Apparatus and method for determining high resolution QoS prediction map
By combining trained artificial deep neural networks with environmental information, low-resolution QoS maps are transformed into high-resolution QoS prediction maps, solving the problem of high-resolution prediction in radio communication networks and achieving accurate coverage estimation and network optimization in the absence of measurement results.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- ROBERT BOSCH GMBH
- Filing Date
- 2021-11-12
- Publication Date
- 2026-07-10
AI Technical Summary
Existing technologies struggle to achieve high-resolution QoS prediction in radio communication networks, especially in the absence of data rate measurements, making it impossible to accurately estimate fine-grained information about network coverage.
By using a trained artificial deep neural network, combined with environmental information such as high-resolution maps and base station locations, and employing upscaling and feature extraction techniques, the low-resolution QoS map is transformed into a high-resolution QoS prediction map. Missing information is filled in using randomized noise maps, and the prediction accuracy is improved by training with a generative adversarial network.
In the absence of high-resolution measurement results, it can generate more accurate high-resolution QoS prediction maps, improve the estimation accuracy of network coverage, and support user equipment to respond appropriately based on prediction conditions.
Smart Images
Figure CN114492725B_ABST
Abstract
Description
Technical Field
[0001] This description relates to enhancements to QoS prediction for radio communication networks. Summary of the Invention
[0002] The problems of the prior art are solved by means of apparatus for determining high-resolution QoS prediction maps, by means of methods for determining high-resolution QoS prediction maps, by means of apparatus for training neural networks, by means of methods for training neural networks, and by the use of the aforementioned apparatus or methods.
[0003] A first aspect of this description relates to an apparatus for determining a high-resolution QoS prediction map of a first radio communication network in a first environment, wherein the apparatus includes: a first input unit configured to determine or provide environmental information characterizing the first environment; a second input configured to determine or provide a low-resolution QoS map associated with the first radio communication network in the first environment or a second radio communication network in the second environment; and a determining unit configured to propagate the low-resolution QoS map and environmental information via a trained artificial deep neural network, wherein the low-resolution QoS map and environmental information are provided as input parameters in an input portion of the trained artificial deep neural network, and wherein a high-resolution QoS prediction map of the first radio communication network is provided in an output portion of the trained artificial deep neural network.
[0004] The QoS prediction map provides improved fine-grained QoS predictions for mobile service applications. These fine-grained QoS predictions represent a reliable cellular coverage map of a first radio communication network. By considering environmental information, artificial neural networks allow for the transmission of coarse measurements of another or the same network, a second radio communication network, and a fine-grained representation of predictions associated with the first radio communication network. Furthermore, available relevant environmental information (such as high-resolution maps, base station locations) can be utilized to accomplish the task. In other words, the determining unit utilizes the correlation between the coverage map and environmental information (such as geographical attributes and infrastructure attributes) to enhance the resolution of the coverage map. The determining unit solves the transformation task, i.e., transforming a low-resolution map of region A in the sense of a low-resolution QoS map into a high-resolution map of a different region B in the sense of a high-resolution QoS prediction map.
[0005] Advantageously, a comprehensive map of the data rate or other quality of service parameters of the first radio communication network can be created even when little or no data rate measurement results for the first radio communication network are available. The trained artificial neural network makes the conversion from low resolution to high resolution dependent on environmental information. The trained model is able to acquire a low-resolution map from region A together with environmental information from region B, and then create a high-resolution map of region B, which may have neither low-resolution nor high-resolution coverage measurements available.
[0006] According to an advantageous example, the apparatus includes: a third input unit configured to determine or provide measurement information characterizing at least one radio channel of a first radio communication network; and a determination unit configured to propagate a low-resolution QoS map, environmental information, and measurement information via a trained artificial deep neural network, wherein the low-resolution QoS map, environmental information, and measurement information are provided as input parameters in the input portion of the trained artificial deep neural network, and wherein a high-resolution QoS prediction map of the first radio communication network is provided in the output portion of the trained artificial deep neural network.
[0007] Advantageously, sparse measurements representing a portion of the current radio communications of the first radio communication network allow for more accurate estimates in the sense of a high-resolution QoS prediction map.
[0008] According to an advantageous example, the device includes an operating unit configured to operate a first radio communication network based on a high-resolution QoS prediction map.
[0009] Advantageously, the UE of the first radio communication network is able to respond appropriately to the predicted radio conditions via the provided high-resolution QoS map.
[0010] According to an advantageous example, the apparatus includes an upscaling portion configured to upscale a low-resolution QoS map to an upscaled representation.
[0011] Advantageously, the upscaling part upsamples the low-resolution input into an upscaled representation with a high-resolution QoS prediction map. Advantageously, the upscaling part extracts the spatial and temporal features of QoS.
[0012] According to an advantageous example, the trained neural network includes a feature extraction section configured to determine a feature map based on the upscaled representation, wherein the feature extraction section includes at least one skip connection between layers of the feature extraction section.
[0013] Advantageously, the feature extraction part extracts features based on an upscaled version of the low-resolution image. Skip connections provide a shortcut to accelerate the determination process.
[0014] According to a favorable example, the trained neural network includes a convolutional part configured to determine a high-resolution prediction map based on the feature map determined by the feature extraction part, wherein the convolutional part does not include skip connections between layers of the convolutional part.
[0015] According to a favorable example, the determining unit includes a preprocessing unit configured to determine a randomized noise map comprising multiple randomized noise values, which may also be referred to as a randomized low-resolution QoS. The preprocessing unit is configured to determine a randomized low-resolution QoS map based on the determined randomized noise map and the low-resolution QoS map, and is configured to provide the randomized low-resolution QoS map to a trained artificial neural network. The randomized noise map is used to fill in missing information from low resolution to high resolution. Since the transmission is not a deterministic mapping, it is governed by random vectors—which we here call noise.
[0016] The trained network takes a given low-resolution QoS map, along with some additional randomized noise input (possibly at the input layer or other hidden layers), and then outputs a high-resolution QoS map. This output varies with the noise. Therefore, it leverages the random relationship between the low-resolution and high-resolution maps that the trained artificial neural network has already learned.
[0017] According to a second aspect of this description, a method is provided for determining a high-resolution QoS prediction map of a first radio communication network in a first environment. The method includes: determining or providing environmental information characterizing the first environment; determining or providing a low-resolution QoS map associated with the first radio communication network in the first environment or with a second radio communication network in a second environment; and propagating the low-resolution QoS map and environmental information via a trained artificial deep neural network, wherein the low-resolution QoS map and environmental information are provided as input parameters in an input portion of the trained artificial deep neural network, and wherein a high-resolution QoS prediction map of the first radio communication network is provided in an output portion of the trained artificial deep neural network.
[0018] According to an advantageous example, the method includes: determining or providing measurement information characterizing at least one radio channel of a first radio communication network; and propagating a low-resolution QoS map, environmental information, and measurement information via a trained artificial deep neural network, wherein the low-resolution QoS map, environmental information, and measurement information are provided as input parameters in the input portion of the trained artificial deep neural network, and wherein a high-resolution QoS prediction map of the first radio communication network is provided in the output portion of the trained artificial deep neural network.
[0019] According to an advantageous example, the method includes operating a first radio communication network based on a high-resolution QoS prediction map.
[0020] According to an advantageous example, the method includes scaling the low-resolution QoS map to an upscaled representation.
[0021] According to an advantageous example, the method includes determining a feature map based on the up-scaled representation, wherein the corresponding feature extraction portion includes at least one skip connection between layers of the feature extraction portion.
[0022] According to an advantageous example, the method includes determining a high-resolution prediction map based on the feature map determined by the feature extraction portion, wherein the convolutional portion does not include skip connections between layers of the convolutional portion.
[0023] According to an advantageous example, the method includes: determining a randomized noise map comprising a plurality of randomized noise values, the determination comprising: determining a randomized low-resolution QoS map depending on the determined randomized noise map and depending on a low-resolution QoS map, and providing the randomized low-resolution QoS map to a trained artificial neural network.
[0024] According to a third aspect of this description, an apparatus for training a deep neural network includes: a preparation unit configured to provide at least one training set, wherein the at least one training set ts includes a low-resolution QoS map associated with a radio communication network, a high-resolution QoS map associated with the radio communication network, and environmental information characterizing the environment of the radio communication network; a determination unit configured to propagate input data including the low-resolution QoS map and environmental information through the deep neural network, wherein input data is provided as input parameters in an input portion of the deep neural network, and wherein at least one neural network-based high-resolution QoS prediction map is provided in an output portion of the deep neural network; a discriminator unit configured to determine a comparison by comparing the neural network-based high-resolution QoS prediction map and the high-resolution QoS map of the training set; and a training unit configured to train the deep neural network using the training set depending on the comparison.
[0025] The model is trained based on a diverse set of input data, including pairs of low-resolution and high-resolution maps, describing the environment. Advantageously, environmental information aids the learning process, thereby requiring a reduced number of training data samples, namely, pairs of low-resolution and high-resolution overlay measurements.
[0026] According to a fourth aspect of this description, a method for training a deep neural network is provided, the method comprising: providing at least one training set, wherein the at least one training set ts includes a low-resolution QoS map associated with a radio communication network, a high-resolution QoS map associated with the radio communication network, and environmental information characterizing the environment of the radio communication network; propagating input data including the low-resolution QoS map and the environmental information through the deep neural network, wherein the input data is provided as input parameters in the input portion of the deep neural network, and wherein at least one neural network-based high-resolution QoS prediction map is provided in the output portion of the deep neural network; determining a comparison by comparing the neural network-based high-resolution QoS prediction map and the high-resolution QoS map of the training set; and training the deep neural network using the training set depending on the comparison. Attached Figure Description
[0027] Figure 1 An apparatus for determining a high-resolution QoS prediction map is described;
[0028] Figure 2 An apparatus for training an artificial neural network is described;
[0029] Figure 3 and 4 The structures of artificial deep neural networks are all described;
[0030] Figure 5 A schematic diagram depicts the arrangement used to determine the high-resolution QoS prediction map;
[0031] Figure 6 The arrangement used to train the artificial neural network is illustrated schematically;
[0032] Figure 7 and Figure 8 Both depict trained artificial neural networks; and
[0033] Figure 9 and 10 Both depict trained networks that function as discriminators. Detailed Implementation
[0034] Figure 1An apparatus 100 is schematically depicted for determining a high-resolution QoS prediction map hrPM1 for a first radio communication network RCN1 in a first environment E1. A first input unit 102 is configured to determine or provide environmental information ei1 characterizing the first environment E1. A second input 104 is configured to determine or provide a predetermined low-resolution QoS map lrM2 associated with the first radio communication network RCN1 in the first environment E1 or associated with a second radio communication network RCN2 in the second environment E2.
[0035] The first environment E1 is an environment including a first radio communication network RCN1, which has a base station BS1 under inspection and two user equipment UE1 and UE2. Therefore, block 101 constitutes the system during operation. On the other hand, the second environment E2, and the second radio communication network RCN2 including a base station BS2 and two user equipment UE3 and UE4, represent the training set generation environment. Therefore, the low-resolution QoS map lrM2 is provided in a predetermined form.
[0036] The determining unit 106 is configured to propagate a predetermined low-resolution QoS map lrM2 and environmental information ei1 through a trained artificial deep neural network GEN, wherein the low-resolution QoS map lrM2 and environmental information ei1 are provided as input parameters in the input portion of the trained artificial deep neural network GEN, and wherein a high-resolution QoS prediction map hrPM1 of the first radio communication network RNC1 is provided in the output portion of the trained artificial deep neural network GEN.
[0037] Environmental information ei1 is used as input to create a more accurate map for QoS prediction. ei1 includes maps with increased levels of detail (e.g., for automated driving purposes), high-resolution satellite imagery, information about building heights, and potentially additional inputs such as infrastructure-based video feeds or 3D scans. By combining base station locations and sample measurements, the correlation between features and data rates is learned.
[0038] At least at the 2D graph scale, high-resolution QoS prediction maps offer higher resolution than low-resolution QoS maps. In other words, compared to low-resolution QoS maps, high-resolution QoS maps provide more equidistant data points in the hypothetical horizontal plane of each graph region.
[0039] According to one example, environmental information ei1 includes at least one of the following: high-resolution image, base station location, vehicle traffic flow information, and weather information (such as rain and humidity).
[0040] In one example, the low-resolution QoS map and the high-resolution QoS prediction map include the same type of QoS parameters. In another example, the low-resolution QoS map and the high-resolution QoS prediction map include different types of QoS parameters.
[0041] The types of QoS parameters include at least one of the following: data rate, packet delay, received / receivable signal strength, packet loss rate, spectrum occupancy, etc.
[0042] According to one example, environmental information ei1 includes a geographic map that represents at least one non-transitory attribute of the environment at a given spatial location. According to one example, this attribute is color. Therefore, the geographic map could be a satellite image indicating the color of each spatial location. According to another example, the geographic attribute is elevation or height, for example, representing the height of a building.
[0043] In yet another example, the geographic attribute is the type of surface, such as metal, concrete, wood, grass, etc. This surface indicates the properties of radio wave reflection and absorption.
[0044] According to one example, environmental information ei1 includes a traffic map, which characterizes at least one flow attribute of traffic at the corresponding spatial location indicated by the traffic map.
[0045] According to one example, environmental information ei1 includes a video stream or perspective photographic representation of a portion of a first environment E1.
[0046] According to one example, environmental information ei1 includes the spatial location of a fixed antenna serving a first radio communication network.
[0047] The at least one traffic attribute includes, for example, the number of vehicles, people, or UEs passing through the spatial location during a time period.
[0048] According to one example, environmental information ei1 includes weather information, which characterizes at least one weather attribute of the first environment E1.
[0049] The third input unit 108 is configured to determine or provide measurement information mi1 characterizing at least one radio channel of the first radio communication network RCN1.
[0050] The determining unit 106 is configured to propagate a low-resolution QoS map lrM, environmental information ei1, and measurement information mi1 through a trained artificial deep neural network GEN, wherein the low-resolution QoS map lrM, environmental information ei1, and measurement information mi1 are provided as input parameters in the input portion of the trained artificial deep neural network GEN, and wherein a high-resolution QoS prediction map hrPM1 of the first radio communication network RCN1 is provided in the output portion of the trained artificial deep neural network GEN.
[0051] According to one example, measurement information mi1 includes at least a plurality of radio measurements of, for example, at least one radio parameter, wherein the corresponding radio measurements are associated with spatial location or with a spatial location and time indicator. According to one example, measurement information mi1 includes at least one of the following: a network quality indicator, such as a UE-based signal strength measurement with location and time; or a QoS parameter, such as a sparse measurement of data rate with location and time.
[0052] Operation unit 110 is configured to operate the first radio communication network RCN1 based on a high-resolution QoS prediction map hrPM1. For example, operation unit 110 identifies an area in the environment that is expected to experience poor quality of service, such as low data rates. Before entering this area, a low data rate indicator warns the UE in preparation for the user plane functions to enter a safe operating state. According to another example, if one of the communication partners resides in an area expected to experience low data rates, the network entity increases transmission power and / or selects one or more of the most appropriate networks and / or disables at least one additional application.
[0053] Preprocessing unit 116 is configured to determine a randomized noise map including multiple randomized noise values. Preprocessing unit 116 is configured to determine a randomized low-resolution QoS map based on the determined randomized noise map and a predetermined low-resolution QoS map lrM2. Preprocessing unit 116 is configured to provide the randomized low-resolution QoS map to a trained artificial neural network GEN, instead of the unrandomized low-resolution QoS map lrM2.
[0054] According to another example, preprocessing unit 116 is configured to preprocess environmental information ei1 (e.g., collected measurement results). An additional machine-trained model determines how to perform the preprocessing.
[0055] According to one example, instead of having a preprocessing unit 116, having both lrM2 and the noise vector as inputs to the GEN may be sufficient. Therefore, the preprocessing unit 116 is optional.
[0056] Figure 2 An apparatus 200 for training a deep neural network GEN is schematically depicted. A preparation unit 202 is configured to provide at least one training set ts. The at least one training set ts includes a low-resolution QoS map lrM2 associated with a radio communication network RCN2, a high-resolution QoS map hrM2 associated with the radio communication network RCN2, and environmental information ei2 characterizing the environment E2 of the radio communication network RCN2.
[0057] The determining unit 204 is configured to propagate input data, including a low-resolution QoS map lrM2 and environmental information ei2, via a deep neural network GEN. The input data is provided as input parameters in the input portion of the deep neural network GEN. At least one high-resolution QoS prediction map hrPM2 based on a neural network is provided in the output portion of the deep neural network GEN.
[0058] The discriminator unit 206 is configured to determine the comparison c by comparing the high-resolution QoS prediction map hrPM2 based on a neural network with the high-resolution QoS map hrM2 of the training set ts. The discriminator unit 206 utilizes a trained deep neural network DSC. The neural networks GEN and DSC represent generative adversarial networks. The deep generator network GEN, combined with the discriminator network DSC, generates a higher-resolution map based on the low-resolution map. The training set ts consists of pairs of low-resolution and high-resolution maps. During the training phase, the generator network GEN takes a low-resolution map of a region of environment E2 and a random noise vector z along with environmental information ei2 as side inputs to generate a high-resolution map estimate of the same region. Here, the random noise vector z serves to create a random relationship between the low-resolution and high-resolution maps.
[0059] The discriminator network (DSC) takes a high-resolution image as input. It is trained to classify whether the input is generated or measured (i.e., the ground truth). The discriminator output in the sense of comparison *c* is used as a reference signal for the generator network (GEN) to improve its output so that the discriminator network (DSC) cannot distinguish between generated and measured images. This training process aims to teach the generator network (GEN) to upscale low-resolution images.
[0060] In one example, conditional environment information ei2 is used. The discriminator DSC compares (hrPM2 and ei2) with (hrm2 and ei2). In addition to this comparison, hrPM2 is downscaled, and the downscaled hrPM2 is compared with the true low-resolution QoS map lrM2.
[0061] Training unit 208 is configured to train the deep neural network GEN using the training set ts, depending on the comparison c. For example, the training performed by training unit 208 is as follows: Figure 6 Perform as illustrated.
[0062] To perform the transformation, the generator network (GEN) acquires a low-resolution QoS map of the second radio communication network (RCN2), random noise z, and environmental information ei2 associated with RCN2. For example, the low-resolution QoS map and random noise z are added element-wise / coordinate-wise. This output is fed into the discriminator network (DSC), which compares its measured high-resolution map of the region of RCN2 with the environmental information ei2. In doing so, the generator network (GEN) learns to transform the low-resolution map based on lateral information. During testing, the generator network (GEN) is used in practice. By drawing z multiple times, the generator network (GEN) can generate multiple high-resolution QoS prediction maps with respect to the input low-resolution map.
[0063] As an example, if a measured low-resolution QoS map and a measured high-resolution QoS map pair associated with a second radio communication network are determined over time, a training set sequence with corresponding timestamps is generated over time. In this case, time information can be utilized.
[0064] Figure 3 An exemplary structure of the neural network GEN is depicted. The upscaling portion 302 is configured to upscale the low-resolution QoS map lrM2 to an upscaled representation ur. The upscaling portion 302 includes multiple blocks 2a-2d, each block comprising a 3D deconvolutional layer, followed by a batch normalization layer, a Leaky ReLU activation layer LReLU, and then a 3D convolutional layer. Each 3D convolutional layer enhances the model's representability. Each batch normalization layer normalizes the output and accelerates training. The LReLU layer improves the model's non-linearity. The LReLU layer can take the form: for x>= 0, LReLux=x, and for x<0, LRELux=ax, where x is the input and a is a positive constant, such as 0.05.
[0065] The neural network GEN includes a feature extraction part 304, which is configured to determine a feature map fm based on an upscaled representation ur, wherein the feature extraction part 304 includes at least one skip connection sc1, sc2, scG between layers of the feature extraction part 304.
[0066] Feature extraction section 304 comprises multiple feature extraction blocks 4a, 4b, 4c, 4d...4x. Each feature extraction block 4a-x includes a convolutional layer, a batch normalization layer, and a Leaky ReLU activation layer. These feature extraction blocks 4a-x are interconnected via interleaved skip connections sc1 and sc2 and a global skip connection gsc. Interleaved skip connections sc1 and sc2 allow skipping one or more subsequent feature extraction blocks 4b and 4c or 4d, while the global skip connection gsc allows skipping feature extraction section 304. Skip connections are shortcut connections and can be referred to as zipper connections.
[0067] The convolutional portion 306 is configured to determine a high-resolution prediction map hrPM1 based on the feature map fm determined by the feature extraction portion 304, wherein the convolutional portion 306 does not include skip connections between layers of the convolutional portion 306. The convolutional portion 306 includes multiple blocks 6a to 6x without skip connections. The respective blocks 6a to 6x include a convolutional layer, followed by a batch normalization layer, and then a Leaky ReLU activation layer.
[0068] Figure 4 An example of the structure of a deep neural network DSC used as a discriminator is depicted. The deep neural network DSC determines the comparison c in the form of the output of the sigmoid function sigm, based on a high-resolution QoS prediction map hrPM according to one of the previous graphs, and a high-resolution QoS map hrM determined via a real-world radio communication network. A corresponding block in the plurality of blocks 40a-40f comprises a convolutional layer, followed by a batch normalization layer, and then a Leaky ReLU activation layer.
[0069] Figure 5 A schematic arrangement for determining the high-resolution QoS prediction map hrPM1 is depicted. Figure 6 The setup for training is shown in the diagram. First, let's discuss the training process.
[0070] The training data eiTrain is provided as input data via input interface 602. This arrangement comprises an artificial neural network (GEN) with an input layer. For time step i, the input tensor of the input data id is passed to the input layer. The input layer is part of the input portion. For the input data id, the output O is determined in the form of a prediction or is known in advance. At time step i, a tensor with observations oitrain is determined from the output O, and these observations oitrain are assigned to the observations of the tensor eitrain. The output O includes a QoS prediction graph. Each time series of the input data id is assigned to one of three input nodes. In the forward path of the artificial neural network GEN, the input layer is followed by at least one hidden layer. In this example, the number of nodes in the at least one hidden layer is greater than the number of input nodes. This number will be considered a hyperparameter. In this example, four nodes are provided in the hidden layer. The neural network GEN is learned, for example, via gradient descent in the form of backpropagation. Therefore, the training of the neural network NN is supervised.
[0071] In the forward path of this example, an output layer 604 is provided after at least one hidden layer. The predicted value is output at output layer 604 in the output portion of the neural network GEN. In this example, an output node is assigned to each predicted value.
[0072] At each time step i, a tensor o'itrain is determined, containing the predicted values for that time step i. In this example, this is fed into training facility 606 along with the column vector of the observations oitrain. In this example, training facility 606 is designed to determine the prediction error using a loss function LOSS, particularly using mean squared error, and to train the model using it and an optimizer, particularly the Adam optimizer. In this example, the loss function LOSS is determined based on the deviations from the values of the tensors of the observations o'itrain and the predicted values oitrain, particularly the mean squared error.
[0073] Once a fixed criterion is reached, the training ends. In this example, if the loss does not decrease over several time steps, specifically if the mean squared error does not decrease, the training is terminated.
[0074] The test data is then fed into the model trained in this manner. This model is generated from the training data. The model is evaluated using the test data, particularly regarding the mean µ and covariance Σ.
[0075] according to Figure 5The arrangement shown uses a trained model to provide a QoS prediction map. The same data preprocessing steps are performed as for the training data. For example, scaling and determination of the input and output data occur. In this example, the determination is... Figure 1 This occurs during the operation of device 100—that is, during the operation of the first radio communication network RCN1.
[0076] Input data is fed into a trained artificial neural network, GEN. Based on this, predicted values are determined. Based on this, a final score is determined.
[0077] Figure 5 A schematic arrangement for determining the QoS prediction map is shown. As described for training, for time step i, the column vector ei is passed to the input layer of the input data id. Then, the column vector is passed to the input layer. After this, in contrast to training, device 502 determines the high-resolution QoS prediction map hrPM1 based on the predicted value y'i.
[0078] Specifically, instructions are provided for a computer program to implement the described neural network GEN, for carrying out the described process. Specialized hardware may also be provided in which the trained model is mapped.
[0079] Figure 7 A trained network GEN in super-resolution mode is described. The trained network GEN maps the low-resolution QoS map lrPM1 of the same network to a high-resolution QoS map hrPM1, based on the environmental information ei1 provided by the network. Super-resolution mode requires the training set to have paired examples; that is, we have both low-resolution and high-resolution measurements from the same network. To train the super-resolution mode, we can add the mean squared error between lrPM1 and the down-scaling (hrPM1 generated by GEN) as an additional term to the training loss of GEN. Essentially, the down-scaled output of GEN should be identical to the input lrPM1 of the super-resolution mode.
[0080] Figure 8 The trained network GEN under the transformation mode is depicted. The trained network GEN maps the low-resolution QoS map lrPM2 of the first network to the high-resolution QoS map hrPM1 of the second network based on the environmental information ei1 provided by the second network. The mode transformation involves a training set containing some low-resolution and high-resolution measurements, which do not need to be paired. Additionally, we should know the environmental map of the high-resolution measurements. To train the transformation mode, we can add a periodic consistency loss as an additional term to the training loss of GEN. Essentially, the GEN output can be mapped back to the low-resolution lrM2 through a downscaling network jointly trained with the transformation network.
[0081] Furthermore, to better perform conversion tasks, it would be beneficial for lrPM2's ei2 to show some similarity to ei1 (e.g., both being urban or rural areas). In short, conversion quality improves as ei2 and ei1 become closer to each other.
[0082] Figure 9 The trained DSC network is described, which acts as a discriminator in a binary classifier. Based on the input, the trained DSC network outputs true / false. To train the DSC network, we can construct the input as follows:
[0083] 1. If the input includes a high-resolution QoS graph hrPM generated by the network GEN, and the provided environmental information, the result is "false".
[0084] 2. If the input includes the measured high-resolution QoS map hrPM and the provided environmental information, the result is "true".
[0085] 3. If the input includes a measured high-resolution QoS map hrPM and incorrect environmental information ei, the result is "false".
[0086] 4. If the input includes a high-resolution QoS graph hrPM generated by the trained network GEN, and incorrect environmental information ei, the result is "false".
[0087] Figure 10 This includes an additional trained network, DSC2, which acts as a further discriminator. To further enhance training performance, the trained DSC2 network distinguishes between true and false on a low-resolution QoS map (lrM). The trained DSC2 network is a binary classifier that outputs true / false based on the input. To train the additional DSC2 network, we can construct the input as follows:
[0088] 1. If the high-resolution QoS map hrPM generated by the neural network GEN is scaled down to a low-resolution QoS map lrPM and the corresponding environmental information, the result is "false".
[0089] 2. If the input includes the measured low-resolution QoS map lrPM and the corresponding environmental information ei, the result is "true".
[0090] On the architecture side, neural networks GEN and DSC can be deep neural networks that include convolutional layers.
Claims
1. An apparatus (100) for determining a high-resolution QoS prediction map (hrPM1) of a first radio communication network (RCN1) in a first environment (E1), wherein the apparatus (100) comprises: The first input unit (102) is configured to determine or provide environmental information (ei1) characterizing the first environment (E1); The second input (104) is configured to determine or provide a low-resolution QoS map (lrM1; lrM2) associated with the first radio communication network (RCN1) of the first environment (E1) or with the second radio communication network (RCN2) of the second environment (E2). as well as The determining unit (106) is configured to propagate a low-resolution QoS map (lrM1; lrM2) and environmental information (ei1) via a trained deep artificial neural network (GEN), wherein the low-resolution QoS map (lrM1; lrM2) and environmental information (ei1) are provided as input parameters in the input portion of the trained deep artificial neural network (GEN), and wherein a high-resolution QoS prediction map (hrPM1) of the first radio communication network (RNC1) is provided in the output portion of the trained deep artificial neural network (GEN).
2. The apparatus (100) according to claim 1, comprising: The third input unit (108) is configured to determine or provide measurement information (mi1) characterizing at least one radio channel of the first radio communication network (RCN1). as well as The determining unit (106) is configured to propagate a low-resolution QoS map (lrM2), environmental information (ei1), and measurement information (mi1) through a trained deep artificial neural network (GEN), wherein the low-resolution QoS map (lrM2), environmental information (ei1), and measurement information (mi1) are provided as input parameters in the input portion of the trained deep artificial neural network (GEN), and wherein a high-resolution QoS prediction map (hrPM1) of the first radio communication network (RCN1) is provided in the output portion of the trained deep artificial neural network (GEN).
3. The apparatus (100) according to claim 1 or 2, comprising: The operating unit (110) is configured to operate the first radio communication network (RCN1) depending on the high-resolution QoS prediction map (hrPM1).
4. The apparatus (100) according to claim 1 or 2, wherein the trained neural network (GEN) comprises: The upscaling portion (302) is configured to upscale the low-resolution QoS map (lrM2) to an upscaled representation (ur).
5. The apparatus (100) according to claim 1 or 2, wherein the determining unit (106) comprises: The preprocessing unit (116) is configured to determine a randomized noise map including multiple randomized noise values, is configured to determine a randomized low-resolution QoS map based on the determined randomized noise map and based on the low-resolution QoS map (lrM2), and is configured to provide the randomized low-resolution QoS map to a trained artificial neural network (GEN).
6. A method for determining a high-resolution QoS prediction map (hrPM1) of a first radio communication network (RCN1) in a first environment (E1), wherein the method comprises: Determine or provide (102) environmental information (ei1) characterizing the first environment (E1); Determine or provide (104) a low-resolution QoS map (lrM1; lrM2) associated with a first radio communication network (RCN1) in the first environment (E1) or with a second radio communication network (RCN2) in the second environment (E2); and (106) Low-resolution QoS maps (lrM1; lrM2) and environmental information (ei1) are propagated through a trained artificial deep neural network (GEN), wherein the low-resolution QoS maps (lrM1; lrM2) and environmental information (ei1) are provided as input parameters in the input portion of the trained artificial deep neural network (GEN), and wherein a high-resolution QoS prediction map (hrPM1) of the first radio communication network (RNC1) is provided in the output portion of the trained artificial deep neural network (GEN).
7. The method of claim 6, comprising: Determine or provide (108) measurement information (mi1) characterizing at least one radio channel of the first radio communication network (RCN1); as well as (106) Low-resolution QoS maps (lrM1; lrM2), environmental information (ei1), and measurement information (mi1) are propagated through a trained deep neural network (GEN), wherein the low-resolution QoS maps (lrM2), environmental information (ei1), and measurement information (mi1) are provided as input parameters in the input portion of the trained deep neural network (GEN), and wherein a high-resolution QoS prediction map (hrPM1) of the first radio communication network (RCN1) is provided in the output portion of the trained deep neural network (GEN).
8. The method according to claim 6 or 7, comprising: The operation of the (110) first radio communication network (RCN1) depends on the high-resolution QoS prediction map (hrPM1).
9. The method according to claim 6 or 7, comprising: Upscale (302) the low-resolution QoS graph (lrM2) to the upscaled representation (ur).
10. The method according to claim 6 or 7, comprising: Determine (116) a randomized noise map comprising multiple randomized noise values, the determination (116) comprising: determining a randomized low-resolution QoS map based on the determined randomized noise map and based on a low-resolution QoS map (lrM2), and providing the randomized low-resolution QoS map to a trained artificial neural network (GEN).
11. An apparatus (200) for training a deep neural network (GEN), wherein the apparatus (200) comprises: The preparation unit (202) is configured to provide at least one training set (ts), wherein the at least one training set (ts) includes a low-resolution QoS map (lrM2) associated with the radio communication network (RCN2), a high-resolution QoS map (hrM2) associated with the radio communication network (RCN2), and environmental information (ei2) characterizing the environment (E2) of the radio communication network (RCN2). The determining unit (204) is configured to propagate input data including a low-resolution QoS map (lrM2) and environmental information (ei2) via a deep neural network (GEN), wherein the input data is provided as input parameters in the input part of the deep neural network (GEN), and wherein at least one high-resolution QoS prediction map (hrPM2) based on a neural network is provided in the output part of the deep neural network (GEN). The discriminator unit (206) is configured to determine a comparison (c) by comparing a high-resolution QoS prediction map (hrPM2) based on a neural network with a high-resolution QoS map (hrM2) of the training set (ts); and The training unit (208) is configured to train a deep neural network (GEN) using the training set (ts) depending on the comparison (c).
12. A method for training a deep neural network (GEN), wherein the method comprises: Provide (202) at least one training set (ts), wherein the at least one training set (ts) includes a low-resolution QoS map (lrM2) associated with the radio communication network (RCN2), a high-resolution QoS map (hrM2) associated with the radio communication network (RCN2), and environmental information (ei2) characterizing the environment (E2) of the radio communication network (RCN2). (204) Input data including a low-resolution QoS map (lrM2) and environmental information (ei2) are propagated through a deep neural network (GEN), wherein the input data is provided as input parameters in the input part of the deep neural network (GEN), and wherein at least one high-resolution QoS prediction map (hrPM2) based on a neural network is provided in the output part of the deep neural network (GEN). The comparison (206) is determined by comparing the high-resolution QoS prediction map (hrPM2) based on the neural network with the high-resolution QoS map (hrM2) of the training set (ts); and Depending on the comparison (c), the training set (ts) is used to train (208) a deep neural network (GEN).
13. Use of the apparatus (100; 200) according to any one of claims 1 to 5 or 11, or the method according to any one of claims 6 to 10 or 12.