A signal recognition method based on prior knowledge integration and result adaptation

Abstract

A signal recognition method based on prior knowledge integration and result adaptation includes a model construction step, a model application step and a knowledge update step. The present invention constructs a Bayesian network to realize preliminary pre-sorting of a large number of radio signals and improves recognition efficiency; constructs a Bayesian neural network to enrich the cognitive degree and scope of signal modulation recognition; finally, the Bayesian network is updated using the result after modulation recognition to realize iterative integration of prior knowledge and result adaptation of signal recognition in a multi-dimensional environment. The present invention forms a two-level network structure, in which the Bayesian network is used for knowledge integration and updating, and plays a role in preliminary screening, and the Bayesian neural network is used to calculate the signal recognition result and its confidence, breaking through the non-binary conclusion of traditional signal recognition, while ensuring the accuracy of signal screening, improving the speed and robustness of screening.

Description

Technical Field

[0001] The present invention relates to the field of wireless communication technology, and in particular to an automated and intelligent signal recognition method for wireless signals based on prior knowledge integration and result adaptation. Background Art

[0002] In the field of wireless communications, signal modulation recognition technology is crucial to improving spectrum utilization and communication system performance. With the continuous advancement of communication technology, signal patterns are becoming more complex and the demand for analysis is growing exponentially. Traditional statistical-based signal modulation recognition methods face great difficulties in terms of effectiveness and efficiency; and the current mainstream machine learning methods face high challenges in terms of robustness when dealing with complex signal environments and non-cooperative situations.

[0003] Existing signal modulation recognition methods mostly perform statistical characteristic analysis and signal feature extraction from known signals, such as spectrum features, time-frequency features, high-order moments, etc., and accumulate to form a signal feature library. When performing signal recognition, feature matching is performed through classifiers such as support vector machines and decision trees, and then sorting and recognition is performed, or modulation recognition is performed through likelihood-based hypothesis testing. The biggest limitations of the above methods are three points. First, the determining factor of modulation recognition quality is determined by the signal feature library, and signals that are not in the feature library cannot be recognized; second, it has a distinct "yes or no" duality and cannot answer ambiguous questions; third, the automatic processing capability is limited when the number is huge.

[0004] Therefore, how to utilize the accumulative nature of signal recognition to improve the possibility of identifying some uncertain signals, while improving the accuracy of signal recognition and taking into account robustness, and improving the modulation recognition capability of massive signals has become a technical problem that needs to be urgently solved in existing technologies. Summary of the invention

[0005] The purpose of the present invention is to propose a signal recognition method based on prior knowledge integration and result adaptation, which utilizes the knowledge learning accumulation of the Bayesian method and instantly updates the signal feature cognition situation, and has the ability of automatic matching and recognition of signal features through the autonomous intelligent learning architecture of the Bayesian neural network, so as to realize the automatic recognition and processing of a large number of wireless signals under complex conditions.

[0006] A signal recognition method based on prior knowledge integration and result adaptation, comprising:

[0007] Model building step S110:

[0008] Preprocess the signal, extract the signal features, construct an initial feature element set of the signal features, respectively construct a Bayesian network and a Bayesian neural network, set a Bayesian neural network confidence characterization function, set the Bayesian neural network hyperparameters according to the input parameter features, fuse the signal features with the signal as training data and test data, and train and test the Bayesian neural network until the model converges;

[0009] Model application step S120:

[0010] When a new signal appears, the features are collected and compared to obtain a list of key signals, and the confidence levels of the key signals are calculated using a trained Bayesian neural network, and modulation recognition is performed using the confidence levels to obtain new signals and their features;

[0011] Knowledge updating step S130:

[0012] The new signal and its features obtained in the model application step S120 are used to improve the initial feature element set of the signal features, and the Bayesian network is updated using the improved signal features.

[0013] Optionally, the model building step S110 is specifically as follows:

[0014] Signal preprocessing sub-step S111:

[0015] The signal is preprocessed by denoising, filtering, frequency conversion sampling and equalization, and the initial feature extraction is used to construct the initial feature element set T [feature 1, feature 2, ..., feature N] of the signal feature;

[0016] Bayesian network construction sub-step S112:

[0017] Based on the initial element set of signal features, a Bayesian network is constructed to intuitively represent the signal features and their relationships, forming an initial cognitive framework for the signal;

[0018] Bayesian neural network construction sub-step S113:

[0019] A Bayesian neural network is used as a confidence calculation network, wherein the network input layer of the Bayesian network is a one-dimensional convolutional layer, and the confidence characterization function is constructed as follows: ,in, represents the number of input training samples, Take values ​​for each parameter in the feature set, Represents the Bayesian neural network according to the input The calculated output result is, is the value label of the data after passing through the classifier, Z1= , represents the aleatoric uncertainty in the Bayesian network. Specifically, is an estimate of the variance of the model output; Z2 Characterize the confidence that the signal comes from device A, Represents the probability value of a certain item in the sample. The model performs T parameter sampling on the same sample. C represents the number of all different categories that the Bayesian neural network needs to distinguish, Represents the probability value of a certain item in the sample; the T-time prediction probability is ,in , indicating that the W tip in the Bayesian neural network obeys the form The distribution of Represents the activation function of the Bayesian neural network;

[0020] Bayesian neural network training sub-step S114:

[0021] The Bayesian neural network hyperparameters are set according to the input data characteristics, and the feature set T [feature 1, feature 2, ..., feature N] is fused with the signal to generate the Bayesian neural network input TB [original signal, feature 1, feature 2, ..., feature N], and the set with true value or calibration value is extracted as training data and test data to form a test set and a training set. The Bayesian neural network with set parameters is trained using the training set, and the test set is used for testing and comparison until the confidence characterization function is detected. The training loss and validation loss tend to be stable, with fluctuation values ​​less than 10 -4 , the training converges and a more mature radio signal modulation recognition model is obtained.

[0022] Optionally, in the signal preprocessing sub-step S111, the feature is selected from a plurality of the following: center frequency, instantaneous frequency, spectrum, square spectrum, fourth power spectrum, initial phase, signal bandwidth, signal amplitude, envelope and signal-to-noise ratio.

[0023] Optionally, in the Bayesian network construction sub-step S112,

[0024] The Bayesian network is a directed acyclic graph constructed with signal features as nodes, wherein the nodes represent random variables and the edges represent the dependencies between variables. The signal features adopt the features in the initial feature element set T, and the element relationships obtained by analysis are used to build the edges between the nodes to form a cognitive network based on radio signal features.

[0025] Optionally, in the Bayesian neural network training sub-step S114,

[0026] The preferred hyperparameters for simulation experiments and tests are network dimension, output dimension, optimizer, and Bachsize.

[0027] Optionally, a key signal acquisition sub-step S121 is included:

[0028] When a new signal recognition requirement arises, the features of the new signal are obtained, and a list of key signals is obtained through preliminary feature matching based on the initial feature element set T of the signal features;

[0029] Uncertainty calculation sub-step S122:

[0030] Using the trained Bayesian neural network to calculate the key signal list respectively, to obtain the corresponding confidence level;

[0031] Modulation identification sub-step S123:

[0032] Modulation identification is performed on the list of key attention signals in the order of confidence from high to low to confirm the requirements, and after the requirements are met, the signals and their characteristics that successfully meet the requirements are recorded.

[0033] Optionally, in the key signal acquisition sub-step S121,

[0034] The initial feature matching is to determine the consistency of the features based on the inner product value between the features of two signals, and those exceeding the threshold are automatically included in the list of key attention signals.

[0035] Optionally, in the modulation identification sub-step S123,

[0036] The requirements include communication reconstruction, sorting and identification, identity confirmation and signal communication.

[0037] Optionally, the knowledge updating step S130 is specifically as follows:

[0038] The signal element set improvement sub-step S131 is included:

[0039] Using the new signal and its features obtained in the model application step S120, the required features are extracted, and combined with the corresponding confidence conclusions, an improved signal element set is obtained;

[0040] Bayesian network update sub-step S132:

[0041] By using the improved set of signal elements, based on a method similar to the Bayesian network construction sub-step S112, the Bayesian network is updated to obtain an updated cognition of the signal.

[0042] In summary, the present invention has the following advantages:

[0043] 1. A two-level network structure is formed by using Bayesian network and Bayesian neural network, in which Bayesian network is used for knowledge integration and updating, and plays a role in preliminary screening, while Bayesian neural network is used to calculate signal recognition results and their confidence. The secondary network improves the speed and robustness of screening while ensuring the accuracy of signal screening.

[0044] 2. Using the Bayesian network as the primary screening network, each knowledge update can be integrated into the existing information database, which provides an effective way to timely update the signal feature library knowledge and effectively identify signals under new conditions (format camouflage, electromagnetic interference, quantity explosion, etc.).

[0045] 3. Bayesian neural network can obtain a more robust model with less data. It can identify radio signals in the absence of some prior knowledge or with very little initial data, for example, in the case of uneven distribution of prior training data, incomplete information, non-cooperative conditions, interference, complex environments, etc. It is of great significance for the effective identification of radio signals in complex environments.

[0046] 4. The Bayesian neural network gives the confidence level of signal recognition, breaking through the non-binary conclusion of traditional signal recognition, avoiding the misjudgment of key signals under adverse conditions, and providing a reasonable solution for further manual judgment and processing, with strong feasibility and scenario applicability.

[0047] 5. The Bayesian neural network introduces Bayesian theory in the convolution layer, adjusts the network input layer to a one-dimensional convolution layer form, modifies the first layer so that the input feature number matches the one-dimensional data feature number, so that it can accept one-dimensional input and is suitable for the data input of the present invention, so that the Bayesian neural network of the present invention has the ability to learn one-dimensional signal input. BRIEF DESCRIPTION OF THE DRAWINGS

[0048] Figure 1 is a flow chart of a signal recognition method based on prior knowledge integration and result adaptation according to a specific embodiment of the present invention;

[0049] Figure 2 is a specific flow chart of the model building steps according to a specific embodiment of the present invention;

[0050] Figure 3 is a specific flow chart of the model updating steps according to a specific embodiment of the present invention;

[0051] Figure 4 is a specific flow chart of the knowledge updating step according to a specific embodiment of the present invention;

[0052] Figure 5 It is a structural comparison diagram of a general neural network and a Bayesian neural network according to a specific embodiment of the present invention. DETAILED DESCRIPTION

[0053] The present invention will be further described in detail below in conjunction with the accompanying drawings and embodiments. It is to be understood that the specific embodiments described herein are only used to explain the present invention, rather than to limit the present invention. It should also be noted that, for ease of description, only parts related to the present invention, rather than all structures, are shown in the accompanying drawings.

[0054] The present invention mainly lies in: constructing a Bayesian network based on the superposition integration of prior knowledge to realize the preliminary pre-sorting of a large number of radio signals and improve the recognition efficiency; constructing a Bayesian neural network, using which the probability distribution can be input and the characteristics of the probability distribution can be output, enriching the cognitive degree and scope of signal modulation recognition; finally, using the result after modulation recognition to update the Bayesian network, realizing the iterative integration of prior knowledge and the adaptive signal recognition results in a multi-dimensional environment.

[0055] For details, see Figure 1 , shows a flow chart of a signal recognition method based on prior knowledge integration and result adaptation according to a specific embodiment of the present invention, comprising the following steps:

[0056] Model building step S110:

[0057] The signal is preprocessed to extract signal features, an initial feature element set of signal features is constructed, a Bayesian network and a Bayesian neural network are constructed respectively, a Bayesian neural network confidence characterization function is set, the Bayesian neural network hyperparameters are set according to the input parameter features, the signal features are fused with the signal as training data and test data, the Bayesian neural network is trained and tested until the model converges, and the trained Bayesian neural network is used as a radio signal modulation recognition model.

[0058] For details, see Figure 2 , including the following sub-steps:

[0059] Signal preprocessing sub-step S111:

[0060] The signal is preprocessed through denoising, filtering, variable frequency sampling and equalization methods, and preliminary feature extraction is performed to construct the initial feature element set T [feature 1, feature 2, ..., feature N] of the signal features.

[0061] Specifically, the feature is selected from a plurality of the following: center frequency, instantaneous frequency, spectrum, square spectrum, fourth power spectrum, initial phase, signal bandwidth, signal amplitude, envelope and signal-to-noise ratio.

[0062] Bayesian network construction sub-step S112:

[0063] A Bayesian network is constructed based on the initial element set of signal features to intuitively represent the signal features and their relationships, forming an initial cognitive framework for the signal and providing a basis for subsequent knowledge integration and iteration.

[0064] Specifically, in the present invention, the Bayesian network is a directed acyclic graph constructed with signal features as nodes, wherein the nodes represent random variables and the edges represent the dependencies between variables, wherein the signal features adopt the features in the initial feature element set T, and the element relationships obtained by analysis are used to build edges between nodes, thereby forming a cognitive network based on radio signal features.

[0065] The important feature of the present invention is to perform rough identification first and then accurate confirmation. This sub-step is used for rough identification and preliminary feature extraction, so that pre-screening can be performed when a large number of signals appear, thereby improving the recognition efficiency.

[0066] Specifically, the present invention adopts a Bayesian network to provide a basic framework for knowledge learning and updating, so that each time a Class A signal appears, the relevant knowledge can be updated, such as A in an ideal environment, A with interference from a solar storm, and A shielded by a signal blocker, and the knowledge base about signal A is continuously enriched. When the knowledge accumulates to a certain level, no matter what A becomes, it can be recognized.

[0067] Because the model has a simple structure, intuitive representation and strong interpretability, as the signal cognition is updated, the nodes and the relationships between nodes can be easily expanded and deleted, realizing the integrated representation and instant updating of relevant knowledge, providing the knowledge foundation and structural framework for subsequent modeling and solving to obtain more in-depth and accurate cognitive information.

[0068] Bayesian neural network construction sub-step S113:

[0069] The Bayesian neural network is selected mainly to utilize its characteristics of being able to input probability distribution and output probability distribution. For example, it can get the conclusion that "a certain signal is A with 67% confidence, or with 78% confidence in the signal-to-noise ratio interval [10,15]", breaking through the traditional binary conclusion of "a certain signal is A" or "a certain signal is not A", enriching the degree and scope of cognition for signal modulation recognition, and thus improving the robustness of signal modulation recognition.

[0070] See also Figure 5 , which shows the difference between a general neural network and a Bayesian neural network, where the left side is a general neural network and the right side is a Bayesian neural network. It can be seen that the Bayesian neural network can output a probability distribution, which can be used to represent a confidence interval in the present invention.

[0071] In the present invention, the Bayesian neural network is based on the Bayesian prior knowledge framework and combines the autonomous learning characteristics of the neural network to achieve an autonomous learning model that supports probability distribution.

[0072] The Bayesian network of the present invention is used to accumulate and update knowledge, while the Bayesian neural network is used to learn knowledge to perform signal recognition and solve, so as to output probability distribution and use the confidence based on probability distribution to perform modulation recognition respectively until the demand is confirmed. Therefore, the structure and function of the Bayesian network and the Bayesian neural network are different.

[0073] Specifically, a Bayesian neural network is used as a confidence calculation network, wherein the network input layer of the Bayesian network, that is, the first convolution layer, is a one-dimensional convolution layer, and the confidence representation function is constructed as follows: ,in, represents the number of input training samples, Take values ​​for each parameter in the feature set, Represents the model for the input The predicted value, specifically The Bayesian neural network is based on the input The calculated output results are the network's response output to these input features; is the value label of the data after passing through the classifier, which can be expressed as 0, 1, 2, 3, 4 in one embodiment; Z1= ,express Accidental uncertainty, that is, the expression formula of accidental uncertainty in the above Bayesian network, specifically, is an estimate of the variance (i.e., the square of the standard deviation) of the model output, which quantifies the model's accuracy in predicting its values. The confidence level can be used to characterize the noise, fluctuation and other non-human characteristics of the data and measure the quality of the data itself. If this value is large, the data quality itself is poor and the recognition conclusion is greatly affected by the data quality itself. Used to characterize the confidence that the signal comes from device A. Represents the probability value of a certain item in the sample. The model performs T parameter sampling on the same sample. C It represents the total number of categories and is often used in classification problems to measure the probability distribution of model predictions. The difference between the probability distribution of the true label and the true label. Here, represents the probability value of a certain item in the sample, then C It is the number of all different categories that the model needs to distinguish. The T-time prediction probability is ,in , indicating that the W tip in the Bayesian neural network obeys the form distribution. represents the activation function of the Bayesian neural network, which is used to convert the output of the model into a probability distribution. Specifically, for a given inputx , the model should actually be a real number vector, but the activation function converts it into a probability vector, generally ,in is the i-th element of the model output vector and C is the total number of categories.

[0074] In this step, the traditional two-dimensional convolutional layer of the Bayesian neural network is improved, the network input layer is adjusted to a one-dimensional convolutional layer form, and the first layer is modified so that the input feature number matches the one-dimensional data feature number, so that it can accept one-dimensional input and is suitable for the data input of the present invention, so that the Bayesian neural network of the present invention has the ability to learn one-dimensional signal input.

[0075] Bayesian neural network training sub-step S114:

[0076] The Bayesian neural network hyperparameters are set according to the input data characteristics, and the feature set T [feature 1, feature 2, ..., feature N] is fused with the signal to generate the Bayesian neural network input TB [original signal, feature 1, feature 2, ..., feature N], and the set with true value or calibration value is extracted as training data and test data to form a test set and a training set. The Bayesian neural network with set parameters is trained using the training set, and the test set is used for testing and comparison until the confidence characterization function is detected. The training loss and validation loss tend to be stable, with fluctuation values ​​less than 10 -4 , the training converges and a more mature radio signal modulation recognition model is obtained.

[0077] In the Bayesian neural network, hyperparameters include: number of iterations, network dimension, network structure, optimizer, learning rate, Bachsize, output dimension, regularization parameter, variational inference parameter, dropout rate, activation function and output dimension. After simulation experiments and tests, the preferred hyperparameters of the present invention are network dimension, output dimension, optimizer and Bachsize.

[0078] In a specific embodiment, the hyperparameters are: network dimension (BBBconv1 2x126, BBBconv2 1x124, Dense1 128, Dense211), output dimension 2x126, Adam optimizer, and Bachsize is 1024.

[0079] Model application step S120:

[0080] When a new signal appears, the features are collected and compared to obtain a list of key focus signals. The trained Bayesian neural network is used to calculate the confidence of each of the key focus signal lists, and the confidence is used for modulation recognition to obtain a new signal and its features.

[0081] For details, see Figure 3 , including the key signal acquisition sub-step S121:

[0082] When a new signal recognition requirement arises, the features of the new signal are acquired, and a list of key signals is obtained through preliminary feature matching based on the initial feature element set T of the signal features.

[0083] In a specific embodiment, the initial feature matching is to determine the consistency of the features based on the value of the inner product of the features between the two signals, and set a threshold value, such as 70% consistency, to automatically include the signals in the list of key concerns.

[0084] Uncertainty calculation sub-step S122:

[0085] The trained Bayesian neural network is used to calculate the key attention signal list respectively to obtain the corresponding confidence level.

[0086] Modulation identification sub-step S123:

[0087] Modulation identification is performed on the list of key attention signals in the order of confidence from high to low to confirm the requirements, and after the requirements are met, the signals and their characteristics that successfully meet the requirements are recorded.

[0088] In this sub-step, the requirements include communication reconstruction, sorting and identification, identity confirmation and signal communication.

[0089] The features include center frequency, instantaneous frequency, spectrum, square spectrum, quartic spectrum, initial phase, signal bandwidth, signal amplitude, envelope and signal-to-noise ratio. Obviously, those skilled in the art can know that the recorded features are consistent with the signal features in the model building step S110.

[0090] Knowledge updating step S130:

[0091] The new signal and its features obtained in the model application step S120 are used to improve the initial feature element set of the signal features, and the Bayesian network is updated using the improved signal features.

[0092] For details, see Figure 4 , including the signal element set improvement sub-step S131:

[0093] By utilizing the new signal and its features obtained in the model application step S120, the required features are extracted, combined with their corresponding confidence conclusions, such as the conclusion that "the confidence of a certain signal is X", and based on a method similar to the signal preprocessing sub-step S110, an improved signal element set is obtained.

[0094] Bayesian network update sub-step S132:

[0095] By using the improved set of signal elements, based on a method similar to the Bayesian network construction sub-step S112, the Bayesian network is updated to obtain an updated cognition of the signal.

[0096] Through this sub-step, the signal is updated with conditional labels such as environment, quantity, identity, and type, thereby realizing iterative integration of prior knowledge and adaptive signal recognition results for the signal in a multi-dimensional environment.

[0097] The above operation ensures that when a signal with incomplete information characteristics is received again, the concentrated information can be improved according to the signal elements to give the recognition confidence.

[0098] The signal recognition method based on prior knowledge integration and result adaptation of the present invention is further illustrated by examples below.

[0099] Example:

[0100] Signal capture and recognition in an unmanned intelligent aerial vehicle cluster game performance competition under a strong solar storm environment.

[0101] In a strong solar storm environment, its X-ray and extreme ultraviolet radiation will rapidly increase the electron density in the Earth's ionosphere, thereby increasing the absorption and scattering of radio signals, causing signal attenuation or even complete interruption; and in the performance competition, especially in the preliminary round, due to the limited resources, numerous signals, unreasonable venue configuration, and complex environment of the organizer's venue, signal frequency band, participating teams, aircraft models, system versions, etc., the communication of the UAV cluster is inevitably interfered. In this case, a signal modulation recognition method that can quickly and autonomously identify one's own signals and establish communication in a complex environment with a large number of signals and interference and blocking, and effectively control one's own UAV cluster, so as to successfully complete the competition, is needed.

[0102] (1) Pre-game preparation - building a model

[0103] 1.1 Signal preprocessing

[0104] When the participating signal A that the team intends to select is received by the receiver, it is preliminarily preprocessed to obtain the initial feature element set of the signal characteristics, that is, the initial feature set T1 [center frequency 1, instantaneous frequency 1, spectrum 1, square spectrum 1, fourth power spectrum 1, initial phase 1, signal bandwidth 1, signal amplitude 1, envelope 1, signal-to-noise ratio 1] is formed.

[0105] 1.2 Building a Bayesian Network

[0106] A Bayesian network is constructed based on the elements in the existing feature set T1 to initially form an understanding of one's own signals.

[0107] 1.3 Building Bayesian Neural Networks and Parameters and Training Settings

[0108] A Bayesian neural network is constructed and used as a confidence calculation network, wherein the network input layer of the Bayesian network, that is, the first convolution layer, is made a one-dimensional convolution layer to construct a confidence characterization function.

[0109] 1.4 Training and Optimization of Bayesian Neural Networks

[0110] According to the signal characteristics and the envisioned scenario, select a suitable probability distribution function, set the network dimension, activation function, output dimension, Bachsize and other parameters. In this case, the settings are network dimension (BBBconv1 2x126, BBBconv2 1x124, Dense1 128, Dense211), output dimension 2x126, Adam optimizer, and Bachsize 1024.

[0111] By preliminarily analyzing the characteristic parameters and data form of the received signal, Gaussian distribution is selected, that is, it is assumed that the parameters of each layer in the Bayesian neural network obey Gaussian distribution. The signal feature set T1 and the signal are used as input to train the constructed Bayesian neural network until the training loss and verification loss in the detection confidence representation function are both stable. In this case, the fluctuation value is set to be less than 10 -4 The training converges and the recognition model D is obtained.

[0112] The above network is used for recognition training, and the parameters are further adjusted according to the recognition results and solution speed until it can better adapt to various preset scene conditions, forming a more mature radio signal recognition model D.

[0113] (2) Application during competition

[0114] After the competition started, the organizers did not reasonably arrange the signal frequency band and the number of signals used, resulting in unstable communication and too narrow bandwidth to meet the needs of group manipulation. In addition, since the day of the competition was during a period of intense solar storm activity, electromagnetic interference was severe, and both uplink and downlink signals were interfered to varying degrees. During the competition, signal interruptions occurred, and then many suspected signals were received but could not be clearly screened and confirmed. At this time, the participating team decided to use the Bayesian neural network in the present invention to identify the signal and establish effective communication.

[0115] 2.1 Obtaining key signals from initial screening

[0116] Based on the initial feature element set T of signal features before the game, a preliminary match is made with signal sources with significantly similar features, automatic tracking is established, the features of the signal sources are analyzed, and an inner product operation is performed with the features of the initial feature element set T of signal features that have been constructed. In this embodiment, signals exceeding 70% are included in the key focus signal list L.

[0117] 2.2 Signal uncertainty calculation and modulation identification

[0118] The recognition model D is used to automatically identify the signals in the focus signal list L, and the corresponding confidence, that is, uncertainty, is output, sorted from high to low. In this order, attempts are made to establish communication until success, and the characteristics of the signal A' are recorded at this time.

[0119] At this point in the competition, the method proposed in this article was used to successfully and quickly identify signals and resume communication in an environment with a large number of signals and obstruction interference.

[0120] (3) Post-match optimization

[0121] The signal A' feature is incorporated into the original initial set of signal features, and the corresponding Bayesian network is updated by adding conditional labels such as environment and quantity to achieve knowledge update of modulation recognition of the signal in a multi-dimensional environment.

[0122] In summary, the present invention has the following advantages:

[0123] 1. A two-level network structure is formed by using Bayesian network and Bayesian neural network, in which Bayesian network is used for knowledge integration and updating, and plays a role in preliminary screening, while Bayesian neural network is used to calculate signal recognition results and their confidence. The secondary network improves the speed and robustness of screening while ensuring the accuracy of signal screening.

[0124] 2. Using the Bayesian network as the primary screening network, each knowledge update can be integrated into the existing information database, which provides an effective way to timely update the signal feature library knowledge and effectively identify signals under new conditions (format camouflage, electromagnetic interference, quantity explosion, etc.).

[0125] 3. Bayesian neural network can obtain a more robust model with less data. It can identify radio signals in the absence of some prior knowledge or with very little initial data, for example, in the case of uneven distribution of prior training data, incomplete information, non-cooperative conditions, interference, complex environments, etc. It is of great significance for the effective identification of radio signals in complex environments.

[0126] 4. The Bayesian neural network gives the confidence level of signal recognition, breaking through the non-binary conclusion of traditional signal recognition, avoiding the misjudgment of key signals under adverse conditions, and providing a reasonable solution for further manual judgment and processing, with strong feasibility and scenario applicability.

[0127] 5. The Bayesian neural network introduces Bayesian theory in the convolution layer, adjusts the network input layer to a one-dimensional convolution layer form, modifies the first layer so that the input feature number matches the one-dimensional data feature number, so that it can accept one-dimensional input and is suitable for the data input of the present invention, so that the Bayesian neural network of the present invention has the ability to learn one-dimensional signal input.

[0128] The present invention can input and output the form of probability distribution through the Bayesian neural network, making it possible to break through the traditional signal recognition method of "yes or no binary theory", making it possible to draw a conclusion such as "the confidence level of signal A is 85%", and avoiding missing key signals.

[0129] Obviously, those skilled in the art should understand that the above-mentioned units or steps of the present invention can be implemented by a general-purpose computing device, they can be concentrated on a single computing device, and optionally, they can be implemented by a program code executable by a computer device, so that they can be stored in a storage device and executed by the computing device, or they can be made into individual integrated circuit modules, or multiple modules or steps therein can be made into a single integrated circuit module for implementation. Thus, the present invention is not limited to any specific combination of hardware and software.

[0130] The above content is a further detailed description of the present invention in combination with a specific preferred embodiment. It cannot be determined that the specific embodiments of the present invention are limited to this. For ordinary technicians in the technical field to which the present invention belongs, several simple deductions or substitutions can be made without departing from the concept of the present invention, which should be regarded as belonging to the protection scope of the present invention determined by the submitted claims.

Claims

1. A signal recognition method based on prior knowledge integration and result adaptation, characterized in that: The steps include: Model building step S110: Preprocess the signal, extract the signal features, construct an initial feature element set of the signal features, respectively construct a Bayesian network and a Bayesian neural network, set a Bayesian neural network confidence characterization function, set the Bayesian neural network hyperparameters according to the input parameter features, fuse the signal features with the signal as training data and test data, and train and test the Bayesian neural network until the model converges; Model application step S120: When a new signal appears, the features are collected and compared to obtain a list of key signals, and the confidence levels of the key signals are calculated using a trained Bayesian neural network, and modulation recognition is performed using the confidence levels to obtain new signals and their features; Knowledge updating step S130: Using the new signal and its features obtained in the model application step S120, the initial feature element set of the signal features is improved, and the Bayesian network is updated using the improved signal features; The model building step S110 includes: Bayesian network construction sub-step S112: Based on the initial element set of signal features, a Bayesian network is constructed to intuitively represent the signal features and their relationships, forming an initial cognitive framework for the signal; Bayesian neural network construction sub-step S113: A Bayesian neural network is used as a confidence calculation network, wherein the network input layer of the Bayesian neural network is a one-dimensional convolutional layer, and the confidence representation function is constructed as follows: ,in, represents the number of input training samples, Take values ​​for each parameter in the feature set, Represents the Bayesian neural network according to the input The calculated output results are: is the value label of the data after passing through the classifier, Z1= , represents the accidental uncertainty in Bayesian neural networks, is an estimate of the variance of the model output; Z2 To characterize the confidence that the signal comes from device A, the model performs T parameter sampling on the same sample. C represents the number of all different categories that the Bayesian neural network needs to distinguish, represents the probability value of a certain item in the sample; the T-time prediction probability is ,in , which means that in the Bayesian neural network The form of compliance is The distribution of Represents the activation function of the Bayesian neural network.

2. The signal recognition method according to claim 1, characterized in that: The model building step S110 further includes: Before the Bayesian network construction sub-step S112, the following steps are included: Signal preprocessing sub-step S111: The signal is preprocessed by denoising, filtering, frequency conversion sampling and equalization, and the initial feature extraction is used to construct the initial feature element set T [feature 1, feature 2, ..., feature N] of the signal feature; After the Bayesian neural network construction sub-step S113, the following steps are included: Bayesian neural network training sub-step S114: The Bayesian neural network hyperparameters are set according to the input data characteristics, the feature set T [feature 1, feature 2, ..., feature N] is fused with the signal to generate the Bayesian neural network input TB [original signal, feature 1, feature 2, ..., feature N], and the set with true value or calibration value is extracted as training data and test data to form a test set and a training set. The Bayesian neural network with set parameters is trained using the training set, and the test set is used for testing and comparison until the confidence characterization function is detected. The training loss and validation loss tend to be stable, with fluctuation values ​​less than 10 -4 , the training completes convergence.

3. The signal recognition method according to claim 2, characterized in that: In the signal preprocessing sub-step S111, the features are selected from a plurality of the following: center frequency, instantaneous frequency, spectrum, square spectrum, quartic spectrum, initial phase, signal bandwidth, signal amplitude, envelope and signal-to-noise ratio.

4. The signal recognition method according to claim 2, characterized in that: In the Bayesian network construction sub-step S112, The Bayesian network is a directed acyclic graph constructed with signal features as nodes, wherein the nodes represent random variables and the edges represent the dependencies between variables. The signal features adopt the features in the initial feature element set T, and the element relationships obtained by analysis are used to build the edges between the nodes to form a cognitive network based on radio signal features.

5. The signal recognition method according to claim 4, characterized in that: In the Bayesian neural network training sub-step S114, The hyperparameters for simulation experiments and tests include network dimension, output dimension, optimizer, and Bachsize.

6. The signal recognition method according to claim 1, characterized in that: The model application step S120 includes: Key signal acquisition sub-step S121: When a new signal recognition requirement arises, the features of the new signal are obtained, and a list of key signals is obtained through preliminary feature matching based on the initial feature element set T of the signal features; Uncertainty calculation sub-step S122: Using the trained Bayesian neural network to calculate the key signal list respectively, to obtain the corresponding confidence level; Modulation identification sub-step S123: Modulation identification is performed on the list of key attention signals in order of confidence from high to low to confirm the requirements, and after the requirements are met, the signals and their characteristics that successfully meet the requirements are recorded.

7. The signal recognition method according to claim 6, characterized in that: In the key signal acquisition sub-step S121, The initial feature matching is to determine the consistency of the features based on the inner product value between the features of two signals, and those exceeding the threshold are automatically included in the list of key attention signals.

8. The signal recognition method according to claim 6, characterized in that: In the modulation identification sub-step S123, The requirements include communication reconstruction, sorting and identification, identity confirmation and signal communication.

9. The signal recognition method according to claim 1, characterized in that: The knowledge updating step S130 is specifically as follows: The signal element set improvement sub-step S131 is included: Using the new signal and its features obtained in the model application step S120, the required features are extracted, and combined with the corresponding confidence conclusions, an improved signal element set is obtained; Bayesian network update sub-step S132: By using the improved set of signal elements and based on the method of the Bayesian network construction sub-step S112, the Bayesian network is updated to obtain an updated cognition of the signal.

Images