, with a total physical hop count of 2 hops, excellent link quality (SNR>20dB), and this path is selected by default by the routing protocol.

[0069] Path 2 (alternative path): A to C to D, with a total physical hop count of 2 hops. However, due to the concealed location of node C, the signal undergoes multiple reflections, resulting in a slightly lower SNR (SNR≈12dB) and slightly higher routing overhead than Path 1.

[0070] Interference intervention and system response process: Initial phase (T=0s): Business data stream is transmitted stably along the "A to B to D" path, and the video is smooth.

[0071] Interference Approach (T=10s): A vehicle-mounted sweeping interference source begins to move towards node B. Although the physical link of node B has not yet been interrupted, discontinuous diagonal textures (sweeping features) begin to appear in the time-frequency spectrum matrix generated by the physical layer feature acquisition module 10.

[0072] Prediction and Decision (T=10.5s): Node B's link state prediction module 20 captures the energy enhancement trend of the texture and predicts "L-VRW (remaining effective time window) = 5.2 seconds". Since the current transmission is a high-priority video service (…),… ), calculated safety threshold Seconds. At this moment. The system is currently maintaining monitoring and has not triggered any actions.

[0073] Preemptive Intervention (T=12s): As the interference source approaches, the predicted L-VRW drops sharply to 2.5 seconds, which is less than the threshold of 4.0 seconds. Node B immediately triggers an intervention command, and the intervention event generation module 30 generates a texture hash and broadcasts it.

[0074] Policy Enforcement and Route Switching (T=12.2s): Node A receives a PIE log containing a risk warning from Node B (or updates the global policy). Node A's policy synchronization and execution module 50 immediately injects virtual costs into the local routing table for the "A to B" link. .

[0075] Lossless handover (T=12.5s): The routing protocol recalculates the path and finds that the total cost of path 1 is now 1+50+1=52, while the total cost of path 2 is 1+1=2. The system smoothly switches the next hop to node C.

[0076] Physical interruption (T=14s): The interference source covers node B, and the signal-to-noise ratio of node B drops to -5dB, resulting in a complete physical link interruption. However, at this time, the service flow had already switched to path 2 1.5 seconds earlier, and the video received by the command and control center D showed no stuttering or screen tearing.

[0077] Experimental verification and effect comparison: Simulation environment settings: Physical layer model: adopts 802.11p standard, carrier frequency 5.9GHz, bandwidth 10MHz, Rayleigh fading channel.

[0078] Network area: 2000m×2000m square area.

[0079] Interference Model: A mobile broadband sweeping interference source is set with an interference power of 30dBm, a sweep period of 1ms, and a speed of 15m / s to cross the central area of ​​the network.

[0080] Service flow: Node 1 sends a constant bitrate (CBR) UDP video stream to Node 50, with a packet size of 1024 bytes and a sending rate of 2Mbps.

[0081] Comparison of options: Baseline: Standard OLSR protocol, which detects link failures based solely on Hello packet loss rate.

[0082] This embodiment (Proposed) enables physical layer texture prediction and virtual metric injection mechanisms.

[0083] Analysis of experimental results: like Figure 4 As shown, comparing the normal channel in Figure (a) with the disturbed channel in Figure (b), the frequency sweeping interference exhibits a diagonal grayscale texture feature in the time-frequency spectrum. This indicates that the physical layer feature acquisition module 10 can effectively convert invisible radio frequency signals into visual images with specific shapes. Even in grayscale mode, the contrast between the interference texture and the background noise remains strong, verifying the feasibility of using computer vision technology to identify physical layer interference.

[0084] like Figure 5 As shown, the dashed line, representing the existing OLSR scheme, causes the throughput to drop to zero for approximately 7 seconds (from T=35s to T=42s) after a physical link interruption at T=35s, during which video services are completely interrupted. In contrast, the solid line, representing the scheme of this invention, performs a route switch in advance based on texture prediction results at T=33s (i.e., 2 seconds before the physical interruption). The curve shows that, except for a slight jitter at the moment of switchover, the throughput remains stable at 2Mbps throughout.

[0085] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A deep learning decision-making system for optimizing shortwave and ultra-shortwave networking protocols, characterized in that, Includes the following steps: The physical layer feature acquisition module is used to capture raw signal data from the baseband of the wireless communication device and generate a time-frequency spectrum matrix representing the current physical state of the channel through time-frequency analysis. The link status prediction module is used to process the time-frequency spectrum matrix based on a deep learning model to predict the remaining effective time window of the link, and to arbitrate the service attributes of the data packets to be transmitted with the risk level parameters set by the strategy synchronization and execution module to output intervention instructions. An intervention event generation module is used to respond to the intervention command, extract the texture features that lead to the intervention, and encapsulate them into a preemptive intervention event log for network distribution; The global policy generation module is used to aggregate the preemptive intervention event logs to identify reproducible link problems and generate a global network policy that includes routing cost adjustment rules. The policy synchronization and execution module is used to inject virtual costs into the local network protocol stack or execute radio frequency control according to the intervention instructions or the network global policy in order to adjust routing behavior.

2. The deep learning decision system for optimizing shortwave and ultra-shortwave networking protocols according to claim 1, characterized in that, The physical layer feature acquisition module includes: The windowing and framing unit is used to receive complex baseband signal streams and perform windowing processing on the signals according to the preset frame length and number of overlap points. The spectrum transformation unit is used to construct a parallel pipeline structure, perform a fast Fourier transform on the windowed signal, and convert the time-domain complex sequence into a frequency-domain complex sequence. The feature map generation unit is used to perform logarithmic mapping and dynamic range quantization on the frequency domain complex sequence, and to downsample the frequency domain dimension through a max pooling strategy, thereby splicing to generate the time-frequency map matrix of a fixed size.

3. The deep learning decision system for optimizing shortwave and ultra-shortwave networking protocols according to claim 1, characterized in that, The link state prediction module includes: The deep feature extraction unit adopts a hybrid architecture of convolutional neural network and recurrent neural network to extract spatiotemporal features reflecting channel evolution from the time-frequency spectrum matrix and capture the morphology and dynamic change patterns of interference texture. The remaining lifetime regression unit maps the extracted spatiotemporal features to a scalar output to obtain the remaining effective time window, which represents the duration for which the link bit error rate remains below the limit under the current texture evolution trend. A multi-dimensional arbitration decision unit is used to calculate the dynamic safety threshold and generate the intervention command when the remaining effective time window is less than the dynamic safety threshold.

4. The deep learning decision-making system for optimizing shortwave and ultra-shortwave networking protocols according to claim 3, characterized in that, When calculating the dynamic security threshold, the multi-dimensional arbitration decision-making unit: Parse the service priority parameters of the data packets to be transmitted, and read the currently effective risk level parameters from the policy synchronization and execution module; Based on the sum of the internal processing latency of the decision-making system and the handshake latency of the routing protocol, and superimposed with the weighted influence of the service priority parameter and the risk level parameter, a dynamic security threshold is calculated to ensure that the switching is triggered in advance in high-priority services or high-risk environments.

5. The deep learning decision system for optimizing shortwave and ultra-shortwave networking protocols according to claim 1, characterized in that, The intervention event generation module includes: The context capture unit is used to record spatiotemporal metadata when the intervention occurs, including timestamps, center frequencies, and predicted duration of the blockade. The texture fingerprint extraction unit is used to extract feature vectors from the intermediate layer of the deep learning model of the link state prediction module and convert them into binary texture hash values ​​for digital classification of physical interference patterns. The event encapsulation and distribution unit is used to combine the spatiotemporal metadata with the texture hash value to construct the preemptive intervention event log, and to piggyback the transmission using the extended area of ​​the routing control message.

6. The deep learning decision system for optimizing shortwave and ultra-shortwave networking protocols according to claim 5, characterized in that, When the texture fingerprint extraction unit generates the texture hash value: Extract the output of the global average pooling layer of the convolutional neural network as a feature vector; The mean of the feature vector elements is calculated as the adaptive quantization threshold; The numerical values ​​of each dimension of the feature vector are compared with the adaptive quantization threshold to generate a binary sequence, which serves as the core identification identifier in the preemptive intervention event log.

7. The deep learning decision system for optimizing shortwave and ultra-shortwave networking protocols according to claim 1, characterized in that, The global policy generation module includes: The log aggregation and cleaning unit is used to maintain a database based on a sliding time window, and to deduplicate and merge the preemptive intervention event logs reported by multiple nodes based on the Hamming distance of the texture hash values. The temporal correlation mining unit is used to cluster the aggregated logs and calculate the global hazard confidence of each interference cluster. The global hazard confidence is determined by the spatial coverage breadth and temporal recurrence frequency of the interference. The policy formulation and distribution unit is used to map interference clusters whose global hazard confidence exceeds a threshold to specific routing penalty costs, generate the global network policy, and distribute it to the entire network.

8. The deep learning decision system for optimizing shortwave and ultra-shortwave networking protocols according to claim 1, characterized in that, The strategy synchronization and execution module includes: The policy incremental synchronization unit is used to listen to and parse the update packets of the network global policy, and write the newly added texture hash and penalty cost rules into the local policy library. The virtual metric injection unit is used to match the texture hash detected in real time with the local policy library during the routing calculation cycle. When the match is successful, the corresponding penalty cost is forcibly added on the basis of physical cost to generate a corrected link cost value and report it to the routing engine to decouple physical connectivity and logical routing priority.

9. The deep learning decision system for optimizing shortwave and ultra-shortwave networking protocols according to claim 8, characterized in that, The strategy synchronization and execution module also includes a passive tracking and switchback unit for use during RF transmit channel shutdown: Keep the receiving channel open, continuously calculate the texture hash of the environmental signal, and update the interference dissipation index using an exponentially weighted moving average algorithm; When the interference dissipation index drops to the recovery threshold and the duration exceeds the safety protection time, it is determined that the interference source has left, and the penalty value in the virtual metric injection unit is gradually reduced until the link overhead returns to the physical true value.

10. The deep learning decision system for optimizing shortwave and ultra-shortwave networking protocols according to claim 3, characterized in that, The physical layer feature acquisition module is configured as a hardware-accelerated IP core in a field-programmable gate array (FPGA), and the link state prediction module runs in the neural processing unit (NPU) of the embedded system-on-a-chip. The two share the time-frequency spectrum matrix data through direct memory access (DMA).

Images