Federal obstacle perception method based on directional terahertz spectrum environment map assistance

Abstract

The present application relates to communication technology, and provides a kind of federal obstacle perception method based on directional terahertz spectrum environment map auxiliary, comprising: several terahertz communication nodes in real communication scene obtain respective local training data set, and then interact with terahertz base station to realize federal learning.Terahertz base station is based on the global deep neural network obtained to build several directional terahertz spectrum environment maps corresponding to different transmitting antenna beam direction.Based on image segmentation algorithm, extract several signal-free coverage areas in directional terahertz spectrum environment map.Intersection operation is carried out to the corresponding position set of several signal-free coverage areas, and the shape and position information of the obstacle in the current real communication scene are obtained.In this way, the problems of user privacy protection, resource / overhead serious consumption, insufficient obstacle perception accuracy and other problems existing in the prior art when perceiving obstacles can be effectively solved.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of communication, in particular to an obstacle perception method based on a terahertz communication signal. BACKGROUND

[0002] Terahertz refers to electromagnetic waves with a frequency of 0.1 THz to 10 THz.

[0003] Terahertz communication is listed as one of the key candidate technologies for 6G due to its extremely high rate, ultra-large bandwidth, and ultra-high perception resolution, and can support application scenarios such as immersive communication and communication perception integration. However, obstacles in real communication scenarios can easily block the free propagation of terahertz communication signals in space because terahertz communication signals only have weak diffraction ability and extremely narrow beams. Therefore, the coverage range of terahertz communication is often limited by environmental obstacles.

[0004] Therefore, researchers have proposed the concept of communication perception integration in order to achieve obstacle perception while performing wireless communication. Thus, in the terahertz communication scenario, the terahertz communication system can perform frequency band switching, beam switching / tracking, and cell switching in a timely manner through the obtained obstacle information, i.e., to avoid the blocking of terahertz communication signals by obstacles. Thus, the coverage range of terahertz communication is enhanced.

[0005] In a communication perception integration system, traditional obstacle perception methods are mainly achieved by signal processing of actively transmitted signals and corresponding received echoes. Further, although the traditional method can identify the existence of obstacles, it has certain limitations in obstacle shape perception. Specifically, the traditional method mainly relies on line-of-sight signals to achieve direct detection of the surface of obstacles within the line-of-sight range, so that it is difficult to obtain obstacle shape information in the non-line-of-sight range, i.e., it fails to analyze the specific shape of the obstacle in depth. Because, to achieve obstacle shape acquisition based on the traditional method, techniques such as simultaneous localization and mapping (SLAM) or multi-base station detection imaging are often used. However, these techniques not only increase the complexity of the wireless communication system, but also consume more hardware resources / communication overhead. In practice, the above-mentioned traditional obstacle perception method ignores the effective information and additional computing power that can be provided by a large number of communication nodes (such as user equipment and other communication receivers) distributed in the communication process. On the one hand, such communication nodes can measure the terahertz signal power spectrum density information such as terahertz signal power when communicating with the base station, and on the other hand, such communication nodes can also perform edge computing with their own processor resources.

[0006] Researchers often use terahertz spectrum environment maps to characterize the power spectral density of terahertz signals measured by terahertz communication nodes. Specifically, a terahertz spectrum environment map characterizes the spatiotemporal distribution of the signal power spectral density within the terahertz frequency band, as well as the distribution of obstacles. These maps include information such as signal strength and terahertz signal coverage, and also reflect the modulation effects of the terahertz signal in the actual communication scenario. These maps can show how signal intensity varies at different locations. Directivity refers to the extremely narrow beam characteristics of terahertz communication, resulting in directional signal transmission. This means that the signal does not propagate uniformly in all directions, but rather is directed towards a specific area. Therefore, directional terahertz spectrum environment maps further refine this concept, not only documenting the distribution of terahertz signals but also emphasizing the directionality of signal propagation. This means that by varying the beam direction, multiple terahertz spectrum environment maps can be generated in different directions. Each map shows how, within a specific beam direction, a line-of-sight terahertz signal is blocked by obstacles and how it propagates through scattering or reflection to reach the communication node.

[0007] When it comes to edge computing in terahertz communication nodes, researchers often use federated learning to implement privacy-preserving computing tasks at edge nodes. Federated learning is a distributed machine learning method that allows multiple participants to jointly train a model while preserving the privacy of their respective data. In this process, each participant trains a model locally using its own data and then sends the model updates (rather than the local raw data) to a central server. The central server aggregates the data to improve the global model and then sends the improved model back to each participant. In this way, federated learning leverages the data resources distributed across different participants while avoiding the direct sharing of sensitive data, thereby reducing the risk of privacy leaks.

[0008] On July 14, 2022, an invention patent with the publication number CN114900234A and the name “A method and device for constructing a terahertz spectrum environment map” was disclosed. It aims at the dynamically changing communication environment based on the terahertz communication perception integrated system to achieve high-precision construction of the terahertz spectrum environment map, thereby avoiding the repeated deployment of terahertz radio monitoring nodes and large computational overhead. However, this patent only focuses on dynamic scenarios where obstacles change their positions. When performing obstacle perception in static scenes, it is often necessary to deploy multiple terahertz monitoring nodes, resulting in additional hardware resource overhead. In other words, the patent ignores the terahertz communication node information that can be used in actual communication scenarios to achieve obstacle perception.

[0009] The invention patent with the publication number CN117498953A and the name "A terahertz spectrum environment map construction method based on mixed active and passive sensing" was published on February 2, 2024. It realizes the acquisition of coarse-grained obstacle information and fine-grained obstacle information by utilizing terahertz communication node information. However, this patent fails to fully protect the privacy information (such as location information) of communication nodes like user equipment during the passive sensing step; on the other hand, it also does not fully consider the directivity of terahertz signals, making it difficult to distinguish between obstacle regions and signal shadow regions blocked by obstacles in certain beam directions, thereby causing errors in obstacle information acquisition. SUMMARY

[0010] The technical problem to be solved by the present application is that the current obstacle sensing technology requires large hardware resources / communication overhead to acquire obstacle shape information in non-line-of-sight range or static scenes, ignores data privacy protection of terahertz communication nodes, and does not consider the directivity of terahertz signals, leading to errors in obstacle information acquisition. Therefore, a method for implementing obstacle sensing using federated learning assisted by directional terahertz signal power information obtained by terahertz communication nodes is provided.

[0011] The technical means adopted by the present application to solve the above technical problems is a method for constructing a directional terahertz spectrum environment map, comprising the following steps:

[0012] Federated learning step: a plurality of terahertz communication nodes in the real communication scene obtain their respective local training data sets, and then interact with the terahertz base station to realize federated learning, thereby obtaining a global deep neural network for estimating signal power by the terahertz base station, and obtaining a local deep neural network by each terahertz communication node;

[0013] Map construction step: the terahertz base station constructs a plurality of directional terahertz spectrum environment maps corresponding to different transmit antenna beam directions based on the obtained global deep neural network;

[0014] Feature extraction step: extract the signal-free coverage areas in the plurality of directional terahertz spectrum environment maps based on an image segmentation algorithm;

[0015] Federated sensing step: obtain the obstacle shape and location information in the current real communication scene by performing intersection operation on the different beam direction corresponding position sets of the plurality of signal-free coverage areas.

[0016] Specifically, in the federated learning step, the terahertz communication nodes first collect signal data in their respective monitoring areas, including but not limited to node coordinates, distances to the terahertz base station, and other information. These data are used as local training data sets to train local deep neural network models that focus on capturing the propagation characteristics of signals in a specific environment. Further, instead of directly sharing their local data sets, the nodes send updates (such as weights and biases) of the local models to the terahertz base station. The terahertz base station aggregates these updates to improve the global deep neural network model, which can estimate the signal power distribution of the entire real-world communication scenario. This approach not only protects the data privacy of each node but also improves the generalization ability and accuracy of the model by integrating the learning achievements of multiple nodes.

[0017] In the map construction step, the directional terahertz spectrum environment map generated by the terahertz base station can reveal the propagation characteristics of terahertz signals in the current real-world communication scenario relative to different directions, showing the coverage range of terahertz signals while reflecting the modulation effect of obstacles on terahertz signals.

[0018] In the feature extraction step, the signal-free coverage area extracted by the terahertz base station includes the actual obstacle-occupied area and the signal shadow area caused by the obstruction of obstacles to terahertz signals, where the actual obstacle-occupied area does not change with the change of beam direction, while the signal shadow area does. Therefore, in the federated perception step, the two types of areas can be distinguished by intersection operation. Through this method, the terahertz base station can accurately map the shape and location of obstacles in the real-world communication scenario, providing important information for subsequent communication system optimization and planning.

[0019] The beneficial effects of the present application are:

[0020] 1.The present application realizes data privacy protection in the model training process through a federal learning framework. Each terahertz communication node does not need to directly share its sensitive raw data set, but only uploads model parameters such as weights and biases. Therefore, this method effectively prevents the leakage of sensitive data while ensuring that the global model can learn useful features from distributed data, which is of great significance for commercial and personal privacy protection. On the one hand, the present application can avoid the leakage of information such as the range of each user's activities, activity trajectory, and activity preferences in the real communication scenario. On the other hand, considering that there may be multiple operators in the real communication scenario, the present application can also realize the indirect sharing of data of each operator under the management of different operators, thereby ensuring the accurate extraction of obstacle information and the ideal implementation of the corresponding terahertz signal coverage enhancement method. That is, as a distributed machine learning method, federal learning allows each user or each operator to train their own model locally, and then only shares the model updates, not the raw data. Therefore, the present application not only protects the privacy of user data and the commercial sensitive information of operators, but also allows all participants to benefit from more accurate obstacle perception and spectrum environment map construction without directly exchanging potentially sensitive user data.

[0021] 2.Compared with traditional obstacle perception methods or other obstacle perception methods based on terahertz spectrum environment maps, the present application greatly reduces hardware resource consumption / communication overhead / computational overhead. On the one hand, the present application does not need to use techniques such as simultaneous localization and mapping (SLAM) or multi-base station detection imaging, i.e., it does not need to consume additional fixed base station resources and mobile base station resources; on the other hand, the present application also does not need to use multiple terahertz radio monitoring nodes for repeated intensive deployment; at the same time, the federal learning technique used by the present application offloads most of the computing tasks to the local terahertz communication nodes, thereby reducing the computational load of the base station server and helping to prolong the service life of the device, which is difficult to achieve in previous methods.

[0022] 3.Compared with other obstacle perception methods based on terahertz spectrum environment maps, the present application effectively distinguishes the actual obstacle occupying area and the signal shadow area by considering the directivity of terahertz signals in the form of federating several directional terahertz spectrum environment maps, further improving the obstacle perception accuracy in the passive perception link.

[0023] Therefore, the present application can effectively solve the problems of user privacy protection, serious resource / overhead consumption, and insufficient obstacle perception accuracy in the prior art during obstacle perception. BRIEF DESCRIPTION OF DRAWINGS

[0024] Figure 1 Flowchart for Example 1;

[0025] Figure 2 A three-dimensional sketch of obstacles in a real communication scenario provided for Embodiment 1;

[0026] Figure 3 A movable range top view of several terahertz communication nodes in a real communication scenario provided for Embodiment 1;

[0027] Figure 4 A corresponding terahertz beam direction set B(t u ) top view at t u 0 to t

[0028] Figure 5 An illustration of a signal-free coverage area when the beam direction is 0 degrees based on image segmentation provided for Embodiment 1;

[0029] Figure 6 An illustration of a signal-free coverage area when the beam direction is 180 degrees based on image segmentation provided for Embodiment 1;

[0030] Figure 7 A structural block diagram of an electronic device provided for Embodiment 3. DETAILED DESCRIPTION

[0031] In order to enable personnel in the technical field to better understand the scheme of the present application, the technical solutions in the embodiments of the present application will be described clearly and completely below in conjunction with the drawings in the embodiments of the present application. Obviously, the described embodiments are only some of the embodiments of the present application, but not all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those of ordinary skill in the art without creative labor fall within the scope of protection of the present application.

[0032] The federated obstacle perception method assisted by the directional terahertz spectrum environment map will be introduced below. It should be noted that the present application does not limit the spatial dimension of the directional terahertz spectrum environment map, i.e., the directional terahertz spectrum environment estimation map obtained by the present application can be characterized as a power spectral density distribution in a two-dimensional communication scenario, or as a power spectral density distribution in a three-dimensional communication scenario. Further, the present application can realize the planar construction of the terahertz spectrum environment map in a simplified two-dimensional communication scenario, or realize the three-dimensional construction of the terahertz spectrum environment map in an actual three-dimensional communication scenario.

[0033] Embodiment 1

[0034] Please refer to Figure 1 , Figure 1A flowchart of a directional terahertz spectrum environment map assisted federated obstacle perception method provided for Embodiment 1 of the present application. The method can include the following steps:

[0035] Step S110: Several terahertz communication nodes in the real communication scene obtain their respective local training data sets, and then interact with the terahertz base station to realize federated learning, so that the terahertz base station and the terahertz communication nodes can respectively obtain a global deep neural network and a local deep neural network for estimating signal power.

[0036] In an exemplary embodiment, the local training data set of the terahertz communication node includes: the terahertz signal power information corresponding to the grid position coordinates of the terahertz communication node at each time from the first time to the t max time.

[0037] Embodiments of the present application give a possible example of a real communication scene, as Figure 2 shown. Figure 2 A three-dimensional diagram of obstacle information in a real communication scene. The real communication scene contains several obstacles of different sizes and positions.

[0038] In an exemplary embodiment, the step S110 includes sub-steps S111 to S114.

[0039] S111: Divide the real communication scene into N grid grids, and the center coordinates of each grid are marked as where p is the grid index; at the same time, the range occupied by several obstacles distributed in the real communication scene will also be divided according to the divided grids.

[0040] Further, the N c terahertz communication nodes distributed in the real communication scene will move randomly in the real communication scene, and the movable range of each terahertz communication node is ensured to be non-intersecting.

[0041] In an embodiment, Figure 3 corresponding to the real communication scene Figure 2 several terahertz communication nodes movable range top view. Specifically, the figure divides 4 regions with dashed lines, and each region has one terahertz communication node, indicating that the terahertz communication node in the region can only move within its region, and there is no cross-region phenomenon, so as to ensure that the movable range of each terahertz communication node is non-intersecting, and the terahertz communication node position set is where t u represents time, and N c = 4 in this embodiment. In addition, the terahertz base station is located at the central grid position of the real communication scene.

[0042] For the division of the grid, the size of each grid is set as Figure 3 Taking the overhead communication scenario as an example, the side length of the overhead communication scenario is L, and thus the size of each terahertz communication node moving area is Each moving area is further divided into N k grids, each of which has a side length equal to The grid side length is related to the positioning accuracy of the terahertz communication node itself, which ensures that the position of the terahertz communication node at any moment is within the grid, and that at the same moment, only one terahertz communication node position corresponds to one grid position, and there is no case of one grid with multiple coordinates.

[0043] S112: Set t0 as the initial moment in the overhead communication scenario, and the terahertz base station starts periodic downlink communication with N a terahertz communication nodes in the overhead communication scenario from the moment t0, T is the downlink communication period in the overhead communication scenario, and t c is the corresponding moment after u downlink communication periods in the overhead communication scenario. u

[0044] S113: Obtain the terahertz signal power information of each node relative to each beam direction at each moment through the following iterative steps, and then realize the acquisition of the local training data set.

[0045] Let u = 0, start iteration;

[0046] Obtain the position of the terahertz communication node at the moment t u = t0 + uT.

[0047] Set the signal beam direction of the terahertz base station about the downlink communication to be performed at the moment t u , and mark the set as where j is the beam direction index of the downlink communication signal of the terahertz base station, the coverage range of each signal beam is mutually exclusive, and the beam width is the same.

[0048] In an embodiment, the radio frequency signal actively emitted by the terahertz base station to the position of the terahertz communication node can be represented as x b (t) = f b s(t)exp(j2πf c t), where f b and f c are the bth beam precoding vector and carrier frequency corresponding to the beamforming of the terahertz base station, respectively. Thus, the received signal y k,b (t) can be represented as y k,b (t) = h k,b ​(t)x b (t)+z(t), where h k,b (t) and z(t) represent the corresponding time-domain channel vector and additive white Gaussian noise with zero mean and variance σ 2

[0049] Thus, the i-th terahertz communication node can measure the terahertz signal power information Ψ u of the terahertz base station relative to the j-th beam direction at time t i,j (t u ). Therefore, the i-th = 1, 2…N c terahertz communication node can obtain the terahertz signal power information of the N a beam directions corresponding to the grid position

[0050] Let u = u + 1, and execute the above iteration steps again until t u = t max , where t max is the corresponding time when all the obtained terahertz signal power information has met the data set capacity requirement;

[0051] Therefore, the local training data set of the i-th terahertz communication node contains the terahertz signal power information and the corresponding node position information

[0052] Figure 4 is the corresponding terahertz beam direction set B(t u ) of the real communication scene at t0 to t u . In this embodiment, the terahertz base station is equipped with a uniform circular array composed of a plurality of isotropic antenna elements, which generates a high-gain directional beam with almost the same beam width θ B for any beam direction, and the coverage of each directional beam does not intersect. Therefore, in the corresponding two-dimensional scene of this embodiment, we can get

[0053] S114: According to the local training data set of a plurality of terahertz communication nodes in the real communication scene obtained in step S113, realize federated learning with the terahertz base station, and further obtain a global deep neural network and a local deep neural network that can be used by the terahertz base station and the terahertz communication node to estimate signal power.

[0054] ​​The local deep neural network architecture in the embodiment is used to predict the terahertz power information corresponding to the terahertz communication node position. By combining traditional neural network layers and specific Tensor Train (TT) layers, the intrinsic features of the terahertz communication node position data can be effectively extracted and learned, providing accurate prediction for the terahertz power information. The local deep neural network architecture is specifically as follows:

[0055] Multi-layer Perceptron (MLP): The preprocessed input data is fed into a multi-layer perceptron composed of multiple fully connected layers and PReLU activation function layers alternately. The MLP can extract high-level features of the position data, providing strong learning ability for position prediction.

[0056] Tensor Train layer: A series of Tensor Train layers are used to further process and learn the intrinsic structure of the position data. These layers use the Tensor Train decomposition technique to effectively reduce the number of model parameters while preserving the key features of the data, reducing the computational resource requirements of subsequent terahertz base stations, and improving computational efficiency.

[0057] Output layer: The output of the MLP is converted into the final prediction result through a fully connected layer.

[0058] In the embodiment, the local deep neural network training process is specifically as follows: for the local training data set of the plurality of terahertz communication nodes in the field communication scenario obtained in step S113, for the i-th terahertz communication node, the maximum number of local deep neural network training rounds N lm and the maximum number of batch training times N bm in the local training round N e are set, the local training round index is set to N b , the batch training index is set to N e , and the local terahertz signal power information set used in the N b -th batch training in the N e -th local training round and the corresponding terahertz communication node position set are recorded as and where the element index is set to k, and

[0059]

[0060] Let N e = 1 and N b = 1; start executing the iteration step.

[0061] Input the terahertz communication node position set into the local deep neural network deployed at the i-th terahertz communication node The network can output a set of local terahertz signal power estimation information corresponding to a plurality of positions and a plurality of beam directions wherein θ i is a network parameter of the local deep neural network G i , t' is any time in the set of times {t0,..., t max} in the real communication scene, t0 is an initial time in the real communication scene, t max is a corresponding time when all obtained terahertz signal power information has met the data set capacity requirement, represents the terahertz signal power information of the i-th terahertz communication node in each beam direction estimated by the local deep neural network at time t'.

[0062] Mean square error oriented obstacle perception The local deep neural network G i is trained, that is, the training target thereof is:

[0063] Let N b = N b + 1, and execute the above iteration steps again until N b = N bm .

[0064] Let N e = N e + 1, and execute the above iteration steps again until N e = N lm , the designed local deep neural network is implemented complete training, and the trained local deep neural network of N c terahertz communication nodes is obtained.

[0065] In an embodiment, the global deep neural network training process is specifically that a maximum value of the global deep neural network training round number is set as N gm , the global deep neural network training round number index is N g , and N g = 1. Each terahertz communication node uploads the parameters of the respective local deep neural network to the terahertz base station through an uplink terahertz communication link or a low-frequency communication link.

[0066] The terahertz base station aggregates all local deep neural network parameters (weights and biases) based on a parameter average aggregation criterion, and takes the aggregated parameters as the parameters of the global deep neural network deployed in the terahertz base station.

[0067] The parameter averaging aggregation principle involves first initializing a temporary variable with the same shape as the global deep neural network model parameters but a value of zero. Then, all local deep neural networks participating in training are traversed, and the corresponding parameters of each local deep neural network are accumulated into this temporary variable. After the traversal is complete, each parameter in the temporary variable is divided by the total number of local deep neural networks to calculate the average of all parameters. Finally, this average is used to update the parameters of the global deep neural network.

[0068] Furthermore, the terahertz base station sends the averaged aggregated parameters to the local deep neural network deployed in each terahertz communication node through a downlink terahertz communication link or a low-frequency communication link.

[0069] Let N g =N g +1, and perform the above iterative steps again until N g =N gm Finally, a global deep neural network model for estimating signal power is obtained.

[0070] Step S120: constructing a number of directional terahertz spectrum environment maps corresponding to different transmitting antenna beam directions according to the terahertz base station global deep neural network obtained in step S110.

[0071] At any time t>t in a real-world communication scenario max , set the signal beam direction set of the terahertz base station for the downlink communication to be carried out at time t as Where j is the beam direction index of the downlink communication signal of the terahertz base station. The coverage range of each signal beam does not intersect with each other and has the same beam width;

[0072] Let j = 1; start the iteration step;

[0073] Use the trained global deep neural network model to estimate the beam direction as b t,j Each grid r in the real-time communication scene p The corresponding terahertz signal power Ψ j (t,r p ). Furthermore, the above power information can be combined according to the grid position corresponding to the actual communication scenario to obtain the beam direction b t,j Corresponding directional terahertz spectrum environment map Therefore, the directional terahertz spectrum environment map can reflect the spatial distribution of signal power affected by obstacles in the actual communication scene, and then the map and power threshold can reveal information such as the location and shape of obstacles in the actual communication scene.

[0074] Let j = j + 1, and perform the above iterative steps again until j = Na ; thereby, a set of directional terahertz spectrum environment maps corresponding to different beam directions at time t in a real communication scenario can be obtained

[0075] Step S130: Based on the set of directional terahertz spectrum environment maps obtained in step S120, a number of signal-free coverage areas in the directional terahertz spectrum environment maps are extracted based on an image segmentation algorithm.

[0076] In an embodiment, the image segmentation method is exemplified by the OTSU method. In addition to the OTSU method, methods such as fixed threshold segmentation, edge detection-based segmentation, and clustering-based segmentation can also be selected according to different application scenarios and different segmentation targets. However, the essence of the corresponding technical solutions does not deviate from the spirit and scope of the technical solutions of the embodiments of the present application.

[0077] The OTSU method is an algorithm for automatically determining the threshold value of image binarization segmentation. This method automatically selects an optimal threshold value by minimizing the variance of the foreground and background after image binarization, or equivalently maximizing the variance between the foreground and background. Specifically, first, the gray level histogram of the image, i.e., the number of pixels for each gray level, and the total number of pixels in the image are calculated. According to the histogram and the number of pixels, the average gray level of the entire image is calculated. Next, for each gray level in the histogram, it is taken as a potential threshold value, and the image is divided into two parts: the background composed of pixels less than or equal to the threshold value and the foreground composed of pixels greater than the threshold value. Further, for each potential threshold value, the proportion of the number of pixels in the background and the foreground to the total number of pixels is calculated, and these two proportions are taken as the weights of the background and the foreground, respectively. Similarly, for each potential threshold value, the average gray level of the background and the foreground is calculated. Using the weights and the average gray levels of the foreground and the background, the inter-class variance under each potential threshold value is calculated. Based on traversing all potential threshold values, the threshold value that maximizes the inter-class variance is selected as the optimal threshold value. This threshold value is the threshold value automatically calculated by the OTSU method.

[0078] Further, the OTSU method described above will be used to calculate the threshold value and subsequent feature extraction for each directional terahertz spectrum environment map in turn. Specifically, first, let j = 1; start the iteration step;

[0079] For the jth directional terahertz spectrum environment map The OTSU image segmentation method described above is used to select the optimal threshold value Ψ for the power information contained in the directional terahertz spectrum environment map j,threshold .

[0080] Compare each grid terahertz signal power Ψ j (t, rp ) and the relative size of the threshold. If the power value Ψ j (t,r p )>Ψ j,threshold , it is believed that the grid can be covered by the terahertz signal, otherwise it is believed that the terahertz signal cannot cover the grid.

[0081] Based on the grid position information that the terahertz signal cannot cover, the signal-free coverage area corresponding to the j-th directional terahertz spectrum environment map is obtained. where sgn(·) is the sign function.

[0082] Let j = j + 1, and perform the above iterative steps again until j = N a ; Thus, the set of signal-free coverage areas corresponding to different beam directions at time t in the field communication scenario can be obtained

[0083] To further illustrate this step, this embodiment Figure 5 The figure shows the area without signal coverage when the beam direction is 0 degrees based on the OTSU image segmentation method. The white area in the figure is the area with signal coverage; the gray area is the area without signal coverage; the black border and the area inside the frame correspond to the area occupied by obstacles in the actual scene; the gray area outside the black border is the signal shadow area caused by the obstruction of the obstacle. Figure 6 The terahertz spectrum environment segmentation map obtained based on the OTSU image segmentation method when the beam direction is 180 degrees is given. Figure 5 and Figure 6 As can be seen, as the terahertz signal beam direction changes, only the signal shadow area in the no-signal coverage area changes, while the area actually occupied by the obstacle remains unchanged. This phenomenon lays the foundation for the proposal and implementation of step S140.

[0084] Step S140: performing an intersection operation on corresponding position sets in the plurality of areas without signal coverage obtained in step S130 to obtain the shape and position information of obstacles in the current field communication scene.

[0085] In an embodiment, the phenomenon that the signal shadow region of the obstacle actually occupied region in the directional terahertz spectrum environment map is fixed and the signal shadow region changes with the beam direction is used to process the several signal-free coverage regions obtained in step S130. Specifically, in view of the grid division processing performed in step S110 and the image segmentation performed in step S130, the several signal-free coverage regions obtained for each directional terahertz spectrum environment map are represented in the form of a matrix, and each matrix element corresponds to each grid coordinate in the spectrum environment map. Therefore, for the jth directional terahertz spectrum environment map, the set of positions where the terahertz signal power is less than the threshold value Ψ j,threshold can be represented as Further, by performing intersection operation on all position sets, the grid positions in all directional terahertz spectrum environment maps where the signal coverage changes with the beam direction and is fixed can be obtained, that is, the obstacle grid set It should be noted that the signal shadow region in the signal-free coverage region is removed by the intersection operation.

[0086] Embodiment 2

[0087] The embodiment provides a federated obstacle perception method assisted by a directional terahertz spectrum environment map. On the basis of embodiment 1, embodiment 2 provides another input form of a local deep neural network based on feature engineering.

[0088] In embodiment 1, the position set passed by each terahertz communication node is input into the deep neural network deployed locally on the node, so that the local terahertz signal power estimation information set corresponding to the position set and the beam direction set can be output by the local network. Essentially, each local deep neural network learns the mapping relationship between the input coordinates and the output power.

[0089] Therefore, in order to further accurately estimate the terahertz signal power, the input design of the deep neural network is realized based on feature engineering in embodiment 2. The feature engineering mainly includes feature construction and feature fusion, and focuses on three types of information, that is, the terahertz communication node position set, the set of Euclidean distances between the terahertz communication node and the terahertz base station, and the set of angles between the vector from the terahertz base station to the terahertz communication node and the positive direction of the horizontal axis.

[0090] Specifically, in embodiment 2, first, an input feature based on the angle information between the terahertz communication node position and the terahertz base station needs to be constructed. First, the coordinates of the terahertz base station and the terahertz communication node position are determined. Based on the position coordinates of the two, the relative position vector between the two points is calculated, as well as the components of the vector on the x-axis and y-axis. According to the components, the angle between the vector and the positive direction of the x-axis is calculated using the arctangent function. The obtained angle value is converted from radians to degrees as an input feature of the neural network. The advantage of this input feature is that it can effectively utilize the direction information of the terahertz communication node position, providing more rich context information for location-based services and applications. By incorporating angle information into the deep neural network model, the processing capability of the model for terahertz communication node position related tasks, i.e. the terahertz signal power estimation task proposed in the present application, can be improved.

[0091] Further, in embodiment 2, an input feature based on the Euclidean distance between the terahertz communication node position (x i ,y i ,z i ) and the terahertz base station (X, Y, Z) also needs to be constructed. According to the terahertz communication node position and the terahertz base station position, the straight-line distance between the two points is calculated using the Euclidean distance formula . This feature can enrich the input of the deep neural network model, thereby improving the performance and accuracy of the model in processing location related tasks.

[0092] Finally, the terahertz communication node position set, the terahertz communication node and the terahertz base station Euclidean distance set, the angle between the vector of the terahertz base station pointing to the terahertz communication node position and the positive direction of the horizontal axis are fused to form a new feature vector.

[0093] The mode of this embodiment has the following effects:

[0094] 1. Input information enrichment: By taking coordinates, distances, angles and other information as inputs, the model can access more context information, which helps to more accurately predict power. For example, distance and angle may directly affect the size of power, and these information can provide a more comprehensive perspective than simply coordinates.

[0095] 2. Improve generalization ability: Fusing multiple features can help the model learn more generalized feature representations, reduce dependence on specific data distribution, and thus improve the model's performance on unseen data.

[0096] 3. Reduce learning difficulty: For some complex mapping relationships, it may be difficult to learn the mapping from coordinates to power directly. Introducing additional information such as distance and angle can serve as an indirect hint to help the model more easily learn this mapping relationship.

[0097] Embodiment 3

[0098] Please refer to Figure 7 , Figure 7 A structural block diagram of an electronic device 900 is provided for Embodiment 3 of the present application. The electronic device 900 in the present application can include one or more of the following components: a processor 910, a memory 920, and one or more application programs, wherein the one or more application programs can be stored in the memory 920 and configured to be executed by the one or more processors 910, and the one or more programs are configured to perform the methods as described in the foregoing method embodiments.

[0099] The processor 910 can include one or more processing cores. The processor 910 connects various parts within the entire electronic device 900 through various interfaces and lines, and performs various functions of the electronic device 900 and processes data by running or executing instructions, programs, code sets or instruction sets stored in the memory 920, and calling data stored in the memory 920. Alternatively, the processor 910 can be implemented in at least one of a hardware form of a digital signal processing (Digital Signal Processing, DSP), a field-programmable gate array (Field-Programmable Gate Array, FPGA), and a programmable logic array (Programmable Logic Array, PLA). The processor 910 can integrate a combination of one or several of a central processing unit (Central Processing Unit, CPU) and a modem, etc. Among them, the CPU mainly processes operating systems and application programs, etc.; the modem is used to process wireless communication. It can be understood that the above-mentioned modem can also not be integrated into the processor 910, but can be realized by a separate communication chip.

[0100] The memory 920 can include a random access memory (Random Access Memory, RAM) and can also include a read-only memory (Read-Only Memory, ROM). The memory 920 can be used to store instructions, programs, codes, code sets or instruction sets. The memory 920 can include a program storage area and a data storage area, wherein the program storage area can store instructions for implementing an operating system, instructions for implementing at least one function, instructions for implementing each of the following method embodiments, etc. The data storage area can also store data created by the electronic device 900 in use, etc.

[0101] Finally, it should be noted that the above examples are only used to illustrate the technical solutions of the present application, and are not intended to limit the same; although the present application has been described in detail with reference to the foregoing examples, those of ordinary skill in the art will understand that they can still modify the technical solutions described in the foregoing examples, or make equivalent replacements for some of the technical features; and these modifications or replacements do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present application.

Claims

1. A federal obstacle perception method based on directional terahertz spectrum environment map assistance, characterized in that: The following steps are involved: Federated learning steps: Several terahertz communication nodes in a field communication scenario obtain their own local training data sets and then interact with the terahertz base station to implement federated learning. The terahertz base station thus obtains a global deep neural network for estimating signal power, and each terahertz communication node obtains a local deep neural network. Map construction steps: The terahertz base station constructs several directional terahertz spectrum environment maps corresponding to different transmitting antenna beam directions based on the obtained global deep neural network; Feature extraction step: extracting the signal-free coverage areas in several directional terahertz spectrum environment maps based on the image segmentation algorithm; Federal perception steps: By performing an intersection operation on the corresponding position sets of different beam directions in several areas without signal coverage, the shape and position information of obstacles in the current field communication scenario are obtained.

2. The method according to claim 1, wherein: The specific steps for a terahertz communication node to obtain a local training data set are: Divide the field communication scene into N grid The center coordinates of each grid are marked as r p , where p is the grid index, p = 1,…,N grid ; Set t0 as the initial time in the field communication scenario, t max The corresponding moment when all the terahertz signal power information obtained meets the data set capacity requirement, the terahertz base station starts from time t0 with N a beam directions and N in real-world communication scenarios c Mobile terahertz communication nodes perform periodic downlink communication, T is the downlink communication period in the field communication scenario, t u is the corresponding time after u downlink communication cycles in the field communication scenario; Let u = 0 and start the iteration step; Get t u = the position of the terahertz communication node at time t0+uT, and mark its set as Where i is the index of the terahertz communication node. The movable ranges of each terahertz communication node do not intersect with each other, and the positions are randomly distributed; i = 1,…,N c , t u =t0,…,t max ; Set t u The direction of each signal beam of the terahertz base station for the downlink communication to be carried out at this moment And mark the collection as Where j is the beam direction index of the downlink communication signal of the terahertz base station. The coverage range of each signal beam does not intersect with each other and has the same beam width; j = 1, ..., N a ; t u At each moment, the terahertz base station sets the beam direction B(t u ) and N c The terahertz communication nodes perform downlink communication, so the i-th terahertz communication node can obtain the corresponding grid position About N a Terahertz signal power information in each beam direction and save them separately; Each terahertz communication node is u Time to t u+1 Random movement within a moment; Let u=u+1 and execute the above iterative steps again until t u =t max ; Therefore, t u The local training data set of the i-th terahertz communication node at time contains the terahertz signal power information and corresponding node location information 3. The method according to claim 1, wherein: The specific steps for the terahertz communication node and the terahertz base station to interact and implement federated learning are as follows: Set the maximum number of global deep neural network training rounds to N gm , the maximum number of local deep neural network training rounds N lm , the global deep neural network training round number index is N g ; Let N g =1; Start executing the iteration steps; Each terahertz communication node performs N training on the local deep neural network based on the local training data set. lm rounds of training; Each terahertz communication node uploads the parameters of its local deep neural network to the terahertz base station through an uplink terahertz communication link or a low-frequency communication link; The terahertz base station aggregates all deep neural network parameters based on the parameter average aggregation criterion, and uses the aggregated parameters as the parameters of the global deep neural network deployed on the terahertz base station; The terahertz base station sends the aggregated parameters to the local deep neural network deployed in each terahertz communication node through a downlink terahertz communication link or a low-frequency communication link; Let N g =N g +1, and perform the above iterative steps again until N g =N gm Federated learning achieves complete training.

4. The method according to claim 3, wherein: The specific implementation steps of the local deep neural network are: Set the maximum number of local deep neural network training rounds N corresponding to the i-th terahertz communication node lm and the maximum number of batch training times N in the local training round bm , set the local training round index to N e , the batch training index is N b , respectively record the Nth e The Nth local training round b The local terahertz signal power information set used for the next batch training is: And its corresponding terahertz communication node position set is The element index is set to k; Let N e =1 and N b =1; start executing the iteration step; Set the location of terahertz communication nodes Input to the local deep neural network deployed in the i-th terahertz communication node j is the beam direction index of the downlink communication signal of the terahertz base station, N a is the total number of beam directions, so the network can output a set of local terahertz signal power estimation information corresponding to several positions and several beam directions Among them, θ i is the local deep neural network G i The network parameters, t′ is the time set {t0,…,t max }, t0 is the initial moment in the field communication scenario, t max is the corresponding moment when the power information of all terahertz signals obtained meets the data set capacity requirement, represents the terahertz signal power information of the i-th terahertz communication node in each beam direction estimated by the local deep neural network at time t′; Mean square error of obstacle perception For the local deep neural network G i For training, the training goal is: minθ i V MSE (G i ); card is the number of elements in the collection; and They are and The kth element in ; Let N b =N b +1, and perform the above iterative steps again until N b =N bm ; Let N e =N e +1, and perform the above iterative steps again until N e =N lm The local deep neural network designed at the time is fully trained.

5. The method according to claim 1, wherein: The specific steps of map construction are: The corresponding time t when all the terahertz signal power information obtained in the field communication scenario has met the data set capacity requirements max At any time t thereafter, t>t max , set the signal beam direction b of the terahertz base station for the downlink communication to be carried out at time t t,j The collection is Where j is the beam direction index of the downlink communication signal of the terahertz base station. The coverage range of each signal beam does not intersect with each other and has the same beam width; Let j = 1; start the iteration step; Use the trained global deep neural network model to estimate the beam direction as b t,j The terahertz signal power corresponding to each grid in the real-time communication scenario Then According to the grid position combination corresponding to the actual communication scene, the beam direction b is obtained t,j Corresponding directional terahertz spectrum environment map collection N grid The total number of grids divided for the field communication scene, the center coordinate of each grid is marked as r p , p is the grid index; Let j = j + 1, and perform the above iterative steps again until j = N a , we get the set of directional terahertz spectrum environment maps corresponding to different beam directions in the field communication scene at time t 6. The method according to claim 5, wherein: The specific steps of feature extraction are: Let j = 1; start the iteration step; For the jth directional terahertz spectrum environment map set The optimal threshold Ψ for the power information contained in the directional terahertz spectrum environment map is selected using the threshold-based OTSU image segmentation method, i.e., by maximizing the between-class variance. j,threshold ; Compare the jth directional terahertz spectrum environment map set The terahertz signal power per grid Ψ j (t,r p ) and the relative size of the threshold; if the power value Ψ j (t,r p )>Ψ j,threshold , it is considered that the grid can be covered by the terahertz signal, otherwise it is considered that the terahertz signal cannot cover the grid; Based on the grid position information that cannot be covered by the terahertz signal, the set of directional terahertz spectrum environment maps at time t and the jth time is obtained. Corresponding no signal coverage area where sgn(·) is the sign function, represents the set of grid positions without signal coverage corresponding to the j-th beam direction at time t; Let j = j + 1, and perform the above iterative steps again until j = N a ; Thus, the set of signal-free coverage areas corresponding to different beam directions at time t in the field communication scenario can be obtained 7. The method according to claim 6, wherein: The specific steps of federated awareness are: For some areas without signal coverage The corresponding position set of different beam directions in Perform intersection operation to get the obstacle grid set Extract obstacle raster collection The shape and position information of obstacles in the actual communication scene can be obtained based on the grid coverage and its center position coordinates.

8. A computer device comprising a memory, a processor, and a computer program stored in the memory, wherein: The processor executes the computer program to implement the steps of the method according to claim 1.

9. A computer program product comprising a computer program / instructions, characterized in that When the computer program / instructions are executed by a processor, the steps of the method according to claim 1 are implemented.

Images