Method and apparatus for estimating radio frequency transmission interference - Patents.com

Abstract

The present disclosure relates to detecting impairments in a received wireless communication signal, which may detect the presence of each of N types (sources) of impairments and, in some cases, the amount of impairment present in the received signal. The detection involves processing the received signal by a trainable module that is trained to distinguish between the N sources of impairments by applying learning, where N is an integer greater than 1. The trainable module outputs, for each source j of the N sources, the contribution of the jth source of impairment to the acquired signal. The contribution may be a binary value indicating the presence or absence of the jth source of impairment, or may also indicate the amount of impairment.

Description

[Technical field]

[0001] The present disclosure relates to estimating impairments associated with transmissions over wireless channels.

[0002] In wireless communication, there are many possible sources of transmission impairments that can cause erroneous signal reception at the receiver. There are several possibilities to deal with the errors. For example, the errors can be at least partially compensated for. Compensation can be done at the transmitter by using predistortion, or at the receiver by compensating for frequency, phase, and / or time shifts. Compensation is possible, at least to some extent, when some characteristics of the wireless transmitter, the channel, and / or the receiver are known.

[0003] Signal impairments can be introduced already in the transmitter, for example by hardware defects. For example, the front end of the transmitter, including filters such as pulse shaping filters, modulators for modulating the signal onto a carrier frequency, clocks for timing software / hardware operations, or power amplifiers, can cause signal distortions such as frequency offsets, phase offsets, or timing offsets. The distortions can be nonlinear. Another source of impairments is the wireless channel, which can suffer from, for example, path loss, multipath propagation, and / or some fading impairments, etc. Finally, the receiver can also be a source of hardware impairments similar to those of the transmitter.

[0004] Future wireless communication systems are expected to provide ever-increasing data rates, which require large transmission bandwidths and high carrier frequencies. These systems are also expected to bring about radio transmitters and receivers with high versatility and reconfigurability. Thus, they can ensure value-added services in diverse applications for seamless quality of service. However, building such small, high-quality radio, low-cost and versatile equipment for future wireless communication systems is a very challenging task. As a result, various defects (impairments) are expected to occur in radio transceivers. If these impairments are not properly estimated and compensated, they can significantly degrade the performance of the wireless communication system.

[0005] To enable compensation for signal transmission errors, it may be desirable to know the characteristics of the errors.

[0006] The present invention relates to a method and apparatus for detecting multiple faults by using model-based learning.

[0007] The invention is defined by the independent claims. Some advantageous embodiments are provided by the dependent claims.

[0008] In particular, some embodiments of the present disclosure relate to detecting multiple faults with the same trained module. Some embodiments also relate to training of the modules.

[0009] According to one embodiment, a method for estimating radio frequency transmission interference is provided, the method comprising the steps of acquiring a signal received through a wireless channel, processing the acquired signal by a trainable module trained to distinguish between N sources of interference, where N is an integer greater than 1, and outputting from the trainable module, for each source j of the N sources, a contribution of the jth source of interference to the acquired signal.

[0010] According to one embodiment, there is provided a method for training a trainable module for estimating radio frequency transmission impairments, the method comprising the steps of obtaining a training set comprising a plurality of training data including an input signal degraded by impairments and by a transmission channel and an impairment indication indicating a type of impairment, the signal; inputting the training set into the trainable module; adapting parameters of the trainable module according to the input training set; and storing the adapted parameters for use in the step of estimating radio frequency transmission impairments.

[0011] According to one embodiment, an apparatus for estimating radio frequency transmission impairments is provided, the apparatus comprising a processing circuit configured to acquire a signal received through a wireless channel, process the acquired signal by a trainable module trained to distinguish between N sources of impairments by applying supervised learning, where N is an integer greater than 1, and output from the trainable module, for each source j of the N sources, a contribution of the jth source of impairments to the acquired signal.

[0012] These and other features and characteristics of the subject matter of the present disclosure, as well as the method and function of operation of the associated elements of construction, and the combination of parts and economies of manufacture, will become more apparent from a consideration of the following description and the appended claims, taken in conjunction with the accompanying drawings, all of which are incorporated herein by reference. It is to be expressly understood, however, that the drawings are for purposes of illustration and description only and are not intended to define the limits of the subject matter of the present disclosure. As used in the specification and claims, the singular forms "a," "an," and "the" include plural referents unless the context clearly indicates a different interpretation. [Brief description of the drawings]

[0013] The features and advantages of various embodiments can be understood with reference to the following figures. [Figure 1] FIG. 1 is a block diagram showing an overall model of a typical communication system. [Diagram 2] Figure 2 is a flow diagram showing the four phases of a general model-based approach. [Diagram 3] FIG. 3 is a functional diagram of multiple RF impairments estimation using a trainable model. [Figure 4] FIG. 4 is a functional diagram of the phase of acquiring training data for multiple RF impairments estimation. [Diagram 5] FIG. 5 is a functional diagram of the output of the training data acquisition phase. [Figure 6] FIG. 6 is a functional diagram of an exemplary training phase. [Figure 7] FIG. 7 is a functional diagram of an exemplary test / inference phase. [Figure 8] FIG. 8 is a functional diagram of multiple RF impairments estimation using a machine learning model with feature extraction. [Figure 9] FIG. 9 is a schematic diagram illustrating an exemplary output format for outputting the contributions of different disturbance sources. [Figure 10] FIG. 10 is a block diagram illustrating an example for single carrier synchronization. [Figure 11] FIG. 11 is a block diagram illustrating an example for multi-carrier synchronization. [Figure 12] FIG. 12 is a block diagram illustrating the physical layer authentication function module in the training phase. [Figure 13] FIG. 13 is a block diagram showing the physical layer authentication function module in the inference phase. [Figure 14] FIG. 14 is a block diagram illustrating an exemplary structure of an apparatus for estimating a fault. [Figure 15] FIG. 15 is a block diagram illustrating an example structure of a memory that may be part of an apparatus for estimating faults. Detailed Description

[0014] For purposes of the following description, the terms "end", "upper", "lower", "right", "left", "vertical", "horizontal", "top", "bottom", "lateral", "longitudinal", and their derivatives shall refer to the subject matter of the present disclosure as oriented in the drawing figures. However, it should be understood that the subject matter of the present disclosure may assume various alternative variations and step sequences unless expressly stated to the contrary. It should also be understood that the specific devices and processes illustrated in the accompanying drawings and described in the following specification are merely exemplary embodiments or aspects of the subject matter of the present disclosure. Thus, specific dimensions and other physical characteristics relating to the embodiments or aspects disclosed herein should not be considered as limiting, unless expressly stated to the contrary.

[0015] As used herein, any aspect, component, element, structure, act, step, function, instruction, and / or the like should not be construed as critical or essential unless expressly stated as such. Additionally, as used herein, the articles "a" and "an" are intended to include one or more items and may be used synonymously with "one or more" and "at least one." Additionally, as used herein, the term "set" is intended to include one or more items (e.g., related items, unrelated items, combinations of related and unrelated items, and / or the like) and may be used synonymously with "one or more" and "at least one." When only one item is intended, "one" or similar language is used. Additionally, as used herein, "has," "hava," "having," or other terms are intended to be open-ended terms. Additionally, the phrase "based on" is intended to mean "based at least in part on," unless expressly stated otherwise.

[0016] There have been some research works in the field of radio frequency (RF) impairment estimation and RF impairment compensation using model-based learning. Research works on RF impairment estimation usually concern the estimation of a single RF impairment type using a model specifically trained for such type of impairment. For example, detection can be done by first detecting the presence of a specific type of RF impairment and then estimating the value of the detected impairment. However, detection can be difficult and does not always yield accurate results. Any error in detection can significantly degrade the performance of RF impairment estimation and possible further compensation. In addition, such methods based only on detecting the presence of a specific type of artifact miss possible correlations between different types of impairments. Exploiting these correlations can improve the estimation quality. This is especially true when the impairments are strongly correlated with each other. In some embodiments of the present disclosure, detection of multiple RF impairments using one model allows for an improvement of the overall performance.

[0017] On the other hand, research on joint RF impairment compensation aims to compensate the received signal for the effects of different kinds of RF impairments. They usually do not identify / recognize the specific RF impairment type. The objective is to train or configure a compensation model to obtain a received signal as similar as possible to the transmitted signal. One of the disadvantages of these techniques is that they do not estimate the impairments. Their only interest is to compensate for the joint effects of the impairments. Thus, the trainable model cannot be used for any other purpose and is useless if the compensation cannot be performed well. According to some embodiments of the present disclosure, the impairments are recognized and estimated separately. Thus, some of the impairments may be estimated incorrectly in some scenarios, but some of them may still be used. Furthermore, if the RF impairments are estimated, they may be used for different purposes. For example, their effects may be compensated for before the signal is sent. Such compensation before transmitting the signal is called predistortion, because the signal to be transmitted is usually distorted so that when it is degraded by the detected / estimated impairments, a signal as similar as possible to the original signal is received.

[0018] Some embodiments of the present disclosure are broadly applicable to any wireless system. In general, RF impairment estimation is an important part of the physical layer. Therefore, standards for the physical layer of the system may take advantage of them. The present disclosure is easily applicable to wireless local area networks (WLANs), such as ZigBee™, WiFi (e.g., IEEE 802.11 family of standards, e.g., the seventh generation of WiFi currently being studied, IEEE 802.11be), LoWPAN (low power wireless personal area networks, such as 6LoWPAN-Ipv6 based LoWPAN), or any other LAN. The present invention may be equally applicable to personal area networks, such as Bluetooth™, Bluetooth LE (low power), or any proprietary network operating, for example, in the 2.4 GHz band. The present invention provides advantages, especially at high frequencies and / or high modulation orders. The present invention may provide improvements in the context of cognitive radio standards (e.g., IEEE 802.22, IEEE 802.15).

[0019] It is noted that the present disclosure is also applicable to mobile systems, such as cellular mobile systems and their different operating modes. For example, some embodiments may be applicable to communication technologies under the Long Term Evolution (LTE) / LTE-Advanced (LTE-A) standard, or in fifth generation (5G) and beyond 5G communication systems. An example of such a 5G system may be New Radio (NR).

[0020] FIG. 1 shows an exemplary functional communication (sub)system CS, where 1 represents a transmitter and 2 represents a receiver. The transmitter 1 has the ability to transmit signals to the receiver 2 through an interface 3. The interface can be, for example, a wireless interface. The interface can be specified with resources that can be used for transmission and reception by the transmitter 1 and the receiver 2. Such resources can be defined by one or more (or all) of the time domain, the frequency domain, the code domain, and the spatial domain. It is noted that in general, the "transmitter" and the "receiver" can also be integrated together in the same device. In other words, the devices 1 and 2 in FIG. 1 can include the functions of 2 and 1, respectively. The transmitter 1 and the receiver 2 can be implemented in any device, such as a base station (eNB, AP), or a terminal (UE, STA), or in any other entity of the communication system CS. A device, such as a base station or a terminal, can implement both 2 and 1. The present disclosure is not limited to any particular implementation of the transmitter 1, the receiver 2, and / or the interface 3. However, the present disclosure may be readily applied to some existing communication systems, as well as to extensions of such systems or to new communication systems. Exemplary existing communication systems may be, for example, 5G New Radio (NR) in current or future releases, and / or IEEE 802.11-based systems such as the recently researched IEEE 802.1be, or systems such as those mentioned above.

[0021] In general, communication systems may suffer from a wide variety of impairments. For example, hardware may introduce some intermodulation and amplifier distortion, local oscillator, and quantization losses. Most designers rely on simplified closed-form models. However, these models do not accurately or comprehensively capture the effects of real-world systems and channels, and they cannot estimate multiple RF impairments simultaneously, as it is a very complex problem. For this complex problem, deep learning (DL)-based methods are very promising. In some embodiments of the present disclosure, a DL-based model can be used as a one-stage method, which can estimate multiple RF impairments together. This approach includes two stages, called training and testing. In the training stage, the received signal is fed as an input, and the known impairments are fed as an output to the DL model to train the DL model. Then, in the testing stage, the received signal is fed to the trained DL model, and the trained DL model finds all the defects.

[0022] In general, with a trainable model, several phases may be involved. Figure 2 illustrates these phases. In a first phase 10, a data set for training the trainable model(s) is obtained. The training set may be obtained by controlling the introduction of impairments into the signal and by measuring the signal received after passing some channel or channel model or noise etc. The introduced impairments (corresponding to the ground truth) and the measured signal then form a training pair for the training set. It is noted that the present disclosure may be used with supervised learning, with unsupervised learning or with reinforcement learning. In a second phase 20, the trainable model(s) is trained by inputting the training data set obtained in phase 10. In a third phase, the trainable model(s) is tested at 30. The testing 30 makes it possible to evaluate the accuracy of the model. If the accuracy is not sufficient, steps 20 and possibly also 10 may be repeated. The fourth phase 40 is the generation phase (also called the inference phase), in which the trained model(s) are used for fault detection.

[0023] Such an approach may allow impairments to be considered as having correlations between them. The presence of impairments should be detected by an expert or some tool, since it is necessary to learn which estimation model should be used. Thus, multiple RF impairments should be estimated using a single model. One of the goals of this disclosure is to estimate impairments. After an impairment is detected / recognized, the information may be used for various purposes. Accordingly, some embodiments include compensation for the detected impairments (at the transmitter or receiver). Some embodiments include application of the detected impairments to authentication of a user or to other security related issues. It is noted that the impairment detection of some embodiments may be used with any of the well-known techniques for compensating for impairments. Furthermore, the impairment detection of some embodiments may be used with any of the well-known techniques for determining whether a user is legitimate (authenticated) based on the detected impairments.

[0024] FIG. 3 shows an overview of multiple RF impairments estimation using a deep learning (DL) model according to one embodiment of the present disclosure. In this embodiment, a one-stage solution is provided for estimating multiple impairments in a system using a DL model. In FIG. 1, a transmit signal Tx is degraded by RF impairments 102 and 103 and transmitted through a channel 110, resulting in a received signal Rx. The received signal Rx is then fed to a DL module 120, which is used to estimate impairments and output an estimated impairment source 125. If the estimation by the DL module 120 is correct, the estimated impairment source 125 corresponds to the RF impairments 102 and 103. These advances make it possible to design a more robust and better performing communication system.

[0025] According to this embodiment, a method for estimating radio frequency transmission interference is provided. The method includes a step of acquiring a signal Rx received through a wireless channel 110. The step of acquiring can be, for example, a step of receiving. However, this step can also simply include a step of acquiring a signal Rx received through an interface to a device that has performed the reception. The method includes a step of processing the acquired signal Rx by a machine learning module 120 trained to distinguish N sources of interference by applying supervised learning, where N is an integer greater than 1. The method further includes a step of outputting, for each source j of the N sources, the contribution of the j-th source of interference to the acquired signal from the machine learning module 120. It is noted that the N sources are not necessarily any possible sources. N is the number of sources for which the trainable module 120 has been trained.

[0026] In general, the machine learning module can be a trained (trainable) module. The contribution can be anything that indicates the presence and / or amount of a particular impairment type (j-th). According to an exemplary implementation, the contribution indicates one of the presence or absence of a contribution from the (j-th) source of impairment to the received signal. In other words, the contribution is indicated by a binary value.

[0027] In the above method, the step of processing the acquired signal may further include a step of obtaining a feature vector including an element for each source j of the N sources, the element being indicative of the degree of contribution of the jth source of impairment to the acquired signal. A subsequent step includes comparing whether each jth element of the feature vector exceeds a threshold. In some embodiments, the threshold may be different for different elements of the feature vector. The elements represent each different type of impairment. Thus, there may be a threshold vector corresponding in dimension to the feature vector. However, the disclosure is not limited thereto, and in some embodiments, the threshold may be the same for all elements. After threshold application, the next step is to set, for each jth element, the contribution of the jth element to TRUE if the jth element exceeds the threshold, and set the jth element to FALSE otherwise. A value TRUE indicates that there is an impairment of the jth type (of the N impairments) in the received signal Rx, and FALSE indicates that there is no impairment of the jth type (of the N impairments) in the received signal Rx.

[0028] However, it is noted that the present disclosure is not limited to outputting a binary value. Rather, in an exemplary implementation, a non-binary value is output, which can be a feature vector. The feature vector can undergo some further processing, such as quantization with more than two possible values ​​(the above-mentioned threshold application can be seen as quantization with two levels). For example, in a possible implementation, the contribution of the j-th source of disturbance to the acquired signal indicates a degree of contribution that can take one of M values, where M>2. The step of processing the acquired signal outputs a feature vector with a j-th element for each source j of the N sources, and the j-th element indicates a degree of contribution of the j-th source of disturbance to the acquired signal. Obtaining the training dataset10

[0029] The above method belongs to the phase 40, inference, shown in FIG. 2. In the following, the training set data preparation phase 10 of FIG. 2 is briefly illustrated. A simple scenario is shown in FIG. 4. In the system mode of FIG. 4, a transmit signal Tx is transmitted and affected by RF impairments 202 and 203. The resulting affected transmit signal is transmitted through the channel 210. The received signal Rx is then captured by the receiver. Afterwards, some specific impairments are detected, for example by some model-based methods or using any single impairment detection method. The detection is illustrated in FIG. 5 by separate models 1 to N, representing the detection module 310. Model 1, model 2, etc. can be conventional models known in the literature, for example for detecting a certain single type of impairment. Also, each model 1 to n can be a deep learning model. In addition, these models can be machine learning models, since they can be specifically designed to solve only one problem, for example to reliably detect one type of impairment. Any of Model 1, Model 2, etc. may estimate two or more RF impairments during the training phase. A model (one of 1 to n) may, for example, estimate the presence of a combination of two or more impairments.

[0030] The impairments can be any impairments such as frequency offset, phase offset, clock offset, power amplifier problems (such as nonlinearity), filter distortion, and IQ modulator problems, etc. These estimated values ​​are then stored as a vector 320. This process is shown in FIG. 5. Correspondingly, the received signal Rx from which these values ​​320 are obtained is stored as an input. These processes are repeated until a sufficient amount of input-output (Rx, 320) pairs of a data set are generated. The desired or appropriate size of the data set can be determined according to system requirements, such as system performance (quality of impairment recognition), computational complexity, and / or memory.

[0031] The training data set obtained in phase 10 may be used for training phase 20. The trained model may then be used for testing 30 or inference 40. Figure 6 illustrates training phase 20, in which a deep learning module 420 is trained.

[0032] In FIG. 5, the determination of the fault indication, indicating the type and / or parameters of the fault, was done by a fault-specific model, or by a model that has already been trained and tested, or by some non-trainable approach. However, the present disclosure is not limited to any particular acquisition of the training data set. In general, acquiring the training set includes the steps of generating an input signal Tx, determining a fault indication, indicating the type and / or parameters of the fault, degrading the input signal by the fault, and acquiring a signal Rx degraded by the fault and the transmission channel by transmitting the degraded input signal through a wireless channel and receiving the transmitted signal. In general, the fault sources can be simulated. For example, there are simulators or simulating software that can generate faults with known parameters (quantities), and then use them to degrade the transmitted signal and perform or simulate the transmission through the channel. In such a case, the training data set pair is formed by the known faults (and possibly their quantities) as the output (ground truth) and the resulting received signal as the input to the trainable model.

[0033] The acquired received signal Rx is then stored together with the impairment indication for use as a training set. The impairment indication may be an index, or a name, or anything that represents an assignment to a particular impairment type that is recognizable by the model to be trained. The impairment indication may include an amount of impairment. As a specific example, the impairment indication may include an impairment type, such as a frequency offset, and an impairment amount, such as a value of the frequency offset. The present disclosure is not limited to the case where both the impairment type and the impairment amount are detected. The present disclosure provides embodiments in which only the impairment type is detected, and embodiments in which both the impairment type and the impairment amount are detected. It is noted that the transmission through the channel 210 may be simulated. In such a simulation, the channel may be represented by a certain mathematical model or may be obtained by simulating actual transmission conditions. However, having data from a real system may help to train the DL model more efficiently for practical use.

[0034] The training set may include pairs of inputs (received signals) and outputs (impairments involved) produced by various combinations of two or more impairments among the distinguishable impairments (impairment sources). This may allow improving the model performance and exploiting correlations between different impairments. training 20

[0035] The DL model training phase 20 is illustrated in Fig. 6. The method may be provided for training a machine learning module 420 to estimate radio frequency transmission impairments 402, 403. The method includes obtaining a training set comprising a plurality of training data including an input signal degraded by impairments and by a transmission channel Rx and an impairment indication indicating a type of impairment 320. Obtaining the training set may be, for example, retrieving the training data from a memory or any type of storage. Obtaining may also correspond to obtaining as described above with reference to Fig. 5. The method may further include inputting the training set to the machine learning module 420. In other words, pairs of the received signal and the ground truth of the detected impairments contained in the received signal are provided to the machine learning module 420. What follows is a step of adapting parameters of the machine learning module according to machine learning using the input training set. Finally, the adapted parameters of the trainable model are stored for use in the step of estimating radio frequency transmission impairments.

[0036] FIG. 6 exemplarily illustrates a deep learning module 420. However, it is noted that in general the module may be a machine learning module, or simply a learning module. In other words, the present disclosure is not limited to deep neural networks for use in fault detection. Any model-based approach may be used. Deep learning refers to a trainable model having two or more layers. Machine learning refers to a trainable model having any structure, including a layer structure, applying one single layer or more layers. The model may be embedded in a functional and / or physical module. Accordingly, when referring to a trainable module or a learning module in this specification, what is meant is a functional module that implements a trainable model, such as a machine learning or deep learning or other type of model. The trainable model may be a neural network or other type of trainable model. Here, for example, a multi-layer perceptron (MLP), a long short-term memory (LSTM), a convolutional neural network (CNN), or a variation or combination thereof may be used as the deep learning 420.

[0037] The machine (trainable model) is trained 20 using the dataset created in the training stage 10. Training can be done, for example, by supervised or unsupervised learning. During training, the trainable parameters of the trainable model 420 are adapted based on the difference between the actual output 430 of the trainable model and the desired output 402, 403 of the trainable model. The desired output can be the associated impairments (ground truth) in the training dataset for a certain received signal Rx. There are many possibilities for learning. In some embodiments, the adaptation of the parameters is done by backpropagation. Backpropagation is known from the literature and is widely used. However, the present disclosure is not limited thereto and alternative techniques for backpropagation exist. For example, alternative approaches such as differential target propagation, HSIC (Hilbert-Schmidt Independence Criterion), online alternating minimization with auxiliary variables, or separated neural interfaces with synthetic differentiation, etc., can be used.

[0038] In the training phase, the model may be configured before it is trained. Parameters of the model or of the learning algorithm that may be set but are not trainable are called hyperparameters. Hyperparameters may be optimized for performance accuracy and complexity. In particular, hyperparameters are parameters whose values ​​are used to control the learning process. In contrast, the values ​​of other (trainable) parameters (usually node weights) are derived by training. Hyperparameters can be classified as model hyperparameters that cannot be inferred during fitting the machine to the training set, because they refer to model selection tasks, or algorithm hyperparameters that do not in principle affect the performance of the model, but affect the speed and quality of the learning process. Examples of model hyperparameters are the topology and size of the neural network. Examples of algorithm hyperparameters are the learning speed and the training batch size.

[0039] It is noted that training may be performed based on a training set having a predefined number of training pairs. However, training may alternatively be performed or stopped based on convergence criteria. For example, training may be stopped if the improvement of the model upon feeding new training data is below a certain threshold. Or, training may be stopped if the difference between the model's output and the desired output (possibly in the mean or median over multiple training data pairs, etc.) is below a predefined threshold. Such thresholds may also be considered as hyperparameters.

[0040] The trained parameters (e.g., weights of a neural network model) are stored and the trainable model is configured with these parameters in a testing and / or inference phase. After training has occurred (training phase), a testing phase may begin. The testing phase 30 is functionally similar to the inference phase 40.

[0041] In Fig. 6, the channel 410 is shown as an additional impairment source. It is noted that the impairments introduced by the channel may also be detected / recognized in some embodiments. For example, during a training phase, the impairments caused by the channel may be trained together with other impairments and thus become recognizable. However, the present disclosure is not limited to such an approach. Rather, the channel may simply be considered as some kind of noise with unknown impairments / characteristics. Test 30 and Inference 40

[0042] The testing phase is shown in FIG. 7. In this phase, a signal Rx is received, which is fed to a trained DL model 550. The DL model 550 can then estimate the RF impairments in a single stage 560. In the testing phase, the trained model can be tested with some Rx signals that are not part of the training set, but for which impairments are known. For example, a testing set can be obtained similar to the training set. The performance of the model 550 can then be tested, for example in terms of accuracy, i.e., how often the trained model 550 correctly identifies the impairments 560. The testing can concern the identification of the presence of impairments, and the detection of the amount of impairments. In the testing phase, if the results (performance of the trained model) are not satisfactory, the training can be continued or some hyperparameters can be modified and the model can be trained again (anew), for example with the same or a different training set.

[0043] In the inference phase, the trained model is used for fault detection (recognizing the presence and / or quantification). Exemplary Implementations and Variations

[0044] As discussed above with reference to FIG. 3 and other figures, embodiments of the present disclosure provide methods and apparatus for detection (estimation) of impairments involved in a transmission chain (transmitter side, channel, receiver side) of a wireless communication system (such as that shown in FIG. 1).

[0045] As mentioned above, the present disclosure is not limited to any particular trainable model. There are a variety of ML and DL techniques that may be used. In some embodiments, the machine learning includes one or more types of machine learning methods, including multi-layer perceptrons, long short-term memory, and convolutional neural networks. For example, at least two of the N sources of impairments may be processed by different types of machine learning methods. For example, since some of the single-carrier and multi-carrier impairments are different, different models may be used for each of them (two different DL models). In other words, the single-carrier signal source impairments may be detected by a different trainable model than the trainable model used to detect impairments from multi-carrier signal sources. Although some of the single-carrier and multi-carrier RF impairments are similar, the RF impairments and their effects may change the waveforms of these two. Thus, the signals in the training dataset may be split into single-carrier and multi-carrier. Training is performed separately for the single-carrier and multi-carrier signals to obtain two separate trained models. The RF impairments are then trained by a deep learning (DL) algorithm to learn the single-carrier RF impairments and another DL algorithm to learn the multi-carrier RF impairments. A received signal in the testing or inference phase is first identified as a single-carrier or multi-carrier signal. If the signal is multi-carrier, a model trained on a multi-carrier waveform is used to estimate multiple RF impairments, and if the signal is a single-carrier signal, a model trained on a single-carrier waveform is used to estimate the RF impairments.

[0046] In the above examples, two different models were used, one for single carrier and one for multi-carrier signals, respectively. The additional models may be preceded by processing with a model dedicated to distinguish whether the received signal is a single carrier signal or a multi-carrier signal. According to the result of the distinction, the signal is input to the corresponding single carrier trained model or multi-carrier trained model for fault detection.

[0047] The (training or testing) data set can be created, for example, using MATLAB or a similar simulator program. In that case, there is no need to estimate the impairments by a model or ML-based methods, since the impairments created in the simulation can be known. These known impairments can be used in the training phase and the trainable model can be trained according to them. In the testing phase, the received signal to be estimated is fed to the DL algorithm, which has been trained as described above.

[0048] The advent of deep learning is to alleviate the need for human-based feature extraction / crafting / engineering. Thus, it is possible to directly apply the above methods to real-world measurements without any feature extraction. Conversely, the "deep" attribute (number of layers) of the machine learning used can be reduced at the expense of adding more feature extraction. In an extreme case, if it is possible to extract distinctive features to estimate multiple RF impairments, it is possible to avoid deep learning altogether and use machine learning. The stages at which feature extraction can be done are shown in Figure 6.

[0049] FIG. 8 shows an overview of such multiple RF impairments estimation using a machine learning model with feature extraction. After obtaining a received signal Rx from a channel 610, a feature extraction module 620 extracts certain features. Some feature(s) of the signal can be learned and these feature(s) can be fed as input to the trainable module. It is noted that in some embodiments, these features can be fed together to the trainable module. For example, the features can include (without limitation) one or more of the following: Error Vector Magnitude (EVM), Complementary Cumulative Distribution Function (CCDF), constellation, etc. An error vector is a vector in the IQ plane between an ideal constellation point and a point received by the receiver. In other words, it is the difference between the actual received symbol and the ideal symbol. The root mean square (RMS) average amplitude of the error vector normalized to the ideal signal amplitude reference is the EVM. The EVM can be expressed in percentage by multiplying the ratio by 100%. The term "constellation" here refers to the modulation constellation.

[0050] For example, it is possible to observe the spectrum before and after the power amplifier, thus observing the spectral regrowth. Such spectral regrowth may represent the features to be extracted. These are just examples, and in general, any other features that may be known from the prior art may be extracted. After feature extraction, the machine learning module 630 recognizes the impairments based on the extracted features.

[0051] The presence (yes or no) of RF impairments can be represented by binary numbers such as 1 and 0, instead of finding their numerical value (amount of impairment). For example, the output becomes represented as 1 if impairments are present and 0 if they are not present. In other words, the regression problem is now changed to a classification problem. By such classification, the presence of impairments in the received signal can be found. After the estimation of the presence of RF impairments, the numerical value of the RF impairments can be found in another way. For example, the amount of impairments can be found by some conventional method or by a module specifically trained for a certain individual impairment. In other words, the trainable module can be a classifier that detects the presence or absence of impairments or combinations of impairments in the received signal.

[0052] Some embodiments of the present disclosure can be used to learn the presence of RF impairments while simultaneously finding their values. In this way, a trainable module is trained to find the amount of each impairment in a received signal. For example, after an impairment is found, if the found value (amount of a particular impairment) is higher than a threshold, it is assumed that the impairment is present. If the found value (amount of a particular impairment) is lower than a threshold, it is assumed that the impairment is not present.

[0053] One advantage of the present disclosure is that the RF impairments are not limited to those mentioned above. It is possible to estimate the impairments for any kind of defects in the system without modifying the architecture, by training the network accordingly, especially in deep learning-based implementations. The present disclosure does not have to be applied to all possible impairments. It can be applied to a certain group of selected RF impairments, which can then be estimated using the proposed method. The remaining impairments can be considered to represent some noise in the system or can be detected by using another method.

[0054] The output of the fault detection can be a vector or a matrix according to the fault to be detected. For example, the channel characteristics can be a vector, while each of the frequency offset and phase offset can be a value. They can be represented together in the output in various ways. For example, the channel vector, the frequency offset, and the phase offset can be stacked in a vector. Such a vector has a length (in terms of number of elements) of the channel vector + 2 (because of the frequency offset and phase offset values). However, the present disclosure is not limited to such an output. FIG. 9 shows some alternative exemplary output formats. Format (a) is a matrix in which there is a column per fault source. The first column corresponds to the channel vector, the second column contains a value f (frequency offset) at a predetermined (here the first) row position, and the last column contains a value p (phase offset) at a predetermined (here the first) row position. There can be further (other) fault sources in between, represented by vectors or values. Format (b) is similar to format (a), with the difference being the predefined row positions for values ​​f and p. Padding may be used to fill unused rows in columns that contain only values ​​such as f and p (or vectors shorter than the number of rows of the matrix). In this example, padding is done by inserting zeros. It is noted that instead of columns, there may be rows per fault source. In general, it is conceivable to produce a tensor representation, e.g., when the fault sources are represented by matrices or tensors. Alternatively, any matrix or tensor may be vectorized.

[0055] The detected impairments may be used for a variety of different purposes, such as receiver side compensation, predistortion, and / or physical layer authentication (PLA).

[0056] Correspondingly, the method applied at the signal receiving side may further comprise a step of compensating the acquired signal based on the outputted contribution of the j-th source of interference to the acquired signal.

[0057] After estimating the multiple RF impairments, the effects of these impairments can be removed from the signal, which is called compensation. It is noted that complete removal may not be possible, but compensation may at least reduce the amount of certain impairments, or combinations of impairments, from the received signal. As a result, such signals can then be decoded with a higher probability of correct decoding. As an example, assume that there is only a phase offset in the system. For a single carrier system, the data is downsampled and the downsampled data is divided by the phase offset to compensate for the effects. When multiple impairments are found using a trained model as described above, the reverse process of introducing impairments in the transmitted signal can be applied for compensation. Compensation of RF impairments can be important for synchronization. For example, FIG. 10 is a block diagram showing an example for single carrier synchronization. FIG. 11 is a block diagram showing an example for multi-carrier synchronization.

[0058] As can be seen in FIG. 10, the initial data is downsampled and then the phase offset is removed. In particular, FIG. 10 shows the steps / functional modules (units) of an exemplary single carrier receiver. After sampling and analog-to-digital conversion, a sampled (digital) received signal is obtained. In step 710, a matched filter may be applied to the received digital signal for one or more channel paths. In step 720, a coarse synchronization may be performed based on the detected channel. The detection in step 710 is usually performed based on some synchronization signal (preamble) that is known (at least in part) to the transmitter and receiver. In step 730, a frequency offset is obtained and in step 735, the frequency offset is compensated. The frequency offset may be obtained by the impairment detection described above. In step 740, symbol timing and sample timing estimation is performed. In step 750, downsampling is performed and the desired time slot is extracted. Phase offset estimation and channel estimation are performed in step 760, and may be performed together with the impairment detection described above and possibly together with the frequency offset estimation in step 730. According to the estimated channel and phase offsets, a correction is performed in step 770. In step 780, the compensated signal is demodulated and in step 790, the demodulated signal is converted to data. Data may be specified in terms of bits, bytes or characters depending on the layer specification (standard), where a byte refers to 8 bits and a character corresponds to one or more bytes depending on, for example, the hardware / software architecture.

[0059] FIG. 11 is a block diagram showing an example for multi-carrier synchronization. The block diagram shows the steps / functional modules of the device. As can be seen in the figure, the received complex valued (IQ) samples are input to the process. Based on that, the start of the packet is detected, a coarse frequency offset estimation is performed, and the frequency offset is corrected accordingly. Then, a timing estimation is performed, and the preamble and symbol timing are adjusted accordingly. Then, the end of the packet is detected, and the packet is extracted. Then, a finer frequency offset is estimated and corrected. Then, subcarrier demodulation is performed. For example, the cyclic prefix can be removed, and a fast Fourier transform (FFT) can be applied to obtain the subcarrier modulation symbols. Then, modulation (IQ) impairments can be detected and compensated. Also, a channel estimation is performed and can be compensated by channel equalization. Based on the pilot tracking, the phase offset, and further the timing offset and gain can be estimated and compensated. Based on the compensated frequency signal values, demodulation can be performed. The present disclosure is not limited to any particular demodulation. For example, the detection output soft information can be used, or the detection can directly output hard values. Further baseband processing may include de-interleaving, decoding, etc. In the example of Figure 11, the frame control header (FCH) is decoded.

[0060] 10 and 11, any one or more (or all) of the impairments may be estimated using the impairment detection described in the embodiments and examples of the present disclosure. For example, any one or more of the following impairments may be estimated in this manner: coarse frequency offset, symbol timing, fine frequency offset, modulation impairments, channel taps, phase offset, power gain, sample timing offset, or any other type of impairment. It is noted that these impairments and receiver chains are merely exemplary. In other receiver architectures, other impairments may be more appropriate.

[0061] From the art, several impairment compensation techniques are known, any of which may be applied. For example, QI, Jian. Analysis and Compensation of Channel and RF Impairments in MIMO Wireless Communication Systems. 2011. PhD Thesis. Universite du Quebec, Institut national de la recherche scientifique (available at http: / / espace.inrs.ca / id / eprint / 2160 / ) provides an overview of some of the techniques that may be used even with the impairment detection of the present disclosure.

[0062] According to one embodiment, a signal received through a wireless channel is received from a transmitting device 1. Then, an impairment detection is performed as described in any of the above examples. The method further comprises a step of transmitting an indication of the contribution to at least one of the N sources of impairment to the transmitting device. In such a case, compensation does not need to be performed at the receiver. For example, the transmitting device may then proceed by receiving an indication of the contribution to at least one of the N sources of impairment at the transmitting device and applying compensation according to the received indication at the transmitting device. Such compensation at the receiver side before the signal is transmitted is called predistortion. There are several predistortion techniques known from the art, any of which may be used. It is noted that the term "predistortion" is used herein in the broader context of any kind of impairment. This may alternatively be called "preprocessing". In general, such predistortion or preprocessing refers to an optimized transmitter and receiver design to compensate for current / upcoming RF impairments according to a previous estimated RF impairment. Predistortion (preprocessing) can be applied to compensate for the PAPR or to compensate for any detected impairment sources.

[0063] The impact of RF impairments on different transceivers has been investigated, for example, in JIN, Yuehai; DAI, Fa Foster. Impact of transceiver RFIC impairments on MIMO system performance; IEEE Transactions on Industrial Electronics, 2011, 59.1:538-549, or KIAYANI, Adnan, et al. Advanced receiver design for mitigating multiple RF impairments in OFDM systems: algorithms and RF measurements; Journal of Electrical and Computer Engineering, 2012. After estimating the RF impairments, if any RF impairments should be removed or their effects should be reduced before they occur in the upcoming signal, a similar approach to the above publication or other publications can be applied. In the above example of predistortion, the impairments were detected at the receiver and compensated at the transmitter. However, the present disclosure is not limited to such an approach. In general, certain types (or combinations) of impairments can be compensated at the receiver side, and other types (or combinations) of impairments can be compensated at the transmitter. The type of compensation may be determined based on the type (origin) of the impairments. Correspondingly, the indication provided by the receiver 2 to the transmitter 1 may include only those impairments that will be addressed at the transmitter side (by predistortion).

[0064] Multiple RF impairments may be estimated using one of the above-mentioned embodiments and examples. The upcoming signal (next signal to be transmitted from the transmitter) is expected to have the same or similar RF impairments. Therefore, a waveform may be designed to mitigate these effects at the receiver. For example, if the PAPR is higher among the estimated RF impairments and if a waveform-shaped (root raised cosine, RRC, etc.) pulsed single carrier is used, the roll-off value may be increased at the transmitter and the PAPR value may be reduced for the upcoming signal. Alternatively, the modulation order may be reduced. In other words, predistortion is not limited to compensating for impairments by inverting this effect. Rather, compensation at transmission may include modifying one or more transmit parameters to improve reception based on the type of impairment detected. The impact of RF impairments may vary according to various waveforms and different scenarios (frequency, mobility). Following this idea, an optimal waveform design may be provided for the upcoming signal according to the estimated RF impairments.

[0065] As mentioned in the above examples, sources of impairments at the transmitter side include one or more of frequency offset, phase offset, clock offset, power amplifier impairment, filter impairment, and modulation impairment, or other impairments.

[0066] RF impairments include, among others, power amplifier nonlinearities (caused by a high peak-to-average power ratio, PAPR, of the signal), IQ modulator impairments (quadrature offset, IQ gain imbalance, and / or DC offset), phase noise, frequency offset, sample clock errors, noise (Additive White Gaussian Noise, AWGN), analog-to-digital and / or digital-to-analog (ADC / DAC) issues, or dynamic range provided by the hardware.

[0067] The present disclosure may further be used to identify, and possibly quantify, channel impairments, including one or more of channel dispersion, fading, interference, and the like. Physical Layer Authentication

[0068] Another application of the detection of impairments according to some embodiments of the present disclosure is physical layer authentication (PLA). Accordingly, after estimating the impairments, the method of some embodiments may further include performing physical layer authentication based on the output contributions of the N sources of impairments to the acquired signal.

[0069] For use in PLA, RF impairments are first detected. These detected impairments are then used for user authentication. Some of the known techniques for physical layer authentication (PLA) may be used. For example, PEI, Chengcheng et al. Channel-based physical layer authentication. 2014 IEEE Global Communications Conference; IEEE, 2014. p.4114-4119, first estimates the channel and then uses machine learning (particularly support vector machine, SVM) for physical layer authentication. Initial channel and / or RF impairments can be estimated by the impairment estimation techniques mentioned above, and then machine learning algorithms can be used for physical layer authentication. In this technique, machine learning is used as a kind of threshold mechanism. The optimal thresholds may also be determined according to the true positive and false positive rates (or learned). The work by PEI et al. mentioned above is merely an example, and the present disclosure may use any other mechanism that relies on estimated impairments to check the legal / illegal status of a particular user. One of the disadvantages of some known PLA works is that they train their models in a training phase according to legitimate signals. Then, according to the received signal, they decide whether the signal is legitimate or not. However, this is disadvantageous in that they cannot provide information about the faults for which they determine that the signal is not legitimate. In other words, the machine, as a black box, gives a decision whether the user is legitimate or not, without detecting RF faults. To improve reliability, in some embodiments of the present disclosure, previously defined faults of legitimate users are collected and compared with the detected actual faults that become identified according to a threshold. Then, a decision is given whether the arriving signal is legitimate or not. In this way, the trainable module can learn which faults give better or worse information about the legitimacy or non-legitimateness. Thus, more reliable results can be obtained in terms of security perspective, since the machine gives information about different (unusual), specific RF faults (or combinations of faults).Such specific faults can be further checked, for example by a human (technician) or using other methods (e.g. model-based single RF fault detection), and their correctness can be doubly confirmed, which increases reliability.

[0070] 12 and 13 illustrate the present invention for the PLA training and testing phases, respectively.

[0071] 12 illustrates an exemplary PLA training stage. In particular, legitimate and possibly illegitimate training pairs are input to the trainable module 860 in order to train it. If the user is legitimate, the legitimate training pairs include possible inputs (received signals) and their respective desired outputs (faults). After training, the trained parameters 870 of the trainable module 869 are output.

[0072] In the training phase, additionally, threshold selection may be added to the process to distinguish whether a user is legitimate or not. First, multiple RF impairments are estimated, and in the threshold training module / phase 850, a threshold is selected according to the true positive rate (TPR) and the false positive rate (FPR). The optimal threshold is learned, for example, in the training phase using legitimate and non-legitimate signals as input to the learning. In the testing phase, the signal to be authenticated is received as input. The result of the threshold learning is a set of thresholds for each set of impairments to be estimated.

[0073] FIG. 13 shows an exemplary PLA in the testing and inference phase. The received signal Rx is input to a trainable module (trained deep learning model) 950, which may correspond to the trained deep learning module 870 and the trained thresholds output from the threshold training module 850. The trainable module 950 estimates the faults as described above. Then, the judgment module 990 determines whether a particular input received signal Rx is valid or invalid based on the estimated faults and by using the determined thresholds. In particular, the judgment module may compare the amount of a particular fault source estimated with the respective thresholds learned for said particular fault source. This may be done for all faults.

[0074] The example of PLA using fault detection described above with reference to Figures 12 and 13 is not limiting of the present disclosure. For example, instead of learning separate thresholds, the detected faults can be input to a deep learning module trained to determine whether the faults indicate that the signal degraded by them is legitimate or non-legitimate. Condition 3 can be designed based on the fault values ​​and their combinations, etc. The output of the PLA can be a binary decision, or it can be the probability that the signal is legitimate / non-legitimate. It is noted that the term non-legitimate can be understood as abnormal compared to an expected (normal, legitimate) signal.

[0075] To summarize, in the above example, multiple RF impairments are estimated and the signal is authenticated according to the estimated RF impairments and thresholds. More specifically, if the difference between the estimated RF impairments is higher than the respective thresholds, the user is deemed to be unauthorized, and if the difference is lower than the thresholds, the user is deemed to be authorized.

[0076] In general, after detecting an RF disturbance, the disturbance may be compared with a certain threshold designed according to some pre-certified value. If the disturbance exceeds the threshold, the user may be considered illegitimate. The threshold(s) may be selected according to the false or true alarm rate. In other words, the selection of the threshold may depend on the scenario. Exemplary Beneficial Effects

[0077] Some embodiments of the present disclosure can be used to detect and compensate for RF interference on the physical layer, and they can be used in physical layer security. Additionally, some embodiments can be related to synchronization. Also, when high order modulation and high frequencies are used, the impact of RF interference increases. Therefore, the present disclosure can be easily used for these applications.

[0078] Some embodiments may allow multiple RF impairments to be detected using a single model. For this model, all given impairments can be estimated, and it is not necessary to find which particular impairment needs to be detected, as is common in the prior art. Additionally, when applying a single model, correlations between impairments are exploited. This may result in higher accuracy in RF impairment estimation.

[0079] Some embodiments may also be used for RF impairment compensation.

[0080] Furthermore, the present disclosure can be used in most of the cognitive radio applications such as spectrum sensing. Moreover, it can be effective in system performance especially at high modulation order and frequency where the impact of RF impairments is increased. In addition, after the channel quality condition and RF impairments are estimated using the present invention, the modulation order can be adjusted according to the estimated value to bring about efficiency. For example, when the channel condition improves, the modulation order can be increased, and when the channel deteriorates, the modulation order can be decreased.

[0081] Some embodiments may be used in security, especially for user authentication. In conventional methods, ML-based models are used to identify whether a user is legitimate or not, without estimating RF impairments. However, when these models are used, it is difficult to control whether the estimation is correct or not. This is because the ML models only give information about the user's legitimacy (as 1 or 0), but not about their impairments. In addition, it has advantages to estimate RF impairments and to justify the user according to these impairments. For example, after estimating the impairments by the proposed method, these impairments can be controlled by a model-based method to verify the estimation, and thus a decision is given with more confidence. In other words, the present disclosure brings higher awareness to authentication. Software and Hardware Implementation

[0082] The methodologies described herein may be implemented by various means depending on the application. For example, these methodologies may be implemented in hardware, an operation system, firmware, software, or any combination of two or all of these. For hardware implementations, any processing circuit may be used, which may include one or more processors. For example, the hardware may include one or more of an application specific integrated circuit (ASIC), a digital signal processor (DSP), a digital signal processing device (DSPD), a programmable logic device (PLD), a field programmable gate array (FPGA), a processor, a controller, any electronic device, or other electronic circuit unit or element designed to perform the functions described above.

[0083] If implemented as program code, the functions performed by the sending device may be stored as one or more instructions or code on a non-transitory computer-readable storage medium, such as memory 610 or any other type of storage. Computer-readable media includes physical computer storage media, which may be any available medium that can be accessed by a computer, or generally by processing circuitry 620. Such computer-readable media may comprise RAM, ROM, EEPROM, optical disk storage, magnetic disk storage, semiconductor storage, or other storage devices. Some specific and non-limiting examples include compact discs (CDs), CD-ROMs, laser discs, optical discs, digital versatile discs (DVDs), Blu-ray (BD) discs, and the like. A combination of different storage media is also possible, in other words, distributed and heterogeneous storage may be used.

[0084] The above-described embodiments and exemplary implementations provide some non-limiting examples. It is understood that various modifications may be made without departing from the claimed subject matter. For example, modifications may be made to adapt the examples to new systems and scenarios without departing from the core concepts described herein.

[0085] FIG. 14 shows an example structure of an apparatus 1050 for estimating multiple faults. The wireless transceiver 1030 can receive the signal Rx and pass it to the processing circuit 1020. The processing circuit 1020 can be configured by a program fetched from the memory 1010 to perform fault estimation, for example, as described in any of the embodiments and examples above. The apparatus can have a user interface 1040 that can be used to control the training and / or inference phases, etc. The wireless transceiver 1030, the memory 1010, the processing circuit 1020, and possibly the user interface 1040 can be interconnected through a bus 1001. FIG. 15 shows an example memory 1010 that includes a trained (trainable) module program 1060 and a learning (training) module program 1080. It is noted that this implementation is merely exemplary. There can be different architectures to implement the inference or testing or training phases, or any combination thereof. Selected embodiments and examples

[0086] In summary, methods and techniques are described for impairment detection for wireless communications. The present disclosure relates to detection of impairments in a received wireless communication signal. It may detect the presence of each of N types (sources) of impairments and possibly the amount of impairment present in the received signal. The detection involves processing the received signal by a trainable module trained to distinguish between the N sources of impairments by applying learning, where N is an integer greater than 1. The trainable module outputs, for each source j of the N sources, the contribution of the jth source of impairment to the acquired signal. The contribution may be a binary value indicating the presence or absence of the jth source of impairment, or may also indicate the amount of impairment.

[0087] According to one embodiment, a method for estimating radio frequency transmission interference is provided, the method comprising the steps of acquiring a signal received through a wireless channel, processing the acquired signal by a trainable module trained to distinguish between N sources of interference, where N is an integer greater than 1, and outputting from the trainable module, for each source j of the N sources, a contribution of the jth source of interference to the acquired signal.

[0088] For example, sources of impairments at the transmitter side include one or more of frequency offset, phase offset, clock offset, power amplifier impairments, filter impairments, and modulation impairments.

[0089] In some embodiments, the contribution indicates one of presence or absence of a contribution from a source of interference to the received signal, and processing the acquired signal further includes obtaining a feature vector comprising an element for each source j of the N sources, the element indicating a degree of contribution of the jth source of interference to the acquired signal; comparing whether each jth element of the feature vector exceeds a threshold; and, for each jth element, if the jth element exceeds the threshold, setting the contribution of the jth element to TRUE, and otherwise setting the jth element to FALSE.

[0090] In some embodiments, the contribution of the jth source of disturbance to the acquired signal indicates a degree of contribution that can take one of M values, where M>2, and the step of processing the acquired signal outputs a feature vector comprising a jth element for each source j of the N sources, the jth element indicating a degree of contribution of the jth source of disturbance to the acquired signal.

[0091] For example, the machine learning includes one or more types of machine learning methods including multi-layer perceptrons, long short-term memory, and convolutional neural networks. In an exemplary implementation, at least two of the N sources of impairment are handled by different types of machine learning methods.

[0092] The method may further include compensating the acquired signal based on the outputted contribution of the j-th source of impairment to the acquired signal.

[0093] In some embodiments, the signal received through the wireless channel is received from a transmitting device, and the method further includes transmitting an indication of the contribution to at least one of the N sources of impairment to the transmitting device. For example, the method further includes receiving, at the transmitting device, an indication of the contribution to at least one of the N sources of impairment, and applying predistortion at the transmitting device according to the received indication.

[0094] The method may include performing physical layer authentication based on the outputted contributions of the N sources of impairments to the acquired signal.

[0095] According to one embodiment, there is provided a method for training a trainable module for estimating radio frequency transmission impairments, the method comprising the steps of obtaining a training set comprising a plurality of training data including an input signal degraded by impairments and by a transmission channel and an impairment indication indicating the type of impairment, the signal, inputting the training set into the trainable module, adapting parameters of the trainable module according to the input training set, and storing the adapted parameters for use in the step of estimating radio frequency transmission impairments.

[0096] In an exemplary implementation, obtaining the training set includes, for each training data in the training set, generating an input signal, determining a fault indication indicating a type and / or parameter of the fault, degrading the input signal with the fault, and obtaining a degraded input signal with the fault and a transmission channel by transmitting the degraded input signal through a wireless channel and receiving the transmitted signal.

[0097] According to one embodiment, an apparatus for estimating radio frequency transmission impairments is provided, the apparatus comprising: a processing circuit configured to acquire a signal received through a wireless channel; process the acquired signal by a trainable module trained to distinguish between N sources of impairments by applying supervised learning, where N is an integer greater than 1; and output from the trainable module, for each source j of the N sources, a contribution of the jth source of impairments to the acquired signal.

[0098] For example, the processing circuitry is further configured to train the learning module by obtaining a training set comprising a plurality of training data including an input signal degraded by impairments and by a transmission channel and an impairment indication indicating a type of impairment, the signal, inputting the training set to the trainable module, adapting parameters of the trainable module according to machine learning using the input training set, and storing the adapted parameters for use in the step of estimating radio frequency transmission impairments.

[0099] According to one embodiment, an apparatus is provided for receiving a signal degraded by a number of impairments, the apparatus comprising a receiver (which is not necessarily a full wireless transceiver, but may be a simple input) for receiving a signal received through a wireless channel, an apparatus (as described above) for estimating radio frequency transmission impairments in the received signal, and a compensation module, the compensation module being configured to compensate the received signal for the estimated radio frequency transmission impairments and to transmit an indication of the estimated radio frequency transmission impairments as feedback to a transmitter from which the signal was received.

[0100] According to one embodiment, an apparatus for training a trainable module for estimating radio frequency transmission impairments is provided, the apparatus comprising: an input for obtaining a training set comprising a plurality of training data including an input signal degraded by impairments and by a transmission channel, and an impairment indication indicating a type of impairment, the signal; a trainable module to which the training set is input via the input; a learning module for adapting parameters of the trainable module according to the input training set; and a memory for storing the adapted parameters for use in the step of estimating radio frequency transmission impairments.

[0101] In an exemplary implementation, obtaining the training set includes, for each training data in the training set, generating an input signal, determining a fault indication indicative of a type and / or parameter of the fault, corrupting the input signal by the fault, and obtaining the corrupted input signal by transmitting the corrupted input signal through a wireless channel and receiving the transmitted signal by a transmission channel. Such obtaining of the training set may be performed by the same device as the device for estimating the faults or in another device, such as a computer or the like.

[0102] As also described above for the method and for the apparatus, the sources of impairments at the transmitter side include one or more of frequency offset, phase offset, clock offset, power amplifier impairment, filter impairment, and modulation impairment. In some embodiments, the contribution indicates one of presence or absence of contribution from the source of impairments to the received signal. The processing circuit may be configured to process the acquired signal by further performing the steps of obtaining a feature vector comprising an element for each source j of the N sources, the element indicating the degree of contribution of the jth source of impairments to the acquired signal, comparing whether each jth element of the feature vector exceeds a threshold, and for each jth element, if the jth element exceeds the threshold, setting the contribution of the jth element to TRUE, and otherwise setting the jth element to FALSE. In some embodiments, the contribution of the jth source of disturbance to the acquired signal indicates a degree of contribution that can take one of M values, where M>2, and the step of processing the acquired signal outputs a feature vector comprising a jth element for each source j of the N sources, the jth element indicating a degree of contribution of the jth source of disturbance to the acquired signal. For example, the machine learning includes one or more types of machine learning methods including multi-layer perceptron, long short-term memory, and convolutional neural network. In an exemplary implementation, at least two of the N sources of disturbance are processed by different types of machine learning methods.

[0103] The apparatus may further include an impairment compensation module (which may also be embodied by a processing circuit) that compensates the acquired signal based on the output contribution of the jth source of impairment to the acquired signal. In some embodiments, the signal received through the wireless channel is received from a transmitting device. The processing circuit of the apparatus may further control the wireless transmitter to transmit an indication of the contribution of the at least one of the N sources of impairment to the transmitting device. The transmitting device may be configured to receive the indication of the contribution of the at least one of the N sources of impairment and apply predistortion at the transmitting device according to the received indication.

[0104] The processing circuitry of the impairment estimation device described above may be further configured to perform physical layer authentication based on the outputted contributions of the N sources of impairments to the acquired signal.

[0105] Additionally, corresponding methods are provided that include steps performed by any of the processing circuit implementations described above.

[0106] Further provided is a computer program stored on a non-transitory medium and comprising code instructions which, when executed by a computer or by a processing circuit, perform the steps of any of the methods described above.

[0107] According to some embodiments, the processing circuitry and / or the transceiver are embedded in an integrated circuit, IC.

[0108] Any of the devices of the present disclosure may be embodied on an integrated chip.

[0109] Any of the above described embodiments and example implementations may be combined.

[0110] Although the subject matter of the present disclosure has been described in detail for illustrative purposes based on what are presently considered to be the most practical and preferred embodiments, it should be understood that such details are for that purpose only, and that the subject matter of the present disclosure is not limited to the disclosed embodiments, but rather is intended to encompass modifications and equivalent arrangements that are within the spirit and scope of the appended "claims." For example, it should be understood that the subject matter of the present disclosure contemplates that, to the extent possible, one or more features of any embodiment can be combined with one or more features of any other embodiment. [Item of invention] [Item 1] 1. A method for estimating radio frequency transmission interference, comprising: acquiring a signal received over a wireless channel; processing the acquired signals by a trainable module trained to distinguish between N sources of impairments, where N is an integer greater than 1; outputting, from the trainable module, for each source j of the N sources, a contribution of the j-th source of impairment to the acquired signal; A method comprising: [Item 2] The source of interference at the transmitter side is Frequency Offset Phase Offset Clock Offset Power Amplifier Fault Filter failure, and Modulation disorder The method according to item 1, comprising one or more of the following: [Item 3] the contribution indicating one of a presence or absence of a contribution from an impairment source to the received signal; The step of processing the acquired signal further comprises: obtaining a feature vector comprising an element for each source j of the N sources, said element being indicative of the degree of contribution of the j-th source of impairment to the obtained signal; comparing whether each j-th element of the feature vector exceeds a threshold; for each jth element, if the jth element exceeds a threshold, setting the contribution of the jth element to TRUE, otherwise setting the jth element to FALSE; 3. The method according to item 1 or 2, further comprising: [Item 4] the contribution of the jth source of disturbance to the acquired signal indicates a degree of the contribution that can take one of M values, where M>2; said step of processing the acquired signal comprises outputting a feature vector comprising a j-th element for each source j of the N sources, said j-th element being indicative of the degree of contribution of the j-th source of impairment to the acquired signal; 3. The method according to item 1 or 2. [Item 5] 5. The method of any one of claims 1 to 4, wherein the machine learning comprises one or more types of machine learning methods, including multi-layer perceptrons, long short-term memory, and convolutional neural networks. [Item 6] 6. The method of claim 5, wherein at least two of the N sources of impairment are processed by different types of machine learning methods. [Item 7] compensating the acquired signal based on the output contribution of the j-th source of impairment to the acquired signal. The method according to any one of items 1 to 6, further comprising: [Item 8] The signal received over a wireless channel is received from a transmitting device; and transmitting, to the transmitting device, an indication of the contribution to at least one of the N sources of impairment. The method according to any one of items 1 to 6. [Item 9] receiving, at the transmitting device, the indication of the contribution to at least one of the N sources of impairment; applying predistortion at the transmitting device in accordance with the received indication; The method according to item 8, comprising: [Item 10] performing physical layer authentication based on the output contributions of the N sources of impairments to the acquired signal. The method according to any one of items 1 to 6, further comprising: [Item 11] 1. A method for training a trainable module for estimating radio frequency transmission interference, comprising: obtaining a training set comprising a plurality of training data including input signals corrupted by impairments and by a transmission channel and impairment indications indicative of the type of impairments, the signals; inputting the training set into the trainable module; adapting parameters of the trainable module according to the input training set; storing said adapted parameters for use in said estimating radio frequency transmission interference step; The method includes: [Item 12] The step of obtaining the training set includes, for each training data in the training set, generating an input signal; determining said fault indication, said fault indication being indicative of a type and / or parameters of said fault; degrading said input signal by said impairment; transmitting the impaired input signal through a wireless channel and receiving the impaired input signal; Item 12. The method according to item 11, comprising: [Item 13] 1. An apparatus for estimating radio frequency transmission interference, comprising: The apparatus includes a processing circuit, the processing circuit comprising: Obtaining a signal received over a wireless channel; processing the acquired signals by a trainable module trained to distinguish between N sources of impairments by applying supervised learning, where N is an integer greater than 1; outputting, from the trainable module, for each source j of the N sources, a contribution of the j-th source of impairment to the acquired signal; The apparatus is configured to: [Item 14] The processing circuitry includes: obtaining a training set comprising a plurality of training data including input signals corrupted by impairments and by a transmission channel and impairment indications indicative of the type of impairments, the signals; inputting the training set into the trainable module; adapting parameters of the trainable module according to the machine learning using the input training set; storing said adapted parameters for use in said estimating radio frequency transmission interference step; Item 14. The apparatus of item 13, further configured to train the learning module by: [Item 15] 1. An apparatus for receiving a signal degraded by a plurality of impairments, comprising: a receiver for receiving a signal received over a wireless channel; Item 15. An apparatus according to item 14 for estimating radio frequency transmission interference in the received signal; A compensation module; The compensation module comprises: compensating the received signal for the estimated radio frequency transmission impairments; transmitting an indication of the estimated radio frequency transmission impairment as feedback to a transmitter from which the signal was received; and The apparatus is configured to:

Claims

1. 1. A method for estimating radio frequency transmission interference, comprising: acquiring a signal received over a wireless channel; processing the acquired signal by a trainable module trained to distinguish between N sources of impairments in a transmitter, where N is an integer greater than 1; outputting, from the trainable module, for each source j of the N sources, the contribution of the j-th source of impairments at the transmitter to the acquired signal; A method comprising:

2. The source of interference at the transmitter is Frequency Offset Phase Offset Clock Offset Power Amplifier Fault Filter failure, and Modulation disorder The method of claim 1 , comprising one or more of:

3. the contribution indicating one of a presence or absence of a contribution from an impairment source to the received signal; The step of processing the acquired signal further comprises: obtaining a feature vector comprising an element for each source j of the N sources, said element being indicative of the degree of contribution of the j-th source of impairment to the obtained signal; comparing whether each j-th element of the feature vector exceeds a threshold; for each jth element, if the jth element exceeds a threshold, setting the contribution of the jth element to TRUE, otherwise setting the jth element to FALSE; The method of claim 1 or 2, further comprising:

4. the contribution of the jth source of disturbance to the acquired signal indicates a degree of the contribution that can take one of M values, where M>2; said step of processing the acquired signal comprises outputting a feature vector comprising a j-th element for each source j of the N sources, said j-th element being indicative of the degree of contribution of the j-th source of impairment to the acquired signal; The method according to claim 1 or 2.

5. The method of any one of claims 1 to 4, wherein the machine learning used in the trainable module comprises one or more types of machine learning methods including multi-layer perceptrons, long short-term memory, and convolutional neural networks.

6. The method of claim 5 , wherein at least two of the N sources of impairment are processed by different types of machine learning methods.

7. compensating the acquired signal based on the output contribution of the j-th source of impairment to the acquired signal. The method of any one of claims 1 to 6, further comprising:

8. The signal received over a wireless channel is received from a transmitting device; and transmitting, to the transmitting device, an indication of the contribution to at least one of the N sources of impairment. The method according to any one of claims 1 to 6.

9. receiving, at the transmitting device, the indication of the contribution to at least one of the N sources of impairment; applying predistortion at the transmitting device in accordance with the received indication; The method of claim 8 , comprising:

10. performing physical layer authentication based on the output contributions of the N sources of impairments to the acquired signal. The method of any one of claims 1 to 6, further comprising:

11. Training a trainable module for estimating radio frequency transmission impairments, obtaining a training set comprising a plurality of training data including input signals corrupted by impairments at the transmitter and by the transmission channel, and impairment indications indicative of the type of impairments, the signals; inputting the training set into the trainable module; adapting parameters of the trainable module according to the input training set; storing the adapted parameters for use in the step of estimating radio frequency transmission interference at the transmitter; The method of any one of claims 1 to 10, comprising a training step comprising:

12. The step of obtaining the training set includes, for each training data in the training set, generating an input signal; determining said fault indication, said fault indication being indicative of a type and / or parameters of said fault; degrading said input signal by said impairment; transmitting the impaired input signal through a wireless channel and receiving the impaired input signal; The method of claim 11 , comprising:

13. 1. An apparatus for estimating radio frequency transmission interference, comprising: The apparatus includes a processing circuit, the processing circuit comprising: Obtaining a signal received over a wireless channel; processing the acquired signal by a trainable module trained to distinguish between N sources of impairments in a transmitter by applying supervised learning, where N is an integer greater than 1; outputting from the trainable module, for each source j of the N sources, a contribution of the j-th source of impairments at the transmitter to the acquired signal; The apparatus is configured to:

14. The processing circuitry includes: obtaining a training set comprising a plurality of training data including input signals corrupted by impairments at the transmitter and by a transmission channel, and impairment indications indicative of the type of impairment, the signal; inputting the training set into the trainable module; adapting parameters of the trainable module according to machine learning using the input training set; storing said adapted parameters for use in said estimating radio frequency transmission interference step; The apparatus of claim 13 , further configured to train the trainable module by:

15. 1. An apparatus for receiving a signal degraded by a plurality of impairments, comprising: a receiver for receiving a signal received over a wireless channel; 15. An apparatus according to claim 14 for estimating radio frequency transmission interference at a transmitter in the received signal; A compensation module; The compensation module comprises: compensating the received signal for the estimated radio frequency transmission impairments; transmitting an indication of the estimated radio frequency transmission impairment as feedback to a transmitter from which the signal was received; and The apparatus is configured to:

Images