Intelligent detection and identification method and system for frequency hopping signals for industrial internet

Abstract

This invention discloses an intelligent detection and recognition method and system for frequency-hopping signals in the industrial internet, relating to the field of signal recognition, to improve the detection rate and type identification accuracy of frequency-hopping signals in communication systems. The invention constructs a Rayleigh channel to simulate multipath effects and shadowing fading, applies predetermined type interference to sample frequency-hopping signals on this channel, performs a short-time Fourier transform on the output signal to generate a time-frequency graph, and constructs a signal detection time-frequency graph dataset. The dataset is labeled and partitioned to train a target detection network model, which incorporates a contextual hierarchical module and a gated aggregation module within the YOLOv5 backbone network. Finally, the trained model is used to identify the time-frequency graph of the frequency-hopping signal to be detected. This invention realistically recreates the electromagnetic environment in industrial internet scenarios, significantly improving the signal detection and recognition performance of the network by hierarchically extracting contextual information and simultaneously fusing long-term and short-term time-frequency features.

Description

Technical Field

[0001] This invention relates to the field of signal recognition, and in particular to a method and system for intelligent detection and recognition of frequency hopping signals for the Industrial Internet. Background Technology

[0002] Wireless communication technology is the core of the industrial internet system. Wired communication often leads to complex wiring and is prone to security risks. Meanwhile, frequency hopping communication technology, as an important anti-interception and anti-interference technology, is widely used in industrial internet scenarios. By rapidly changing its carrier frequency in a wide frequency band, it avoids signal interruption caused by interference, while also improving security and increasing the difficulty for potential eavesdroppers to track and decode the signal.

[0003] However, the time-varying characteristics of the carrier frequency of frequency-hopping signals make their detection and identification exceptionally difficult compared to traditional signals. Furthermore, considering the complexity of the electromagnetic environment in industrial internet scenarios and the impact of various electromagnetic interferences, the detection and identification of frequency-hopping signals become even more challenging. In harsh industrial environments, many large building structures and equipment obstruct signal propagation, and metal surfaces cause signal reflection and scattering, resulting in multipath effects. In addition, due to numerous devices and complex operating conditions, a large number and variety of interference signals exist between frequency-hopping channels, primarily in the form of single-tone interference, multi-tone interference, and pulse interference, severely affecting normal communication and degrading the performance of frequency-hopping signal detection and identification.

[0004] Currently, methods for frequency-hopping signal detection and identification can be broadly categorized into two types: feature engineering-based and deep learning-based methods. Feature engineering-based methods utilize expert knowledge to analyze and extract time-frequency features to identify the signal modulation type. This type of method is highly dependent on expert experience and has limited applicability. With the rapid development of artificial intelligence, deep learning has gradually become the mainstream method for frequency-hopping signal modulation identification. Using time-frequency analysis techniques, it can obtain signal features in different domains, such as time-frequency characteristics and multi-time constellation diagrams. Then, deep neural networks are used to automatically analyze and extract signal features to identify the modulation type. This method does not rely on expert features, significantly improving both recognition speed and accuracy, and is applicable to various scenarios. However, for industrial internet environments with significant interference, targeted improvements to the target detection network are still needed.

[0005] Chinese patent document CN116010798A discloses a method for fast identification and classification of complex typical communication signals based on YOLOv5. It uses a YOLOv5 network to predict and classify time-frequency maps, solving the problem of accurate and rapid identification and classification of overlapping complex communication signals. However, the algorithm has significant errors when the signal-to-interference ratio (JSR) is high, and it only applies to three types of signals, which limits its effectiveness.

[0006] Chinese patent document CN116318249A discloses a shortwave frequency hopping signal sorting method based on an improved YOLOv5. This method adds a CA mechanism to the backbone network based on YOLOv5, replaces NMS with Soft-SIoU_NMS, and uses NWD instead of NMS and the Intersection of Union (IoU) metric in the regression loss function for small target detection. This allows for the sorting of various frequency hopping signals from a large number of frequency hopping stations, achieving frequency hopping signal sorting under low signal-to-noise ratio (SNR). However, this method only considers background noise, resulting in a single dimension of interference signal characteristics, narrow application scenarios, and the accuracy of the identification algorithm needs improvement at low SNR levels. Summary of the Invention

[0007] The purpose of this invention is to provide a method and system for intelligent detection and identification of frequency hopping signals for the Industrial Internet, addressing the aforementioned problems, thereby solving the impact of complex industrial environments on signal detection and significantly improving the detection rate and type identification accuracy of frequency hopping signals in communication systems.

[0008] The technical solution adopted in this invention is as follows:

[0009] A method for intelligent detection and identification of frequency-hopping signals for the Industrial Internet, comprising:

[0010] S1. Construct a Rayleigh channel to simulate multipath effects and shadowing fading, apply predetermined type of interference to the sample frequency hopping signals on the channel, perform short-time Fourier transform on the output signal to generate a time-frequency graph, and construct a signal detection time-frequency graph dataset.

[0011] S2. Label the modulation category of all signal detection time-frequency maps in the signal detection time-frequency map dataset; train the pre-built target detection network model using the labeled signal detection time-frequency map dataset to obtain the signal detection model;

[0012] The target detection network model is constructed as follows: the backbone network uses a context-layering module to extract multi-level linear features from the input time-frequency map, and then uses a gated aggregation module to aggregate the multi-level linear features into a fused feature vector; the fused feature vector is then passed through the neck network and the head network to obtain the recognition result.

[0013] S3. Input the time-frequency diagram of the frequency hopping signal to be detected into the signal detection model to identify the modulation category.

[0014] Furthermore, the method for multi-level linear feature extraction in the context-layered module includes:

[0015] The contextual layering module uses multiple cascaded deep separable convolutional layers to extract features from the input time-frequency map step by step. Each level of deep separable convolutional layer is activated by GeLU to obtain a first-level linear feature, and the linear features of the last level of deep separable convolutional layer are obtained by average pooling to obtain the last-level linear feature.

[0016] Furthermore, the gating aggregation module's methods for multi-level linear feature fusion include:

[0017] The gated aggregation module uses a gated network to access multi-level linear features and employs a weight matrix to perform weighted aggregation of the outputs of each level of the gated network.

[0018] Furthermore, an object detection network model is trained using the labeled signal detection time-frequency graph dataset, including:

[0019] Convert the labeled signal detection time-frequency graph dataset into a format suitable for object detection network models;

[0020] Time-frequency maps are randomly extracted from the signal detection time-frequency map dataset according to a predetermined ratio to divide the dataset into training set, validation set and test set;

[0021] Based on the set training parameters, the object detection network model is trained using the training set, and the object detection network model is validated and tested using the validation set and test set, respectively.

[0022] Furthermore, the predetermined type of interference includes at least one of single-tone interference, linear frequency modulation interference, noise frequency modulation interference, radio frequency noise interference, and pulse discontinuity noise interference.

[0023] To address the above problems, this invention also provides an intelligent detection and identification system for frequency hopping signals for the Industrial Internet, comprising:

[0024] The dataset generation module is configured to: receive a configuration for a Rayleigh channel that simulates multipath effects and shadowing fading; receive configured interference and apply a predetermined type of interference to the sample frequency-hopping signals on the channel; and perform a short-time Fourier transform on the output signal to generate a time-frequency graph, thereby constructing a signal detection time-frequency graph dataset.

[0025] The signal detection model training module is configured to: receive a configured target detection network model; receive annotations of modulation categories for all signal detection time-frequency maps in the signal detection time-frequency map dataset; and train the target detection network model using the annotated signal detection time-frequency map dataset to obtain the signal detection model.

[0026] The target detection network model includes a backbone network, a neck network, and a head network. The backbone network includes a contextual layering module and a gated aggregation module. The contextual layering module is configured to sequentially extract multi-level linear features from the input time-frequency map. The gated aggregation module is configured to aggregate the multi-level linear features output by the contextual layering module into a fused feature vector. The fused feature vector is then processed sequentially by the neck network and the head network to output the recognition result.

[0027] The input module is configured to receive the frequency hopping signal to be detected and perform a short-time Fourier transform to generate a time-frequency diagram of the frequency hopping signal to be detected.

[0028] The output module is configured to: use the signal detection model trained by the signal detection model training module to detect and identify the time-frequency graph of the frequency hopping signal to be detected, and identify the modulation category.

[0029] Furthermore, the contextual layering module includes multiple cascaded depthwise separable convolutional layers, each configured with a GeLU activation function; the last depthwise separable convolutional layer is followed by an average pooling layer; the input time-frequency map is subjected to feature extraction by the multi-level depthwise separable convolutional layers step by step, each level of depthwise separable convolutional layer is activated by GeLU to obtain a first-level linear feature, and the linear feature of the last level of depthwise separable convolutional layer is averaged by an average pooling layer to obtain the last-level linear feature.

[0030] Furthermore, the gating aggregation module includes a gating network and an aggregation module, the aggregation module being configured with a weight matrix; the multi-level linear features output by the context layering module are connected to the gating network, the aggregation module is connected to each level of the gating network output, and the weight matrix is ​​used to perform weighted aggregation of each level of the gating network output.

[0031] Furthermore, the signal detection model training module trains the target detection network model according to the following configuration:

[0032] Convert the labeled signal detection time-frequency graph dataset into a format suitable for object detection network models;

[0033] Time-frequency maps are randomly extracted from the signal detection time-frequency map dataset according to a predetermined ratio to construct training set, validation set and test set respectively;

[0034] The system receives the configured training parameters, trains the object detection network model using the training set based on these parameters, and validates and tests the object detection network model using the validation set and test set, respectively.

[0035] Furthermore, the predetermined type of interference includes at least one of single-tone interference, linear frequency modulation interference, noise frequency modulation interference, radio frequency noise interference, and pulse discontinuity noise interference.

[0036] In summary, due to the adoption of the above technical solution, the beneficial effects of the present invention are:

[0037] 1. This invention uses Rayleigh channels to simulate large-scale fading caused by obstacle obstruction and small-scale fading caused by signal reflection and scattering due to mechanical effects of metal structures. This can effectively simulate and analyze the propagation characteristics of wireless signals in complex industrial environments.

[0038] 2. This invention investigates various equipment in different factory workshops, not limited to a single factory, and summarizes the various types of interference signals generated by these devices during operation or startup. It then simulates these signals to make frequency hopping signals and interference coexist, so that the dataset covers different scenarios. Both interference and channel characteristics are considered simultaneously, which can realistically reproduce the electromagnetic environment in the industrial Internet scenario.

[0039] 3. This invention extracts contextual information hierarchically and simultaneously fuses long-term and short-term time-frequency features to expand the receptive field while preserving the refined perceptual capabilities of the signal detection model. By utilizing feature maps generated through contextual hierarchy, the receptive field is enlarged, capturing more contextual information at a coarser granularity. A gated aggregation module selectively learns coarse-grained spatial information simultaneously for different detection targets, making it easier for the network to capture the time-frequency features of transitional signals, distinguishing signals from interference, reducing the probability of false detections and missed detections, and thus significantly improving the network's signal detection and recognition performance. Attached Figure Description

[0040] The present invention will be described by way of example and with reference to the accompanying drawings, wherein:

[0041] Figure 1 This is a flowchart of a frequency hopping signal intelligent detection and recognition method in one embodiment of the present invention;

[0042] Figure 2 This is a comparison diagram of the simulated complex electromagnetic environment of the Industrial Internet in this invention;

[0043] Figure 3 This is a schematic diagram of the YOLOv5 network structure;

[0044] Figure 4 This is a schematic diagram of error detection in an embodiment of the present invention;

[0045] Figure 5 This is a schematic diagram of missed detection in an embodiment of the present invention;

[0046] Figure 6 This is a schematic diagram of multi-detection in an embodiment of the present invention;

[0047] Figure 7 This is a schematic diagram of false detection in an embodiment of the present invention;

[0048] Figure 8 This is a schematic diagram of the structure of the YOLO-ad network in an embodiment of the present invention;

[0049] Figure 9 This is a schematic diagram of the context layering module structure in an embodiment of the present invention;

[0050] Figure 10 This is a schematic diagram of the gated aggregation module structure in an embodiment of the present invention;

[0051] Figure 11 This is the frequency hopping signal identification confusion matrix diagram based on the embodiments of the present invention;

[0052] Figure 12 This is a comparison chart of the F1 score curves of the two network models in this embodiment of the invention;

[0053] Figure 13 and Figure 14 These are two examples comparing the detection results of two different network models. Detailed Implementation

[0054] All features disclosed in this specification, or all steps in all disclosed methods or processes, may be combined in any way, except for mutually exclusive features and / or steps.

[0055] Any feature disclosed in this specification (including any appended claims and abstract) may be replaced by other equivalent or similar features, unless specifically stated otherwise. That is, unless specifically stated otherwise, each feature is merely one example of a series of equivalent or similar features.

[0056] Example 1

[0057] like Figure 1 As shown, the intelligent detection and identification method for frequency hopping signals for the Industrial Internet in this embodiment includes the following steps:

[0058] Step 1: Reproduce the complex electromagnetic environment of the Industrial Internet, construct a Rayleigh channel to simulate multipath effects and shadow fading, and apply common types of interference in the industrial environment (i.e., the predetermined type of interference in this embodiment) to the sample frequency hopping signal on the Rayleigh channel. Perform short-time Fourier transform on the (channel) output signal to generate a time-frequency graph and construct a signal detection time-frequency graph dataset.

[0059] The specific process for this step includes:

[0060] For multipath effects and shadow fading, this embodiment considers common anti-scattering devices and obstructions in industrial environments. In most cases, multipath effects follow a Rayleigh distribution. Shadow fading can be achieved by adding fading compensation to the Rayleigh fading channel. In obstructed areas, the signal strength is weakened due to the presence of obstacles; this fading can be modeled as an attenuation factor.

[0061] For electromagnetic interference, this embodiment considers one or more of the following interference signals: single-tone interference, linear frequency modulation interference, noise frequency modulation interference, radio frequency noise interference, and pulse discontinuity noise interference. Since the frequency of frequency-hopping signals changes irregularly over time, they are typical non-stationary signals, making comprehensive analysis in both the time and frequency domains difficult. Therefore, to more effectively detect frequency-hopping signals, a time-frequency analysis method is used. The Short-Term Fourier Transform (STFT) can effectively extract the transient characteristics of frequency-hopping signals; therefore, this embodiment uses STFT analysis to extract the time-frequency characteristics of the signal and generate a time-frequency map of the frequency-hopping signal. Multiple samples of the frequency-hopping signal yield multiple time-frequency maps, which together construct the required signal detection time-frequency map dataset. Considering a signal-to-interference-plus-noise ratio (SNR) of 5dB, the number of frequency-hopping signal samples is 300, with 90% selected as the training set, 5% as the validation set, and 5% as the test set.

[0062] The short-time Fourier transform is defined as:

[0063]

[0064] Where s(t) is the frequency hopping signal and w(t) is the window function.

[0065] The inverse short-time Fourier transform is:

[0066]

[0067] When the window functions h(t) and w(t) satisfy the following relationship, the signal s(t) can be reconstructed from STFT(τ, ω) using the above equation:

[0068]

[0069] The discrete STFT definition of a sequence is:

[0070]

[0071] Where N is the length of the sequence s[n], and the width of the window function w[n] is controllable. For time-frequency analysis of a signal, its amplitude is usually analyzed, and its spectrum can be defined as:

[0072] SPEC[m,k]=|DSTFT[m,k]| 2

[0073] The corresponding amplitude diagram is the time-frequency diagram.

[0074] Figure 2 (a) shows the performance of the frequency hopping signal in the absence of interference. Figure 2 (b) The performance under the influence of interference and scale fading shows that the presence of interference and scale fading weakens the signal features and adds redundant waveform interference recognition, which greatly affects the network's learning of signal features and reduces the detection success rate.

[0075] Step 2: Construct an object detection network model.

[0076] Figure 3 The diagram shows the YOLOv5 network structure. In this embodiment, the object detection network model is based on the YOLOv5 network, with a contextual layering module and a gating aggregation module added to the backbone network.

[0077] In the complex electromagnetic spectrum environment of the Industrial Internet scenario, it is difficult to achieve a high signal-to-interference ratio (SIR). Strong interference signals coexisting with the signal will severely affect the detection and identification performance of frequency-hopping signals, specifically including the following:

[0078] 1)Reference Figure 4 Interference signals such as single-tone and linear frequency modulation have very similar characteristics to frequency hopping signals using 8PSK modulation in the time-frequency diagram, making it difficult for target detection networks to distinguish them and easily leading to false detections.

[0079] 2)Reference Figure 5 In the time-frequency domain, interference signals that overlap with the frequency-hopping signal will cover the characteristics of the frequency-hopping signal, causing the signal to be missed.

[0080] 3)Reference Figure 6 In situations where signals are densely distributed, the YOLOv5 network may place the background and signal in the same detection frame, causing confusion and resulting in multiple detections.

[0081] 4)Reference Figure 7 When a target detection network learns the characteristics of signals with multiple modulation types, some modulation signals exhibit highly similar time-frequency characteristics. Therefore, relying solely on the time-frequency characteristics of the signal can easily lead to false detections of modulation types.

[0082] Therefore, it is necessary to comprehensively consider the time-frequency characteristics of interference and frequency hopping signals, and optimize the target detection network structure to enable it to effectively distinguish between interference and frequency hopping signals, thereby improving its detection rate and recognition rate.

[0083] Therefore, this embodiment proposes to improve the recognition rate of frequency-hopping signals by fusing the characteristics of the target signal itself and its surrounding background features. This is because background features can reflect the changing patterns of the signal and effectively distinguish frequency-hopping signals from interference. Based on this, this embodiment expands the network's field of vision and enhances the network's ability to fuse signal and background features by employing a hierarchical context module and a gated aggregation module, thereby improving the signal detection and recognition rate.

[0084] The backbone of the object detection network model uses focused modulation, first aggregating the contextual information surrounding the target signal, and then obtaining the query results based on the aggregated information. Its mathematical model is shown below:

[0085] y i =Γ(M(i,X),x i )

[0086] In the formula, i represents the position in the time-frequency plot, X is the time-frequency plot within the entire analysis frequency band, M(i, X) operation represents the time-frequency aggregation features around position i, and Γ represents the signal feature x at position i in the time-frequency plot. i Interacting with the aggregated feature M(i, X) yields fused features of the signal and background. M(i, X) employs contextual hierarchical and gated aggregation to expand the receptive field of the neural network and extract contextual information from the time-frequency plot. Furthermore, Γ is a lightweight operator implemented using a query mapping function and the Hadamard product, resulting in low computational complexity.

[0087] Since the target detection network model in this embodiment is designed based on the YOLOv5 network, for ease of description later, the target detection network model in this embodiment will be referred to as the YOLO-ad network.

[0088] Figure 8 The diagram shown is a schematic representation of the YOLO-ad network in this embodiment. Figure 3As can be seen from the comparison, the YOLO-ad network in this embodiment directly adopts the neck network of the YOLOv5 network. In fact, it also adopts its head network. The main difference between the YOLO-ad network and the YOLOv5 network is that the backbone network of the YOLOv5 network uses convolutional modules and C3 modules to be stacked alternately, while the backbone network (i.e. the feature extraction network) of the YOLO-ad network is designed with a contextual hierarchical module and a gated aggregation module. The time-frequency map is input into the linear layer, and depthwise separable convolution is used to extract hierarchical contextual information. Then, a gated aggregation mechanism is introduced, and finally, feature weighted fusion is used to obtain fused features and pass them to the SPPF layer.

[0089] Thus, it can be seen that the target detection network model in this implementation is constructed as follows: the backbone network uses the context layering module to extract multi-level linear features from the input time-frequency map, and then uses the gating aggregation module to aggregate the multi-level linear features into a fused feature vector; the fused feature vector is then passed through the neck network and the head network to obtain the recognition result.

[0090] The methods for multi-level linear feature extraction in context-layered modules include:

[0091] The contextual layering module uses multiple cascaded deep separable convolutional layers to extract features from the input time-frequency map step by step. Each level of deep separable convolutional layer is activated by GeLU to obtain a first-level linear feature, and the linear features of the last level of deep separable convolutional layer are obtained by average pooling to obtain the last-level linear feature.

[0092] Figure 9 This is a schematic diagram of the context layering module structure in this embodiment. See also... Figure 9 First, the time-frequency map feature map X is input, and it is projected into a new feature space through a linear layer. Contextual information is extracted using L depthwise separable convolutional layers. By simultaneously fusing long and short-term time-frequency features, the network's fine-grained perception capability is preserved while expanding the receptive field.

[0093] The gating aggregation module's method for fusing multi-level linear features includes: the gating aggregation module uses a gating network to access multi-level linear features, and uses a weight matrix to perform weighted aggregation of the outputs of each level of the gating network.

[0094] Figure 10 This is a schematic diagram of the gated aggregation module structure in this embodiment. See also... Figure 10 In this module, the (L+1) linear feature maps obtained from the context layering module are compressed into a single feature vector. In the time-frequency graph, different levels are set according to different detection targets. Here, a spatially and level-aware weight matrix G is obtained using a linear layer. Finally, the fused feature vector 2 is calculated using the weight matrix G. out .

[0095] Step 3: Label the modulation categories of all signal detection time-frequency maps in the signal detection time-frequency map dataset; train the pre-built target detection network model using the labeled signal detection time-frequency map dataset to obtain the signal detection model. For comparison purposes, this step uses the labeled signal detection time-frequency map dataset to train the YOLOv5 network and the YOLO-ad network respectively.

[0096] The training environment is configured as follows: CPU model is Intel Core i7-12700H, memory is 16GB, graphics card type is RTX3060 with 8GB of video memory, and Python version is 3.8.

[0097] The network training process includes:

[0098] Convert the labeled signal detection time-frequency graph dataset into a format suitable for object detection network models.

[0099] Time-frequency maps are randomly extracted from the signal detection time-frequency map dataset according to a predetermined ratio to divide the dataset into training set, validation set and test set.

[0100] Based on the set training parameters, the object detection network model is trained using the training set, and the object detection network model is validated and tested using the validation set and test set, respectively.

[0101] Specifically, there are:

[0102] 1) Tag all images using the labelImg tool and save the generated XML file.

[0103] 2) Convert the XML file into a TXT file suitable for the YOLOv5 network (and also for the YOLO-ad network), and randomly generate training, test, and validation sets according to the specified proportions.

[0104] 3) The initial learning rate of the YOLO-ad network was set to 0.01, and after multiple training rounds, the learning rate was reduced to 0.0001. The training batch size was set to 16, and 300 training rounds were performed.

[0105] Step 4: Input the time-frequency graph of the frequency hopping signal to be detected into the signal detection model (and the trained YOLOv5 network) to identify the modulation category.

[0106] This embodiment also verifies the performance of the designed method. Precision (P) and recall (R) are used to evaluate the experimental results, and the specific calculation process is shown in the following formula.

[0107]

[0108] TP is the number of correctly detected targets; FP is the number of incorrectly detected targets; FN is the number of missed targets.

[0109] The F1 score is a weighted average of precision and recall. The best F1 score is 1, and the worst is 0. The formula for calculating the F1 score is:

[0110]

[0111] Recall and precision are plotted on the x and y axes, respectively, to create a Recall-Precision curve. The integral of this curve yields the average precision (AP). Since multiple categories are considered, the average of these values ​​is denoted as mean average precision (mAP), used to measure the model's detection performance across all categories. The calculation formula is as follows:

[0112] AP=∫0 1 PdR

[0113]

[0114] In the formula, N represents the number of target categories detected in the dataset, and i is the current category number. This embodiment sets N=8 and considers eight modulation types—8PSK, 16QAM, 32QAM, 64QAM, BPSK, FSK, MSK, and QPSK—to analyze the network's detection and recognition performance. The mAP50 metric is calculated only when the Intersection over Union (IoU) threshold is 0.5. mAP50 primarily focuses on coarser detection performance because it allows for a certain degree of deviation between the predicted bounding box and the ground truth bounding box. The mAP50-95 metric considers multiple IoU thresholds, increasing by 0.05 from 0.5 to 0.95. This means the algorithm needs to perform well at different IoU thresholds to achieve a high mAP50-95 score.

[0115] After determining the performance metrics, the network's recognition performance is further illustrated by the following simulation results:

[0116] like Figure 11 The diagram shown is a confusion matrix for frequency hopping signal identification based on a YOLO-ad network. Figure 11 As can be seen, frequency-hopping signals of all four modulation types can be accurately detected. Specific experimental data are shown in Table 1 below:

[0117] Table 1 Comparison of YOLOv5s and YOLO-ad metrics

[0118]

[0119] As can be seen, the introduction of the YOLO-ad network improves detection accuracy by 12%, recall by 1.8%, mAP50 by 8.8%, and mAP50-95 by 12.3%. The model's improvements in accuracy, mAP50, and mAP50-95 are significant, demonstrating the effectiveness and robustness of the signal recognition network designed for the Industrial Internet.

[0120] Precision (P) and recall (R) represent the false positive rate and false negative rate of the model, respectively. While the YOLOv5 network has a high recall rate, its precision is relatively low, indicating a significant problem with false positives. In contrast, the YOLO-ad network achieves near-perfect precision and recall, demonstrating a balance between the two. It avoids sacrificing accuracy for false negatives and avoids sacrificing complete recall for false positives.

[0121] Figure 12 This is a comparison of the F1 score curves for the two network models in this invention. Since recall and precision are equally important in the detection and identification of frequency-hopping signals, the relative contributions of precision and recall to the F1 score are equal, and this value is a weighted average of precision and recall; therefore, it is introduced for comparison. It can be seen that YOLO-ad has a 23% improvement in this value.

[0122] Figure 13 This is a comparison of the detection capabilities of the two network models in this embodiment. Under the same hardware conditions, the trained models were tested on the time-frequency spectrum. The left image shows the unmodified YOLOv5s network, and the right image shows the YOLO-ad network. It can be seen that the confidence level has been significantly improved, with a large improvement of nearly 20% near interference areas and an improvement of about 8% in areas not affected by interference. However, for situations with small frequency changes resulting in a denser spectrum and interference, the YOLOv5 network exhibits both over-detection and under-detection issues. Figure 13 (a) False detections may occur when signal characteristics are unclear, such as... Figure 13 (b). YOLO-ad can better distinguish this situation. For example... Figure 14 (a) The YOLOv5 network experienced widespread false detections, such as... Figure 14 In (b), this situation did not occur in the YOLO-ad network.

[0123] Example 2

[0124] This embodiment corresponds to the method in Embodiment 1, and proposes a frequency-hopping signal intelligent detection and identification system for the Industrial Internet. The system includes:

[0125] The dataset generation module is configured to: receive a configuration for a Rayleigh channel that simulates multipath effects and shadowing fading; receive configured interference and apply a predetermined type of interference to the sample frequency-hopping signal on the channel, such as one or more of single-tone interference, linear frequency modulation interference, noise frequency modulation interference, radio frequency noise interference, and pulse discontinuity noise interference; and perform a short-time Fourier transform on the output signal to generate a time-frequency graph, thereby constructing a signal detection time-frequency graph dataset.

[0126] The signal detection model training module is configured to: receive a configured target detection network model; receive annotations of modulation categories for all signal detection time-frequency maps in the signal detection time-frequency map dataset; and train the target detection network model using the annotated signal detection time-frequency map dataset to obtain the signal detection model.

[0127] In some embodiments, the signal detection model training module trains the target detection network model according to the following configuration: converting the labeled signal detection time-frequency map dataset into a format suitable for the target detection network model; randomly sampling time-frequency maps from the signal detection time-frequency map dataset at a predetermined ratio to construct training sets, validation sets, and test sets respectively; receiving configured training parameters; training the target detection network model using the training set based on the training parameters; and validating and testing the target detection network model using the validation set and test set respectively.

[0128] See Figure 8 As shown, the target detection network model includes a backbone network, a neck network, and a head network. The backbone network includes a contextual layering module and a gating aggregation module. The contextual layering module is configured to extract multi-level linear features from the input time-frequency map sequentially. The gating aggregation module is configured to aggregate the multi-level linear features output by the contextual layering module into a fused feature vector. The fused feature vector is then processed by the neck network and the head network sequentially to output the recognition result.

[0129] like Figure 9 As shown, in some embodiments, the context layering module includes multiple cascaded depthwise separable convolutional layers, each of which is configured with a GeLU activation function; the last depthwise separable convolutional layer is followed by an average pooling layer; the input time-frequency map is subjected to feature extraction by the multi-level depthwise separable convolutional layers step by step, each level of depthwise separable convolutional layer is activated by GeLU to obtain a first-level linear feature, and the linear feature of the last level of depthwise separable convolutional layer is averaged by an average pooling layer to obtain the last-level linear feature.

[0130] like Figure 10As shown, in some embodiments, the gating aggregation module includes a gating network and an aggregation module, the aggregation module being configured with a weight matrix; the multi-level linear features output by the context layering module are connected to the gating network, the aggregation module is connected to each level of the gating network output, and the weight matrix is ​​used to perform weighted aggregation of each level of the gating network output.

[0131] The system also includes:

[0132] The input module is configured to receive the frequency hopping signal to be detected and perform a short-time Fourier transform to generate a time-frequency diagram of the frequency hopping signal to be detected.

[0133] The output module is configured to use the signal detection model trained by the signal detection model training module to detect and identify the time-frequency graph of the frequency hopping signal to be detected, and identify the modulation category.

[0134] This invention is not limited to the specific embodiments described above. The invention extends to any new feature or combination disclosed in this specification, as well as any new method or process step or combination disclosed herein.

Claims

1. An industrial internet-oriented frequency hopping signal intelligent detection and identification method, characterized in that, include: S1. Construct a Rayleigh channel to simulate multipath effects and shadowing fading, apply predetermined type of interference to the sample frequency hopping signals on the channel, perform short-time Fourier transform on the output signal to generate a time-frequency graph, and construct a signal detection time-frequency graph dataset. S2. Label the modulation category of all signal detection time-frequency maps in the signal detection time-frequency map dataset; train the pre-built target detection network model using the labeled signal detection time-frequency map dataset to obtain the signal detection model; The target detection network model is constructed as follows: the backbone network uses a context-layering module to extract multi-level linear features from the input time-frequency map, and then uses a gated aggregation module to aggregate the multi-level linear features into a fused feature vector; the fused feature vector is then passed through the neck network and the head network to obtain the recognition result. S3. Input the time-frequency diagram of the frequency hopping signal to be detected into the signal detection model to identify the modulation category.

2. The industrial internet of things oriented frequency hopping signal intelligent detection and identification method of claim 1, wherein, The methods for multi-level linear feature extraction in context-layered modules include: The contextual layering module uses multiple cascaded deep separable convolutional layers to extract features from the input time-frequency map step by step. Each level of deep separable convolutional layer is activated by GeLU to obtain a first-level linear feature, and the linear features of the last level of deep separable convolutional layer are obtained by average pooling to obtain the last-level linear feature. 3.The industrial internet of things oriented frequency hopping signal intelligent detection and identification method of claim 2, wherein, The gating aggregation module's methods for multi-level linear feature fusion include: The gated aggregation module uses a gated network to access multi-level linear features and employs a weight matrix to perform weighted aggregation of the outputs of each level of the gated network.

4. The intelligent detection and identification method for frequency hopping signals for the Industrial Internet as described in claim 1, characterized in that, Training an object detection network model using an annotated signal detection time-frequency graph dataset includes: converting the annotated signal detection time-frequency graph dataset into a format suitable for the object detection network model; Time-frequency maps are randomly extracted from the signal detection time-frequency map dataset according to a predetermined ratio to divide the dataset into training set, validation set and test set; Based on the set training parameters, the object detection network model is trained using the training set, and the object detection network model is validated and tested using the validation set and test set, respectively.

5. The intelligent detection and identification method for frequency hopping signals for the Industrial Internet as described in claim 1, characterized in that, The predetermined type of interference includes at least one of single-tone interference, linear frequency modulation interference, noise frequency modulation interference, radio frequency noise interference, and pulse intermittent noise interference.

6. A frequency-hopping signal intelligent detection and recognition system for the Industrial Internet, characterized in that, include: The dataset generation module is configured to receive a configuration for a Rayleigh channel that simulates multipath effects and shadowing fading. In addition to receiving the configured interference, a predetermined type of interference is applied to the sample frequency hopping signal on the channel; And perform short-time Fourier transform on the output signal to generate a time-frequency graph, and construct a signal detection time-frequency graph dataset; The signal detection model training module is configured to: receive a configured target detection network model; receive annotations of modulation categories for all signal detection time-frequency maps in the signal detection time-frequency map dataset; and train the target detection network model using the annotated signal detection time-frequency map dataset to obtain a signal detection model; wherein, the target detection network model includes a backbone network, a neck network, and a head network, the backbone network includes a contextual layering module and a gated aggregation module, the contextual layering module is configured to sequentially perform multi-level linear feature extraction on the input time-frequency map; the gated aggregation module is configured to aggregate the multi-level linear features output by the contextual layering module into a fused feature vector; the fused feature vector is sequentially processed by the neck network and the head network to output the recognition result; The input module is configured to receive the frequency hopping signal to be detected and perform a short-time Fourier transform to generate a time-frequency diagram of the frequency hopping signal to be detected. The output module is configured to: use the signal detection model trained by the signal detection model training module to detect and identify the time-frequency graph of the frequency hopping signal to be detected, and identify the modulation category.

7. The frequency hopping signal intelligent detection and recognition system for the Industrial Internet as described in claim 6, characterized in that, The contextual layering module includes multiple cascaded depthwise separable convolutional layers, each configured with a GeLU activation function; the last depthwise separable convolutional layer is followed by an average pooling layer; the input time-frequency map is subjected to feature extraction by the multi-level depthwise separable convolutional layers, each level of depthwise separable convolutional layer is activated by GeLU to obtain a first-level linear feature, and the linear feature of the last level of depthwise separable convolutional layer is averaged by an average pooling layer to obtain the last-level linear feature.

8. The frequency hopping signal intelligent detection and recognition system for the Industrial Internet as described in claim 7, characterized in that, The gated aggregation module includes a gated network and an aggregation module, the aggregation module being configured with a weight matrix; the multi-level linear features output by the context layering module are connected to the gated network, the aggregation module is connected to each level of the gated network output, and the weight matrix is ​​used to perform weighted aggregation of each level of the gated network output.

9. The frequency hopping signal intelligent detection and recognition system for the Industrial Internet as described in claim 6, characterized in that, The signal detection model training module trains the target detection network model according to the following configuration: converting the labeled signal detection time-frequency map dataset into a format suitable for the target detection network model; randomly extracting time-frequency maps from the signal detection time-frequency map dataset according to a predetermined ratio to construct training sets, validation sets, and test sets respectively; The system receives the configured training parameters, trains the object detection network model using the training set based on these parameters, and validates and tests the object detection network model using the validation set and test set, respectively.

10. The frequency hopping signal intelligent detection and recognition system for the Industrial Internet as described in claim 6, characterized in that, The predetermined type of interference includes at least one of single-tone interference, linear frequency modulation interference, noise frequency modulation interference, radio frequency noise interference, and pulse intermittent noise interference.

Images