Signal intelligent model knowledge graph construction method and device

Abstract

The invention discloses a signal intelligent model knowledge graph construction method and device, and the method comprises the steps: S1, employing data sets of two recognition tasks, modulation recognition and individual recognition, and adding three kinds of interference to the data sets of the two recognition tasks, such as single-tone interference, linear sweep frequency interference and partial frequency band interference; s2, a plurality of deep learning models are used for pre-training the data sets containing the two tasks and the three interferences, and the corresponding relation between the deep learning models and the three interferences of the two tasks is obtained according to the parameters, the calculated amount, the task precision and other attributes of the deep learning models; s3, converting the used data into nodes in the knowledge graph, generating model nodes and task nodes under specific interference application, and endowing corresponding attributes and relationships; and converting the association between the model and the task into a connecting edge of a graph structure to form a knowledge graph. And S4, constructing and maintaining the atlas by using a Neo4j database through a Cypher query language and a Python API (Application Program Interface). According to the method, various interferences are added to the data sets of different tasks, a complex scene in an actual communication environment is simulated, and the data sets added with the interferences are used for training the existing deep learning model, so that intelligent selection and optimization of the signal recognition model are realized.

Description

Technical Field

[0001] The present invention relates to a signal processing and intelligent model management technology, in particular to a method and device for modeling and constructing a knowledge graph for signal processing model attributes. The invention belongs to the field of artificial intelligence and data mining, and specifically relates to a technology for efficiently constructing, managing and optimizing signal processing models using knowledge graphs. Background Art

[0002] Signal recognition technology has been widely used in wireless communications, electronic countermeasures, radio spectrum management and other fields. With the rapid development of wireless communication technology, the task of signal recognition has become increasingly complex. In particular, in complex and changing wireless communication environments, existing signal recognition methods face many challenges. For example, the model has poor adaptability in different signal scenarios and cannot flexibly respond to signal interference or task changes; there is a lack of effective knowledge sharing and synergy between different signal recognition models, resulting in insufficient performance in multi-task processing; when encountering various interference factors (such as noise, interference signals, etc.), the model has poor robustness and the recognition accuracy drops significantly (Reference [1]: Hu S, Pei Y, Liang PP, et al. Deep neural network for robust modulation classification under uncertain noise conditions [J]. IEEE Transactions on Vehicular Technology, 2019, 69(1): 564-577.).

[0003] Currently, traditional signal recognition methods are mostly based on pre-set models for specific tasks and environments. This approach cannot flexibly adapt to complex and dynamically changing communication environments. For example, the performance of many traditional signal recognition algorithms degrades dramatically when faced with high noise, low signal-to-noise ratio (SNR), spectrum overlap, or signal variations (Reference [2]: O'Shea TJ, Corgan J, Clancy T C. Convolutional radio modulation recognition networks [C] / / Engineering Applications of Neural Networks: 17th International Conference, EANN 2016, Aberdeen, UK, September 2-5, 2016, Proceedings 17. Springer International Publishing, 2016: 213-226.). Although signal recognition methods based on deep learning have achieved good results in some standardized environments, their universality and robustness are still limited. Especially when facing complex and dynamic radio spectrum environments, deep learning models often lack sufficient flexibility to effectively adapt to various changes (Reference [3]: Wu Q, Ruan T, Zhou F, et al. Aunified cognitive learning framework for adapting to dynamic environments and tasks [J]. IEEE Wireless Communications, 2021, 28 (6): 208-216.).

[0004] In addition, most current signal recognition systems are isolated models that operate independently, making it difficult to share and transfer knowledge. This means that when faced with new scenarios or unseen signals, existing models are often unable to adjust or transfer quickly, resulting in low efficiency and high error rates in multi-task environments (Reference [4]: ​​Zhou R, Liu F, Gravelle C W. Deeplearning for modulation recognition: Asurvey with a demonstration [J]. IEEE Access, 2020, 8: 67366-67376.). Therefore, signal recognition technology urgently needs a new method that can overcome the problems of poor adaptability, knowledge isolation between models, and insufficient interference response in current methods, and improve overall performance.

[0005] With the continuous development of knowledge graph technology, academia and industry have begun to explore its application in various intelligent systems, especially in the fields of automated reasoning, model recommendation and optimization (Reference [5]: Chen Z, Wang Y, Zhao B, et al. Knowledge graph completion: A review [J]. IEEE Access, 2020, 8: 192435-192456.). Knowledge graphs provide the system with the ability to store and reason knowledge by representing and storing the relationships between different entities, and can assist the system in effectively selecting and optimizing models when encountering new tasks (Reference [6]: Sun R, Cao X, Zhao Y, et al. Multi-modal knowledge graphs for recommender systems [C] / / Proceedings of the 29th ACM international conference on information & knowledge management. 2020: 1405-1414.). However, the application of existing knowledge graph technology in the field of signal recognition is still in its early stages. How to build a dynamic, adaptable and scalable knowledge graph for signal recognition is still an urgent issue to be solved.

[0006] Therefore, building a dynamic adjustment and recommendation mechanism for signal recognition models based on knowledge graphs has become an important research direction. This mechanism can not only enhance the model's adaptability in complex environments, but also effectively improve the ability to share knowledge between models, enabling intelligent and dynamic model selection, thereby achieving more efficient and accurate signal recognition. Summary of the Invention

[0007] In order to overcome the challenges currently faced in the field of signal recognition, such as the poor adaptability of models in complex scenarios, the difficulty in sharing knowledge between models, and the insufficient response to interference and task changes, the present invention proposes a method and device for constructing a knowledge graph of a signal intelligent model.

[0008] The present invention systematically represents the relationship between signal recognition tasks, interference factors and models by integrating and managing multiple existing signal recognition models. By constructing a dynamically updated knowledge graph, the present invention can automatically retrieve and recommend the most suitable signal recognition model according to different scenario requirements in actual applications. This method can automatically adjust the parameters of the model or select different sub-models according to task requirements and interference conditions in a complex wireless communication environment to optimize the effect of signal recognition. The construction of the knowledge graph not only improves the robustness of the model, but also can effectively deal with interference factors in signal recognition, improve the adaptability of the system, and thus provide the signal recognition system with greater flexibility and scalability. In addition, the model recommendation and completion functions based on the knowledge graph also provide a convenient solution for the rapid iteration and application of the model. Through the accumulation and analysis of historical signal recognition tasks, the system can continuously learn and optimize the knowledge graph, gradually achieve self-improvement and knowledge sharing, and thus achieve more efficient and accurate signal recognition in actual complex scenarios.

[0009] The technical solutions adopted by the present invention to solve the technical problems are as follows:

[0010] A method for constructing a signal intelligence model knowledge graph includes the following steps:

[0011] S1: Using datasets of two recognition tasks, modulation recognition and individual recognition, three types of interference are added to the datasets of the two recognition tasks, such as single-tone interference, linear sweep interference, and partial frequency band interference.

[0012] S2: Use multiple deep learning models to pre-train the datasets containing two tasks and three interferences. According to the parameters, computational complexity, task accuracy and other properties of the deep learning models, the corresponding relationship between the deep learning models and the two tasks and three interferences is obtained.

[0013] S3: Convert the used data into nodes in the knowledge graph, generate model nodes, task nodes under specific interference, and assign corresponding attributes and relationships; convert the association between the model and the task into edges in the graph structure to form a knowledge graph.

[0014] S4: Use the Neo4j database to build and maintain graphs through the Cypher query language and Python API.

[0015] Furthermore, the step S1 includes the following contents:

[0016] S1.1: Add single-tone interference to the dataset.

[0017] Determine the added interference frequency f based on the center frequency, bandwidth, and amplitude of the data set signal j , amplitude A and initial phase The expression of single-tone interference is:

[0018]

[0019] Generate a single-tone interference signal of the same length as the original signal. The signal after adding partial frequency band interference is:

[0020] s add (t)=s(t)+I tonal (t) (2)

[0021] where s(t) is the real part of the original IQ signal.

[0022] S1.2: Add partial-band interference to the dataset.

[0023] Determine the added interference center frequency f based on the center frequency, bandwidth, and amplitude of the data set signal c , bandwidth B and amplitude A, assuming the interference signal is I band (t), which is expressed as:

[0024]

[0025] Where rect is a rectangular function used to limit the frequency band. Generate a partial frequency band interference signal of the same length as the original signal. s(t) is the original signal. The signal after adding partial frequency band interference is:

[0026] s add (t)=s(t)+I band (t) (4)

[0027] S1.3: Add linear frequency sweep interference to the dataset.

[0028] Determine the added sweep start frequency f based on the center frequency, bandwidth, and amplitude of the data set signal start and the sweep stop frequency f end , sweep period T, amplitude A and initial phase Assume that the interference signal is I sweep (t), which is expressed as:

[0029]

[0030] Generate a partial frequency band interference signal of the same length as the original signal. s(t) is the original signal. The signal after adding partial frequency band interference is:

[0031] s add (t)=s(t)+I sweep (t) (6)

[0032] Furthermore, step S2 includes the following contents:

[0033] S2.1: For each model, different network structures and hyperparameter configurations are designed, and pre-trained under two tasks (such as modulation recognition and interference recognition) and three interference conditions (such as noise interference, spectrum overlap, and multipath interference). During the training process, each model is first trained independently to obtain its performance under different interference conditions. The key to this stage is to ensure that each model fully learns the signal characteristics through sufficient training and exhibits a certain degree of robustness in the face of changes in tasks and interference.

[0034] After S2.2 training is completed, evaluate the training results of each model, focusing on the following indicators:

[0035] Task accuracy: Each model is evaluated on its classification accuracy in a specific task (e.g., modulation recognition).

[0036] Computational complexity: Computational complexity is evaluated by measuring the computing resources consumed by the model during training and inference (such as computing power, memory usage, time consumption, etc.).

[0037] Model parameter count: Count the number of parameters of each model to measure the model size and training complexity.

[0038] Furthermore, step S3 includes the following contents:

[0039] S3.1: Establish a model node and retain attributes such as model name, parameter quantity, computational complexity, task accuracy, and model type.

[0040] Model name: A unique identifier for the model, for example, "CNN1D_001".

[0041] Parameters: The total number of parameters in the model, which can be calculated by the model's layer structure. The total number of parameters can be obtained by adding the parameters of each layer.

[0042] Computational Amount (FLOPs): The computational amount refers to the number of floating-point operations required for the model to perform one forward propagation. The calculation formula varies depending on the model architecture.

[0043] Task accuracy: The recognition accuracy of a model for a specific task. It is calculated by the model's performance on the test dataset.

[0044] Model type: The category of the model, such as "Convolutional Neural Network (CNN)" or "Recurrent Neural Network (RNN)".

[0045] S3.2: Establish a task node for applying specific interference to represent the type of signal recognition task and the interference scenario, and retain attributes such as task type, interference type, interference intensity, and bandwidth.

[0046] Task name: For example, "Modulation Recognition" or "Individual Recognition".

[0047] Interference Type: The name of the interference type, such as "Single Tone Interference" and "Linear Sweep Interference".

[0048] Interference strength: The intensity of interference is expressed by the signal-to-interference ratio (SIR), which is usually calculated by measuring the average signal power and the average interference power. The SIR is expressed as:

[0049]

[0050] Among them, P Signal is the average signal power, P Interference is the average interference power.

[0051] Furthermore, step S4 includes the following contents:

[0052] S4.1: Store nodes and their attributes, edges, and their relationships in a graph database. Use the Neo4j graph database to create nodes, add attributes, and establish edges between nodes using the Cypher language.

[0053] S4.2: Graph query and application, providing corresponding interference scenarios, task types, model parameter quantities and computational requirements, and using the query language of the graph database for intelligent recommendation and analysis.

[0054] The second aspect of the present invention relates to a signal intelligent model knowledge graph construction device, including a memory and one or more processors, wherein the memory stores executable code, and when the one or more processors execute the executable code, they are used to implement the signal intelligent model knowledge graph construction method of the present invention.

[0055] The third aspect of the present invention relates to a computer-readable storage medium having a program stored thereon, which, when executed by a processor, implements the signal intelligent model knowledge graph construction method of the present invention.

[0056] The innovative features of the present invention are:

[0057] The present invention constructs the signal recognition model, task type (modulation recognition / individual recognition) and interference scenario (single tone / sweep frequency / partial band interference) into a unified knowledge graph node, quantitatively defines the model attributes (parameter quantity, computational complexity, accuracy) and interference parameters (type, intensity, bandwidth), and uses the graph structure to dynamically associate the three relationships. This realizes intelligent model recommendation and parameter adaptive adjustment based on actual scenario requirements (such as signal-to-interference ratio and hardware resource limitations), breaking through the limitation of traditional static models that cannot adapt to complex environments.

[0058] The working principle of this invention is to simulate the complex scenarios found in real-world communication environments by adding various interferences to datasets for different tasks. This interference-added dataset is then used to train an existing deep learning model. By analyzing the training results, the relationships between the model, tasks, and interferences are established, and a knowledge graph containing nodes and attributes is constructed. Finally, this knowledge graph is constructed and maintained using the Neo4j graph database, enabling intelligent selection and optimization of signal recognition models.

[0059] The advantages of the present invention are:

[0060] The present invention improves system robustness and decision-making efficiency in complex radio environments by constructing a three-dimensional knowledge graph that integrates signal models, mission scenarios, and interference parameters. It utilizes the dynamic association mechanism of the graph to achieve cross-model knowledge sharing and adaptive interference matching, enhancing recognition stability in extreme scenarios. It intelligently recommends the optimal model combination based on quantitative attributes, and simultaneously optimizes computing resource allocation and real-time response. Its scalable architecture continuously absorbs new interference and advanced models, providing self-evolving and efficient decision-making support for dynamic communication environments. BRIEF DESCRIPTION OF THE DRAWINGS

[0061] Figure 1 Flowchart of the method of the present invention.

[0062] Figure 2 It is a schematic diagram of the device of the present invention. DETAILED DESCRIPTION

[0063] The technical solution of the present invention is further described below with reference to the accompanying drawings.

[0064] Example 1

[0065] like Figure 1 This embodiment relates to a wireless signal recognition method using the signal intelligent model knowledge graph construction method of the present invention, which is specifically divided into the following steps:

[0066] S1: Using datasets of two recognition tasks, modulation recognition and individual recognition, three types of interference are added to the datasets of the two recognition tasks, such as single-tone interference, linear sweep interference, and partial frequency band interference.

[0067] S2: Use multiple deep learning models to pre-train the datasets containing two tasks and three interferences. According to the parameters, computational complexity, task accuracy and other properties of the deep learning models, the corresponding relationship between the deep learning models and the two tasks and three interferences is obtained.

[0068] S3: Convert the used data into nodes in the knowledge graph, generate model nodes, task nodes under specific interference, and assign corresponding attributes and relationships; convert the association between the model and the task into edges in the graph structure to form a knowledge graph.

[0069] S4: Use the Neo4j database to build and maintain graphs through the Cypher query language and Python API.

[0070] S5: Input the graph query based on the specific parameters formulated for the actual recognition scenario to obtain the model with the best performance under the parameter scenario. Call the model according to the node attributes in the graph to perform signal recognition for the corresponding task.

[0071] Furthermore, the step S1 includes the following contents:

[0072] S1.1: Add single-tone interference to the dataset.

[0073] Determine the added interference frequency f based on the center frequency, bandwidth, and amplitude of the data set signal j , amplitude A and initial phase The expression of single-tone interference is:

[0074]

[0075] Generate a single-tone interference signal of the same length as the original signal. The signal after adding partial frequency band interference is:

[0076] s add (t)=s(t)+I tonal (t) (2)

[0077] where s(t) is the real part of the original IQ signal.

[0078] S1.2: Add partial-band interference to the dataset.

[0079] Determine the added interference center frequency f based on the center frequency, bandwidth, and amplitude of the data set signal c , bandwidth B and amplitude A, assuming the interference signal is I band (t), which is expressed as:

[0080]

[0081] Where rect is a rectangular function used to limit the frequency band. Generate a partial frequency band interference signal of the same length as the original signal. s(t) is the original signal. The signal after adding partial frequency band interference is:

[0082] s add (t)=s(t)+I band (t) (4)

[0083] S1.3: Add linear frequency sweep interference to the dataset.

[0084] Determine the added sweep start frequency f based on the center frequency, bandwidth, and amplitude of the data set signal start and the sweep stop frequency f end , sweep period T, amplitude A and initial phase Assume that the interference signal is I sweep (t), which is expressed as:

[0085]

[0086] Generate a partial frequency band interference signal of the same length as the original signal. s(t) is the original signal. The signal after adding partial frequency band interference is:

[0087] s add (t)=s(t)+I sweep (t) (6)

[0088] Furthermore, step S2 includes the following contents:

[0089] S2.1: For each model, different network structures and hyperparameter configurations are designed, and pre-trained under two tasks (such as modulation recognition and interference recognition) and three interference conditions (such as noise interference, spectrum overlap, and multipath interference). During the training process, each model is first trained independently to obtain its performance under different interference conditions. The key to this stage is to ensure that each model fully learns the signal characteristics through sufficient training and exhibits a certain degree of robustness in the face of changes in tasks and interference.

[0090] After S2.2 training is completed, evaluate the training results of each model, focusing on the following indicators:

[0091] Task accuracy: Each model is evaluated on its classification accuracy in a specific task (e.g., modulation recognition).

[0092] Computational complexity: Computational complexity is evaluated by measuring the computing resources consumed by the model during training and inference (such as computing power, memory usage, time consumption, etc.).

[0093] Model parameter count: Count the number of parameters of each model to measure the model size and training complexity.

[0094] Furthermore, step S3 includes the following contents:

[0095] S3.1: Establish a model node and retain attributes such as model name, parameter quantity, computational complexity, task accuracy, and model type.

[0096] Model name: A unique identifier for the model, for example, "CNN1D_001".

[0097] Parameters: The total number of parameters in the model, which can be calculated by the model's layer structure. The total number of parameters can be obtained by adding the parameters of each layer.

[0098] Computational Amount (FLOPs): The computational amount refers to the number of floating-point operations required for the model to perform one forward propagation. The calculation formula varies depending on the model architecture.

[0099] Task accuracy: The recognition accuracy of a model for a specific task. It is calculated by the model's performance on the test dataset.

[0100] Model type: The category of the model, such as "Convolutional Neural Network (CNN)" or "Recurrent Neural Network (RNN)".

[0101] S3.2: Establish a task node for applying specific interference to represent the type of signal recognition task and the interference scenario, and retain attributes such as task type, interference type, interference intensity, and bandwidth.

[0102] Task name: For example, "Modulation Recognition" or "Individual Recognition".

[0103] Interference Type: The name of the interference type, such as "Single Tone Interference" and "Linear Sweep Interference".

[0104] Interference strength: The intensity of interference is expressed by the signal-to-interference ratio (SIR), which is usually calculated by measuring the average signal power and the average interference power. The SIR is expressed as:

[0105]

[0106] Among them, P Signal is the average signal power, P Interference is the average interference power.

[0107] Furthermore, step S4 includes the following contents:

[0108] S4.1: Store nodes and their attributes, edges, and their relationships in a graph database. Use the Neo4j graph database to create nodes, add attributes, and establish edges between nodes using the Cypher language.

[0109] S4.2: Graph query and application, providing corresponding interference scenarios, task types, model parameter quantities and computational requirements, and using the query language of the graph database for intelligent recommendation and analysis.

[0110] Furthermore, step S5 includes the following contents:

[0111] S5.1: Specify specific parameters based on the actual recognition scenario, including: task type, interference type, maximum parameter quantity, maximum computational complexity, etc. Search the constructed graph based on these parameters to find the model node that meets the requirements and has the best performance in the scenario.

[0112] S5.2: Based on the node attributes of the query result, the model structure of the model and the model parameters trained in step 2 are locally called to perform the specified signal recognition task.

[0113] In step 1, the specific operation process is as follows: Figure 1 The interference type, such as single-tone interference, linear frequency sweep interference, or partial-band interference, was selected based on the actual application scenario. Interference of varying intensities was added to the original modulation recognition and individual identification datasets, controlling the signal-to-interference ratio (SIR) between -20 and 10 dB in 2 dB increments. Signal generation and simulation tools were used to process the original signals to simulate various interference conditions in real-world communication environments. The resulting interfered datasets were split into training and test sets with a 4:1 ratio.

[0114] In step 2, the specific operation process is as follows: Figure 1 Select a variety of implemented deep learning models for signal recognition, such as CNN1D, LSTM, GRU, and Transformer. Based on the task type and interference type, configure each deep learning model's hyperparameters, including the network structure, learning rate, and optimization algorithm. Train the model on a dataset with added interference. During training, evaluate each model's performance by calculating metrics such as accuracy, computational complexity, and parameter count.

[0115] In step 3, the specific operation process is as follows: Figure 1Based on different combinations of models, tasks, and interference, nodes in the graph are constructed. Each node represents an entity, and relevant attributes are added to each node. For example, the model node contains attributes such as model name, parameter quantity, computational complexity, task accuracy, and model type; the task node includes attributes such as task type, interference type, interference intensity, and bandwidth. Based on the training results, relationships between nodes are established. The relationship can be directed, for example, from the "model" node to the "task" node, indicating that the model can complete a certain task. The training results are integrated with relevant information about the task and interference to form a complete knowledge graph structure.

[0116] In step 4, the specific operation process is as follows: Neo4j is selected as the knowledge graph database system. The nodes and relationships in the knowledge graph are constructed and stored using the Neo4j API and Cypher query language in Python. During this process, a Python script reads the training results of the model, task, and interference from the data table and inserts this information into the Neo4j database. Using the Cypher query language, a query statement is written to obtain information such as the best model for a specific task and the optimal parameters under specific interference conditions. For example, given the task type, interference type, model parameter requirements, and computational requirements, the best performing model for a task under a specific interference condition is found.

[0117] In step 5, the specific operation process is as follows: specific parameters are specified based on the actual recognition scenario, including: task type, interference type, maximum parameter value, maximum computational load, etc. Based on these parameters, a node query is performed in the constructed graph to find the model node that meets the requirements and has the best performance for the scenario. Based on the node attributes of the query result, including the model structure path and model training parameter path, the model structure and model parameters trained in step 2 are locally called to perform the specified signal recognition task. The recognition result of the model for the current task and signal data is output, including the recognition result, recognition inference time, model computational load, etc.

[0118] The above examples illustrate the practical application of the present invention using modulation recognition and individual identification datasets, as well as three specific interference scenarios. By adding various interferences to datasets for different tasks, the present invention simulates the complex scenarios found in real-world communication environments. These datasets, with added interference, are then used to train existing deep learning models. By analyzing the training results, the relationships between the model, tasks, and interference are established, and a knowledge graph containing nodes and attributes is constructed. Finally, this knowledge graph is constructed and maintained using the Neo4j graph database, the Cypher query language, and the Python API, enabling intelligent selection and optimization of signal recognition models.

[0119] Example 2

[0120] like Figure 2 This embodiment relates to a signal intelligent model knowledge graph construction device, including a memory and one or more processors, wherein the memory stores executable code, and when the one or more processors execute the executable code, they are used to implement the signal intelligent model knowledge graph construction method of Example 1.

[0121] Example 3

[0122] This embodiment relates to a computer-readable storage medium having a program stored thereon. When the program is executed by a processor, the signal intelligence model knowledge graph construction method of embodiment 1 is implemented.

[0123] It is only illustrative and not restrictive of the invention. Those skilled in the art will appreciate that many changes, modifications, and even equivalents may be made within the spirit and scope of the invention claims, and all of these will fall within the scope of protection of the present invention.

Claims

1. A method for constructing a knowledge graph of a signal intelligence model, characterized by: The method comprises the following steps: S1: Using datasets of two recognition tasks, modulation recognition and individual recognition, three types of interference are added to the datasets of the two recognition tasks, such as single-tone interference, linear sweep interference, and partial-band interference; S2: Use multiple deep learning models to pre-train the dataset containing two tasks and three interferences. Based on the parameters, computational complexity, task accuracy and other properties of the deep learning models, the corresponding relationship between the deep learning models and the two tasks and three interferences is obtained; S3: Convert the used data into nodes in the knowledge graph, generate model nodes and task nodes under specific interference, and assign corresponding attributes and relationships; convert the associations between models and tasks into edges in the graph structure to form a knowledge graph; S4: Use the Neo4j database to build and maintain graphs through the Cypher query language and Python API.

2. The method for constructing a signal intelligence model knowledge graph according to claim 1, characterized in that: The step S1 includes the following contents: S1.1: Add single-tone interference to the dataset; Determine the added interference frequency f based on the center frequency, bandwidth, and amplitude of the data set signal j , amplitude A and initial phase The expression of single-tone interference is: Generate a single-tone interference signal of the same length as the original signal. The signal after adding partial frequency band interference is: s add (t)=s(t)+I tonal (t) (2) Where s(t) is the real part of the original IQ signal; S1.2: Add partial frequency band interference to the dataset; Determine the added interference center frequency f based on the center frequency, bandwidth, and amplitude of the data set signal c , bandwidth B and amplitude A, assuming the interference signal is I band (t), which is expressed as: Where rect is a rectangular function used to limit the frequency band range; a partial-band interference signal with the same length as the original signal is generated, s(t) is the original signal, and the signal after adding the partial-band interference is: s add (t)=s(t)+I band (t) (4) S1.3: Add linear frequency sweep interference to the dataset; Determine the added sweep start frequency f based on the center frequency, bandwidth, and amplitude of the data set signal start and the sweep stop frequency f end , sweep period T, amplitude A and initial phase Assume that the interference signal is I sweep (t), which is expressed as: Generate a partial frequency band interference signal of the same length as the original signal. s(t) is the original signal. The signal after adding partial frequency band interference is: s add (t)=s(t)+I sweep (t) (6) 3. The method for constructing a signal intelligence model knowledge graph according to any one of claims 1 to 3, characterized in that: The step S3 includes the following contents: S3.1: Create a model node and retain attributes such as model name, parameter quantity, computational complexity, task accuracy, and model type; Model name: unique identifier of the model; Parameters: The total number of model parameters, which can be calculated through the model's layer structure; the total number of parameters can be obtained by adding the parameters of each layer; FLOPs: The number of floating-point operations required for the model to perform one forward pass. The calculation formula varies depending on the model architecture. Task accuracy: The recognition accuracy of the model under a specific task; calculated by the performance of the model on the test dataset; Model type: category of the model; S3.2: Establish a task node for applying specific interference, which is used to represent the type of signal recognition task and the interference scenario, and retain attributes such as task type, interference type, interference intensity, and bandwidth; Task name: including "modulation recognition" and "individual recognition"; Interference Type: the name of the interference type; Interference strength: The interference strength is expressed by the signal-to-interference ratio (SIR), which is usually calculated by measuring the average signal power and the average interference power. The SIR is expressed as: Among them, P Signal is the average signal power, P Interference is the average interference power; S3.4: Establish an edge between the model and the task, representing the situation when a specific model is used to perform a specific interference applied to a specific task; the attribute of the edge is the execution accuracy; the execution accuracy is the average recognition accuracy of the model in the task, which is directly obtained from the model test data.

4. The method for constructing a signal intelligence model knowledge graph according to any one of claims 1 to 4, characterized in that: The step S4 comprises the following contents: S4.1: Store nodes, their attributes, edges, and their relationships in a graph database. Use the Neo4j graph database to create nodes, add attributes, and establish edges between nodes using the Cypher language. S4.2: Graph query and application, providing corresponding interference scenarios, task types, model parameter quantities and computational requirements, and using the query language of the graph database for intelligent recommendation and analysis.

5. A signal intelligence model knowledge graph construction device, characterized in that: It includes a memory and one or more processors, wherein the memory stores executable code, and when the one or more processors execute the executable code, it is used to implement the signal intelligent model knowledge graph construction method described in any one of claims 1-4.

6. A computer-readable storage medium, characterized in that A program is stored thereon, and when the program is executed by a processor, the signal intelligent model knowledge graph construction method described in any one of claims 1-4 is implemented.

Images