Multi-sensor joint signal detection and time-frequency positioning method and device
By employing a multi-sensor joint signal detection method using distributed processing and deep neural networks, the problem of low signal-to-noise ratio in wireless sensor networks is solved, achieving high-precision signal detection and time-frequency positioning, and improving detection accuracy and computational efficiency.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- CHINESE PEOPLES LIBERATION ARMY UNIT 63891
- Filing Date
- 2022-06-29
- Publication Date
- 2026-06-12
AI Technical Summary
In wireless sensor networks, individual sensor nodes have low signal-to-noise ratios and are sensitive to environmental noise and channel fading. Existing attention models struggle to effectively utilize signal information from multiple sensors, resulting in insufficient detection robustness and difficulty in meeting practical needs.
A distributed processing strategy is adopted, which uses deep neural networks to calculate the probability of signal presence and bounding box parameters through multiple independent sensors. A fusion center is used for signal detection and time-frequency localization. Confidence-based soft information fusion is used to avoid error propagation.
It improves the accuracy of multi-sensor signal detection and time-frequency positioning precision, reduces the burden of large sample data transmission, is suitable for distributed receiving scenarios in wireless sensor networks, and improves computing speed and processing capabilities.
Smart Images
Figure CN115310476B_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of radio signal detection technology, specifically to a multi-sensor joint signal detection and time-frequency positioning method and apparatus. Background Technology
[0002] In applications such as radio signal monitoring, classification and identification, demodulation and decoding, and target localization, it is necessary to first determine the existence of signals within the target frequency band and measure parameters such as the spectral location and start and end times of potential unknown signals. In wireless sensor networks, signals acquired by a single node often have a low signal-to-noise ratio and are highly sensitive to environmental noise and channel fading. Using a single sensor node to independently complete signal detection is not robust enough and usually fails to meet practical needs. Therefore, designing robust and efficient detection algorithms to effectively fuse signals from multiple sensors is crucial. Due to differences in receiver unit types, receiving antenna apertures, and distances from the signal source, coupled with the varying characteristics of each independent receiving channel, there are differences in the quality of the observed signals. Fusion of multiple sensor signals requires considering these differences in signal quality and assigning different weights to each received signal. The calculation of these weights has always been a key focus and challenge in research.
[0003] Within the framework of deep neural networks, attention models are commonly used weighted aggregation methods. However, existing attention models are based on the assumption that only a few observations are relevant to the task objective, thus tending to assign near-zero weights to the majority of observations. Currently, in the field of wireless sensor networks, there is a lack of a scheme that can fully utilize the signal information received by each array element to improve detection performance. Summary of the Invention
[0004] In view of this, the present invention provides a multi-sensor joint signal detection and time-frequency positioning method and apparatus. It adopts a distributed processing strategy, in which multiple independent sensors use deep neural networks to calculate the probability of the existence of a target frequency band signal (confidence soft information) and the bounding box parameters such as the start and end times and start and end frequencies of each signal. Then, the data are sent to the fusion center to calculate the joint confidence and correct the bounding box parameters of each signal, so as to finally realize the joint detection of multi-sensor signals and the determination of time-frequency position.
[0005] To achieve the above objectives, the multi-sensor joint signal detection and time-frequency positioning method provided by the present invention includes the following steps:
[0006] First, a deep neural network model is constructed. The network input is a signal waterfall plot, and the network output is the confidence level of the signal's existence in the waterfall plot and the bounding box parameters of the corresponding signal.
[0007] The target frequency band signal data is acquired, and the signal waterfall plot is obtained after preprocessing the target frequency band signal data as a training sample to train the constructed deep neural network model.
[0008] Multiple sensors independently acquire signals, and then a trained deep neural network is used to process the signals acquired by the sensors to obtain the processing results of the signals acquired by a single sensor, including the confidence level of the acquired signal and the bounding box parameters of the signal.
[0009] The processing results of individual sensor signals are fused to obtain the joint confidence of the signal's existence, and the signal bounding box parameters are corrected to ultimately achieve joint detection of multi-sensor signals and time-frequency positioning.
[0010] Furthermore, a deep neural network model is constructed, specifically: the network structure directly uses existing classic object detection networks, including CNN, RCNN, YOLO, and SSD.
[0011] Furthermore, acquiring target frequency band signal data specifically involves: acquiring target frequency band signal data by collecting actual signals or through simulation; the target frequency band of the target frequency band signal data is not unique; the target frequency band width of the target frequency band signal data is not fixed; the target frequency band signal data directly uses radio frequency data, or intermediate frequency or baseband data after frequency conversion.
[0012] Furthermore, after preprocessing the target frequency band signal data, a signal waterfall plot is obtained as a training sample. Specifically, the target frequency band signal data is processed through data segmentation, normalization, and short-time Fourier transform to obtain a signal waterfall plot as a training sample.
[0013] Furthermore, multiple sensors independently acquire signals, specifically: multiple sensors are randomly arranged in space; the receiving antenna types, apertures, and receiver channel characteristics of multiple sensors are the same or different; and the signals acquired by multiple sensors are acquired synchronously or asynchronously.
[0014] Another embodiment of the present invention provides a multi-sensor joint signal detection and time-frequency positioning device, including a deep neural network model module, a training sample acquisition module, multiple independent sensors, and a fusion module;
[0015] The deep neural network model module contains a deep neural network model. The network input is a signal waterfall plot, and the network output is the confidence level of the signal in the waterfall plot and the bounding box parameters of the corresponding signal.
[0016] The training sample acquisition module is used to acquire target frequency band signal data. After preprocessing the target frequency band signal data, a signal waterfall plot is obtained as a training sample to train the constructed deep neural network model.
[0017] Multiple independent sensors acquire signals, which are then fed into a trained deep neural network model for processing. This yields the processing results of individual sensor signals, including the confidence level of the acquired signal and the bounding box parameters of the signal. The processing results of the sensor signals are then fed into a fusion module.
[0018] The fusion module fuses the processing results of individual sensor signals to obtain the joint confidence level of the signal's existence, and corrects the signal bounding box parameters, ultimately achieving joint detection of multi-sensor signals and time-frequency positioning.
[0019] Beneficial effects:
[0020] 1. This method can fuse signals from multiple sensors, which can effectively improve detection accuracy and signal time-frequency positioning accuracy compared to single-channel reception.
[0021] 2. This method employs a distributed processing strategy. Multiple sensor nodes utilize their own observed signals to achieve signal detection and time-frequency position determination. The processed results are then sent to a fusion center for fusion. No signal sampling waveform transmission is required between the sensor nodes and the fusion center, avoiding the communication burden caused by large-sample data transmission. Simultaneously, the distributed processing architecture ensures computational speed and processing capacity, making it suitable for distributed reception applications in wireless sensor networks.
[0022] 3. In this method, confidence fusion is a soft information fusion, which, compared to traditional decision (hard decision) fusion, can effectively avoid the problem of system performance degradation caused by error propagation. In addition, neural networks trained with a large number of samples usually provide higher sensitivity than traditional methods, and can better meet the processing needs of wireless sensor networks receiving low signal-to-noise ratio signals. Attached Figure Description
[0023] Figure 1 This is a schematic diagram of the signal bounding box parameters of the present invention;
[0024] Figure 2 This is a schematic diagram of the network training structure of the present invention;
[0025] Figure 3 This is a schematic diagram of the multi-sensor joint detection structure of the present invention. Detailed Implementation
[0026] The present invention will now be described in detail with reference to the accompanying drawings and embodiments.
[0027] This invention provides a confidence-weighted multi-sensor joint signal detection and time-frequency localization method, comprising the following specific steps:
[0028] Step 1: Construct a deep neural network model. The network input is a signal waterfall plot, and the network output is the confidence level of each signal in the waterfall plot, along with its corresponding time and frequency coordinates. The expression is as follows:
[0029] [p1,x 1min ,y 1min ,x 1max ,y 1max ,L,p i ,x imin ,y imin ,x imax ,y imax ,L] T
[0030] Where, x imin y imin x imax and y imax Let x be the bounding box parameter of the i-th target signal in the waterfall plot sample. imin and y imin The minimum value of the x-coordinate (corresponding to frequency) and y-coordinate (corresponding to time) of the signal bounding box, x imax and y imax This represents the maximum value of the x-coordinate and y-coordinate, p. i This represents the detection confidence level of the i-th signal in the corresponding sample, which characterizes the probability of the signal's existence.
[0031] Step 2: Construct training signal samples. Acquire target frequency band signal data by collecting actual signals or through simulation. Perform data segmentation, normalization, and short-time Fourier transform operations sequentially to obtain a signal waterfall plot as training samples. Each sample contains multiple signals, each with differences in center frequency, bandwidth, signal-to-noise ratio, start and end times, etc.
[0032] Step 3, Deep Neural Network Model Training. The deep neural network model constructed in Step 1 is trained using the training dataset obtained in Step 2 to obtain the trained network parameters;
[0033] Step 4, Multi-sensor signal acquisition. K receiving units acquire signals respectively, and perform preprocessing operations such as data segmentation, normalization, and short-time Fourier transform on each acquired signal to generate waterfall plot samples suitable for deep neural network processing;
[0034] Step 5, Distributed Processing. The waterfall plots of the multiple sensor signals obtained in Step 4 are fed into the deep neural network model trained in Step 3. The deep neural network is then used to calculate the confidence level and bounding box parameters of each signal in the data samples collected by each receiving unit. For the k-th receiving unit, the neural network output is:
[0035]
[0036] Step 6, Multi-sensor fusion processing. The processing results of multiple received signals from Step 5 are sent to the fusion center to calculate the joint confidence score. The signal bounding box parameters are then corrected using the confidence scores of each sensor signal. The specific calculation is as follows:
[0037] Joint confidence level:
[0038]
[0039] Signal bounding box parameter correction:
[0040]
[0041]
[0042]
[0043]
[0044] Joint confidence level P i That is, the probability X that the algorithm gives of the existence of the i-th signal. imin and X imax Y represents the signal start frequency and stop frequency, respectively. imin and Y imax These represent the start and end times of the signal within the observed data segment, respectively. This enables joint detection of signals from multiple sensors and time-frequency positioning.
[0045] Another embodiment of the present invention provides a multi-sensor joint signal detection and time-frequency positioning device, the structure of which is as follows: Figure 3 As shown, it includes a deep neural network model module, a training sample acquisition module, multiple independent sensors, and a fusion module;
[0046] The deep neural network model module contains a deep neural network model. The network input is a signal waterfall plot, and the network output is the confidence score of the signal's presence in the waterfall plot and the corresponding bounding box parameters. The network structure can directly use existing classic object detection networks such as RCNN, Fast RCNN, and YOLO. The network input is the signal waterfall plot, and the output is the detection confidence score of each signal in the waterfall plot and the corresponding time and frequency location coordinates (e.g., ...). Figure 1 (As shown).
[0047] The training sample acquisition module is used to acquire target frequency band signal data. After preprocessing the target frequency band signal data, a signal waterfall plot is obtained as a training sample to train the constructed deep neural network model. In this embodiment of the invention, target frequency band signal data is acquired by acquiring actual signals or through simulation. Data segmentation, normalization, and short-time Fourier transform are performed sequentially to obtain a signal waterfall plot as a training sample. Each sample contains multiple signals, and the center frequency, bandwidth, signal-to-noise ratio, start and end times of each signal are different. The acquired training samples are used to train the constructed deep neural network model using optimization algorithms such as stochastic gradient descent and Adam. The training process is as follows: Figure 2 As shown, the final trained network parameters θ are obtained. * θ * It can be viewed as the maximum a posteriori estimate of the network parameter θ given the training dataset, i.e.
[0048] Multiple independent sensors acquire signals, which are then fed into a trained deep neural network model for processing. This yields the processing results for each sensor's signal, including the confidence level of the acquired signal and its bounding box parameters. These processing results are then fed into a fusion module. In this embodiment, four receiving units are used as an example. Assuming there is only one signal within the target frequency band, each receiving unit acquires a signal and performs preprocessing operations such as data segmentation, normalization, and short-time Fourier transform on its acquired signal to generate a waterfall plot sample suitable for deep neural network processing. Different receiving units have different distances from the signal source, resulting in varying receiving channel characteristics and consequently, differences in the quality of the observed signal.
[0049] The waterfall plots of signals received from multiple sensors are fed into a trained deep neural network model. The deep neural network is then used to calculate the confidence level and bounding box parameters of each signal in the data samples collected by each receiving unit. For the k-th (k = 1, 2, ..., 4) receiving unit, the neural network output is:
[0050] The fusion module fuses the processing results of individual sensor signals to obtain the joint confidence level of the signal's existence, and corrects the signal bounding box parameters, ultimately achieving joint detection of multi-sensor signals and time-frequency positioning.
[0051] In summary, the above are merely preferred embodiments of the present invention and are not intended to limit the scope of protection of the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.
Claims
1. A multi-sensor joint signal detection and time-frequency positioning method, characterized in that, Includes the following steps: First, a deep neural network model is constructed. The network input is a signal waterfall plot, and the network output is the confidence level of the signal in the waterfall plot and the bounding box parameters of the corresponding signal. Acquire target frequency band signal data, preprocess the target frequency band signal data to obtain a signal waterfall plot as a training sample, and train the constructed deep neural network model. Multiple sensors independently acquire signals, and then a trained deep neural network is used to process the signals acquired by the sensors to obtain the processing results of the signals acquired by a single sensor, including the confidence level of the acquired signal and the bounding box parameters of the signal; The processing results of individual sensor signals are fused to obtain the joint confidence of the signal's existence, and the signal bounding box parameters are corrected to ultimately achieve joint detection of multi-sensor signals and time-frequency localization. The multiple sensors independently acquire signals, specifically as follows: Multiple sensors are randomly arranged in space; Multiple sensors may have the same or different antenna types, apertures, and receiver channel characteristics. The signals collected by multiple sensors can be acquired synchronously or asynchronously.
2. The multi-sensor joint signal detection and time-frequency positioning method according to claim 1, characterized in that: The construction of the deep neural network model specifically involves using existing classic object detection networks directly for the network structure.
3. The multi-sensor joint signal detection and time-frequency positioning method according to claim 1 or 2, characterized in that: The acquisition of target frequency band signal data specifically involves acquiring target frequency band signal data by collecting actual signals or through simulation. The target frequency band of the target frequency band signal data is not unique; The target frequency band width of the target frequency band signal data is not fixed; The target frequency band signal data can be directly obtained from radio frequency data, or from intermediate frequency or baseband data after frequency conversion.
4. The multi-sensor joint signal detection and time-frequency positioning method according to claim 1, characterized in that: The process of obtaining a signal waterfall plot as a training sample after preprocessing the target frequency band signal data specifically involves: obtaining a signal waterfall plot as a training sample by performing data segmentation, normalization, and short-time Fourier transform operations on the target frequency band signal data.
5. A multi-sensor combined signal detection and time-frequency positioning device, characterized in that, It includes a deep neural network model module, a training sample acquisition module, multiple independent sensors, and a fusion module; The deep neural network model module includes a deep neural network model. The network input is a signal waterfall plot, and the network output is the confidence level of the signal in the waterfall plot and the bounding box parameters of the corresponding signal. The training sample acquisition module is used to acquire target frequency band signal data, and after preprocessing the target frequency band signal data, obtain a signal waterfall plot as a training sample to train the constructed deep neural network model. Multiple sensors independently acquire signals, which are then fed into a trained deep neural network model for processing. This yields the processing results of individual sensor signals, including the confidence level of the acquired signal and the bounding box parameters. The processing results of the sensor signals are then fed into a fusion module. The fusion module fuses the processing results of individual sensor signals to obtain the joint confidence of the signal's existence, and corrects the signal bounding box parameters, ultimately achieving joint detection of multi-sensor signals and time-frequency localization. The multiple sensors independently acquire signals, specifically as follows: Multiple sensors are randomly arranged in space; Multiple sensors may have the same or different antenna types, apertures, and receiver channel characteristics. The signals collected by multiple sensors can be acquired synchronously or asynchronously.