Artificial intelligence-based wireless communication data processing method

Abstract

The present application relates to the technical field of DOA estimation, and particularly relates to a wireless communication data processing method based on artificial intelligence. The method comprises the following steps: obtaining first probabilities corresponding to each pixel point in a connected domain of a target object in an acquired factory ground image according to the coordinates of the pixel points; constructing a semantic distribution probability graph according to the first probabilities corresponding to each pixel point in the connected domain in the factory ground image; constructing an RSSI sequence according to RSSI values corresponding to each RSSI receiver; inputting correlation coefficients of left array element RSSI differential sequences and right array element RSSI differential sequences corresponding to the RSSI sequence into a DOA range estimation network to construct an angle range probability graph corresponding to the target object; and obtaining a DOA corresponding to the target object according to a semantic distribution probability graph, an angle range probability graph and an RSSI sequence corresponding to the target object. The present application uses computer vision to assist wireless communication to realize DOA estimation, so that the obtained DOA is more accurate.

Description

[0001] The present application is a divisional application of Chinese Patent Application No. 202211543977.9, filed on December 3, 2022, entitled "Wireless Communication Data Processing Method Based on Artificial Intelligence". TECHNICAL FIELD

[0002] The present application relates to the technical field of DOA estimation, in particular to a wireless communication data processing method based on artificial intelligence. BACKGROUND

[0003] Research on DOA estimation methods has been ongoing for many years, during which many classic algorithms have been proposed. Common DOA estimation algorithms are generally based on array signal processing theory, and most of them habitually use the phase or frequency difference information of the array element received signal as the positioning parameter. In practical engineering, such algorithms rely on sufficient snapshot numbers and accurate time correction. However, when the environment is complex, the accuracy of the algorithm may be affected by multiple factors in the environment, resulting in low DOA precision. SUMMARY

[0004] In order to solve the problem of low precision of existing DOA estimation algorithms, the purpose of the present application is to provide a wireless communication data processing method based on artificial intelligence, and the technical solution adopted is as follows:

[0005] The present application provides a wireless communication data processing method based on artificial intelligence, which comprises the following steps:

[0006] Obtaining a factory ground image;

[0007] According to the coordinates of each pixel point in the connected domain corresponding to the target object in the factory ground image, obtaining the first probability corresponding to each pixel point in the connected domain, the first probability corresponding to each pixel point being the probability of each pixel point being the centroid of the connected domain; according to the first probability corresponding to each pixel point in the connected domain in the factory ground image, constructing a semantic distribution probability graph corresponding to the target object;

[0008] Obtaining the RSSI value of the RSSI transmitter of the target object to each RSSI receiver, constructing an RSSI sequence according to the RSSI value corresponding to each RSSI receiver; dividing the RSSI sequence according to the center point of the antenna array, obtaining a left array element RSSI sequence and a right array element RSSI sequence;

[0009] Obtaining the correlation coefficient of the left array element RSSI difference sequence and the right array element RSSI difference sequence and inputting it into the trained DOA range estimation network to obtain the probability of the DOA range corresponding to the target object being different angle ranges, constructing an angle range probability graph corresponding to the target object;

[0010] According to the semantic distribution probability graph corresponding to the target object, the corresponding angle range probability graph and the RSSI sequence, the DOA corresponding to the target object is obtained.

[0011] Preferably, according to the coordinates of each pixel point in the connected domain corresponding to the target object in the factory ground image, the first probability corresponding to each pixel point in the connected domain is obtained, comprising:

[0012] The connected domain of the target object in the factory ground image is extracted;

[0013] The coordinates of each pixel point in the connected domain corresponding to the target object are fitted by a Gaussian function, and a Gaussian distribution probability density function corresponding to the connected domain is obtained.

[0014] The coordinates of each pixel point in the connected domain are brought into the corresponding Gaussian distribution probability density function to obtain the first probability corresponding to each pixel point in the connected domain.

[0015] Preferably, the method for obtaining the correlation coefficient of the left array element RSSI differential sequence and the right array element RSSI differential sequence comprises:

[0016] According to the center point of the antenna array, the RSSI sequence is divided, and the RSSI sequence in the first half of the RSSI sequence is recorded as the left array element RSSI sequence, and the RSSI sequence in the second half of the RSSI sequence is recorded as the right array element RSSI sequence;

[0017] The difference between each adjacent two elements in the left array element RSSI sequence is obtained, and the left array element RSSI differential sequence is obtained.

[0018] The difference between each adjacent two elements in the right array element RSSI sequence is obtained, and the right array element RSSI differential sequence is obtained.

[0019] The correlation coefficient of the left array element RSSI differential sequence and the right array element RSSI differential sequence is calculated.

[0020] Preferably, the correlation coefficient of the left array element RSSI differential sequence and the right array element RSSI differential sequence corresponding to the RSSI sequence is input into the trained DOA range estimation network to construct the angle range probability graph corresponding to the target object, comprising:

[0021] The correlation coefficient of the left array element RSSI differential sequence and the right array element RSSI differential sequence corresponding to the RSSI sequence is input into the trained DOA range estimation network to obtain the probability of the DOA range corresponding to the target object being different angle ranges.

[0022] The probability of different angle ranges is taken as the second probability corresponding to the pixel points in the corresponding range of the factory ground image.

[0023] The second probability of each pixel point in the factory ground image is taken as a pixel value of a corresponding position pixel point to obtain an angle range probability map.

[0024] Preferably, the DOA corresponding to the target object is obtained according to the semantic distribution probability map corresponding to the target object, the angle range probability map corresponding to the target object and the RSSI sequence, and the DOA corresponding to the target object comprises:

[0025] The semantic distribution probability map corresponding to the target object and the angle range probability map corresponding to the target object are added to obtain a probability map of the target object, and a pixel value corresponding to each pixel point in the probability map is a sum of a first probability and a second probability corresponding to each pixel point;

[0026] A DOA estimation network is constructed, and the DOA estimation network comprises a convolutional neural network and a fully connected network.

[0027] The probability map of the target object is input into the convolutional neural network to obtain a feature map, then the feature map is flattened by a Flatten operation to obtain a feature vector, then the feature vector and the RSSI sequence of the target object are subjected to a Concatenate joint operation to obtain a final feature vector, and the final feature vector is input into the fully connected network for fitting to obtain the DOA corresponding to the target object.

[0028] The training set of the DOA estimation network is a historical probability matrix, and label data of the historical probability matrix is a DOA of a target object corresponding to a historical probability map.

[0029] Preferably, the method for obtaining the factory ground image comprises:

[0030] A plurality of local factory images are obtained.

[0031] The plurality of local factory images are processed by image stitching to obtain a factory ground panoramic image.

[0032] The factory ground panoramic image is processed by using the trained semantic segmentation network to obtain the factory ground image.

[0033] The embodiments of the present application have the following beneficial effects:

[0034] The application fuses a semantic distribution probability graph corresponding to a target object in a factory ground image and an angle range probability graph corresponding to the target object to obtain a DOA corresponding to the target object, wherein a first probability corresponding to each pixel point in the semantic distribution probability graph is a probability that each pixel point is a mass point of a connected domain corresponding to the target object, and a probability value corresponding to each pixel point in the angle range probability graph is a probability that the target object is in different angle ranges, and the probability value corresponding to each pixel point in the angle range probability graph is obtained by inputting a correlation coefficient of a left array element RSSI differential sequence and a right array element RSSI differential sequence corresponding to an RSSI sequence of the target object into a trained DOA range estimation network.

[0035] The application realizes DOA estimation by computer vision and wireless communication cooperation, first represents a probability that a pixel point is a position corresponding to a target by a possibility that the pixel point is a mass center of a connected domain, analyzes the position of the target object from the perspective of computer vision, then obtains RSSI values of an RSSI transmitter of the target object to each RSSI receiver in an antenna array to form an RSSI sequence, divides the RSSI sequence into a left array element RSSI differential sequence and a right array element RSSI differential sequence, obtains a DOA range corresponding to the target object according to the correlation of the two sequences, considers that there is a spacing between array elements in the antenna array and the array elements have a certain arrangement order, estimates the DOA range by the correlation of the two differential sequences, and then represents a probability that the target object is in an angle range by a possibility that a DOA range corresponding to the target object is in different angle ranges, fully utilizes arrangement characteristics of the RSSI sequence to analyze the position of the target object, further fuses the two to more accurately reflect the direction of the target object, and thus more accurately obtains the DOA of the target object. The application uses artificial intelligence to assist wireless communication to realize DOA estimation, makes the obtained DOA more accurate, and combines the RSSI sequence with the DOA estimation network to avoid the problem of prediction error caused by multiple target objects in the same angle range. BRIEF DESCRIPTION OF DRAWINGS

[0036] In order to more clearly illustrate the technical solutions in the embodiments of the present application or the prior art, and the advantages thereof, below, a brief introduction will be given to the drawings needed in the embodiments or prior art description. Obviously, the drawings in the following description are only some embodiments of the present application, and for those skilled in the art, other drawings can be obtained without creative labor based on these drawings.

[0037] Figure 1 A flowchart of the wireless communication data processing method based on artificial intelligence provided by the present application. DETAILED DESCRIPTION

[0038] In order to further illustrate the technical means and functional effects taken by the present application to achieve the predetermined inventive objectives, the following detailed description of the artificial intelligence-based wireless communication data processing method according to the present application is provided in conjunction with the accompanying drawings and preferred embodiments.

[0039] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this application belongs.

[0040] The specific scheme of the artificial intelligence-based wireless communication data processing method provided by the present application is described in detail below in conjunction with the accompanying drawings.

[0041] Embodiment of the artificial intelligence-based wireless communication data processing method:

[0042] As shown in Figure 1 The artificial intelligence-based wireless communication data processing method of the present embodiment includes the following steps:

[0043] Step S1, obtaining a factory ground image.

[0044] The estimation method of the present embodiment is described taking the scenario of factory robot automated transportation as an example, and the present embodiment is also applicable to environments similar to the scenario.

[0045] In order to estimate the DOA (direction of arrival) of the wireless communication signal of the target object in the factory, the present embodiment installs an RSSI transmitter at the center position of each target object in the factory, and deploys an RSSI receiver at a fixed position in the factory to receive the information emitted by the RSSI transmitter on the target object, which needs to be within the monitoring range of the camera. The target object in the present embodiment is a robot in the factory.

[0046] The RSSI is the received signal strength indication, which is the value after RSS conversion through artificial processing; the present embodiment uses RSSI, and as other implementation manners, RSS can also be used directly.

[0047] The physical quantity of RSSI in the present embodiment does not need to rely on close time synchronization and high sampling rate when measuring, so it can avoid the limitations existing in phase-based algorithms and solve the problem of DOA estimation in a low-cost manner.

[0048] In order to accurately map the target object and the receiver in the factory to an image, the embodiment obtains a ground panoramic image of the factory; using a single camera to obtain the ground panoramic image of the factory will cause an angle difference in the camera monitoring picture far from the camera optical center, so the embodiment deploys several cameras at the top of the factory to obtain multiple local factory images of different areas in the field of view corresponding to the cameras, and then uses image stitching technology to obtain the ground panoramic image of the factory to avoid such errors and improve the accuracy of target object projection. In order to ensure that image stitching can be realized, the monitoring pictures of adjacent cameras should have overlapping parts, and the images captured by each camera should be consistent in time, and the captured pictures should not differ greatly in the vertical direction to ensure the real-time and accuracy of imaging.

[0049] The process of image stitching is as follows: first, image registration is performed, that is, a certain matching strategy is adopted to find the corresponding positions of the feature points in the to-be-stitched image (i.e., the local factory image) in the reference image, and then the transformation relationship between the two images is determined. This step is to perform image registration on the local factory images collected by two adjacent cameras. There are various methods for extracting image feature points, such as SIFT, SURF, ORB, etc., which can all be used for extracting image feature points in the image stitching process.

[0050] Then, according to the corresponding relationship between the image features, the values of each parameter in the mathematical model are calculated, and a mathematical transformation model of the two images is established, that is, a homography matrix is solved. Further, according to the established mathematical conversion model, the to-be-stitched image is converted into the coordinate system of the reference image, thereby completing the transformation of the unified coordinates. Finally, the overlapping area of the to-be-stitched image is fused to obtain a stitched and reconstructed ground panoramic image of the factory, and the fusion method is, for example, a weighted fusion method. The image stitching process described in the embodiment is a prior art and can be realized through the OpenCV library, which will not be described here.

[0051] In order to estimate the DOA of the target object, the embodiment performs target perception on the obtained ground panoramic image of the factory to extract the target object. The embodiment processes the ground panoramic image of the factory using a trained semantic segmentation network to obtain a ground image of the factory, specifically:

[0052] First, the semantic segmentation network is trained, and the embodiment first obtains a plurality of historical factory ground panoramic images, takes the obtained plurality of historical factory ground panoramic images as a training set of the semantic segmentation network, then labels the label data of each historical factory ground panoramic image, and the label data is divided into two categories of others and robots. Then, according to each historical factory ground panoramic image and the label data corresponding to each historical factory ground panoramic image, the semantic segmentation network is trained, and a trained semantic segmentation network is obtained, and the training process is as follows:

[0053] First, each historical factory ground panoramic image is normalized to have a value range in the interval [0, 1], which is helpful for network convergence. Then, the normalized historical factory ground panoramic image and the corresponding label data (the label data is processed by one-hot encoding, and the corresponding label image needs to be labeled by a person, the pixel value of the pixel point corresponding to the robot is 1, that is, the pixel value of the pixel point corresponding to the target object is 1; the pixel value of the pixel point of other categories is 0.) are sent into the network for training. The network includes an encoder and a decoder, wherein the encoder is used for feature extraction of the input image data, and outputs a corresponding feature map; the input of the decoder is the feature map output by the encoder, and then the input feature map is up-sampled and fitted, and the output is a target segmentation map. In this embodiment, the loss function for training the semantic segmentation network is a cross-entropy loss function, and the optimization method of the network parameters can be SGD, Adam, etc. The specific network model can be Unet, Enet, etc. Finally, the argmax operation is performed on the target segmentation map output by the network to obtain a factory ground image, which is a binary image, wherein the pixel value of the pixel point corresponding to the robot is 1, that is, the pixel value of the pixel point corresponding to the target object is 1; the pixel value of the pixel point of other categories is 0.

[0054] In step S2, according to the coordinates of each pixel point in the connected domain corresponding to the target object in the factory ground image, a first probability corresponding to each pixel point in the connected domain is obtained, and the first probability corresponding to each pixel point is the probability that each pixel point is the centroid of the connected domain; and according to the first probability corresponding to each pixel point in the connected domain in the factory ground image, a semantic distribution probability map corresponding to the target object is constructed.

[0055] In order to more accurately determine the position of the target object in the factory ground image, the embodiment first extracts the connected domain of the factory ground image, that is, the connected domain corresponding to the target object, whose pixel value is 1. The method of connected domain extraction is, for example, stroke-based labeling, contour-based labeling, etc.

[0056] Since the connected domain corresponding to the target object is composed of multiple pixel points, the embodiment obtains the coordinates of each pixel point in the connected domain corresponding to the target object, for example, a pixel point is in the xth row and yth column in the figure, and the corresponding coordinates are (x, y). Then the pixel point coordinates in the connected domain corresponding to the target object are fitted by using a Gaussian function fitting method to obtain a Gaussian distribution probability density function of the connected domain corresponding to the target object. The greater the function value corresponding to the coordinate point in the Gaussian distribution probability density function, the greater the possibility that the corresponding pixel point is the centroid of the corresponding connected domain.

[0057] The coordinates of each pixel point in the connected domain corresponding to the target object are substituted into the Gaussian distribution probability density function corresponding to the connected domain to obtain the first probability value of each pixel point in the connected domain. The first probability value is a normalized probability value, and the first probability corresponding to each pixel point is the probability that each pixel point is the centroid of the connected domain. Because the probability value of the centroid in the two-dimensional Gaussian distribution is the largest, and the probability becomes smaller and smaller outward, that is, the greater the first probability of the pixel point, the greater the probability that the corresponding pixel point is the centroid of the connected domain. Finally, the first probability calculation is performed for each target object in the panoramic image. Then, the embodiment sets the first probability of the pixel point corresponding to other categories to 0, that is, the first probability of the pixel point with a pixel value of 0 in the factory floor image is set to 0, which means that these pixel points cannot be the position of the target object.

[0058] Finally, the first probability corresponding to each pixel point in the factory floor image is mapped to a semantic distribution probability map which is the same size as the factory floor image. The pixel value of each pixel point in the semantic distribution probability map is the first probability corresponding to each pixel point in the factory floor image. That is, the embodiment reflects the probability of the centroid position of the target object by the first probability. The greater the first probability, the more likely it is the centroid position of the target object.

[0059] In step S3, the RSSI values of the RSSI transmitter of the target object to each RSSI receiver are obtained, and an RSSI sequence is constructed according to the RSSI values corresponding to each RSSI receiver. The RSSI sequence is divided according to the center point of the antenna array to obtain a left array element RSSI sequence and a right array element RSSI sequence. The correlation coefficients of the left array element RSSI differential sequence and the right array element RSSI differential sequence are obtained and input into the trained DOA range estimation network to obtain the probability that the DOA range corresponding to the target object is a different angle range, and an angle range probability map corresponding to the target object is constructed.

[0060] Common antenna arrays include uniform linear arrays, uniform circular arrays and cross arrays. Generally, based on the DOA estimation of the antenna array, in order to accurately determine the incident angle of the signal, it is necessary to improve the array antenna of the receiver and enhance the directivity of the antenna array. The embodiment does not need to improve the antenna array, but uses a common uniform linear array to realize the DOA estimation through computer vision and wireless communication cooperation, and obtains a more accurate signal incident direction, i.e. the DOA corresponding to the target object. The embodiment first estimates the DOA range of the target object, specifically:

[0061] In the embodiment, for each target object, a wireless communication signal is sent by the RSSI transmitter to all RSSI receivers in the antenna array. The antenna array includes M sensors, i.e. there are M RSSI receivers, so when the RSSI transmitter of the target object sends a signal, there will be M RSSI receivers to receive it. Since there is a gap between the receivers, the RSSI values obtained will be different, i.e. each RSSI receiver corresponds to an RSSI value. Therefore, when the RSSI transmitter of a target object sends a signal, it will correspond to M RSSI values.

[0062] The embodiment first obtains the RSSI values of the RSSI transmitter of the target object to each RSSI receiver in the antenna array, and then constructs an RSSI sequence according to the obtained RSSI values of each RSSI receiver in the antenna array. The RSSI sequence is arranged in a specified order, for example, from sensor 1 to M, and each target object corresponds to an RSSI sequence.

[0063] Since there is a gap between the elements in the antenna array, the distances between the target object and each element in the antenna array will be different, so it is difficult to determine the direction of the target object. The embodiment divides the RSSI sequence into a left element RSSI difference sequence and a right element RSSI difference sequence, and obtains the DOA range corresponding to the target object according to the correlation of the two sequences, specifically:

[0064] The embodiment takes the center point of the antenna array as the center to divide the RSSI sequence, and records the first half of the RSSI sequence in the RSSI sequence as the left array element RSSI sequence, and records the second half of the RSSI sequence in the RSSI sequence as the right array element RSSI sequence. Because the distance between sensors is small, the RSSI values of the sensors are small, and the correlation coefficient of the left array element RSSI sequence and the right array element RSSI sequence cannot determine the direction of the target object. Therefore, the embodiment obtains the left array element RSSI difference sequence and the right array element RSSI difference sequence corresponding to the RSSI sequence, specifically: the difference between each adjacent two elements in the left array element RSSI sequence is obtained to obtain the left array element RSSI difference sequence; the difference between each adjacent two elements in the right array element RSSI sequence is obtained to obtain the right array element RSSI difference sequence. The length of the left array element RSSI difference sequence is one less than the length of the left array element RSSI sequence; similarly, the length of the right array element RSSI difference sequence is one less than the length of the right array element RSSI sequence.

[0065] In order to use the left array element RSSI difference sequence and the right array element RSSI difference sequence corresponding to the target object to roughly determine the direction of the target object, the embodiment uses the Pearson correlation coefficient method to measure the left array element RSSI difference sequence and the right array element RSSI difference sequence, and obtains the Pearson correlation coefficient of the two difference sequences, specifically:

[0066] The embodiment first forms a [2, N] matrix, that is, a 2-row N-column matrix, where N is the length of the difference sequence, and then uses the Pearson correlation coefficient method to obtain the Pearson correlation coefficient matrix. The values in the Pearson correlation coefficient matrix are between -1 and 1, where -1 represents strong negative correlation, +1 represents strong positive correlation, and 0 represents no relationship. Because there are negative values in the Pearson correlation coefficient matrix, in order to facilitate calculation, the embodiment adds one to all values in the matrix. According to the obtained Pearson correlation coefficient matrix, the embodiment can obtain the Pearson correlation coefficient of the left array element RSSI difference sequence and the right array element RSSI difference sequence, which is recorded as the correlation coefficient corresponding to the RSSI difference sequence. The correlation coefficient corresponding to the RSSI difference sequence can effectively reflect the correlation of the left array element and the right array element of the RSSI sequence.

[0067] Then, the embodiment utilizes the trained DOA range estimation network to roughly judge the DOA of the target object, specifically: input the correlation coefficients corresponding to the RSSI difference sequence of the target object into the trained DOA range estimation network, and then output the probability of the DOA range corresponding to the target object being different angle ranges through feature fitting. In the embodiment, the different angle ranges are: 0°-5° is angle range 1, 5°-10° is angle range 2, and so on, and the largest angle range is 175°-180°. In the embodiment, the angle ranges are represented by sequential Arabic numerals for easy classification.

[0068] In the embodiment, the DOA range estimation network is a fully connected neural network; the training set is the correlation coefficients corresponding to the RSSI difference sequence of the historical target object; the corresponding label data is the number of the angle range where the target object is located, and the data label is obtained through artificial measurement and classification; the loss function of the DOA range estimation network is a cross-entropy loss function, and the classification function is a Softmax function.

[0069] In the embodiment, the center of the antenna array in the factory floor image is taken as the signal incidence point, and then the probability of the different angle ranges corresponding to the target object is taken as the second probability of each pixel point in the corresponding range of the factory floor image, which represents the probability that the angle range where the pixel point is located is the angle range where the target object is located. The greater the second probability of each pixel point in a certain angle range, the greater the possibility that the target object is in the corresponding angle range. Finally, the second probability of each pixel point in the factory floor image is taken as the pixel value of the corresponding position pixel point, and an angle range probability map equal in size to the factory floor image is obtained.

[0070] It should be noted that the angle range probability map is corresponding to each target object, and the semantic distribution probability map is corresponding to all target objects.

[0071] Step S4: obtaining the DOA corresponding to the target object according to the semantic distribution probability map corresponding to the target object, the angle range probability map corresponding to the target object, and the RSSI sequence.

[0072] In order to more accurately estimate the DOA information of the target object, the embodiment adds the semantic distribution probability map obtained in step S2 and the angle range probability map obtained in step S3 to obtain a probability map, wherein the pixel value corresponding to each pixel point in the probability map is the sum of the pixel value (first probability) of each pixel point in the semantic distribution probability map and the pixel value (second probability) of the corresponding pixel point in the angle range probability map. Then, the probability map is input into the trained DOA estimation network to obtain a more accurate DOA value of the target object.

[0073] The pixel value of each pixel point in the semantic distribution probability graph in the embodiment reflects the position of the target object, and the position of the target object is the position of the centroid point of the connected domain where the target object is located. Therefore, the position of the pixel point with the maximum pixel value is the centroid point position of the connected domain, that is, the position of the target object. The pixel value of each pixel point in the angle range probability graph reflects the possibility of the angle range where the target object is located, and the angle range where the pixel point with the maximum pixel value is located is the angle range where the target object is located. In the embodiment, the pixel value corresponding to each pixel point in the probability graph is the sum of the first probability and the second probability. Therefore, the position of the pixel point with the greater pixel value in the probability graph is the position where the target object in the RSSI sequence received by the current RSSI receiver is located.

[0074] In the embodiment, the DOA estimation network includes a convolutional neural network and a fully connected network. The convolutional neural network can adopt a two-dimensional network model such as ResNet18 or MobileNet.

[0075] The input of the convolutional neural network is the probability graph corresponding to the target object obtained in the above process. After the probability graph is subjected to feature extraction through two-dimensional convolution, a feature map is obtained. Then, the feature map is subjected to a Flatten operation to obtain a feature vector. Finally, the obtained feature vector and the RSSI sequence of the target object are subjected to a Concatenate joint operation to obtain a final feature vector, which is input into the fully connected network for fitting to obtain the DOA corresponding to the target object. In the embodiment, the function of the last fully connected layer is preferably a Relu activation function.

[0076] The reason why the RSSI sequence of the target object is input into the DOA estimation network in the embodiment is to avoid the prediction error caused by the existence of multiple target objects in the same angle range. Because the RSSI value has a corresponding relationship with the distance, the DOA of the target object can be accurately predicted through the RSSI sequence.

[0077] The training set of the DOA estimation network is a historical probability graph, and the label data corresponding to the historical probability graph is the accurate position of the target object in the historical probability graph, which is represented by an arc. In the embodiment, the label data is obtained by manual measurement. In the embodiment, the loss function of the DOA estimation network is a mean square error loss function.

[0078] In the embodiment, the Gaussian distribution probability is used in the process of obtaining the semantic distribution probability graph. The Gaussian distribution probability can enable the network to learn more accurate semantic coordinate information, and then combine the probability of the corresponding angle range to achieve more accurate estimation of the DOA information of the received signal, that is, more accurate estimation of the DOA information of the target object.

[0079] The embodiment fuses a semantic distribution probability graph corresponding to a target object in a factory ground image and an angle range probability graph corresponding to the target object to obtain a DOA corresponding to the target object. A first probability corresponding to each pixel point in the semantic distribution probability graph is a probability that each pixel point is a centroid point of a connected domain corresponding to the target object. A probability value corresponding to each pixel point in the angle range probability graph is a probability that the target object is in different angle ranges. The probability value corresponding to each pixel point in the angle range probability graph is obtained by inputting a correlation coefficient corresponding to an RSSI differential sequence of the target object into a trained DOA range estimation network. The embodiment fuses the two to more accurately reflect the direction of the target object, and combines the RSSI sequence to more accurately obtain the DOA of the target object. The embodiment uses artificial intelligence to assist wireless communication to implement DOA estimation, so that the obtained DOA is more accurate.

[0080] It should be noted that the above only describes the preferred embodiments of the present application, and is not intended to limit the present application. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present application shall be included in the protection scope of the present application.

Claims

1. A wireless communication data processing method based on artificial intelligence, characterized in that, The method includes the following steps: Acquire images of the factory grounds; Based on the coordinates of each pixel in the connected component corresponding to the target object in the factory ground image, a first probability corresponding to each pixel in the connected component is obtained, wherein the first probability corresponding to each pixel is the probability that each pixel is the centroid of the connected component; based on the first probability corresponding to each pixel in the connected component in the factory ground image, a semantic distribution probability map corresponding to the target object is constructed. Obtain the RSSI values ​​from the RSSI transmitter of the target object to each RSSI receiver, construct an RSSI sequence based on the RSSI values ​​corresponding to each RSSI receiver, and divide the RSSI sequence according to the center point of the antenna array to obtain the left array element RSSI sequence and the right array element RSSI sequence. The correlation coefficients of the left-element RSSI difference sequence and the right-element RSSI difference sequence are obtained and input into the trained DOA range estimation network to obtain the probability that the DOA range of the target object is different angle ranges, and construct the angle range probability map corresponding to the target object. The semantic distribution probability map and the corresponding angle range probability map of the target object are added together to obtain the probability map. Then, the probability map is input into the trained DOA estimation network to obtain the DOA of the target object. Each target object has a corresponding angle range probability map, and all target objects correspond to the same semantic distribution probability map.

2. The artificial intelligence-based wireless communication data processing method according to claim 1, characterized in that, Based on the coordinates of each pixel in the connected component corresponding to the target object in the factory ground image, the first probability corresponding to each pixel in the connected component is obtained, including: Connectivity extraction of target objects in factory ground images; Gaussian function fitting is performed on the coordinates of each pixel in the connected component corresponding to the target object to obtain the Gaussian probability density function of the connected component. The first probability corresponding to each pixel in the connected component is obtained by substituting the coordinates of each pixel in the connected component into the corresponding Gaussian distribution probability density function.

3. The artificial intelligence-based wireless communication data processing method according to claim 1, characterized in that, Methods for obtaining the correlation coefficient between the left-element RSSI difference sequence and the right-element RSSI difference sequence include: The RSSI sequence is divided according to the center point of the antenna array. The first half of the RSSI sequence is denoted as the left element RSSI sequence, and the second half of the RSSI sequence is denoted as the right element RSSI sequence. The difference between any two adjacent elements in the left matrix RSSI sequence is obtained by taking the difference between the left matrix RSSI sequences. The difference between any two adjacent elements in the right matrix RSSI sequence is obtained by taking the difference between the right matrix RSSI sequences. Calculate the correlation coefficient between the RSSI difference sequences of the left and right matrix elements.

4. The artificial intelligence-based wireless communication data processing method according to claim 1, characterized in that, Construct a probability map of the angle range corresponding to the target object, including: The probabilities of different angle ranges are used as the second probabilities corresponding to the pixels within the corresponding range of the factory ground image; The second probability of each pixel in the factory ground image is used as the pixel value of the corresponding pixel to obtain the angle range probability map.

5. The artificial intelligence-based wireless communication data processing method according to claim 1, characterized in that, The method further includes: Based on the semantic distribution probability map, the corresponding angular range probability map, and the RSSI sequence of the target object, the DOA corresponding to the target object is obtained.

6. The artificial intelligence-based wireless communication data processing method according to claim 5, characterized in that, Based on the semantic distribution probability map, the corresponding angular range probability map, and the RSSI sequence of the target object, the DOA corresponding to the target object is obtained, including: The probability map of the target object is obtained by adding the semantic distribution probability map corresponding to the target object and the angle range probability map corresponding to the target object. The pixel value corresponding to each pixel in the probability map is the sum of the first probability and the second probability corresponding to each pixel. Construct a DOA estimation network, which includes a convolutional neural network and a fully connected network; The probability map of the target object is input into a convolutional neural network to obtain a feature map. The feature map is then flattened to obtain a feature vector. The feature vector is then concatenated with the RSSI sequence of the target object to obtain a final feature vector. The final feature vector is then input into a fully connected network for fitting to obtain the DOA corresponding to the target object. The training set of the DOA estimation network is a historical probability matrix, and the label data of the historical probability matrix is ​​the DOA of the target object corresponding to the historical probability map.

7. The artificial intelligence-based wireless communication data processing method according to claim 1, characterized in that, Methods for obtaining factory ground images include: Acquire multiple local factory images; Multiple partial factory images are processed using image stitching to obtain a panoramic image of the factory grounds; The trained semantic segmentation network is used to process the panoramic view of the factory ground to obtain the factory ground image.

Images