Raw wireless signal target recognition method and device based on model fusion decision

Abstract

This invention provides a method and apparatus for identifying raw wireless signal targets based on model fusion decision-making, relating to the field of intelligent signal sensing technology. The method involves inputting raw wireless signal data into the signal sensing and preprocessing layer of a signal target identification model to obtain dimensional representation information output by the signal sensing and preprocessing layer. Each signal representation data in the dimensional representation information is then input into the parallel detection and feature extraction layer of the signal target identification model to obtain structured detection evidence corresponding to each signal representation data output by the parallel detection and feature extraction layer. All structured detection evidence is then input into the intelligent fusion and decision-making layer of the signal target identification model to obtain the fusion decision result output by the intelligent fusion and decision-making layer. Finally, the fusion decision result is input into the result output layer of the signal target identification model to obtain the target identification result of the raw wireless signal data output by the result output layer. The technical solution of this invention improves the accuracy of signal identification.

Description

Technical Field

[0001] This invention relates to the field of intelligent sensing technology for wireless signals, and in particular to a method and apparatus for identifying raw wireless signal targets based on model fusion decision-making. Background Technology

[0002] In practical engineering, real-time signal identification and analysis are particularly important.

[0003] Currently, in existing technologies, signal recognition often employs a single deep neural network to detect the representation of the input signal. For example, it could be... The received in-phase quadrature (IQ) data is converted into a time-frequency graph via short-time Fourier transform. This graph is then input into a YOLO (a single-stage real-time object detection algorithm), convolutional neural network, or Transformer (a self-attention deep learning model architecture) network to output the bounding box, category label, frequency band location, and confidence score. In practical deployments, signal morphology is significantly affected by bandwidth variations, modulation scheme changes, power fluctuations, occlusion, multipath propagation, device state switching, and environmental noise. Therefore, relying on a single model for signal recognition is too one-sided, often only covering a portion of the signal distribution and failing to provide comprehensive signal identification, thus reducing the accuracy of signal recognition.

[0004] Therefore, there is an urgent need for a method for identifying raw wireless signal targets based on model fusion decision-making to improve the accuracy of signal identification. Summary of the Invention

[0005] This invention provides a method and apparatus for target recognition of raw wireless signals based on model fusion decision-making. It addresses the shortcomings of existing technologies where a single model for signal recognition is too one-sided, unable to comprehensively identify signals, and thus reduces the accuracy of signal recognition. By performing multi-dimensional transformation on the raw wireless signal and parallel detection processing on the transformed multi-dimensional signal representation data, highly robust signal detection is achieved. Then, based on the structured detection evidence from the parallel detection, probabilistic fusion decision-making is performed to finally obtain the target recognition result of the raw wireless signal, thereby improving the accuracy of signal recognition.

[0006] This invention provides a method for identifying raw wireless signal targets based on model fusion decision-making, comprising the following steps.

[0007] Acquire raw wireless signal data; The raw wireless signal data is input into the signal perception and preprocessing layer of the signal target recognition model to obtain the dimensional representation information output by the signal perception and preprocessing layer; wherein, the signal perception and preprocessing layer is a layer that transforms and processes the raw wireless signal data; the dimensional representation information includes signal representation data of at least one dimension; The signal representation data of each signal in the dimensional representation information is input into the parallel detection and feature extraction layer in the signal target recognition model to obtain the structured detection evidence corresponding to each signal representation data output by the parallel detection and feature extraction layer; wherein, the parallel detection and feature extraction layer is a layer that performs basic detection and feature extraction on the dimensional representation information. All structured detection evidence is input into the intelligent fusion and decision layer of the signal target recognition model to obtain the fusion decision result output by the intelligent fusion and decision layer; whereby the intelligent fusion and decision layer is the layer that performs fusion decision on the signal detection results; The fused decision results are input into the output layer of the signal target recognition model to obtain the target recognition results of the original wireless signal data output by the output layer.

[0008] According to the present invention, a method for identifying raw wireless signal targets based on model fusion decision-making includes at least one of the following transformation processes: short-time Fourier transform processing, power spectral density estimation processing, normalization processing, rotation processing, resampling processing, and tensor splicing processing.

[0009] According to the present invention, a method for identifying raw wireless signal targets based on model fusion decision-making determines the state prediction result based on historical state variables, including: The signal representation data from each dimension representation information are input into the parallel detection and feature extraction layer of the signal target recognition model to obtain the structured detection evidence corresponding to each signal representation data output by the parallel detection and feature extraction layer, including: The signal representation data of each signal in the dimension representation information is input into the basic detector pool in the parallel detection and feature extraction layer of the signal target recognition model to obtain the signal detection results corresponding to each signal representation data output by the basic detector pool; wherein, the signal detection results include at least one of the following: candidate target box, category label, confidence, center frequency, bandwidth range, time position, intensity score and protocol type.

[0010] The signal detection results corresponding to each signal representation data are input into the evidence extraction module in the parallel detection and feature extraction layer of the signal target recognition model to obtain the structured detection evidence corresponding to each signal representation data input to the evidence extraction module.

[0011] According to the present invention, a method for identifying raw wireless signal targets based on model fusion decision-making includes a basic detector pool in the parallel detection and feature extraction layer comprising at least one signal detector; each signal detector corresponds one-to-one with each signal representation data in the dimensional representation information; each signal representation data in the dimensional representation information is input into the basic detector pool in the parallel detection and feature extraction layer of the signal target identification model to obtain the signal detection results corresponding to each signal representation data output by the basic detector pool, including: Each signal representation data in the dimension representation information is input in parallel into each signal detector in the basic detector pool of the parallel detection and feature extraction layer of the signal target recognition model, and the signal detection result corresponding to each signal representation data output by each signal detector is obtained.

[0012] According to the original wireless signal target recognition method based on model fusion decision provided by the present invention, the signal detection results corresponding to each signal representation data are input into the evidence extraction module in the parallel detection and feature extraction layer of the signal target recognition model to obtain the structured detection evidence corresponding to each signal representation data input to the evidence extraction module, including: The signal detection results corresponding to each signal characterization data are input into the evidence extraction module of the evidence extraction module of the parallel detection and feature extraction layer of the signal target recognition model to obtain the initial structured detection evidence corresponding to each signal detection result output by the evidence recognition unit. The initial structured detection evidence corresponding to each signal detection result is input into the evidence extraction module of the parallel detection and feature extraction layer of the signal target recognition model, and the structured detection evidence corresponding to each signal detection result is obtained from the output of the evidence normalization unit.

[0013] According to the original wireless signal target recognition method based on model fusion decision provided by the present invention, before obtaining the fusion decision result output by the intelligent fusion and decision layer of the signal target recognition model by inputting all structured detection evidence into the intelligent fusion and decision layer, the method further includes: All structured detection evidence is uniformly standardized and mapped to obtain a set of standardized message structures.

[0014] According to the present invention, a method for identifying raw wireless signal targets based on model fusion decision-making is provided, wherein all structured detection evidence is input into the intelligent fusion and decision layer of the signal target identification model to obtain the fusion decision result output by the intelligent fusion and decision layer, including: The candidate target set and the detector responsibility set are determined based on each standardized message structure in the standardized message structure set. The candidate target set includes at least one candidate target, and the detector responsibility set includes the detector responsibility vectors of the signal detectors corresponding to each standardized message structure detected. The detector responsibility vectors are used to characterize the relative interpretability of the signal detectors for the candidate targets. The standardized message result is the result after uniformly standardizing and mapping the structured detection evidence. Obtain the posterior responsibility of the historical detector and the responsibility transition matrix of the historical detector; Based on the posterior responsibility of historical detectors and the responsibility transition matrix of historical detectors, the responsibility vector of each detector in the set of detector responsibility is predicted in advance to obtain the predicted responsibility. The soft association weights are obtained by performing likelihood matching between each detector responsibility vector in the detector responsibility set and each candidate target in the candidate target set. The target state vector and confidence information of structured detection evidence for each candidate target are obtained; wherein, the target state vector includes at least one of the following: frequency position, bandwidth, power, category, existence probability, motion information and orientation information of the candidate target.

[0015] The target state vector is fused and a decision is made based on the predictive responsibility, soft association weight, and confidence information to obtain the fusion decision result.

[0016] According to the present invention, a method for identifying raw wireless signal targets based on model fusion decision-making is provided, which performs likelihood matching between each detector responsibility vector in the detector responsibility set and each candidate target in the candidate target set to obtain soft correlation weights, including: The target observation correlation graph is obtained by performing likelihood matching between each detector responsibility vector in the detector responsibility set and each candidate target in the candidate target set. The soft correlation between each detector responsibility vector and each candidate target is calculated based on the target observation correlation diagram, and the soft correlation weight is obtained.

[0017] According to the original wireless signal target recognition method based on model fusion decision provided by the present invention, after performing likelihood matching between each detector responsibility vector in the detector responsibility set and each candidate target in the candidate target set to obtain soft correlation weights, the method further includes: The responsibility vectors of each detector in the detector responsibility set are updated in reverse according to the soft correlation weights. Based on the updated responsibility vectors, the steps of prior prediction of the responsibility vectors of each detector in the detector responsibility set according to the historical detector posterior responsibility and historical detector responsibility transition matrix are continued to obtain the predicted responsibility.

[0018] The present invention also provides a raw wireless signal target identification device based on model fusion decision, comprising the following modules: The data acquisition module is used to acquire raw wireless signal data; The conversion processing module is used to input the original wireless signal data into the signal perception and preprocessing layer in the signal target recognition model to obtain the dimensional representation information output by the signal perception and preprocessing layer; wherein, the signal perception and preprocessing layer is a layer that performs conversion processing on the original wireless signal data; the dimensional representation information includes signal representation data of at least one dimension; The evidence detection module is used to input the signal representation data of each signal in the dimensional representation information into the parallel detection and feature extraction layer in the signal target recognition model, and obtain the structured detection evidence corresponding to each signal representation data output by the parallel detection and feature extraction layer; wherein, the parallel detection and feature extraction layer is a layer that performs basic detection and feature extraction on the dimensional representation information. The fusion decision module is used to input all structured detection evidence into the intelligent fusion and decision layer of the signal target recognition model, and obtain the fusion decision result output by the intelligent fusion and decision layer; wherein, the intelligent fusion and decision layer is the layer that performs fusion decision on the signal detection results; The result recognition module is used to input the fused decision results into the result output layer of the signal target recognition model to obtain the target recognition result of the original wireless signal data output by the result output layer.

[0019] The present invention also provides an electronic device, including a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, it implements any of the above-described methods for identifying raw wireless signal targets based on model fusion decision-making.

[0020] The present invention also provides a non-transitory computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements any of the above-described methods for identifying raw wireless signal targets based on model fusion decision-making.

[0021] The present invention also provides a computer program product, including a computer program that, when executed by a processor, implements any of the above-described methods for identifying raw wireless signal targets based on model fusion decision-making.

[0022] This invention provides a method and apparatus for identifying raw wireless signal targets based on model fusion decision-making. The method involves: acquiring raw wireless signal data; inputting the raw wireless signal data into a signal perception and preprocessing layer in a signal target identification model to obtain dimensional representation information output by the signal perception and preprocessing layer; wherein the signal perception and preprocessing layer is a layer that transforms and processes the raw wireless signal data; the dimensional representation information includes signal representation data of at least one dimension; inputting each signal representation data in the dimensional representation information into a parallel detection and feature extraction layer in the signal target identification model to obtain structured detection evidence corresponding to each signal representation data output by the parallel detection and feature extraction layer; wherein the parallel detection and feature extraction layer is a layer that performs basic detection and feature extraction on the dimensional representation information; inputting all structured detection evidence into an intelligent fusion and decision-making layer in the signal target identification model to obtain a fusion decision result output by the intelligent fusion and decision-making layer; wherein the intelligent fusion and decision-making layer is a layer that performs fusion decision-making on the signal detection results; and inputting the fusion decision result into a result output layer in the signal target identification model to obtain the target identification result of the raw wireless signal data output by the result output layer. The technical solution of this invention addresses the shortcomings of existing technologies where a single model for signal recognition is too one-sided, unable to comprehensively identify signals, and thus reduces the accuracy of signal recognition. By performing multi-dimensional transformation on the original wireless signal and parallel detection processing on the transformed multi-dimensional signal representation data, highly robust signal detection is achieved. Then, based on the structured detection evidence from the parallel detection, probabilistic fusion decision-making is performed to finally obtain the target recognition result of the original wireless signal, thereby improving the accuracy of signal recognition. Attached Figure Description

[0023] To more clearly illustrate the technical solutions in this invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of this invention. For those skilled in the art, other drawings can be obtained from these drawings without creative effort.

[0024] Figure 1 This is one of the flowcharts of the original wireless signal target recognition method based on model fusion decision provided by the present invention.

[0025] Figure 2 This is the second flowchart of the original wireless signal target recognition method based on model fusion decision provided by the present invention.

[0026] Figure 3 This is the third flowchart of the original wireless signal target recognition method based on model fusion decision provided by the present invention.

[0027] Figure 4This is a schematic diagram of the original wireless signal target identification device based on model fusion decision provided by the present invention.

[0028] Figure 5 This is a schematic diagram of the structure of the electronic device provided by the present invention. Detailed Implementation

[0029] To make the objectives, technical solutions, and advantages of this invention clearer, the technical solutions of this invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of this invention. All other embodiments obtained by those skilled in the art based on the embodiments of this invention without creative effort are within the scope of protection of this invention.

[0030] The following is combined with Figure 1 This invention describes the original wireless signal target recognition method based on model fusion decision-making provided by the present invention. The original wireless signal target recognition method based on model fusion decision-making provided by the present invention can be applied to the detection and tracking of non-cooperative targets and the recognition of original wireless signal targets in the intelligent sensing of wireless signals. The execution subject of this method can be an electronic device or an original wireless signal target recognition device based on model fusion decision-making installed in the electronic device. The original wireless signal target recognition device based on model fusion decision-making can be implemented by software, hardware or a combination of both. Figure 1 This is one of the flowcharts illustrating the original wireless signal target recognition method based on model fusion decision provided by the present invention, such as... Figure 1 As shown, the method includes the following steps: 101, 102, 103, 104 and 105.

[0031] Step 101: Obtain raw wireless signal data.

[0032] In this step, the raw wireless signal data refers to the raw signal data collected from multiple channels or antenna groups; this embodiment does not limit this.

[0033] Specifically, during the acquisition process, the raw wireless signal data can be obtained through front-end devices such as high-speed serial computer extended bus standard (peripheral component interconnect express, PCIe) acquisition boards, software radio, and spectrum receivers within a continuous time window. At the same time, the current center frequency, channel number, and sampling time of the raw wireless signal data are recorded synchronously. This embodiment does not limit this.

[0034] Step 102: Input the original wireless signal data into the signal perception and preprocessing layer of the signal target recognition model to obtain the dimensional representation information output by the signal perception and preprocessing layer.

[0035] In this step, the signal sensing and preprocessing layer is a layer that transforms and processes the raw wireless signal data; the dimensional representation information includes signal representation data of at least one dimension.

[0036] The transformation process includes at least one of the following: short-time Fourier transform processing, power spectral density estimation processing, normalization processing, rotation processing, resampling processing, and tensor splicing processing. This embodiment does not limit the specific processing method.

[0037] Dimensional representation information may include, for example, time-frequency plots, amplitude plots, complex tensors, or multi-channel feature maps, as limited in this embodiment.

[0038] Specifically, after acquiring the original wireless signal data, the original wireless signal data is preprocessed. The preprocessing specifically involves transforming the original wireless signal. The transformation process includes at least one of the following: short-time Fourier transform, power spectral density estimation, normalization, rotation, resampling, and tensor splicing, thereby obtaining dimensional representation information of at least one dimension after transformation.

[0039] For example, if the original wireless signal data acquired from multiple channels is dimensionally transformed to obtain a time-frequency diagram, and then the time-frequency diagram is reorganized into a two-dimensional matrix, and the two-dimensional matrix is ​​transformed to obtain the transformed signal characterization data, the signal characterization data can be image or tensor data that can be used for signal target recognition models. This embodiment does not limit this.

[0040] Step 103: Input the signal representation data of each signal in the dimension representation information into the parallel detection and feature extraction layer in the signal target recognition model to obtain the structured detection evidence corresponding to each signal representation data output by the parallel detection and feature extraction layer.

[0041] In this step, the parallel detection and feature extraction layer is a layer that performs basic detection and feature extraction on the dimensional representation information.

[0042] Specifically, multiple signal representation data from the dimensional representation information at the same time are input into the parallel detection and feature extraction layer in the signal target recognition model. Parallel detection processing is performed based on the parallel detection and feature extraction layer in the signal target recognition model to obtain the structured detection evidence corresponding to each signal representation data output by the parallel detection and feature extraction layer.

[0043] In one specific embodiment, the signal representation data of each signal in the dimensional representation information is input into the parallel detection and feature extraction layer of the signal target recognition model to obtain the structured detection evidence corresponding to each signal representation data output by the parallel detection and feature extraction layer. This includes: inputting the signal representation data of each signal in the dimensional representation information into the basic detector pool of the parallel detection and feature extraction layer of the signal target recognition model to obtain the signal detection result corresponding to each signal representation data output by the basic detector pool; wherein, the signal detection result includes at least one of candidate target box, category label, confidence level, center frequency, bandwidth range, time position, intensity score, and protocol type; and inputting the signal detection result corresponding to each signal representation data into the evidence extraction module of the parallel detection and feature extraction layer of the signal target recognition model to obtain the structured detection evidence corresponding to each signal representation data input by the evidence extraction module.

[0044] In this step, the signal detection results include at least one of the following: candidate target box, category label, confidence level, center frequency, bandwidth range, time location, strength score, and protocol type. This embodiment does not limit these.

[0045] Specifically, the signal representation data of each signal representation data in the dimensional representation information is input into the basic detector pool in the parallel detection and feature extraction layer of the signal target recognition model. Based on the basic detector pool, basic detection is performed on each signal representation data to obtain the signal detection results corresponding to each signal representation data output by the basic detector pool. The signal detection results corresponding to each signal representation data are then input into the evidence extraction module in the parallel detection and feature extraction layer of the signal target recognition model. Based on the evidence extraction module, evidence is extracted from the signal detection results corresponding to each signal representation data to obtain the structured detection evidence corresponding to each signal representation data input by the evidence extraction module.

[0046] In one specific embodiment, the basic detector pool in the parallel detection and feature extraction layer includes at least one signal detector; the signal detector corresponds one-to-one with each signal representation data in the dimensional representation information; inputting each signal representation data in the dimensional representation information into the basic detector pool in the parallel detection and feature extraction layer of the signal target recognition model to obtain the signal detection result corresponding to each signal representation data output by the basic detector pool includes: inputting each signal representation data in the dimensional representation information into each signal detector in the basic detector pool of the parallel detection and feature extraction layer of the signal target recognition model in parallel to obtain the signal detection result corresponding to each signal representation data output by each signal detector.

[0047] In this step, the signal detector may be, for example, the YOLO (You Only Look Once, a deep learning framework for object detection) series, convolutional neural networks, attention-based Transformer (a deep learning model architecture based on self-attention mechanism for processing sequential data) network, temporal convolutional network, recurrent neural network, and contrastive learning detection network, etc. This embodiment does not limit the type of detector.

[0048] The signal detector is not limited to processing UAV radio frequency signals, but can also be used to process industrial wireless network signals, abnormal electromagnetic signals, remote control link signals, wireless telemetry signals or other signals that can be constructed into image / tensor representations. This embodiment does not limit this.

[0049] Specifically, the signal representation data of each dimension representation information are input in parallel into each signal detector in the basic detector pool of the parallel detection and feature extraction layer of the signal target recognition model. Each signal detector performs detection on each input signal representation data one by one, and obtains the signal detection result corresponding to each signal representation data output by each signal detector.

[0050] For example, for a signal detector whose output signal detection result is a candidate target box, the frequency band range and duration of the signal characterization data can be inferred from the coordinates of the detected candidate target box on the time-frequency graph. This embodiment does not limit this.

[0051] In one specific embodiment, the signal detector is further subjected to half-precision inference, batch inference, asynchronous inference, or multi-device distributed inference.

[0052] Specifically, different signal detectors can process the same input simultaneously, or they can process inputs from different channels or with different representations. This embodiment does not limit this.

[0053] The advantage of this setup is that it balances real-time performance.

[0054] In one specific embodiment, the signal detection results corresponding to each signal characterization data are input into the evidence extraction module in the parallel detection and feature extraction layer of the signal target recognition model to obtain the structured detection evidence corresponding to each signal characterization data input to the evidence extraction module. This includes: inputting the signal detection results corresponding to each signal characterization data into the evidence extraction module in the parallel detection and feature extraction layer of the signal target recognition model to obtain the initial structured detection evidence corresponding to each signal detection result output by the evidence identification unit; and inputting the initial structured detection evidence corresponding to each signal detection result into the evidence normalization unit in the evidence extraction module in the parallel detection and feature extraction layer of the signal target recognition model to obtain the structured detection evidence corresponding to each signal detection result output by the evidence normalization unit.

[0055] Specifically, after obtaining the signal detection results corresponding to each signal representation data, the signal detection results corresponding to each signal representation data are input into the evidence extraction module of the evidence extraction module in the parallel detection and feature extraction layer of the signal target recognition model. Based on the evidence identification unit, structured detection evidence is extracted from each signal detection result to obtain the initial structured detection evidence corresponding to each signal detection result output by the evidence identification unit. The initial structured detection evidence corresponding to each signal detection result is then input into the evidence normalization unit in the evidence extraction module of the parallel detection and feature extraction layer of the signal target recognition model. Based on the evidence normalization unit, all the extracted initial structured detection evidence is normalized to obtain the structured detection evidence corresponding to each signal detection result output by the evidence normalization unit.

[0056] For example, if the signal detection result output by the signal detector is a target box on the time-frequency graph, the corresponding region data is extracted according to the position of the target box, and the center frequency, start and end frequency bands, duration and intensity characteristics corresponding to the region are calculated, so as to obtain initial structured detection evidence based on the center frequency, start and end frequency bands, duration and intensity characteristics. This embodiment does not limit this.

[0057] The intensity features are estimated using robust statistical methods to reduce the impact of spike noise and extreme values ​​on the results. For example, the sample values ​​of the target bounding box are sorted, and a portion of low-end and high-end samples are removed before calculating the truncated mean, which is used as an estimate of the power or intensity of the candidate target. This embodiment does not limit this approach.

[0058] For example, a structured detection evidence can be expressed as structured detection evidence e={t, c, l, b, s, q, o}, where t represents time information, c represents channel / antenna / sensor identifier, l represents category or protocol label, b represents frequency band or time-frequency region parameter, s represents intensity feature, q represents detection confidence, and o represents optional azimuth angle, position or other auxiliary observations. This embodiment does not limit this.

[0059] Step 104: Input all structured detection evidence into the intelligent fusion and decision layer of the signal target recognition model to obtain the fusion decision result output by the intelligent fusion and decision layer.

[0060] In this step, the intelligent fusion and decision-making layer is the layer that performs fusion decisions on the signal detection results; Specifically, after obtaining structured detection evidence, all structured detection evidence is input into the intelligent fusion and decision layer of the signal target recognition model. Based on the probabilistic fusion decision of the structured detection evidence, the intelligent fusion and decision layer outputs the fusion decision result.

[0061] In one specific embodiment, before inputting all structured detection evidence into the intelligent fusion and decision layer of the signal target recognition model and obtaining the fusion decision result output by the intelligent fusion and decision layer, the method further includes: performing a unified standardized mapping on all structured detection evidence to obtain a standardized message structure set.

[0062] In this step, the standardized message structure set includes at least one standardized message structure. The standardized message structure may include, for example, task type, task version, global alarm valid bit, global timestamp, center frequency or global frequency information, as well as the number of detections organized by channel / antenna, frequency band list, intensity list, category list, device type identifier and reserved fields. This embodiment does not limit these.

[0063] Specifically, before inputting all structured detection evidence into the intelligent fusion and decision layer of the signal target recognition model and obtaining the fusion decision result output by the intelligent fusion and decision layer, all structured detection evidence is uniformly standardized and mapped to obtain a set of standardized message structures. The set of standardized message structures includes at least one standardized message structure, which may include, for example, task type, task version, global alarm valid bit, global timestamp, center frequency or global frequency information, as well as the number of detections organized by channel / antenna, frequency band list, intensity list, category list, device type identifier and reserved fields. This embodiment does not limit these.

[0064] In one specific embodiment, channel mapping can also be performed on the structured detection evidence of multiple channels, corresponding the logical channel number to the actual antenna number; for each channel, the detection quantity, frequency band start and end position, signal strength, category label and auxiliary attributes are maintained respectively, thereby providing a unified input for subsequent intelligent fusion and decision-making layer. This embodiment does not limit this.

[0065] In one specific embodiment, after obtaining the standardized message structure set, all standardized message structures in the standardized message structure set can be written to a local file, shared memory, network packet, or message queue.

[0066] In one specific embodiment, preferably, JSON (JavaScript Object Notation, a lightweight data exchange format mainly used for transmitting and storing data between different systems) can also be used as an intermediate expression and sent to the intelligent fusion and decision-making layer or external business system via User Datagram Protocol (UDP) or other lightweight communication protocols. UDP is a connectionless transport layer protocol, and this embodiment does not limit it.

[0067] In one specific embodiment, all structured detection evidence is input into the intelligent fusion and decision layer of the signal target recognition model to obtain the fusion decision result output by the intelligent fusion and decision layer, including: determining a candidate target set and a detector responsibility set according to each standardized message structure in the standardized message structure set; wherein, the candidate target set includes at least one candidate target, and the detector responsibility set includes the detector responsibility vector of the signal detector corresponding to each standardized message structure detected, and the detector responsibility vector is used to characterize the relative interpretability of the signal detector for the candidate target; the standardized message result is the result after uniformly standardizing and mapping the structured detection evidence; and obtaining the historical detector posterior responsibility and historical... The detector responsibility transfer matrix is ​​used. Based on the posterior responsibility of historical detectors and the historical detector responsibility transfer matrix, prior prediction is performed on the responsibility vectors of each detector in the detector responsibility set to obtain the predicted responsibility. Likelihood matching is performed between the responsibility vectors of each detector in the detector responsibility set and the candidate targets in the candidate target set to obtain soft association weights. Target state vectors and confidence information of structured detection evidence for each candidate target are obtained. The target state vectors include at least one of the following: frequency position, bandwidth, power, category, existence probability, motion information, and orientation information of the candidate target. A fusion decision is made on the target state vectors based on the predicted responsibility, soft association weights, and confidence information to obtain the fusion decision result.

[0068] In this step, the target state vector includes at least one of the following: frequency position, bandwidth, power, category, existence probability, motion information, and orientation information of the candidate target. This embodiment does not limit this.

[0069] Here, the posterior responsibility of the historical detector refers to the detector responsibility vector of the previous time step relative to the detector responsibility vector of the current time step, and the historical detector responsibility transition matrix refers to the detector responsibility transition matrix of the previous time step. This embodiment does not limit these aspects.

[0070] Specifically, after obtaining the standardized message structure set, for each standardized message structure in the standardized message structure set, a candidate target set and a detector responsibility set are established respectively according to each standardized message structure. The candidate target set includes at least one candidate target, and the detector responsibility set includes the detector responsibility vector of the signal detector corresponding to each standardized message structure. The detector responsibility vector is used to characterize the relative interpretability of the signal detector for the candidate target. The standardized message result is the result after uniformly standardizing and mapping the structured detection evidence. Then, the posterior responsibility of historical detectors and the historical detector responsibility transition matrix are obtained. Based on these, prior predictions are made for each detector responsibility vector in the detector responsibility set. Specifically, this is done using a matrix recursive approach, where the predicted responsibility at the current time is calculated using the posterior responsibility of historical detectors and the historical detector responsibility transition matrix from the previous time step. Next, likelihood matching is performed between each detector responsibility vector in the current time step and each candidate target in the candidate target set to obtain soft association weights. The target state vectors of each candidate target and the confidence information of the structured detection evidence are obtained. Based on the predicted responsibility, soft association weights, and confidence information, a fusion decision is made for the target state vectors, i.e., updating the existence probability, category posterior, fusion frequency band, fusion strength, fusion orientation, and lifecycle state of the candidate targets, thereby obtaining the fusion decision result.

[0071] In one specific embodiment, for each candidate target, there are a target state vector and a detector responsibility vector for the dimensional candidate target. The target state vector includes at least frequency position, bandwidth, power, category, existence probability, and optional motion or orientation information. The detector responsibility vector is used to characterize the relative interpretability of multiple signal detectors for the candidate target, and this embodiment does not limit this.

[0072] In one specific embodiment, the soft association weight is obtained by performing likelihood matching between each detector responsibility vector in the detector responsibility set and each candidate target in the candidate target set. This includes: performing likelihood matching between each detector responsibility vector in the detector responsibility set and each candidate target in the candidate target set to obtain a target observation association graph; and calculating the soft association relationship between each detector responsibility vector and each candidate target based on the target observation association graph to obtain the soft association weight.

[0073] Specifically, the detector responsibility vectors in the detector responsibility set are likely matched with the candidate targets in the candidate target set to obtain the target observation association graph. Then, the soft association relationship between each detector responsibility vector and each candidate target in the target observation association graph is calculated by message passing or iterative normalization method, thereby obtaining the soft association weight between each candidate target and the detector trust vector.

[0074] For example, likelihood matching is performed between each detector responsibility vector in the detector responsibility set and each candidate target in the candidate target set to obtain a target observation association graph. Then, the soft association relationship between each detector responsibility vector and each candidate target in the target observation association graph is calculated by message passing or iterative normalization method, so as to obtain the soft association weight between each candidate target and the detector trust vector. In this process, many-to-many weighted relationships can be established, and closed-loop feedback can be further realized based on the soft association weight after weighted relationship, that is, the detector responsibility vector is corrected, thereby realizing the automation of the signal target recognition model.

[0075] The advantage of this setup is that soft associations do not discard conflict evidence directly in a single update, but instead retain the probabilistic contribution of the standardized message structure.

[0076] In one specific embodiment, after performing likelihood matching between each detector responsibility vector in the detector responsibility set and each candidate target in the candidate target set to obtain soft correlation weights, the method further includes: performing reverse updating of each detector responsibility vector in the detector responsibility set according to the soft correlation weights, and continuing to perform prior prediction of each detector responsibility vector in the detector responsibility set based on the reverse-updated detector responsibility vectors to obtain predicted responsibility.

[0077] Specifically, after performing likelihood matching between each detector responsibility vector in the detector responsibility set and each candidate target in the candidate target set to obtain soft correlation weights, the detector responsibility vectors in the detector responsibility set are updated in reverse according to the soft correlation weights. Based on the updated detector responsibility vectors, the process continues to perform prior prediction on each detector responsibility vector in the detector responsibility set according to the historical detector posterior responsibility and the historical detector responsibility transition matrix to obtain the predicted responsibility.

[0078] For example, if the observation output by a signal detector at the current moment has a higher consistency with a candidate target in terms of frequency, intensity, category, orientation, and temporal continuity, then the posterior responsibility of the signal detector for that candidate target is increased; otherwise, the responsibility is decreased. This forms a closed-loop linkage between the signal detector and the candidate target, but this embodiment does not limit this.

[0079] In one specific embodiment, the technical means for updating the existence probability, category posterior, fusion frequency band, fusion strength, fusion orientation, and life cycle state of candidate targets can be, for example, Gaussian approximation for linear or near-linear scenarios, or particle method, sampling method, or other approximate inference methods for nonlinear scenarios. This embodiment does not limit these methods.

[0080] Step 105: Input the fusion decision result into the result output layer of the signal target recognition model to obtain the target recognition result of the original wireless signal data output by the result output layer.

[0081] In this step, the fusion decision results may include, for example, the probability of the target's existence, the posterior probability of the category, the duration threshold, the spatial / frequency domain continuity, and business rules, etc. This embodiment does not limit these.

[0082] The target identification results may include, for example, whether a target exists, the target category, the confidence level, the frequency band range, the intensity, the source channel, the timestamp, and optional trajectory or orientation information, etc. This embodiment does not limit these.

[0083] Specifically, after obtaining the fusion decision result, the fusion decision result is input into the result output layer of the signal target recognition model. Target-level alarm recognition is performed based on the result output layer, thereby obtaining the target recognition result of the original wireless signal data output by the result output layer.

[0084] For example, regarding the target existence probability in the fusion decision result, when the target existence probability is determined to be higher than a set threshold (the set threshold is a pre-set threshold used to judge the target existence probability, which is not limited in this embodiment) and the duration meets the business conditions (the business conditions are also a pre-set business duration, which is not limited in this embodiment), the target identification result is output. At this time, the target identification result includes the alarm validity bit, alarm time, target category, frequency band range, power value, source channel and version number, and is sent to the host computer via UDP. This embodiment does not limit this.

[0085] In one specific embodiment, candidate targets can also be managed through generation, confirmation, maintenance, merging, splitting, and termination. Candidate targets that appear briefly and lack sustained support are suppressed to reduce false alarms; candidate targets that exist stably across time periods and continuously receive multi-model support are confirmed to improve alarm reliability.

[0086] In one specific embodiment, when some signal detectors are missing, communication is interrupted, or computing resources are insufficient, the target recognition model is allowed to degenerate into a few-model mode or a single-model mode to continue operating. That is, the remaining signal detectors detect one-dimensional dimensional representation information, which continues to serve as input to the signal target recognition model. At this time, the intelligent fusion and decision layer automatically adjusts the dimension and update path of the detector responsibility vector to ensure that the target recognition model has fault tolerance and availability.

[0087] In one specific embodiment, Figure 2 This is the second flowchart of the original wireless signal target recognition method based on model fusion decision provided by the present invention, as shown below. Figure 2 As shown, it includes steps 201, 202, 203, 204, 205, 206 and 207.

[0088] Step 201: Obtain raw wireless signal data.

[0089] Step 202: Input the original wireless signal data into the signal perception and preprocessing layer of the signal target recognition model to obtain the dimensional representation information output by the signal perception and preprocessing layer.

[0090] Specifically, after acquiring the original wireless signal data, the original wireless signal data is preprocessed. The preprocessing specifically involves transforming the original wireless signal. The transformation process includes at least one of the following: short-time Fourier transform, power spectral density estimation, normalization, rotation, resampling, and tensor splicing, thereby obtaining dimensional representation information of at least one dimension after transformation.

[0091] Step 203: Input the signal representation data of each signal in the dimension representation information into the basic detector pool in the parallel detection and feature extraction layer of the signal target recognition model, and obtain the signal detection results corresponding to each signal representation data output by the basic detector pool.

[0092] Specifically, the signal representation data of each dimension representation information are input in parallel into each signal detector in the basic detector pool of the parallel detection and feature extraction layer of the signal target recognition model. Each signal detector performs detection on each input signal representation data one by one, and obtains the signal detection result corresponding to each signal representation data output by each signal detector.

[0093] For example, the basic detector pool includes three independent signal detectors, which are used to perform protocol saliency detection, weak signal supplementation detection and abnormal pattern recognition on each input signal representation data, respectively. This embodiment does not limit this.

[0094] The basic detector pool may also include three independent signal detectors, namely a first type signal detector, a second type signal detector, and a third type signal detector. The first type detector is used to detect known protocol categories, the second type detector is used to discover unknown or abnormal signal segments, and the third type detector is used to estimate auxiliary attributes, such as direction, position, modulation method, or transmission source type. This embodiment does not limit this.

[0095] Step 204: Input the signal detection results corresponding to each signal representation data into the evidence extraction module in the parallel detection and feature extraction layer of the signal target recognition model to obtain the structured detection evidence corresponding to each signal representation data input to the evidence extraction module.

[0096] Specifically, after obtaining the signal detection results corresponding to each signal representation data, the signal detection results corresponding to each signal representation data are input into the evidence extraction module of the evidence extraction module in the parallel detection and feature extraction layer of the signal target recognition model. Based on the evidence identification unit, structured detection evidence is extracted from each signal detection result to obtain the initial structured detection evidence corresponding to each signal detection result output by the evidence identification unit. The initial structured detection evidence corresponding to each signal detection result is then input into the evidence normalization unit in the evidence extraction module of the parallel detection and feature extraction layer of the signal target recognition model. Based on the evidence normalization unit, all the extracted initial structured detection evidence is normalized to obtain the structured detection evidence corresponding to each signal detection result output by the evidence normalization unit.

[0097] Step 205: Perform a unified standardization mapping on all structured detection evidence to obtain a standardized message structure set.

[0098] Specifically, before obtaining the fusion decision result output by the intelligent fusion and decision layer in the signal target recognition model by inputting all structured detection evidence into the intelligent fusion and decision layer, all structured detection evidence is uniformly standardized and mapped to obtain a standardized message structure set.

[0099] Step 206: Input all structured detection evidence into the intelligent fusion and decision layer of the signal target recognition model to obtain the fusion decision result output by the intelligent fusion and decision layer.

[0100] Specifically, after obtaining the standardized message structure set, for each standardized message structure in the standardized message structure set, a candidate target set and a detector responsibility set are established respectively according to each standardized message structure. The candidate target set includes at least one candidate target, and the detector responsibility set includes the detector responsibility vector of the signal detector corresponding to each standardized message structure. The detector responsibility vector is used to characterize the relative interpretability of the signal detector for the candidate target. The standardized message result is the result after uniformly standardizing and mapping the structured detection evidence. Then, the posterior responsibility of historical detectors and the historical detector responsibility transition matrix are obtained. Based on these, prior predictions are made for each detector responsibility vector in the detector responsibility set. Specifically, this is done using a matrix recursive approach, where the predicted responsibility at the current moment is calculated using the posterior responsibility of historical detectors and the historical detector responsibility transition matrix from the previous moment. Next, likelihood matching is performed between each detector responsibility vector in the current moment's detector responsibility set and each candidate target in the candidate target set to obtain soft association weights. The target state vectors of each candidate target and the confidence information of the structured detection evidence are obtained. Based on the predicted responsibility, soft association weights, and confidence information, a fusion decision is made for the target state vectors, i.e., updating the existence probability, category posterior, fusion band, fusion strength, fusion orientation, and lifecycle state of the candidate targets, thereby obtaining the fusion decision result.

[0101] Step 207: Input the fusion decision result into the result output layer of the signal target recognition model to obtain the target recognition result of the original wireless signal data output by the result output layer.

[0102] Specifically, after obtaining the fusion decision result, the fusion decision result is input into the result output layer of the signal target recognition model. Target-level alarm recognition is performed based on the result output layer, thereby obtaining the target recognition result of the original wireless signal data output by the result output layer.

[0103] Figure 3 This is the third flowchart of the original wireless signal target recognition method based on model fusion decision provided by the present invention, as shown below. Figure 3As shown, obtaining the fusion decision result includes steps 301, 302, 303, 304, 305, 306, and 307.

[0104] Step 301: For each standardized message structure in the standardized message structure set, establish a candidate target set and a detector responsibility set according to each standardized message structure.

[0105] Specifically, after obtaining the standardized message structure set, for each standardized message structure in the standardized message structure set, a candidate target set and a detector responsibility set are established respectively according to each standardized message structure. The candidate target set includes at least one candidate target, and the detector responsibility set includes the detector responsibility vector of the signal detector corresponding to each standardized message structure. The detector responsibility vector is used to characterize the relative interpretability of the signal detector for the candidate target. The standardized message result is the result after uniformly standardizing and mapping the structured detection evidence.

[0106] Step 302: Obtain the posterior responsibility of the historical detector and the historical detector responsibility transition matrix.

[0107] Specifically, the posterior responsibility of the historical detector refers to the detector responsibility vector of the previous time step relative to the detector responsibility vector of the current time step, and the historical detector responsibility transition matrix refers to the detector responsibility transition matrix of the previous time step. This embodiment does not limit these aspects.

[0108] Step 303: Based on the posterior responsibility of historical detectors and the responsibility transition matrix of historical detectors, perform prior prediction on the responsibility vector of each detector in the detector responsibility set to obtain the predicted responsibility.

[0109] Specifically, the posterior responsibility of historical detectors and the historical detector responsibility transition matrix are obtained; based on the posterior responsibility of historical detectors and the historical detector responsibility transition matrix, the responsibility vector of each detector in the detector responsibility set is predicted in advance. Specifically, the update is performed in a matrix recursive form, that is, the predicted responsibility of the current moment is calculated using the posterior responsibility of historical detectors and the historical detector responsibility transition matrix of the previous moment.

[0110] Step 304: Perform likelihood matching between each detector responsibility vector in the detector responsibility set and each candidate target in the candidate target set to obtain the soft association weights.

[0111] Specifically, the detector responsibility vectors in the detector responsibility set are likely matched with the candidate targets in the candidate target set to obtain the target observation association graph. Then, the soft association relationship between each detector responsibility vector and each candidate target in the target observation association graph is calculated by message passing or iterative normalization method, thereby obtaining the soft association weight between each candidate target and the detector trust vector.

[0112] Step 305: Update the detector responsibility vector in the detector responsibility set in reverse according to the soft correlation weight.

[0113] Specifically, after performing likelihood matching between each detector responsibility vector in the detector responsibility set and each candidate target in the candidate target set to obtain soft correlation weights, the detector responsibility vectors in the detector responsibility set are updated in reverse posteriorly based on the soft correlation weights. Then, based on the updated detector responsibility vectors, the process continues to perform prior prediction on each detector responsibility vector in the detector responsibility set according to the historical detector posterior responsibility and the historical detector responsibility transition matrix to obtain the predicted responsibility step 303.

[0114] Step 306: Obtain the target state vector and confidence information of structured detection evidence for each candidate target.

[0115] Step 307: Perform a fusion decision on the target state vector based on the prediction responsibility, soft association weight, and confidence information to obtain the fusion decision result.

[0116] Specifically, the target state vector is fused based on the predictive responsibility, soft association weight, and confidence information. This involves updating the existence probability, category posterior, fusion frequency band, fusion strength, fusion orientation, and lifecycle state of the candidate target to obtain the fusion decision result.

[0117] The advantage of this setup is that, during the process of obtaining the fusion decision results, responsibility prediction, soft association, and posterior update are performed, which makes the target recognition model have good transferability and scalability.

[0118] This invention provides a method for identifying raw wireless signal targets based on model fusion decision-making. The method involves: acquiring raw wireless signal data; inputting the raw wireless signal data into a signal perception and preprocessing layer of a signal target identification model to obtain dimensional representation information output by the signal perception and preprocessing layer; wherein the signal perception and preprocessing layer is a layer that transforms and processes the raw wireless signal data; the dimensional representation information includes signal representation data of at least one dimension; inputting each signal representation data in the dimensional representation information into a parallel detection and feature extraction layer of the signal target identification model to obtain structured detection evidence corresponding to each signal representation data output by the parallel detection and feature extraction layer; wherein the parallel detection and feature extraction layer is a layer that performs basic detection and feature extraction on the dimensional representation information; inputting all structured detection evidence into an intelligent fusion and decision-making layer of the signal target identification model to obtain a fusion decision result output by the intelligent fusion and decision-making layer; wherein the intelligent fusion and decision-making layer is a layer that performs fusion decision-making on the signal detection results; and inputting the fusion decision result into a result output layer of the signal target identification model to obtain the target identification result of the raw wireless signal data output by the result output layer. The technical solution of this invention addresses the shortcomings of existing technologies where a single model for signal recognition is too one-sided, failing to comprehensively identify signals and thus reducing accuracy. By performing multi-dimensional transformation on the original wireless signal and parallel detection processing on the transformed multi-dimensional signal representation data, highly robust signal detection is achieved. Then, probabilistic fusion decision-making is performed based on the structured detection evidence from the parallel detection, ultimately obtaining the target identification result of the original wireless signal, thereby improving signal recognition accuracy. In the above embodiments, the technical solution of this invention addresses the shortcomings of existing technologies where a single model for signal recognition is too one-sided, failing to comprehensively identify signals and thus reducing accuracy. By performing multi-dimensional transformation on the original wireless signal and parallel detection processing on the transformed multi-dimensional signal representation data, highly robust signal detection is achieved. Then, probabilistic fusion decision-making is performed based on the structured detection evidence from the parallel detection, ultimately obtaining the target identification result of the original wireless signal, thereby improving signal recognition accuracy.

[0119] The original wireless signal target recognition device based on model fusion decision provided by the present invention will be described below. The original wireless signal target recognition device based on model fusion decision described below can be referred to in correspondence with the original wireless signal target recognition method based on model fusion decision described above.

[0120] Figure 4 This is a schematic diagram of the original wireless signal target identification device based on model fusion decision provided by the present invention, with reference to... Figure 4As shown, the raw wireless signal target recognition device 400 based on model fusion decision-making includes: a data acquisition module 401, a conversion processing module 402, an evidence detection module 403, a fusion decision-making module 404, and a result recognition module 405; wherein, The data acquisition module 401 is used to acquire raw wireless signal data.

[0121] The conversion processing module 402 is used to input the original wireless signal data into the signal perception and preprocessing layer in the signal target recognition model to obtain the dimensional representation information output by the signal perception and preprocessing layer; wherein, the signal perception and preprocessing layer is a layer that performs conversion processing on the original wireless signal data; the dimensional representation information includes signal representation data of at least one dimension.

[0122] The evidence detection module 403 is used to input the signal representation data of each signal in the dimensional representation information into the parallel detection and feature extraction layer in the signal target recognition model to obtain the structured detection evidence corresponding to each signal representation data output by the parallel detection and feature extraction layer; wherein, the parallel detection and feature extraction layer is a layer that performs basic detection and feature extraction on the dimensional representation information.

[0123] The fusion decision module 404 is used to input all structured detection evidence into the intelligent fusion and decision layer of the signal target recognition model to obtain the fusion decision result output by the intelligent fusion and decision layer; wherein, the intelligent fusion and decision layer is the layer that performs fusion decision on the signal detection results.

[0124] The result recognition module 405 is used to input the fused decision result into the result output layer of the signal target recognition model to obtain the target recognition result of the original wireless signal data output by the result output layer.

[0125] In one example embodiment, the transformation process includes at least one of short-time Fourier transform processing, power spectral density estimation processing, normalization processing, rotation processing, resampling processing, and tensor splicing processing.

[0126] In one example embodiment, the evidence detection module 403 is specifically used to: input the signal representation data of each signal in the dimension representation information into the basic detector pool in the parallel detection and feature extraction layer of the signal target recognition model, and obtain the signal detection result corresponding to each signal representation data output by the basic detector pool; input the signal detection result corresponding to each signal representation data into the evidence extraction module in the parallel detection and feature extraction layer of the signal target recognition model, and obtain the structured detection evidence corresponding to each signal representation data input by the evidence extraction module.

[0127] In one example embodiment, the base detector pool in the parallel detection and feature extraction layer includes at least one signal detector; the signal detector corresponds one-to-one with each signal representation data in the dimensional representation information.

[0128] In one example embodiment, the evidence detection module 403 inputs each signal representation data in the dimensional representation information into the basic detector pool in the parallel detection and feature extraction layer of the signal target recognition model to obtain the signal detection results corresponding to each signal representation data output by the basic detector pool. Specifically, it is used to: input each signal representation data in the dimensional representation information into each signal detector in the basic detector pool of the parallel detection and feature extraction layer of the signal target recognition model in parallel to obtain the signal detection results corresponding to each signal representation data output by each signal detector; wherein, the signal detection results include at least one of candidate target boxes, category labels, confidence scores, center frequencies, bandwidth ranges, time positions, intensity scores, and protocol types.

[0129] In one example embodiment, the evidence detection module 403 inputs the signal detection results corresponding to each signal representation data into the evidence extraction module in the parallel detection and feature extraction layer of the signal target recognition model to obtain the structured detection evidence corresponding to each signal representation data input by the evidence extraction module. Specifically, it is used to: input the signal detection results corresponding to each signal representation data into the evidence recognition unit in the evidence extraction module of the parallel detection and feature extraction layer of the signal target recognition model to obtain the initial structured detection evidence corresponding to each signal detection result output by the evidence recognition unit; and input the initial structured detection evidence corresponding to each signal detection result into the evidence normalization unit in the evidence extraction module of the parallel detection and feature extraction layer of the signal target recognition model to obtain the structured detection evidence corresponding to each signal detection result output by the evidence normalization unit.

[0130] In one example embodiment, the device further includes a standardization mapping module. The standardization mapping module is used to: perform a unified standardization mapping on all structured detection evidence before obtaining the fusion decision result output by the intelligent fusion and decision layer in the signal target recognition model, thereby obtaining a standardized message structure set.

[0131] In one example embodiment, the fusion decision module 404 is specifically configured to: determine a candidate target set and a detector responsibility set based on each standardized message structure in the standardized message structure set; wherein, the candidate target set includes at least one candidate target, and the detector responsibility set includes detector responsibility vectors of the signal detectors corresponding to each standardized message structure detected, the detector responsibility vectors being used to characterize the relative interpretability of the signal detectors for the candidate targets; the standardized message result is the result after uniformly standardizing and mapping the structured detection evidence; obtain the historical detector posterior responsibility and the historical detector responsibility transition matrix; and, based on the historical detector posterior responsibility... The responsibility degree of the detectors and the historical detector responsibility degree transition matrix are used to make prior predictions on the responsibility degree vectors of each detector in the detector responsibility degree set to obtain the predicted responsibility degree. The responsibility degree vectors of each detector in the detector responsibility degree set are then subjected to likelihood matching with each candidate target in the candidate target set to obtain the soft association weight. The target state vectors of each candidate target and the confidence information of the structured detection evidence are obtained. The target state vector includes at least one of the following: frequency position, bandwidth, power, category, existence probability, motion information and orientation information of the candidate target. The target state vectors are then fused and decided based on the predicted responsibility degree, the soft association weight and the confidence information to obtain the fusion decision result.

[0132] In one example embodiment, the fusion decision module 404 performs likelihood matching between each detector responsibility vector in the detector responsibility set and each candidate target in the candidate target set to obtain soft association weights. Specifically, it performs likelihood matching between each detector responsibility vector in the detector responsibility set and each candidate target in the candidate target set to obtain a target observation association graph; and calculates the soft association relationship between each detector responsibility vector and each candidate target based on the target observation association graph to obtain soft association weights.

[0133] In one example embodiment, the apparatus further includes a reverse update module. The reverse update module is configured to: after performing likelihood matching between each detector responsibility vector in the detector responsibility set and each candidate target in the candidate target set to obtain soft correlation weights, perform a reverse update on each detector responsibility vector in the detector responsibility set based on the soft correlation weights, and continue to perform a step of prior prediction on each detector responsibility vector in the detector responsibility set based on the reverse-updated detector responsibility vectors to obtain predicted responsibility.

[0134] The apparatus of this embodiment can be used to execute the method of any embodiment in the side embodiment of the original wireless signal target recognition method based on model fusion decision. Its specific implementation process and technical effects are similar to those in the side embodiment of the original wireless signal target recognition method based on model fusion decision. For details, please refer to the detailed description in the side embodiment of the original wireless signal target recognition method based on model fusion decision, which will not be repeated here.

[0135] Figure 5 This is a schematic diagram of the structure of the electronic device provided by the present invention, such as... Figure 5 As shown, the electronic device may include: a processor 510, a communications interface 520, a memory 530, and a communications bus 540, wherein the processor 510, the communications interface 520, and the memory 530 communicate with each other through the communications bus 540. The processor 510 can call logic instructions in the memory 530 to execute a raw wireless signal target recognition method based on model fusion decision-making. This method includes: acquiring raw wireless signal data; inputting the raw wireless signal data into a signal perception and preprocessing layer in a signal target recognition model to obtain dimensional representation information output by the signal perception and preprocessing layer; wherein the signal perception and preprocessing layer is a layer that performs transformation processing on the raw wireless signal data; the dimensional representation information includes signal representation data of at least one dimension; inputting each signal representation data in the dimensional representation information into a parallel detection and feature extraction layer in the signal target recognition model to obtain structured detection evidence corresponding to each signal representation data output by the parallel detection and feature extraction layer; wherein the parallel detection and feature extraction layer is a layer that performs basic detection and feature extraction on the dimensional representation information; inputting all structured detection evidence into an intelligent fusion and decision layer in the signal target recognition model to obtain a fusion decision result output by the intelligent fusion and decision layer; wherein the intelligent fusion and decision layer is a layer that performs fusion decision-making on the signal detection results; and inputting the fusion decision result into a result output layer in the signal target recognition model to obtain a target recognition result of the raw wireless signal data output by the result output layer.

[0136] Furthermore, the logical instructions in the aforementioned memory 530 can be implemented as software functional units and, when sold or used as independent products, can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention, in essence, or the part that contributes to the prior art, or a part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods of the various embodiments of the present invention. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0137] On the other hand, the present invention also provides a computer program product, which includes a computer program that can be stored on a non-transitory computer-readable storage medium. When the computer program is executed by a processor, the computer is able to execute the original wireless signal target recognition method based on model fusion decision provided by the above methods. The method includes: acquiring original wireless signal data; inputting the original wireless signal data into a signal perception and preprocessing layer in a signal target recognition model to obtain dimensional representation information output by the signal perception and preprocessing layer; wherein, the signal perception and preprocessing layer is a layer that transforms and processes the original wireless signal data; the dimensional representation information includes signal representation data of at least one dimension; and the dimensional representation information is further processed by the signal perception and preprocessing layer. The signal representation data of each dimension representation information is input into the parallel detection and feature extraction layer in the signal target recognition model to obtain the structured detection evidence corresponding to each signal representation data output by the parallel detection and feature extraction layer; wherein, the parallel detection and feature extraction layer is the layer that performs basic detection and feature extraction on the dimension representation information; all structured detection evidence is input into the intelligent fusion and decision layer in the signal target recognition model to obtain the fusion decision result output by the intelligent fusion and decision layer; wherein, the intelligent fusion and decision layer is the layer that performs fusion decision on the signal detection results; the fusion decision result is input into the result output layer in the signal target recognition model to obtain the target recognition result of the original wireless signal data output by the result output layer.

[0138] In another aspect, the present invention also provides a non-transitory computer-readable storage medium storing a computer program thereon, which, when executed by a processor, implements the original wireless signal target recognition method based on model fusion decision provided by the above methods. The method includes: acquiring original wireless signal data; inputting the original wireless signal data into a signal perception and preprocessing layer in a signal target recognition model to obtain dimensional representation information output by the signal perception and preprocessing layer; wherein the signal perception and preprocessing layer is a layer that performs transformation processing on the original wireless signal data; the dimensional representation information includes signal representation data of at least one dimension; and converting each signal representation data in the dimensional representation information... The parallel detection and feature extraction layer in the input signal target recognition model is used to obtain structured detection evidence corresponding to each signal representation data output by the parallel detection and feature extraction layer. The parallel detection and feature extraction layer is a layer that performs basic detection and feature extraction on the dimensional representation information. All structured detection evidence is input into the intelligent fusion and decision layer in the signal target recognition model to obtain the fusion decision result output by the intelligent fusion and decision layer. The intelligent fusion and decision layer is a layer that performs fusion decision on the signal detection results. The fusion decision result is input into the result output layer in the signal target recognition model to obtain the target recognition result of the original wireless signal data output by the result output layer.

[0139] The device embodiments described above are merely illustrative. The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of this embodiment according to actual needs. Those skilled in the art can understand and implement this without any creative effort.

[0140] Through the above description of the embodiments, those skilled in the art can clearly understand that each embodiment can be implemented by means of software plus necessary general-purpose hardware platforms, and of course, it can also be implemented by hardware. Based on this understanding, the above technical solutions, in essence or the part that contributes to the prior art, can be embodied in the form of a software product. This computer software product can be stored in a computer-readable storage medium, such as ROM / RAM, magnetic disk, optical disk, etc., and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute the methods described in the various embodiments or some parts of the embodiments.

[0141] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method for identifying raw wireless signal targets based on model fusion decision-making, characterized in that, include: Acquire raw wireless signal data; The original wireless signal data is input into the signal perception and preprocessing layer of the signal target recognition model to obtain the dimensional representation information output by the signal perception and preprocessing layer; wherein, the signal perception and preprocessing layer is a layer that performs transformation processing on the original wireless signal data; the dimensional representation information includes signal representation data of at least one dimension; The signal representation data in the dimensional representation information are input into the parallel detection and feature extraction layer in the signal target recognition model to obtain the structured detection evidence corresponding to each signal representation data output by the parallel detection and feature extraction layer; wherein, the parallel detection and feature extraction layer is a layer that performs basic detection and feature extraction on the dimensional representation information; All the structured detection evidence is input into the intelligent fusion and decision layer of the signal target recognition model to obtain the fusion decision result output by the intelligent fusion and decision layer; wherein, the intelligent fusion and decision layer is the layer that performs fusion decision on the signal detection result; The fusion decision result is input into the result output layer of the signal target recognition model to obtain the target recognition result of the original wireless signal data output by the result output layer.

2. The original wireless signal target identification method based on model fusion decision-making according to claim 1, characterized in that, The transformation process includes at least one of the following: short-time Fourier transform, power spectral density estimation, normalization, rotation, resampling, and tensor splicing.

3. The original wireless signal target identification method based on model fusion decision-making according to claim 1, characterized in that, The step of inputting each of the signal representation data in the dimensional representation information into the parallel detection and feature extraction layer in the signal target recognition model to obtain the structured detection evidence corresponding to each of the signal representation data output by the parallel detection and feature extraction layer includes: The signal representation data in the dimensional representation information are input into the basic detector pool in the parallel detection and feature extraction layer of the signal target recognition model to obtain the signal detection result corresponding to each signal representation data output by the basic detector pool; wherein, the signal detection result includes at least one of candidate target box, category label, confidence, center frequency, bandwidth range, time position, intensity score and protocol type; The signal detection results corresponding to each of the signal characterization data are input into the evidence extraction module in the parallel detection and feature extraction layer of the signal target recognition model to obtain the structured detection evidence corresponding to each of the signal characterization data input by the evidence extraction module.

4. The original wireless signal target identification method based on model fusion decision-making according to claim 3, characterized in that, The basic detector pool in the parallel detection and feature extraction layer includes at least one signal detector; each signal detector corresponds one-to-one with each signal representation data in the dimensional representation information; the step of inputting each signal representation data in the dimensional representation information into the basic detector pool in the parallel detection and feature extraction layer of the signal target recognition model to obtain the signal detection result corresponding to each signal representation data output by the basic detector pool includes: Each of the signal representation data in the dimensional representation information is input in parallel into each of the signal detectors in the basic detector pool of the parallel detection and feature extraction layer of the signal target recognition model, so as to obtain the signal detection result corresponding to each of the signal representation data output by each of the signal detectors.

5. The original wireless signal target identification method based on model fusion decision-making according to claim 3, characterized in that, The step of inputting the signal detection results corresponding to each of the signal characterization data into the evidence extraction module in the parallel detection and feature extraction layer of the signal target recognition model to obtain the structured detection evidence corresponding to each of the signal characterization data input by the evidence extraction module includes: The signal detection results corresponding to each of the signal characterization data are input into the evidence extraction module of the evidence extraction unit in the parallel detection and feature extraction layer of the signal target recognition model to obtain the initial structured detection evidence corresponding to each of the signal detection results output by the evidence recognition unit. The initial structured detection evidence corresponding to each of the signal detection results is input into the evidence normalization unit in the evidence extraction module of the parallel detection and feature extraction layer of the signal target recognition model to obtain the structured detection evidence corresponding to each of the signal detection results output by the evidence normalization unit.

6. The original wireless signal target identification method based on model fusion decision-making according to any one of claims 1-5, characterized in that, Before inputting all the structured detection evidence into the intelligent fusion and decision layer of the signal target recognition model and obtaining the fusion decision result output by the intelligent fusion and decision layer, the method further includes: All the structured detection evidence is uniformly standardized and mapped to obtain a set of standardized message structures.

7. The original wireless signal target recognition method based on model fusion decision-making according to claim 6, characterized in that, The step of inputting all the structured detection evidence into the intelligent fusion and decision layer of the signal target recognition model to obtain the fusion decision result output by the intelligent fusion and decision layer includes: A candidate target set and a detector responsibility set are determined based on each standardized message structure in the standardized message structure set; wherein, the candidate target set includes at least one candidate target, and the detector responsibility set includes detector responsibility vectors of the signal detectors that detected each standardized message structure, the detector responsibility vectors being used to characterize the relative interpretability of the signal detectors for the candidate targets; the standardized message result is the result of uniformly standardizing and mapping the structured detection evidence; Obtain the posterior responsibility of the historical detector and the responsibility transition matrix of the historical detector; Based on the posterior responsibility of the historical detectors and the responsibility transition matrix of the historical detectors, a priori prediction is performed on each of the detector responsibility vectors in the detector responsibility set to obtain the predicted responsibility. The detector responsibility vectors in the detector responsibility set are subjected to likelihood matching with the candidate targets in the candidate target set to obtain soft association weights; Obtain the target state vector of each candidate target and the confidence information of the structured detection evidence; wherein, the target state vector includes at least one of the candidate target's frequency position, bandwidth, power, category, existence probability, motion information and orientation information; The target state vector is fused and a decision is made based on the predicted responsibility, the soft association weight, and the confidence information to obtain the fusion decision result.

8. The original wireless signal target identification method based on model fusion decision-making according to claim 7, characterized in that, The step of performing likelihood matching between each detector responsibility vector in the detector responsibility set and each candidate target in the candidate target set to obtain soft association weights includes: The detector responsibility vectors in the detector responsibility set are likely matched with the candidate targets in the candidate target set to obtain a target observation association graph; The soft association relationship between each detector responsibility vector and each candidate target is calculated based on the target observation association graph to obtain the soft association weight.

9. The method for identifying raw wireless signal targets based on model fusion decision-making according to claim 7, characterized in that, After performing likelihood matching between each detector responsibility vector in the detector responsibility set and each candidate target in the candidate target set to obtain soft association weights, the method further includes: The detector responsibility vectors in the detector responsibility set are updated in reverse according to the soft correlation weights, and the step of performing prior prediction on the detector responsibility vectors in the detector responsibility set based on the historical detector posterior responsibility and the historical detector responsibility transition matrix is ​​continued based on the updated detector responsibility vectors to obtain the predicted responsibility.

10. A raw wireless signal target identification device based on model fusion decision-making, characterized in that, include: The data acquisition module is used to acquire raw wireless signal data; The conversion processing module is used to input the original wireless signal data into the signal perception and preprocessing layer of the signal target recognition model to obtain the dimensional representation information output by the signal perception and preprocessing layer; wherein, the signal perception and preprocessing layer is a layer that performs conversion processing on the original wireless signal data; the dimensional representation information includes signal representation data of at least one dimension. The evidence detection module is used to input the signal representation data of each of the dimensional representation information into the parallel detection and feature extraction layer in the signal target recognition model to obtain the structured detection evidence corresponding to each of the signal representation data output by the parallel detection and feature extraction layer; wherein, the parallel detection and feature extraction layer is a layer that performs basic detection and feature extraction on the dimensional representation information; The fusion decision module is used to input all the structured detection evidence into the intelligent fusion and decision layer of the signal target recognition model to obtain the fusion decision result output by the intelligent fusion and decision layer; wherein, the intelligent fusion and decision layer is the layer that performs fusion decision on the signal detection result; The result recognition module is used to input the fusion decision result into the result output layer of the signal target recognition model to obtain the target recognition result of the original wireless signal data output by the result output layer.

Images