Radio frequency digital signal classification

Abstract

The invention relates to the field of Radio Frequency (RF) signal classification and recognition, specifically a method for training of Convolutional Neural Networks (CNN) and applying the trained CNN for the classification or identification of digitally modulated RF signals. The method (100) comprises providing a Deep Learning CNN, providing input digital signals (102), generating an input data set from those signals (103), processing the input data set to provide a training set (104), training the CNN (105). The training (105) comprises the steps of providing time-based signals comprising at least a first symbol timing and generating IQ representations of those signals. The method is suitable for classifying digital signals, and is particularly suited to digitally modulated communications signals.

Description

[0001] RADIO FREQUENCY DIGITAL SIGNAL CLASSIFICATION

[0002] Technical Field of the Invention

[0003] The invention relates to the field of Radio Frequency (RF) signal classification and recognition, specifically a method for training of Convolutional Neural Networks (CNN) and applying the trained CNN for the classification or identification of digitally modulated RF signals.

[0004] Background to the Invention

[0005] The ability to detect, analyse, classify and respond to signals within the modern Contested ElectroMagnetic Environment (CEME) is a key requirement for many communications applications.

[0006] Future, highly adaptive, Software Defined Radio (SDR) will have significant processing power at their disposal which enables the use of more sophisticated processing techniques to help detect, discrimination and classify digital signals.

[0007] Over recent years technological advancement has enabled Artificial Intelligence (Al) to be exploited by a huge number of application areas. One of those application areas is RF signal classification, where Deep Learning (DL) CNNs have successfully been applied to many binary and multi-class classification examples.

[0008] This combination of developments in SDR and data processing techniques, including the use of DL CNNs, provide an opportunity to greatly improve the performance of classification systems working in near real-time.

[0009] Modern RF communication digital signals are often complex in nature. Amplitude, frequency and phase components may be modulated separately or, in combination, to provide a diverse range of signals which are difficult to distinguish from each other without sophisticated signal processing or prior knowledge. An approach to classifying such digital signals is to use DL CNNs. These models are typically trained, validated and / or tested on data sets which may be simulated, real-world or a combination of. The training data may take different forms, for example being simply constituent parts of the digital signals or data or may be represented in another way such as an image or plot, where, for example, complex RF digital signals (IQ data sets) are represented in diagrammatic forms for example as IQ diagrams or plots. In these plots T is the in- phase signal component (cosine) and ‘Q’ is the Quadrature signal component (sine). When combined, I and Q signal components form a vector whose magnitude is given by the Root-Sum-Square (RSS) and whose argument (or phase) is given by the arc tangent. Such images may be used to train a model which subsequently may be used to classify the simulated or real world digital signals.

[0010] Existing approaches often require that input signals are transformed into the frequency domain using methods such as Fast Fourier Transform (FFT). To produce effective models multiple FFTs are undertaken, for example to produce a plurality of spectrogram images which can then be used for training or within a DL CNN based classification. These existing approaches for classifying or recognising signals are often too computationally intensive to deliver prediction outputs in in near real-time and I or do not capture the full range of signal perturbations resulting in poor classification performance.

[0011] Therefore the current invention is directed toward a method for producing a DL CNN, which overcomes these issues.

[0012] Summary of the Invention

[0013] According to a first aspect, the invention provides a computer implemented method for Radio Frequency digitally modulated signal classification, comprising the steps of, providing a Deep Learning Convolution Neural Network, training a Convolutional Neural Network comprising the steps of, providing a plurality of time based digitally modulated signals comprising an at least first symbol timing, generating an input data set comprising an IQ representation of the plurality of digitally modulated signals, processing the input data set to provide a training data set, using the training data set to train the Convolutional Neural Network, applying the trained Convolutional Neural Network to digitally modulated signals to perform Radio Frequency digitally modulated signal classification.

[0014] Radio Frequency (RF) portion of the electromagnetic spectrum covers the frequency range 3 kHz to 3 THz. Within this range, the Very High Frequency (VHF) and Ultra High Frequency (UHF) bands encompass many of the wireless communications systems which are in use today. Digitally modulated signals are signals which are modulated, modulation refers to the process by which information is mapped or imparted onto a RF carrier signal by varying the carrier’s amplitude / frequency / phase. This process is well known and examples of communications systems which exploit this include 5G, Wi-Fi, Bluetooth, DAB, GNSS). The method of the invention is a computer implemented method for digital signal recognition or classification. Classification, in this context, means the ability to classify or distinguish between certain types of digital signal timing or modulations. The digital signals may be complex in makeup, potentially having being corrupted or distorted by noise or interference and other degradations (e.g. fading, shadowing) whilst traversing a channel, such as air, or a cable. As such it can be difficult to distinguish one digital signal from another as the boundaries blur and I or key characteristics may overlap, therefore detection or classification models need to be trained on data covering a variety of signal conditions. The method of the invention has been shown to be effective at successfully classifying such digital signals.

[0015] The first step of the method involves providing a Deep Leaning (DL) Convolutional Neural Network(CNN). DL CNNs are well understood and include models such as AlexNet, GoogLeNet, ResNet, and may further include Sequential Networks, such a pre-trained model or alternatively a model created from scratch is provided for use within the method. The next step of the method is to train the DL CNN, again the requirement to train these models on sample data is also well understood. The models are trained or exposed to data or images, such that they “learn” to recognise or classify the data of interest. Typically the performance of these models is assessed by training on a subset of data, often termed training data and then validating and I or testing them against a different set of validation data. Performance is generally determined by the ability of the model to successfully classify signals of interest, but other metrics of performance may also be based on the level of training required for a given level of performance. The next step of training the DL CNN in the method for the invention comprises the step of providing a plurality of time based digital signals comprising an at least first symbol timing. The plurality of signals are time based. This means they have not been transformed for example into the frequency domain (e.g. via the FFT) and retain timing information, this has been shown to advantageously allows for the CNN when in use, to operate more quickly as the processing requirements are reduced as no conversion or transformation is required. Furthermore the use of time based signals allows for better discrimination of signals as time or timing information is maintained, whereas any frequency transformation process will degrade or obscure signal timing information. The signals may have been pre-processed, for example to be time and or frequency synchronised. The digital signals comprise an at least first symbol timing, these terms are well understood and may include symbol timings or modulations such as Quadrature phase shift keying (QPSK), Quadrature Amplitude Modulation (QAM) and Chirp Spread Spectrum (CSS). The plurality of digitally modulated signals may be modulated using the same techniques but having different timings or may equally include more than one such modulation and / or symbol timing. The next step in the method comprises generating an input data set, the input data set is the IQ data or signal components or representation of each of the plurality of signals with time. The process for generating the IQ data from such signals is well understood and may be done as a processing step for example within the method of the invention or may be done by hardware or software on signal capture prior to being provided to the model.

[0016] The next step in the method is to process the input data set to provide the training data set as a plurality of training data or images. This step comprises modifying the input data set or signals, such as to be suitable for training the CNN. This may include creating large data sets from the input set, subdividing out specific data sets or combining data. In preferred embodiments of the method the processing includes the addition of “noise” and / or ‘IQ impairments’ to the input data. “Noise” is a well understood concept in that it introduces random amplitude and / or phase variations to a signal. Graded noise may be applied to all or part of the input data, for example it may be applied to just the I or Q component or both to achieve different levels of variability within the input data. ‘IQ impairments are caused by mismatches between the analogue components of in-phase and quadrature paths in the RF chain. Impairments (e.g. gain imbalance, frequency offset, quadrature skew, timing skew) may also be applied to ‘perfect’ signals to deliver graded real-world device effects. Even more preferably the IQ data are processed such as to form images to provide the training data set. The processing of the data has advantageously been shown to improve the training of the CNN resulting in a more robust model with improved classification performance and I or reduced training burden.

[0017] The next step is to train the CNN using the training data set. This is well understood and comprises providing the training data set to the CNN as an input and validating the trained model. The input training data set may be split into a training data set and a validation data set. The DL CNN is then trained against the input training data set, this is understood and depends on the CNN in use, but may include typical processing steps such as using the convolutional operations to recognise patterns in the input data set, the application of activation functions (e.g. Sigmoid, ReLu), to support the network to “learn” non-linear relationships between the features in the input data set, pooling of layers to capture the most significant features from the convolution process and reduce the spatial dimensions of the applied layer. It may also include the modification of other aspects of the model including hyper-parameter values. This is understood and, for example, may include the application of different activation functions and / or optimisers tuning the learning rate and / or batch size and / or number of epochs to improve accuracy. The training step generates probability output values to label or classify the input data set.

[0018] The model is then validated against validation input data set, typically these data have not been used in the training phase, so have not been “seen” by the of the DL CNN, but have been similarly processed as the training data set. Validation may include the evaluation of the DL CNN model performance. Should validation of the model not provide the desired performance level, then further training may be undertaken with additional training data until an acceptable performance level is achieved.

[0019] Finally the resultant trained DL CNN is applied to digitally modulated signals to perform RF digital signal classification. Applied means the trained model is used to classify input signals of an unknown digital modulation. The use of a DL CNN for signal classification is well understood, here digitally modulated signals are provided, for example by capture of real-world signals using suitable hardware such as an antenna, receiver and signal processor, or “playback” of pre-captured or recorded signals. The signal processing steps are the same as that for the previous method steps during the training of the model and may include time and or frequency synchronisation, and further comprise turning the provided digitally modulated signals into their constituent IQ components with time. In some embodiments of the method the input signals are further processed into images or diagrams as described for the training aspects of the model, for example being an intensity and or vector diagram or a k-means cluster diagram with or without vector information. The processing steps applied in the application or use of the model match those that were used during the training element of the method. The DL CNN then provides the classification of the input signals, typically by identifying the specific features of the input signal, for example the modulation technique and timing. This advantageously allows for relatively quick processing of time based input signals or samples as the processing steps are reduced when compared to other models, for example those operating on signals which have to be converted into the frequency domain, as fewer signals or data points are required and the processing is less computationally burdensome as no transformation into the frequency domain is required. The trained model has been shown to have good classification performance, and due to the nature of the relatively low burden signal or data processing, is suitable for use in a variety of applications.

[0020] In certain embodiments of the method, the step of providing the plurality of time based signals comprises recording real-world digitally modulated signals and I or the generation of digital signals synthetically. Real-world digital signals may be captured and recorded using RF test instrumentation or devices including systems which can emulate said signals. The plurality of time based signals may be pre-processed, such as to time I frequency synchronise the signals, these are well understood processing steps and may include techniques such as the use of phase-lock loops. Providing real world signals advantageously includes system and I or environmental parameters (e.g. additive noise) within those signals, this has been shown to improve training robustness. Alternatively, signals may be synthetically generated using modelling and simulation techniques, this advantageously, allows large data sets with controlled levels of variation, for example distortion, to be produced corresponding to the full range of signals (including edge cases) which may be encountered under real world conditions. Real-world or synthetic signals may be used independently in combination when forming input data sets. This, advantageously, allows time based signals to be highly representative of any signal condition which may be encountered in the real world, whilst massively increasing the amount of input data available to train the model. This has been shown to, advantageously, further improve the robustness of training of the CNN, resulting in improved performance of signal classification.

[0021] In some embodiments of the method, the step of processing the input data further comprises the generation of intensity images from the input data set. The processing step comprises, dividing the input data set comprising IQ data into subsets or frames, and plotting the I data against Q data. Here “plotted” takes its normal meaning in that the IQ data represent two coordinates in time which may be drawn or plotted onto a plane, resulting in an image, the image may be coloured or greyscale. This process may be repeated for every Ith and Qth sample from the input data. The data may then be plotted as an intensity plot or image, these are also well understood ways of representing data, which distribute the input data into two dimensional bins or groups to create a two-dimensional histogram, the plot area may then be divided into a number of cells of equal size (like a chess board). The count of the number of data points that fall within each region is then assigned a colour or intensity. These have advantageously been shown to be effective as training images for the CNN, particularly for digital modulation schemes and for reducing the effects of additive noise.

[0022] In certain embodiments the input data may be plotted as a vector diagram. These diagrams are well known, due to the fact that the input data are time-based they plot or trace a path between subsequent IQ data pairs with time, such vector plots subsequently show a characteristic pattern or trace of the digital signal modulation. This advantageously provides a training data set which preserve the timing or modulation information, further improving the performance of the model when used to classify digitally modulated signals.

[0023] Additionally the image types may be combined such as to produce an intensity image with vector information, which may have different colours or intensities. A DL CNN trained using both image types in combination has been shown to provide improved classification performance over vector or intensity images when used alone. In certain embodiments of the method the processing step further comprises k-means clustering of the input data set. K-means clustering is a known method of grouping data. In this context it groups data into K clusters where K is an integer number. In this use the number of clusters is dependent on the modulation scheme or timing. For example, there would be eight for Differential Quadrature Phase Shift Keying (DQPSK) based upon the similarity of features or characteristics of the data. The optimised position of each set of data points is known as the centroid. K-means clustering only plots a constellation like trace around these centroids, however this has been shown to improve classification performance within the trained model as the approach reduces the randomness in the IQ representation.

[0024] In some embodiments of the invention comprising K-means clustering the processing further comprises the addition of vector information. Typically during processing steps such as K-means clustering the point-to-point transitions (vector or timing information is lost), however the inventor has shown that by optionally preserving the index of each realigned data point, the mapping or transition between data points can be reinserted and vector information reinstated along with the centroids. Thus a K-means cluster image with vector information may be produced and used as an input training set for the DL CNN. This allows simpler and cleaner input training images to be applied to DL CNNs which, in turn, allows simpler networks to be used and / or improves signal classification performance. This has advantageously been shown, that for certain modulation schemes which produce identical constellation diagrams (e.g. Quadrature Phase Shift Keying (QPSK) and Offset QPSK (OQPSK), 8 level Phase Shift Keying (8PSK) and pi / 4 DQPSK), these can be successfully classified using the trained DL CNN of the invention. An additional advantage is that the addition of vector information to the image, either enhances DL CNN classification or enables DL CNN performance to be undertaken where it otherwise could not.

[0025] In some embodiments of the method the input data set comprises 100-1000 IQ data points per image. The inventor has shown that the method of the invention may allow for a relatively small number of data points (e.g. 100-1000) to be used in preparing the training data as images, whilst still providing good classification performance. The data points are IQ representation of the time based signals. This, advantageously, speeds up the training of the DL CNN and reduces the associated data storage and processing burden associated with other methods.

[0026] In some embodiments of the method wherein one or more of plurality of time based digitally modulated signals comprising an at least first symbol timing comprising, quadrature phase shift keying, offset quadrature phase shift keying, 8 phase shift keying or pi / 4 differential quadrature phase shift keying. The symbol timings or modulation techniques of the input signals may include those listed above. Either individually or in combination or alongside any other modulation technique. Certain modulation techniques, for example QPSK and OQPSK are indistinguishable when plotted on the IQ plane for example as a scatter or constellation plot. Advantageously the method of the invention has been shown to classify or distinguish modulation techniques or symbol timings such as these.

[0027] In certain embodiments of the method where the input data are processed as images the plurality of training images are greyscale images. Greyscale takes its normal meaning in that the image data is a single channel or layer having intensity only. The inventor has shown that using these images reduces the data storage requirement and training and processing burden however, for these simple image types, surprisingly delivers the same classification or recognition performance as a similar DL CNN model trained using coloured images.

[0028] In some embodiments of the method the Deep Learning Convolutional Neural Network is a pretrained deep neural networks. Pretrained deep learning neural networks are well understood. They have been trained to extract powerful and informative features from natural images for a variety of uses. The inventor has shown that using a pretrained network with transfer learning is, typically, much faster and simpler than training a completely new network. It has also been shown that the method of the present invention can advantageously be used within simple, and computationally light, networks for example SqueezeNet or GoogLeNet, whilst still showing good classification performance. In certain embodiments of the Deep Learning Convolutional Neural Network is a Sequential Network having no more than five layers. Sequential Networks are a known type of DL CNN, having sequential or linearly stacked layers to process the input data. The inventor has shown that the approach of processing the input data of the invention allows for a relatively simple model to be used. These simple CNNs have very low computational burden allowing it to be implemented or run on systems with relatively low processing power, or equally allow it to be run concurrently or multiple times within more complex or sophisticated computer systems.

[0029] In some embodiments of the method where the training data are images, the plurality of images are fewer than 10,000. The method of the invention has been shown to provide good classification with fewer than 10,000 training images. The inventor has shown that in creating image-based training data sets of certain embodiments of the invention, the model can be trained more effectively then when compared with other training data types. As such this advantageously provides a model with reduced processing burden when compared to other approaches.

[0030] In certain embodiments the method is implemented on a computer Central Processing Unit (CPU). Implementation of DL CNNs on CPUs is known. However the method of the present invention provides sufficiently small or light data sets such that the trained model is able to be run on or operate on a CPU. This advantageously allows for the model to be easily deployed with relatively low-cost hardware. Alternatively, the model may be run on a Graphics Processor Unit (GPU). Advantageously, due to method being computationally light, the GPU allows concurrent operations to be performed and therefore may be suited to processing of large data sets concurrently.

[0031] In certain embodiments of the method, the DL CNN is implemented on a Software Defined Radio (SDR). As the method of the invention has low computational burden it can be run or used directly on relatively simple hardware such a low cost SDR. This advantageously has been shown to further increase the speed of processing and classifying digitally modulated signals, as there is no need to pass data or information to an external computer or processor. In some embodiments of the method the signals are classified in near real-time. In this context near real-time means that the captured or input signals may be classified within milliseconds of capture. For example, the inventor has shown that when signals are streamed from a Software Defined Radio the trained model of the first aspect accurately classifies signals using a host personal computer in less than 250ms. This advantageously means the system can be used in many more applications requiring rapid processing.

[0032] Any feature in one aspect of the invention may be applied to any other aspects of the invention, in any appropriate combination. In particular method aspects may be applied to device or use aspects and vice versa. The invention is further described with reference to the accompanying drawings.

[0033] In all aspects, the invention may comprise, consist essentially of, or consist of any feature or combination of features.

[0034] Brief Description of the

[0035] The invention will now be described, purely by way of example, with reference to the accompanying drawings, in which;

[0036] Figure 1 is a block diagram of the method steps of the invention; and

[0037] Figure 2a is a block diagram of the method step of processing the input data as a intensity and vector diagram; and

[0038] Figure 2b is a block diagram of the method step of processing the input data as a k- means cluster and vector diagram;

[0039] Figures 3a and 3b are example training images;

[0040] Figure 4 is a figure indicating the performance of the CNN in classifying signals.

[0041] Figure 5 illustrates the use or application of trained CNN.

[0042] The drawings are for illustrative purposes only and are not to scale.

[0043] Figure 1 is a block diagram of the first aspect of the invention (100). The preparation of training data and training of the process (101 ) of the invention is shown. The input digital signals (102) are a plurality of input digitally modulated signals having a modulation timing of pi / 4 which utilise the pi / 4 DQPSK modulation technique. The signals are time based digital signals captured from a real-world environment using an Ettus USRP E310 Software Defined Radio (SDR). The plurality of input digital signals (102) are processed into the IQ components to use as input data set (103). The input data set are then further processed (104) using suitable software to add signal impairments and graded noise. The processed signals, the training set, are then used to train the DL CNN (105). The training data are split into a subset of training data and validation data, and are then passed through the DL CNN, SqueezeNet which is a network pre-trained for general image classification. The training process includes the steps of retraining the pretrained model using the processed signals (104) using transfer learning. The model performance is then validated using a subset of the training data. Following validation the model is then implemented on a GPU (not show) such as to classify digitally modulated signals (106) in the real world.

[0044] Figure 2a shows a more detailed block diagram of the processing steps (200) of the first aspect of the invention, showing the signal processing (201 ) which is the preparation of the IQ data and addition of signal impairments and noise. The data are then further processed (202) to produce intensity data sets using a mathematical software tool. A two dimensional histogram is created from IQ data points, comprising the binning of data. The data are then interpolated and low pass filtered to generate intensity vector (203) and intensity data (202) used in the greyscale training images (204). 1000 such images are produced to form the training set. 800 of these images are then used to train the DL CNN (205). The pretrained CNN utilises convolutional layers to recognise patterns in the input data sets, the ReLu activation functions is applied after each convolution operation, pooling layers capture the most significant features from the convolution process and reduce the dimensions of the applied layer. Fully connected layer inputs correspond to the flattened one-dimensional matrix generated by the last pooling layer, a softmax prediction layer then generates probability output values for each label or signal classification (the highest probability value is the predicted label). The DL CNN is then validated using 800 training images, providing a validation accuracy of 80 % to 90 % . Figure 2b shows a more detailed block diagram of the processing steps (250) of the first aspect of the invention, showing the signal processing (251 ) which is the preparation of the IQ data and addition of signal impairments and graded noise. The data are then further processed to produce k-means cluster data sets using a mathematical software tool. The optimised position of each set of data points is determined (centroids) and the data around these centroids are plotted. The input digital signals are pi / 4 DQPSK modulated, as such there are 8 clusters I centroids. During this processing the index in time of the IQ data pairs used to create the clusters are preserved, The index is used to generate the vector data. The k-means data and vector data are then plotted together to produce greyscale training images (254). 1000 such images are produced to form the image set. 800 of these images are then used to train the DL CNN (255). The pretrained CNN utilises convolutional layers to recognise patterns in the input data sets, the Sigmoid activation function is applied after each convolution operation, pooling layers capture the most significant features from the convolution process and reduce the dimensions of the applied layer. Fully connected layer inputs correspond to the flattened one-dimensional matrix generated by the last pooling layer, a softmax prediction layer then generates probability output values for each label, the highest probability value being the predicted label. The DL CNN is then validated using 100 validations images. The performance is then assessed using a confusion matrix (not shown) which summarises the performance of the model.

[0045] Figure 3a is an example of a training image (300), the intensity diagram (301 ) shows a pi / 4 DQPSK digital modulation IQ plot wherein each of the plurality of input signals (not show) have been processed and represented in the plot (301 ). The images (300) are then used to train the model.

[0046] Figure 3b shows a training image (350) used within the first aspect of the invention. The image shows the input signals having been processed using k-means clustering, having 8 centroids (351 ). Furthermore the index in time for the underlying IQ data have been preserved allowing the vector information (352) to also be represented in the image. Figure 4 shows a figure (400) of the performance of the model for training output for the pi / 4 DQPSK modulation technique using four different, but closely aligned, modulation timings (symbol rates). Network performance is optimised by varying the training parameters including the solver, learning rate, validation frequency (revalidation interval), batch size and number of epochs. The validation performance is shown (401 ) and is similar for 80 (402) and 800 (403) training images.

[0047] Figure 5 illustrates the application or use of the trained DL CNN (500). A digitally modulated signal (501 ) is received by an antenna (502), connected to a Software Defined Radio (SDR) (503). The signals are pre-processed to frequency and time synchronise the signals. The trained DL CNN, trained on time based IQ data represented as k-means clustered images, including vector information, is run on the SDR, the digitally modulated signals (501 ) are similarly processed into IQ data represented as k-means clustered images with vector information, the signals are then classified, with the classification label displayed on the screen (504) for the user (not shown).

[0048] It will be understood that the present invention has been described above purely by way of example, and modification of detail can be made within the scope of the invention.

[0049] Moreover, the invention has been described with specific reference to characterising digital signals. It will be understood that this is not intended to be limiting and the invention may be used more generally. Additional applications of the invention will occur to the skilled person.

Claims

CLAIMS1 . A computer implemented method for Radio Frequency digitally modulated signal classification, comprising the steps of; a. providing a Deep Learning Convolution Neural Network; b. training the Deep Learning Convolutional Neural Network comprising the steps of; i. providing a plurality of time based digitally modulated signals comprising an at least first symbol timing; ii. generating an input data set comprising an IQ representation of the plurality of time based digitally modulated signals ; iii. processing the input data set to provide a training data set, wherein the processing comprises;I. generating k-means clustering for the input data set;II. preserving the index of each IQ representation used in the k- means clustering to generate vector information;III. generating a combined k-means cluster and vector image; iv. using the training data set to train the Convolutional Neural Network; c. Applying the trained Convolutional Neural Network to digitally modulated signals to perform Radio Frequency digitally modulated signal classification.

2. The method of claim 1 wherein step 1 .b.i of providing the plurality of time based digitally modulated signals comprises recording real-world digital signals3. The method of claim 1 or 2 wherein step 1 .b.i of providing the plurality of time based digitally modulated signals further comprises the generation of digital signals synthetically.

4. The method of any preceding claim wherein the input data set of step l .a.ii comprises 100-1000 IQ data points per image.

5. The method of any preceding claim wherein one or more of the plurality of time based digitally modulated signals comprising an at least first symbol timingcomprising, quadrature phase shift keying, offset quadrature phase shift keying, 8 phase shift keying or pi / 4 differential quadrature phase shift keying.

6. Method of any preceding claim wherein the images are greyscale images.

7. The method of any preceding claim wherein the Deep Learning Convolutional Neural Network is a pre-trained neural network.

8. The method of any proceeding claim wherein the Deep Learning Convolutional Neural Network is a Sequential Network having no more than five layers.

9. The method of any preceding claim implemented on a computer Central Processing Unit.

10. The method of proceeding claims 1 -8 implemented on a Software Defined Radio.11 . The method of claim 10 wherein the digitally modulated signals are classified in near real-time, the classification being performed in less than 250ms.

Images