Exploiting Spectral Diversity for Machine Learning-Based Estimation of Radio Frequency Signal Parameters
By employing frequency spectrum diversity and machine learning techniques, the method addresses the challenge of multipath interference in RF signal estimation, improving accuracy and reducing complexity in devices with limited size and time constraints.
Patent Information
- Application Number
- JP2022504170
- Authority / Receiving Office
- JP · JP
- Patent Type
- Patents
- Current Assignee / Owner
- Priority Date
- 2019-09-26
- Filing Date
- 2020-07-14
- Publication Date
- 2024-11-13
- Estimated Expiration
- 2040-07-14
AI Technical Summary
Accurate estimation of radio frequency (RF) signal parameters, such as angle-of-arrival (AoA), is challenging due to multipath propagation caused by reflectors in the environment, which leads to signal interference and fading, particularly in devices with limited physical size and observation time, like mobile or wearable devices.
Utilizing frequency spectrum diversity and sequential transmission of RF signals across multiple channels, combined with machine learning techniques, to infer AoA and other parameters by analyzing RF signal power levels in the spatial domain, employing neural networks like CNNs and RNNs to process data sequentially and reduce dimensionality.
Enhances estimation accuracy of RF signal parameters with reduced computational complexity, enabling effective AoA estimation in devices with limited resources by leveraging frequency spectral diversity and machine learning models.
Smart Images

Figure 0007583024000001 
Figure 0007583024000002 
Figure 0007583024000003
Abstract
Description
[Technical field]
[0001] Related Applications This application is an international application of U.S. Nonprovisional Patent Application No. 16 / 584,545, filed September 26, 2019, and claims the benefit of U.S. Provisional Patent Application No. 62 / 878,192, filed July 24, 2019, all of which are incorporated herein by reference. [Background technology]
[0002] Accurate estimation of the angle-of-arrival (AoA) and / or other parameters of a radio frequency (RF) signal may be desirable to perform various tasks, such as source tracking, including determining the location of a transmitter relative to a receiving antenna array. In certain implementations, source tracking can be performed based on multiple measurements of AoA obtained by one or more receiving devices.
[0003] The present disclosure is illustrated by way of example, and not by way of limitation, in the figures of the accompanying drawings. [Brief description of the drawings]
[0004] [Figure 1] 1 is a diagram illustrating a plot showing an example distribution of Received Signal Strength Indicator (RSSI) for spatial characteristics and frequency spectrum diversity of a receive antenna array; [Diagram 2] FIG. 1 is a line diagram depicting an example visualization of transformed input data plotted in a reduced-dimensional space. [Diagram 3] FIG. 1 is a diagram illustrating the structure of an exemplary neural network. [Figure 4] 1 is a diagram illustrating multiple plots showing successive RF signal power level measurements on multiple RF channels. [Diagram 5]FIG. 1 is a simplified diagram of a sequence of models in which each successive model is trained with a training data set generated in a signal propagation environment that includes a greater number of signal paths than the number of signal paths in the signal propagation environment that generated the training data set used to train the previous model in the sequence. [Figure 6] FIG. 6 is a diagram illustrating the error in Angle of Arrival (AoA) estimation according to the model sequence of FIG. 5. [Figure 7] FIG. 1 is a block diagram of a system that implements estimation of AoA and / or other parameters of an RF signal received by an antenna array. [Figure 8] 4 is a flow diagram of an example method for estimating AoA and / or other parameters of RF signals received by an antenna array in multiple RF channels. [Figure 9] 1 is a flow diagram of an example method for iteratively estimating AoA and / or other parameters of an RF signal based on RF signal level measurements received sequentially by an antenna array on multiple RF channels. [Figure 10] 1 is a flow diagram of an example method for iteratively estimating AoA and / or other parameters of RF signals received by an antenna array in multiple RF channels with a set of machine learning models.
[0005] Detailed Description The embodiments described herein relate to systems and methods for estimating the angle of arrival (AoA) and other parameters of a radio frequency (RF) signal received by an antenna array. Such parameters may include the line-of-sight (LoS) angle of arrival, the reflected AoA, the attenuation of the reflection, the relative delay of the reflection, the relative phase, and / or other parameters. In an illustrative example, the estimated AoA and / or other parameters can be used for source tracking, i.e., to identify the location of a transmitter relative to a receiving antenna array. In an illustrative example, the location of the transmitter can be estimated based on three AoA measurements obtained by three receivers located at different spatial points.
[0006] The propagation of an RF signal can be affected by various factors, including the presence of reflectors in the environment surrounding the transmitter and receiver that can create multiple paths that the RF signal can travel. As a result, the receiver sees a superposition of multiple copies of the transmitted RF signal, each traveling a different path. Multipath propagation of an RF signal can cause phase shifts and signal interference, which in itself can cause signal fading, thus adversely affecting the estimation accuracy of the AoA and other RF signal parameters.
[0007] Various common implementations attempt to compensate for multipath signal propagation by increasing the physical size of the receiving antenna and / or increasing the observation time, however, the antenna's physical size and / or observation time may be limited by antenna design requirements, e.g., by the intended use of a mobile or wearable device.
[0008] The systems and methods described herein address the above and various other deficiencies of typical implementations by relying on frequency spectrum diversity, which exists in various wireless communication protocols and implementations (e.g., Bluetooth Low Energy (BLE) channels or Wi-Fi subcarriers) and involves simultaneous and / or sequential transmission of RF signals in multiple RF channels. As illustrated diagrammatically by FIG. 1 and described in more detail herein below, multiple RF channels can be utilized to compensate for a limited number of antenna array elements, since the propagation of the RF signal in each channel responds differently to the presence of a reflector in the signal path. Thus, by analyzing signal parameter values received in the frequency spectrum domain, the antenna response in the spatial domain can be inferred.
[0009] While theoretically inference can be made analytically, i.e., by using curve fitting and surface extrapolation, the complexity of such functions will depend exponentially on the dimensionality of the task, and in particular on the number of reflectors in the signal path. Conversely, the systems and methods described herein utilize machine learning techniques to build and train models to infer the AoA and / or other parameters (e.g., LoS AoA, reflected AoA, reflection attenuation, relative delay of reflections, relative phase, and / or other parameters) of a received RF signal from RF signal power levels (e.g., represented by RF signal magnitude and phase values) observed in multiple RF channels. In particular, the machine learning-based systems and methods of the present disclosure are highly scalable since the same set of models can be utilized for a wide range of input dimensions.
[0010] Various aspects of the above methods and systems are described herein by way of example, rather than limitation. The methods described herein may be implemented by hardware (e.g., general-purpose and / or special-purpose processing devices, and / or other devices and associated circuitry), software (e.g., instructions executable by a processing device), or a combination thereof.
[0011] As described above in this specification, the physical dimensions of the antenna array can be increased and / or the observation time can be increased to improve the estimation accuracy of the AoA and various other parameters of the RF signal received by the antenna array in the presence of RF signal multipath. However, the applicability of such a "brute force" approach can be significantly limited by various design requirements that may limit the physical size and / or maximum observation time of the antenna array (e.g., for mobile or wearable devices). Conversely, the systems and methods of the present disclosure rely on the frequency spectrum diversity inherent in various aspects of wireless communication (e.g., BLE (Bluetooth Low Energy) channels or Wi-Fi subcarriers) and involves the transmission of simultaneous and / or sequential RF signals in multiple RF channels. As shown diagrammatically by FIG. 1, the propagation of the RF signal in each channel responds differently to the presence of a reflector in the signal path, so that a limited number of antenna array elements can be compensated by utilizing multiple RF channels.
[0012] 1 depicts a plot 100 of a distribution of received signal strength indicators (RSSIs) with respect to spatial characteristics of a receiving antenna array (i.e., the number of antenna array elements) and frequency spectrum diversity (i.e., the number of RF channels utilized for signal transmission). In the illustrative example of FIG. 1, RSSIs are utilized, but in various other implementations, other RF signal parameters (phase, group delay, etc.) can be utilized.
[0013] As can be seen in Fig. 1, the RSSI response plot 100 includes multiple repeating patterns, whereby the plotted peaks and valleys have similar sizes and occur periodically. Thus, the RSSI values in the spatial domain can be inferred by analyzing the RSSI values in the frequency spectrum domain. In the illustrative example of Fig. 1, the spatial dimension may be truncated at a particular value (e.g., at four antenna array elements), and the RSSI values for a larger number of antenna array elements can be extrapolated by RSSI values obtained from RF signals received on multiple RF channels.
[0014] As discussed herein above, the systems and methods described herein employ machine learning techniques to build and train models to infer the AoA and / or other parameters of a received signal (e.g., LoS AoA, reflected AoA, attenuation of reflections, relative delay of reflections, relative phase, and / or other parameters) from RF signal power levels observed in multiple RF channels. In some implementations, multiple pairs of RF signal magnitude and phase for each antenna array element can represent the observed RF signal power levels.
[0015] Machine learning based models can perform feature extraction to reduce the dimensionality of the input data and thus significantly reduce the computational complexity of the task to be solved. In FIG. 2, an exemplary visualization of the transformed input data plotted in a reduced dimensional space of in-phase and quadrature (I and Q) values is depicted. In FIG. 2, each circular plot 210A-210K visualizes the AoA of the RF signal received by a respective antenna array element, while each dot in the plots 210A-210K corresponds to an observation in a single frequency channel. Although the example described herein operates on AoA values, the systems and methods of the present disclosure are equally applicable to other parameters of the received RF signal (e.g., LoS AoA, reflected AoA, attenuation of reflections, relative delay of reflections, relative phase and / or other parameters).
[0016] A machine learning based model can utilize the extracted features (i.e., a representation of the input data in a reduced dimensional space) to perform a regression task to estimate values of RF signal parameters. As illustrated diagrammatically by FIG. 2, the regression process can be visualized by rotating the initial circle plots 210A-210K relative to a center of rotation 215 that is external to any of the plots 210A-210K until they merge with a single combined plot 220, which may have a known set of parameters. Thus, one or more parameters (e.g., rotation angle) of a transformation applied to each initial circle plot 210A-210K will represent the difference between the known set of parameters of the combined plot 220 and the parameters of the initial circle plots 210A-210K. Thus, the AoA and / or other parameters of the RF signal depicted by the initial circle plots 210A-210K can be calculated by modifying the known set of RF signal parameters represented by the combined plot 220 based on the parameters of the transformation applied to each initial circle plot 210A-210K.
[0017] In the illustrative example of Figure 2, the transformation applied to each initial circle plot 210A-210K is a rotation transformation, although other types of transformations (e.g., scaling along one or more axes) may be applied by the machine learning model when performing the regression task. Thus, parameters of these transformations may include, for example, rotation angles, aspect ratios, scale factors, etc.
[0018] In some implementations, the model used to perform feature extraction and regression can be implemented by a neural network, such as a convolutional neural network (CNN) or a recurrent neural network (RNN). A neural network is a computational model that implements a large number of connected nodes called "artificial neurons", whereby each artificial neuron processes one or more input signals (including bias signals) and sends an output signal to one or more neighboring artificial neurons. The output of an artificial neuron can be calculated by applying its activation function to a linear combination of its inputs. A neural network can be trained by processing examples ("training data sets") to perform feature extraction, regression and / or classification tasks, generally without being programmed with any task-specific rules, as described in more detail herein below.
[0019] As shown diagrammatically in FIG. 3, the neural network used by the disclosed system and method can be represented by a multilayer perceptron (MLP) 300, whose artificial neurons are grouped into multiple layers including an input layer 310, one or more hidden layers 320A-320L, and an output layer 330. The input layer 310 includes one or more neurons 340A-340N, which are connected to one or more neurons 350A-350K of a first hidden layer 320A. The first hidden layer neurons 350A-350K are themselves connected to one or more neurons 360A-360M of a second hidden layer 320L. The second hidden layer neurons 360A-360M are themselves connected to one or more neurons 370A-370Z of the output layer 330. At least some of the nodes of the artificial neural network 300 may utilize non-linear activation functions, while the remaining nodes (e.g., nodes in the output layer) may utilize linear activation functions. Although FIG. 3 illustrates a single hidden layer 350 diagrammatically, in various implementations of the systems and methods of the present disclosure, the number of hidden layers may vary. In some embodiments, the number of hidden layers is a hyperparameter of the model, i.e., a parameter whose value is specified before the training process. Other hyperparameters of the model may include the number of nodes in each layer, activation function type, etc.
[0020] Each edge in Figure 3 represents a connection for transmitting a signal from one artificial neuron to another, in a manner similar to how synapses work in the human brain. Edge weights that boost or attenuate the signal transmitted through each connection are defined in a network training stage based on a training dataset that includes multiple labeled inputs (i.e., inputs with known classifications).
[0021] The neural network training procedure can begin with initializing all edge weights and neuron bias values to random or predefined values. Forward propagation involves sequential processing of labeled data items of the training data set by the neural network (i.e., sequential processing of multiple vectors of RF signal power levels such that each vector containing magnitude and phase values of RF signals measured by multiple antenna elements at a given frequency channel is labeled with a corresponding AoA value and / or other RF signal parameter value). The observed output of the neural network is compared to the desired output specified by the label associated with the vector being processed, and the error is propagated back to the previous layer of the neural network, where the weights and bias values are adjusted to minimize a predefined loss function (e.g., the difference between the observed output and the desired output specified by the label). This process can be repeated until the output error falls below a predefined threshold.
[0022] As described herein above with reference to FIG. 2, a machine learning based model implementing the systems and methods of the present disclosure can process multiple parameters of RF signals received in multiple frequency channels by first performing feature extraction to reduce the dimensionality of the raw data and then performing regression to estimate values of the RF signal parameters. In some implementations, both feature extraction and regression can be performed by the same neural network, where one or more layers perform feature extraction while the remaining layers perform regression tasks. In other implementations, two neural networks trained separately can be used to perform feature extraction and regression tasks.
[0023] As described herein above, the systems and methods of the present disclosure exploit frequency spectrum diversity to infer the distribution of signal power levels in the spatial domain. In practice, however, the number of available RF channels may be limited by various design and / or implementation specific constraints. As shown diagrammatically by the series of plots 410A-410D in FIG. 4, raw data measurements on a limited number of available frequency channels may not show closed loop curves similar to the circular plots shown in FIG. 2, but only certain portions of each curve.
[0024] Additionally, the RF signal may be transmitted sequentially rather than simultaneously on multiple frequency channels. In the illustrative example of FIG. 4, each plot 410A-410D represents RF signal measurements on a corresponding set of RF channels. Although various common techniques may not be suitable for processing sequential data, the systems and methods of the present disclosure advantageously use a recurrent neural network (RNN) to sequentially process batches of raw data in real time.
[0025] An RNN can maintain its own state, which represents information about inputs previously processed by the network. The RNN will take the state information into account to process each subsequent input. In other words, the network output is defined not only by the current input, but also by the inputs previously processed by the network.
[0026] Some RNNs can be susceptible to the gradient attenuation effect, which makes the network virtually unable to process long input sequences. Gradient decay can be avoided by utilizing multiple long short-term memory (LSTM) layers, which utilize a gating mechanism that allows the network to choose between its own state and the input to execute the next layer of processing. Because LSTM neural networks exhibit very low gradient decay, such networks are able to process longer input sequences.
[0027] Thus, the RNN-based model can sequentially process multiple batches of parameters of a received RF signal in multiple frequency channels, and can improve the accuracy of estimating the AoA and / or other parameters of the received RF signal after processing each batch. In the illustrative example of Figure 4, plot 420 represents the sequential processing of the received RF signal in four batches depicted by plots 410A-410D.
[0028] The neural networks utilized by the disclosed systems and methods can be deployed on hardware (such as mobile or wearable devices) with limited computing power and / or limited available memory, since raw data representing thousands of measurement points are processed sequentially, and each batch of raw data can be discarded after it has been processed by the neural network, since the internal state of the neural network stores necessary information extracted from the input data processed by the network.
[0029] In some implementations, the raw input data can be processed by a sequence of two or more models, all of which infer the AoA and / or other parameters (e.g., LoS AoA, reflected AoA, attenuation of reflections, relative delay of reflections, relative phase, and / or other parameters) of a received RF signal from RF signal power levels (e.g., represented by RF signal magnitude and phase values) observed in multiple RF channels, with each subsequent model trained with a training data set generated in a signal propagation environment that includes a greater number of reflectors (and thus a greater number of signal paths) than the number of reflectors in the signal propagation environment that generated the training data set utilized to train the preceding model in the sequence. In the illustrative example of FIG. 5, model 510A is trained with a first training data set generated in a first signal propagation environment that includes a single signal path (i.e., no multipath), and model 510A infers the AoA from the observed RF signal power levels. Model 510B is trained on a second training data set generated in a second signal propagation environment including at least two signal paths, where model 510B infers the AoA and characteristics of the first reflection (reflection AoA, reflection attenuation, relative delay of reflection, relative phase) from the observed RF signal power levels. Model 510C is trained on a third training data set generated in a third signal propagation environment including at least three signal paths, where model 510C infers the AoA and characteristics of the first and second reflections (reflection AoA, reflection attenuation, relative delay of reflection, relative phase) from the observed RF signal power levels.
[0030] Although a set of three models is shown in FIG. 5, in various implementations of the systems and methods of this disclosure, the number of models used sequentially may vary.
[0031] Estimating the AoA error may involve reconstructing the input (i.e., the magnitude and phase values of the RF signal) based on the estimated AoA output and comparing the reconstructed input with the observed input, such that the AoA error can be estimated even if the true AoA value is unknown.
[0032] The output of every model 510 is analyzed by comparing the estimated error value with a predefined threshold (which may be specified by a hyperparameter of the corresponding model 510). If the estimated error value is above the predefined threshold, the next model 510 may be activated. This iteration may continue until the estimated error value is below the predefined threshold, at which point the method may be terminated and the output generated by the current level of model 510 may be used as the final output.
[0033] 5 can be trained separately using a training data set generated in a signal propagation environment that includes a required number of signal paths. In some implementations, the models can be further trained with another training data set generated in a signal propagation environment that includes fewer than the required number of signal paths, thus ensuring backward compatibility of the models 510A-510K, i.e., ensuring that higher-level models produce satisfactory results for raw data measured in a signal propagation environment that has fewer signal paths than the number of signal paths in the signal propagation environment utilized to train the models.
[0034] FIG. 6 shows a schematic of the error in AoA estimation in degrees by the set of models 510A-510K of FIG. 5. In FIG. 6, plot 610A plots the AoA estimation error (in degrees) for a 0 multipath environment by a model trained on a training data set obtained from the 0 multipath environment, plot 610B plots the AoA estimation error (in degrees) for a 0 multipath environment by a model trained on a training data set obtained from the 1 multipath environment, plot 610C plots the AoA estimation error (in degrees) for a 1 multipath environment by a model trained on a training data set obtained from the 1 multipath environment, and plot 610D plots the AoA estimation error (in degrees) for a 1 multipath environment by a model trained on a training data set obtained from the 0 multipath environment. As can be seen in FIG. 6, the pair of plots 610A-610B show similar AoA estimation errors, thus demonstrating the backward compatibility of higher order models in lower order multipath environments.
[0035] In particular, models 510A-510K can be deployed on hardware (such as mobile or wearable devices) that has limited computing power and / or limited available memory. Since only one model needs to be operational at any given time, the hardware footprint should simply be sufficient to deploy the largest model in the series, since each model can be discarded after use and the next model loaded onto the same hardware.
[0036] FIG. 7 is a block diagram of a system implementing estimation of AoA and / or other parameters of RF signals received by an antenna array. The system 700 can include a processing unit 706, which itself includes a front-end circuit 708 connected to an antenna array 707. The front-end circuit 708 can include a transceiver 712 and an analog-to-digital converter (ADC) 714. The transceiver 712 connected to the antenna array 707 can transmit and receive RF signals via the antenna array 707. The processing unit can implement the AoA estimation tool 120 as described herein. The processing unit can include one or more application processors, one or more host processors, one or more microcontrollers, and / or other processing components. In some embodiments, the system 700 can be implemented as a system-on-chip (SoC) that includes a transceiver and a processor that processes digital values representing RF signals received or transmitted by the transceiver. The transceiver and the processor may be provided on a common carrier substrate or may be implemented in separate integrated circuits.
[0037] In some implementations, the processing device may receive a plurality of pairs of RF signal magnitudes and phases measured by each antenna array element in each of a plurality of RF channels. In this case, the processing device may utilize a machine learning model to calculate the AoA and / or other parameters 721A-721N (e.g., LoS AoA, reflected AoA, attenuation of reflections, relative delay of reflections, relative phase and / or other parameters) of the received RF signals. The machine learning model may perform feature extraction with the aim of reducing the dimensionality of the input data, as described in further detail herein above, and may then utilize the extracted features (i.e., a representation of the input data in a reduced dimensional space) to perform a regression task with the aim of estimating RF signal parameter values.
[0038] FIG. 8 is a flow diagram of an exemplary method for estimating AoA and / or other parameters of an RF signal received by an antenna array in multiple RF channels. Method 800 and / or each of its individual functions, routines, subroutines or operations may be performed by processing logic including hardware (circuitry, dedicated logic, etc.), software (such as those executed on a general-purpose computing system or a dedicated machine), firmware (embedded software), or any combination thereof. Two or more functions, routines, subroutines or operations of method 800 may be performed in parallel or in an order that may differ from the order described below. In a particular implementation, method 800 may be performed by a single processing thread. Method 800 may alternatively be performed by two or more processing threads, each thread performing one or more individual functions, routines, subroutines or operations of the method. In an illustrative example, the processing threads implementing method 800 may be synchronized (e.g., using semaphores, critical sections and / or other thread synchronization mechanisms). Alternatively, the processing threads implementing method 800 may be executed asynchronously with respect to one another. In one embodiment, the operations of method 800 may be performed by the processing unit 706 of the example system 700 of FIG.
[0039] At block 810, a processing unit implementing the above method may receive multiple radio frequency (RF) signal power measurements over multiple RF channels and by multiple antenna elements. In some implementations, as described in further detail herein above, multiple pairs of RF signal magnitude and phase for each antenna array element may represent the observed RF signal power level.
[0040] At block 820, the processing device may calculate estimated RF signal parameter values by applying a machine learning model to the multiple RF signal power measurements. The RF signal parameters may be represented by the AoA and / or other parameters of the received signal (e.g., LoS AoA, reflected AoA, attenuation of the reflections, relative delay of the reflections, relative phase and / or other parameters). The machine learning model may perform feature extraction to reduce the dimensionality of the input data. The machine learning may then utilize the extracted features (i.e., a representation of the input data in a reduced dimensional space) to perform a regression task with the goal of estimating the parameter values of the RF signal, as described in further detail herein above.
[0041] At block 830, the processing device may output the RF signal parameter values (e.g., by displaying the RF signal parameter values and / or transmitting the RF signal parameter values to one or more recipients over one or more communications networks) and the method may end.
[0042] 9 is a flow diagram of an exemplary method for iteratively estimating AoA and / or other parameters of an RF signal based on RF signal level measurements received sequentially by an antenna array on multiple RF channels. Method 900 and / or each of its individual functions, routines, subroutines or operations may be performed by processing logic including hardware (circuitry, dedicated logic, etc.), software (such as those executed on a general-purpose computing system or dedicated machine), firmware (embedded software), or any combination thereof. Two or more functions, routines, subroutines or operations of method 900 may be performed in parallel or in an order that may differ from the order described below. In certain implementations, method 900 may be performed by a single processing thread. Method 900 may alternatively be performed by two or more processing threads, each thread performing one or more individual functions, routines, subroutines or operations of the method. In an illustrative example, the processing threads implementing method 900 may be synchronized (e.g., using semaphores, critical sections and / or other thread synchronization mechanisms). Alternatively, the processing threads implementing the method 900 may execute asynchronously with respect to one another. In one embodiment, the operations of the method 900 may be performed by the processing unit 706 of the example system 700 of FIG.
[0043] At block 910, a batch of measurements of radio frequency (RF) signal power levels may be received by a processing unit implementing the above method over multiple RF channels and multiple antenna elements. In some implementations, as described in further detail herein above, multiple pairs of RF signal magnitude and phase for each antenna array element may represent the observed RF signal power level.
[0044] At block 920, the processing device may calculate estimated RF signal parameter values by applying a machine learning model to the batch of received RF signal power measurements. The RF signal parameters may be represented by the AoA and / or other parameters of the received signal (e.g., LoS AoA, reflected AoA, attenuation of reflections, relative delay of reflections, relative phase and / or other parameters). The machine learning model may perform feature extraction to reduce the dimensionality of the input data. The machine learning may then utilize the extracted features (i.e., a representation of the input data in a reduced dimensional space) to perform a regression task with the goal of estimating parameter values of the RF signal, as described in further detail herein above.
[0045] In response to determining at block 930 that the parameter estimation error is above the predetermined threshold error value, the method may loop back to block 910 where the next batch of measurements may be received for a different set of RF channels. Otherwise, i.e., if the parameter estimation error is less than or equal to the predetermined threshold error value, the method may continue to block 940.
[0046] At block 940, the processing device may output the RF signal parameter values (e.g., by displaying the RF signal parameter values and / or transmitting the RF signal parameter values to one or more recipients over one or more communications networks) and the method may end.
[0047] FIG. 10 is a flow diagram of an exemplary method for iteratively estimating AoA and / or other parameters of an RF signal received by an antenna array in multiple RF channels by a set of machine learning models. Method 1000 and / or each of its individual functions, routines, subroutines or operations may be performed by processing logic including hardware (circuitry, dedicated logic, etc.), software (such as those executed on a general-purpose computing system or dedicated machine), firmware (embedded software), or any combination thereof. Two or more functions, routines, subroutines or operations of method 1000 may be performed in parallel or in an order that may differ from the order described below. In a particular implementation, method 1000 may be performed by a single processing thread. Method 1000 may alternatively be performed by two or more processing threads, each thread performing one or more individual functions, routines, subroutines or operations of the method. In an illustrative example, the processing threads implementing method 1000 may be synchronized (e.g., using semaphores, critical sections and / or other thread synchronization mechanisms). Alternatively, the processing threads implementing the method 1000 may execute asynchronously with respect to one another. In one embodiment, the operations of the method 1000 may be performed by the processing unit 706 of the example system 700 of FIG.
[0048] At block 1010, a processing unit implementing the method may receive multiple RF signal power measurements on multiple RF channels and by multiple antenna elements. In some implementations, as described in further detail herein above, multiple pairs of RF signal magnitude and phase for each antenna array element may represent the observed RF signal power level.
[0049] At block 1020, the processing unit can select a next machine learning model from a sequence of two or more models, where each subsequent model is trained with a training data set generated in a signal propagation environment that includes a greater number of reflectors (and therefore a greater number of signal paths) than the number of reflectors in the signal propagation environment that generated the training data set used to train the previous model in the sequence. Each machine learning model can perform feature extraction to reduce the dimensionality of the input data. The machine learning can then utilize the extracted features (i.e., a representation of the input data in a reduced dimensional space) to perform a regression task to estimate parameter values of the RF signal, as described in more detail herein above.
[0050] At block 1030, the selected machine learning model may be applied by a processing unit to the plurality of RF signal power measurements to calculate an estimated RF signal parameter value. As described in further detail herein above, the RF signal parameter may be represented by the AoA and / or other parameters of the received signal (e.g., LoS AoA, reflected AoA, attenuation of reflections, relative delay of reflections, relative phase, and / or other parameters).
[0051] In response to determining at block 1040 that the parameter estimation error is above the predetermined threshold error value, the method may loop back to block 1010 where the next model from the sequence of models is selected. Otherwise, i.e., if the parameter estimation error is less than or equal to the predetermined threshold error value, the method may continue to block 1050.
[0052] At block 1050, the processing device may output the RF signal parameter value (e.g., by displaying the RF signal parameter value and / or transmitting the RF signal parameter value to one or more recipients over one or more communications networks) and the method may end.
[0053] The embodiments described herein may be implemented by an electronic system including a processing device that may perform the same or similar functions as described with respect to the figures above, or vice versa. In another embodiment, the processing device may be a microprocessor or a microcontroller. The AoA estimation tool 120 may be implemented as a microcontroller or firmware running on a microprocessor. The microcontroller may report the estimations described herein to an application processor. The electronic system may include a host processor for a computer system that utilizes a microcontroller to provide the estimations described herein. The host processor may include one or more processing devices, memory, and other hardware or software components that perform operations for the electronic system.
[0054] The processing unit may include analog and / or digital general purpose input / output ("GPIO") ports. The GPIO ports may be programmable. The GPIO ports may be connected to a programmable interconnect and logic ("PIL"), which serves as an interconnect between the GPIO ports and the digital block array of the processing unit. The processing unit may include analog blocks that are programmable and reprogrammable in some cases to implement various analog functions. The processing unit may also include a digital block array. In one embodiment, the digital block array can be configured to implement various digital logic circuits (e.g., DACs, digital filters, or digital control systems) using configurable user modules ("UM"). The digital block array may be connected to a system bus. The processing unit may also include memory devices such as random access memory (RAM) and program flash. The RAM may be static RAM (SRAM) and the program flash may be a non-volatile storage device that can be used to store firmware (e.g., control algorithms executable by the processor to implement the operations described herein). The processing device may also include a memory controller unit ("MCU") connected to the memory and the processor. The processor may be a processing element (e.g., a processor core) configured to execute instructions or perform operations. The processor may include other processing elements, as will be understood by those skilled in the art having the benefit of this disclosure. It should also be noted that the memory may be internal or external to the processing device. If the memory is internal, the memory can be connected to a processing element, such as a processing core.If the memory is external to the processing unit, then the processing unit is connected to another device in which the memory is located, as will be understood by those skilled in the art having the benefit of this disclosure.
[0055] In one embodiment, the processing device further includes processing logic. Some or all of the operations of the processing logic can be implemented in firmware, hardware, or software, or some combination thereof. The processing logic can receive signals from the antenna array as described herein. The processing device may also include an analog block array (e.g., a field programmable analog array). The analog block array can also be connected to the system bus. In one embodiment, the analog block array can be configured to implement various analog circuits (e.g., ADCs or analog filters) using configurable UMs. The analog block array can also be connected to a GPIO port. The processing device may include an internal oscillator / clock and communication block ("COM"). In another embodiment, the processing device includes a spread spectrum clock. The oscillator / clock block provides clock signals to one or more components of the processing device. The communication block can be used to communicate with external components, such as a host processor (also referred to as an application processor), via application interface ("I / F") lines.
[0056] The processing device may be provided on a common carrier substrate, such as, for example, an integrated circuit ("IC"), a die substrate, or a multi-chip module substrate. Alternatively, the components of the processing device may be one or more separate integrated circuits and / or discrete components. In one exemplary embodiment, the processing device is a Programmable System on a Chip (PSoC®) processing device developed by Cypress Semiconductor Corporation of San Jose, Calif. The processing device may alternatively be one or more other processing devices known to those skilled in the art, such as a microprocessor or a central processing unit, a controller, a special purpose processor, a digital signal processor (DSP), an application specific integrated circuit (ASIC), a field programmable gate array (FPGA), or the like.
[0057] It should also be noted that the embodiments described herein are not limited to configurations having a processing device connected to an application processor, but may include systems that measure RF signals and transmit raw data to a host computer where the raw data is analyzed by an application, and in fact the processing performed by the processing device may occur in the application processor.
[0058] The AoA estimation tool 120 may be incorporated into the processor IC or alternatively into another IC. Alternatively, a description of the AoA estimation tool 120 may be generated and compiled for incorporation into another integrated circuit. For example, a behavioral level code describing parameters for the AoA estimation tool 120, or a portion thereof, may be generated using a hardware description language such as VHDL or Verilog and stored on a machine-accessible medium (e.g., CD-ROM, hard disk, floppy disk, etc.). Additionally, the behavioral level code may be compiled into register transfer level ("RTL") code, a netlist, or even a circuit layout and stored on a machine-accessible medium. The behavioral level code, RTL code, netlist, and circuit layout may represent various levels of abstraction describing the AoA estimation tool 120. It should be noted that the components of an electronic system may include all, some, or none of the components described above.
[0059] In one embodiment, the electronic system may be used in a tablet computer. The electronic device may alternatively be used in other applications, such as notebook computers, mobile phones, personal data assistants (PDAs), keyboards, televisions, remote controls, monitors, handheld multimedia devices, handheld media (audio and / or video) players, handheld gaming devices, signature entry devices for point of sales (POS) transactions, eBook readers, medical devices, etc., warehouse tracking devices, such as scanners used by shipping companies, automotive devices, such as car keys and electronic devices in vehicles, global position systems (GPS), or control panels, etc. Certain embodiments may be implemented as a computer program product, which may include instructions stored on a machine-readable medium. These instructions may be used to program a general-purpose or special-purpose processor to perform the described operations. A machine-readable medium includes any mechanism for storing or transmitting information in a form (e.g., software, processing applications) readable by a machine (e.g., a computer). Machine-readable media may include, but are not limited to, magnetic storage media (e.g., floppy disks), optical storage media (e.g., CD-ROMs), magneto-optical storage media, read-only memory (ROM), random-access memory (RAM), erasable programmable memory (e.g., EPROM and EEPROM), flash memory, or another type of medium suitable for storing electronic instructions.
[0060] Additionally, some embodiments may be practiced in a distributed computing environment in which the machine-readable medium is stored and / or executed by two or more computer systems. Additionally, information transferred between computer systems may be pulled or pushed across the communications medium connecting the computer systems.
[0061] Although the operations of the methods herein are shown and described in a particular order, the order of the operations of each method can be changed, such that certain operations can be performed in a reverse order, or certain operations can be performed, at least in part, in parallel with other operations. In alternative embodiments, different operation instructions or sub-operations can be intermittent and / or alternating. Terms such as "first," "second," "third," "fourth," etc., as used herein are intended as labels to distinguish between different elements and are not necessarily meant to indicate an order according to their numbering. The term "connected to" as used herein can mean directly connected or indirectly connected through one or more intermediate components. Any signals provided through the various buses described herein may be time multiplexed with other signals and provided through one or more common on-die buses. Additionally, the interconnects and interfaces between circuit components or blocks can be viewed as buses or as single signal lines. Each bus may alternatively be one or more single signal lines, and each single signal line may alternatively be a bus.
[0062] In the above description, numerous specific details are set forth, such as examples of specific systems, components, methods, and the like, to provide an understanding of multiple embodiments of the present invention. However, it will be apparent to one of ordinary skill in the art that at least some embodiments can be practiced without these specific details. In other cases, well-known components or methods have not been described in detail or have been shown in simplified block diagram form to avoid unnecessarily obscuring the embodiments of the present invention. Thus, the specific details described are merely exemplary. A particular implementation may differ from these exemplary details and still be considered to be within the scope of the embodiments of the present invention.
[0063] Embodiments of the claimed subject matter include, but are not limited to, various operations described herein that may be performed by hardware components, software, firmware, or any combination thereof.
[0064] The above description provides numerous specific details, such as examples of specific systems, components, methods, and the like, to provide an understanding of embodiments of the claimed subject matter. However, it will be apparent to those skilled in the art that at least some embodiments of the above method can be practiced without these specific details. In other cases, well-known components or methods have not been described in detail or have been shown in simple block diagram form. Thus, the specific details described are merely exemplary. A particular implementation may differ from these example details and still be considered to be within the scope of the claimed subject matter.
Claims
1. receiving, by a processing device, a plurality of radio frequency (RF) signal power measurements by a plurality of antenna elements over a plurality of RF channels; calculating, by the processing device, estimated RF signal parameter values applying a machine learning model to the plurality of RF signal power measurements; outputting, by said processing unit, said RF signal parameter values; responsive to determining that the estimated error value exceeds a predetermined threshold, recalculating the estimated RF signal parameter value by applying a second machine learning model to the plurality of RF signal power measurements; having the second machine learning model corresponds to a greater number of RF signal multipaths than the first machine learning model; method.
2. the plurality of RF signal power measurements include a plurality of RF signal magnitude values and a plurality of RF signal phase values. The method of claim 1.
3. the RF signal parameter value is one of a line-of-sight (LoS), angle-of-arrival (AoA), a reflection AoA, an attenuation of a reflection, a relative delay or a relative phase of a reflection; The method of claim 1.
4. the machine learning model performs feature extraction to convert the plurality of RF signal power measurements into an intermediate representation in a reduced dimensional space. The method of claim 1.
5. the machine learning model performs a regression task to infer estimated values of the RF signal parameters from the intermediate representation in the reduced dimensional space. The method of claim 4.
6. The method further comprises training the machine learning model by processing a training data set including a plurality of RF signal power measurements from a plurality of antenna elements in a plurality of RF channels, the RF signal power measurements being labeled with corresponding RF signal parameter values. The method of claim 1.
7. An apparatus, comprising: a transceiver configured to be coupled to an antenna having a plurality of antenna elements to receive a plurality of radio frequency (RF) signal power measurements on a plurality of frequency channels; a processor connected to the transceiver; having The processor, configured to select a machine learning model from a sequence of machine learning models, a first machine learning model of the sequence being trained with a first training data set generated in a first signal propagation environment including a first number of signal paths, and a second machine learning model of the sequence being trained with a second training data set generated in a second signal propagation environment including a second number of signal paths, the second number of signal paths being greater than the first number of signal paths; The processor, calculating estimated RF signal parameter values by applying the selected machine learning model to the plurality of RF signal power measurements; and in response to determining that the parameter estimation error is less than or equal to a predetermined error threshold, outputting the RF signal parameter value. [0023] system.
8. The processor, and in response to determining that a parameter estimation error exceeds a predetermined error threshold, the method is further configured to iteratively repeat the selecting and calculating operations until the estimated RF signal parameter value falls below the predetermined error threshold. The system of claim 7.
9. the plurality of RF signal power measurements include a plurality of RF signal magnitude values and a plurality of RF signal phase values. The system of claim 7.
10. the RF signal parameter value is one of a line-of-sight (LoS), angle-of-arrival (AoA), a reflection AoA, an attenuation of a reflection, a relative delay or a relative phase of a reflection; The system of claim 7.
11. each machine learning model of the sequence of machine learning models performs feature extraction to convert the plurality of RF signal power measurements into an intermediate representation in a reduced dimensional space; The system of claim 7.
12. the machine learning model performs a regression task to infer estimated values of the RF signal parameters from the intermediate representation in the reduced dimensional space. The system of claim 11.
Citation Information
Patent Citations
Errancy direction-of-arrival estimation method based on sparse Bayesian learning
CN109490819A
Object identifying device
JP1995160665A
Diagnosis processing device, system, method and program, and recording medium readable by computer and classification processing device
JP2012235796A
Direction estimation device
JP2017106834A
Information processing device, information processing method and program
JP2017215272A
