Time interleaved analog-to-digital converter with ADC cell correction
By introducing neural networks and ADC marking generators into the time interleaved analog-to-digital converter, combining PVT parameters to identify and correct the error of each ADC unit, the time domain error problem caused by gain and offset error of individual ADC units is solved. Improves signal-to-noise distortion ratio and signal accuracy.
Patent Information
- Application Number
- CN202410595418.5
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Priority Date
- 2023-05-23
- Filing Date
- 2024-05-14
- Publication Date
- 2024-11-26
AI Technical Summary
Gain and offset errors between individual ADC units in time interleaved analog-to-digital converters (ADCs) lead to time-domain errors in the output signal, which are difficult to effectively correct for the prior art, especially at high sampling frequency, where error noise has Fixed mode, affecting signal-to-noise distortion ratio (SNDR) and bandwidth.
Introducing machine learning systems, especially neural networks, combine PVT parameters in the training stage to identify and correct specific errors of each ADC unit. ADC mark generator is used to provide unit identifiers. The machine learning system applies a customized error correction scheme to compensate for gain. and offset error to improve the accuracy of digital signals.
It effectively reduces noise and error in the time-interleaved ADC output signal, improves signal-to-noise distortion ratio (SNDR), eliminates the need for additional mathematical modeling, reduces costs, and maintains efficient error correction capabilities at high sampling frequency.
Smart Images

Figure CN119030537A_ABST
Abstract
Description
Technical Field
[0001] Errors in analog-to-digital converters (ADCs) that rely on time-interleaving outputs from multiple independent ADCs are compensated by using a machine learning system that may include a neural network. The machine learning system may be trained based on simulated or measured data in combination with data from an external timing source. The effects of process-voltage-temperature (PVT) parameters on analog-to-digital conversion errors may be incorporated into the training of the machine learning system. Background Art
[0002] In electronics, an analog-to-digital converter ("ADC") is a system that converts analog signals, such as sound picked up by a microphone or light entering a digital camera, into digital signals. An ADC can also provide isolated measurements when implemented as an electronic device that converts an input analog voltage or current into a digital number that represents the magnitude of the input voltage or current.
[0003] More specifically, an ADC converts a continuous-time and continuous-amplitude analog signal into a discrete-time and discrete-amplitude digital signal. This conversion involves quantization of the input, so the process necessarily introduces a small amount of error or noise. In addition, instead of performing the conversion continuously, the ADC performs the conversion periodically, thereby sampling the input signal at a certain sampling rate, thus limiting the allowable bandwidth of the input signal.
[0004] The performance of an ADC is characterized primarily by its bandwidth and signal-to-noise-and-distortion ratio ("SNDR"). The bandwidth of an ADC is characterized primarily by its sampling rate. The SNDR of an ADC is affected by many factors, including resolution, linearity and accuracy (how well the quantization levels match the true analog signal), aliasing, and jitter.
[0005] The effective sampling rate of a time-interleaved (TI) ADC device can be increased by operating several individual ADC units in parallel. By time-interleaving the output of each individual ADC unit, the input signal can be processed at a higher sampling rate than the sampling rate of the individual ADC units.
[0006] While this approach can be used to increase the effective sampling rate of a TI ADC, since the individual ADC cells may each exhibit small gain and offset differences, the final processed output signal may include time domain error patterns that require correction. Summary of the invention
[0007] In some aspects, the technology described herein relates to a system comprising: a time-interleaved analog-to-digital converter (ADC) configured to operate at a sampling rate fs, the time-interleaved ADC comprising: an input terminal configured to receive an analog input signal, an output terminal configured to output a digital signal, wherein the digital signal is a digital approximation of the analog input signal, and a plurality of ADC units, wherein each ADC unit of the plurality of ADC units is configured to: measure an analog value of the analog input signal, convert the analog value into a digital value, and send the digital value to the output terminal of the time-interleaved ADC; a controller configured to: receive a clock signal having a frequency determined by fs, and sequentially enable the plurality of ADC units at a rate determined by the clock signal an ADC unit in the time-interleaved ADC; an ADC tag generator configured to: receive the clock signal and use the clock signal to generate an ADC unit identifier output signal, the ADC unit identifier output signal identifying the ADC unit currently enabled among the multiple ADC units; and a machine learning system configured to: receive the digital signal from the output end of the time-interleaved ADC, receive the ADC unit identifier signal from the ADC tag generator, and modify the digital signal to generate a corrected digital signal by applying an error correction scheme associated with the ADC unit identified by the ADC unit identifier signal to compensate for analog-to-digital conversion errors occurring in the ADC unit, wherein the machine learning system uses the ADC unit identifier signal to determine the error correction scheme.
[0008] In some aspects, the system further includes a sensor configured to provide a parameter to the machine learning system, wherein the output of the ADC unit varies with the value of the parameter.
[0009] In some aspects, the machine learning system has been configured through a training phase to compensate for variations in the output of the ADC unit using the parameters.
[0010] In some aspects, the parameter includes at least one of semiconductor manufacturing related process variations, voltage, and temperature.
[0011] In some aspects, the analog-to-digital conversion errors include gain and DC offset errors.
[0012] In some aspects, the machine learning system includes a neural network having a plurality of neurons arranged into a network having fewer than three layers and fewer than ten neurons in each layer.
[0013] In some aspects, the ADC flag generator includes at least one of a modulo-N counter and an N-bit shift register, where N is equal to the number of ADC units in the plurality of ADC units.
[0014] In some aspects, the technology described herein relates to a system comprising: a time-interleaved analog-to-digital converter (ADC), which includes a plurality of ADC cells connected in parallel between an input and an output of the time-interleaved ADC; an ADC tag generator, which is configured to output a first identifier of a first ADC cell among the plurality of ADCs when the first ADC cell is enabled; and an error correction system, which is configured to: receive a first digital signal from the output of the time-interleaved ADC, receive the first identifier of the first ADC cell from the ADC tag generator, modify the first digital signal by applying a first error correction scheme associated with the first ADC cell to compensate for a first analog-to-digital conversion error occurring within the first ADC cell to generate a first corrected digital signal, and output the first corrected digital signal.
[0015] In some respects,
[0016] The ADC flag generator is configured to output a second identifier of a second ADC unit among the plurality of ADCs when the second ADC unit is enabled; and
[0017] The error correction system is configured to:
[0018] receiving a second digital signal from the output of the time-interleaved ADC,
[0019] receiving the second identifier of the second ADC unit from the ADC signature generator,
[0020] modifying the second digital signal to produce a second corrected digital signal by applying a second error correction scheme associated with the second ADC unit to compensate for a second analog-to-digital conversion error occurring within the second ADC unit, and
[0021] The second corrected digital signal is output.
[0022] In some aspects, the error correction system includes a machine learning system that has been configured through a training phase to compensate for the analog-to-digital conversion error.
[0023] In some aspects, the machine learning system includes a neural network having a plurality of neurons arranged into a network having fewer than three layers and fewer than ten neurons in each layer.
[0024] In some aspects, the system further includes a sensor configured to provide a parameter to the error correction system, wherein the outputs of the plurality of ADC units vary as a function of a value of the parameter.
[0025] In some aspects, the parameter includes at least one of semiconductor manufacturing related process variations, voltage, and temperature.
[0026] In some aspects, the first analog-to-digital conversion error includes gain and DC offset errors.
[0027] In some aspects, the time-interleaved ADC includes a controller configured to:
[0028] The receiving device has s determines the frequency of the clock signal, and
[0029] ADC units in the plurality of ADC units are enabled sequentially at a rate determined by the clock signal.
[0030] In some aspects, the ADC signature generator is configured to:
[0031] receiving the clock signal; and
[0032] An ADC unit identifier output signal is generated using the clock signal to sequentially identify the ADC units among the plurality of ADC units when the ADC units are enabled, and the ADC unit identifier output signal is sent to the error correction system.
[0033] In some aspects, the technology described herein relates to a method comprising: receiving a first digital output signal from a time-interleaved analog-to-digital converter (ADC), wherein the time-interleaved ADC comprises a plurality of ADC units; receiving a first identifier from an ADC tag generator, wherein the first identifier identifies a first ADC unit that is currently enabled among the plurality of ADC units; and modifying the first digital output signal by applying a first error correction scheme associated with the first ADC to compensate for a first analog-to-digital conversion error occurring within the first ADC to produce a first corrected digital output signal.
[0034] In some aspects, the method further comprises:
[0035] receiving a second digital output signal from the time-interleaved ADC;
[0036] receiving a second identifier from the ADC tag generator, wherein the second identifier identifies a second ADC unit among the plurality of ADC units that generates the second digital output signal; and
[0037] The second digital output signal is modified by applying a second error correction scheme associated with the second ADC to compensate for a second analog-to-digital conversion error occurring within the second ADC to produce a second corrected digital output signal.
[0038] In some aspects, the method further includes receiving parameters including at least one of semiconductor manufacturing related process variations, voltage, and temperature, and wherein the first error correction scheme is configured to compensate for variations in the output of the first ADC unit using the parameters.
[0039] In some aspects, the time-interleaved ADC has a sampling rate f s , and the method further comprises:
[0040] The receiving device has s a clock signal of a certain frequency; and
[0041] An ADC unit identifier output signal is generated using the clock signal, the ADC unit identifier output signal identifying a currently enabled ADC unit among the plurality of ADC units. BRIEF DESCRIPTION OF THE DRAWINGS
[0042] A more complete understanding of the subject matter may be derived by referring to the detailed description and claims when considered in conjunction with the following figures, wherein like reference numerals refer to like elements throughout the various figures.
[0043] Figure 1 A TIADC device is depicted that includes a certain number (N) of individual ADC units numbered ADC 1, ADC 2, ..., ADC N.
[0044] Figure 2 Several graphs are shown depicting the errors that may be introduced by ADC conversions performed by different ADC units within a TI ADC device.
[0045] Figure 3A Depicted is a system configured to implement training of a neural network based on an input target signal to correct ADC unit specific errors within a TI ADC device.
[0046] Figure 3B Depiction Figure 3A A system configured in a production or active configuration for incorporation into an end-use application.
[0047] Figure 4 A system including a neural network configured to generate an error correction signal configured to correct ADC unit specific errors within a TI ADC device is described.
[0048] Figure 5A and5B is a schematic diagram depicting an example implementation of an ADC signature generator.
[0049] Fig. 6A is a graph depicting the frequency spectrum of the output signal of a conventional TI ADC device.
[0050] Figure 6B is a graph depicting the frequency spectrum of an output signal of a system configured in accordance with the present disclosure.
[0051] Figure 7 is a flow chart depicting a method for correcting the output of a TI ADC device by applying an ADC unit specific correction. DETAILED DESCRIPTION
[0052] The following detailed description is merely illustrative in nature and is not intended to limit the embodiments of the subject matter of the application and the use of such embodiments. As used herein, the words "exemplary" or "example" mean "serving as an example, instance, or illustration". Any embodiment or embodiment described herein as exemplary or example is not necessarily to be construed as being preferred or advantageous over other embodiments. In addition, it is not intended to be bound by any explicit or implicit theory presented in the aforementioned technical field, background technology, or the following detailed description.
[0053] The effective sampling rate of an analog-to-digital converter (ADC) system can be increased by time-interleaving the outputs of several individual ADC units. In this configuration, the outputs of the individual ADC units can be sampled at a sampling frequency higher than the sampling rate of the individual ADC units to produce an output signal having an effective sampling rate at the higher sampling frequency.
[0054] Although time interleaving is a useful method for increasing the effective sampling rate of an ADC device, variations in the gain and offset errors of each individual ADC unit within the TI ADC device may generate noise in the output signal of the TI ADC device. The noise will generally have a fixed pattern in the time domain because the errors are introduced by the various individual ADC units while their output signals are being actively sampled.
[0055] During the design phase of a TI ADC device, the design of individual ADC cells may be optimized (e.g., via improved integrated circuit (IC) design and other design factors) in an attempt to minimize the effects of such gain and offset errors. However, such efforts may significantly increase the cost of the TI ADC device. Additionally, even with this additional cost, process variations between individual ADC cells may still introduce ADC cell specific gain and offset errors.
[0056] In an embodiment of the system of the present invention, a machine learning engine (MLE), such as a neural network, is incorporated into the TI ADC device to analyze and adjust the output of the TI ADC to compensate for variations in gain and offset errors between individual ADC units. Such a machine learning system can be used to provide such error correction without mathematically modeling these error signals. Alternatively, the machine learning system can be trained to receive a distorted TI ADC device output signal and apply a learned mathematical transformation to correct the distortion to produce a more accurate output signal. As described herein, in addition to being used to correct gain and offset errors, such a neural network can also be further utilized to compensate and correct other types of errors that may be generated by the TI ADC device or its individual ADC units.
[0057] When a machine learning system is implemented to provide this improved functionality without additional input, the machine learning system does not know which individual ADC unit is producing the current output of the TI ADC device. Therefore, the machine learning system is generally unable to characterize the gain and offset errors of a specific ADC unit within the TI ADC device. In order to enable the machine learning system to develop an error correction scheme customized for each individual ADC unit, the present disclosure encompasses an additional input to the machine learning system that identifies, for each output value generated by the TI ADC device, the specific ADC unit responsible for producing that output value. By inputting this ADC tag or identification value into the machine learning system, the machine learning system can learn and apply the appropriate correction scheme based on the individual ADC unit.
[0058] Accordingly, the present disclosure provides an improved TI ADC device implementation in which a machine learning system, knowing the identity of a specific ADC unit responsible for producing the current output of the TI ADC device, applies an ADC unit-specific correction scheme to account for the gain and offset errors (and, possibly other errors) of that specific ADC unit.
[0059] When the machine learning system is implemented as a neural network, the neural network can use various machine learning techniques (e.g., artificial neural networks, deep neural networks, convolutional neural networks, recurrent neural networks, autoencoders, reinforcement learning, etc.), fuzzy logic, artificial intelligence ("AI"), deep learning algorithms, deep structured learning hierarchical learning algorithms, support vector machines ("SVM") (e.g., linear SVM, nonlinear SVM, SVM regression, etc.), decision tree learning (e.g., classification and regression trees ("CART"), ensemble methods (e.g., ensemble learning, random forest, sampling with replacement and without replacement (Bagging and Pasting), patch and subspace, boosting, stacking, etc.), dimensionality reduction (e.g., projection, manifold learning (Manifold Learning), principal component analysis, etc.) and / or deep machine learning algorithm implementation. Non-limiting examples of publicly available machine learning algorithms, software, and libraries that can be utilized within embodiments of the present disclosure include Python, OpenCV, Inception, Theano, Torch, PyTorch, Pylearn2, Numpy, Blocks, TensorFlow, MXNet, Caffe, Lasagne, Keras, Chainer, Matlab Deep Learning, CNTK, MatConvNet (MATLAB toolbox for implementing convolutional neural networks for computer vision applications), DeepLearnToolbox (Matlab toolbox for deep learning (from Rasmus Berg Palm)), BigDL, Cuda-Convnet (fast C++ / CUDA implementation of convolutional (or more generally feed-forward) neural networks), Deep Belief Networks (Deep Belief Networks), RNNLM, RNNLIB-RNNLIB, matrbm, deeplearning4j, Eblearn.lsh, deepmat, MShadow, Matplotlib, SciPy, CXXNET, Nengo-Nengo, Eblearn, cudamat, Gnumpy, 3-way factored RBM and mcRBM, mPoT, ConvNet, Elektronn, OpenNN, NeuralDesigner, Theano generalized Hebbian learning, Apache Singa, Lightnet, and SimpleDNN.
[0060] Various embodiments disclosed herein are described with respect to the use of artificial neural networks (also referred to herein simply as "neural networks"), however, one of ordinary skill in the art will be able to implement any suitable machine learning system or other logic network for compensating for individual ADC unit errors in the time-interleaved TI ADC device of the embodiments of the present disclosure.
[0061] Neural networks are typically based on a collection of connected units or nodes called artificial neurons, which loosely model neurons in biological brains. Each connection, like a synapse in a biological brain, can send signals to other neurons. The artificial neuron that receives the signal can then process the signal and can signal the neurons connected to it. In a neural network implementation, the "signal" at the connection is a real number, and the output of the neuron is calculated by a nonlinear function of the sum of the inputs of each neuron (e.g., sigmoid activation). The connection is called an edge. Neurons and edges typically have weights that are adjusted as the training of the neural network proceeds. The weight increases or decreases the strength of the signal at the connection. Neurons may have a threshold so that the signal is sent only when the collective signal crosses the threshold. Typically, neurons are aggregated into layers. Different layers can perform different transformations on their inputs. The signal travels from the first layer (input layer) to the last layer (output layer), and may need to traverse the layers multiple times. The neural network architecture can be configured as a feedforward network with one or more hidden layers and a back-propagation learning algorithm. The implementation of a neural network may involve three stages: a training or learning stage; a validation stage; and an inference or output stage. In the training phase, the neural network generally learns by comparing its actual output to known correct outputs (or at least outputs that are closer to the intended output) to find errors. The neural network then modifies the model accordingly. In the validation phase, the trained neural network is validated with the aid of data (a "validation set") that may be different from the data used in the training phase. In the inference or production phase, the trained and validated neural network is now configured and able to provide outputs corresponding to any input.
[0062] The result of the foregoing is a TI ADC device that utilizes an error correction neural network that has been trained to "know" what the analog-to-digital conversion errors of each individual ADC unit of the TI ADC device are, and also "know" how to compensate for those errors by applying ADC unit specific correction schemes.
[0063] As mentioned herein, an analog-to-digital conversion error may be any type of distortion (e.g., gain and offset errors) caused by the circuitry within an individual ADC unit that causes the digital output signal of the TI ADC device to not be a true or "ideal" digital conversion of the input analog signal of the TI ADC device. Such analog-to-digital conversion errors may be caused by nonlinearities within the individual ADC units or any other defects or processing parameters (which may be in the form of noise, distortion, harmonics, etc.) that may produce ADC unit errors. Each individual ADC unit of the TI ADC device may be any type of ADC implemented within any type of device or circuitry that utilizes an ADC. The neural network described herein may be implemented in any type of hardware and / or software (as further disclosed herein), such as within a suitable set of digital circuitry suitable for implementing a neural network or other type of machine learning system. The illustrations of the neural network within the accompanying drawings are not intended to limit the scope of the embodiments of the present disclosure. Within the various drawings, "(A)" indicates that the signal is an analog signal, and "(D)" indicates that the signal is a digital signal.
[0064] Figure 1 A TI ADC device 100 is depicted that includes a certain number (N) of individual ADC cells 102 (numbered ADC 1, ADC 2, ..., ADC N). The TI ADC device 100 includes an input 104 configured to receive an input analog signal (A). In various embodiments, the TI ADC device 100 includes four to six individual ADC cells, but the present disclosure may be used to correct the output of a TI ADC device 100 having two or more individual ADC cells. The TI ADC device 100 is configured to generate an analog signal at a specific frequency (f s ) samples the input analog signal and converts the sampled analog value into a corresponding digital output value (e.g., a digital bit stream) that is output as a digital output signal (D) at output terminal 106.
[0065] In the TI ADC device 100, each ADC unit 102 has an input terminal 108 configured to receive an input analog signal A. Each ADC unit 102 is configured to sample the analog input signal A and convert the sampled analog value into a digital value output as a digital signal D at a terminal 110. The ADC unit 102 is configured to generate a digital signal at a frequency less than f. s The frequency f u The input analog signal is sampled. In the embodiment of the TI ADC device 100, f u Equal to f s / N.
[0066] Each ADC unit 102 has an input switch 112 and an output switch 114, so that different individual ADC units 102 can be connected to the input terminal 104 and the output terminal 106, so that the output of the TI ADC device 100 at the terminal 106 is produced by different ADC units 102 at different times. In essence, each input switch 112 and output switch 114 provide a mechanism by which a specific ADC unit 102 among the number of ADC units 102 can be "enabled" so that only the output of that ADC unit 102 is output at the output terminal 106. It will be apparent to those of ordinary skill in the art that other methods can be implemented to allow a specific ADC unit 102 to be enabled (i.e., to become the ADC unit 102 that produces the output value of the TI ADC device 100) while disabling other ADC units 102 in the TI ADC device 100. In various embodiments of the TI ADC device 100, the disabled ADC unit 102 is still operational (i.e., produces an output signal), but the output of the disabled ADC unit 102 is not sent as the output of the TI ADC device 100.
[0067] The outputs of the individual ADC units 102 are time-interleaved to achieve a higher sampling frequency f of the TI ADC device 100 s . When the corresponding sampling period T s =1 / f s The frequency f s When implementing time interleaving, the TI ADC device 100 is initially configured (e.g., by a controller configured to operate the input switch 112 and the output switch 114) so that the first ADC unit 102 (ADC1) is enabled and receives an input analog signal at time t0 and converts the analog signal into an n-bit digital value output at the output terminal 106. At this time, the controller disables the other ADC units 102. s Seconds later, the controller adjusts the input switch 112 and the output switch 114 to enable the second ADC unit 102 (ADC2) with the connection to the input terminal 104 and the output terminal 106, and the second ADC unit 102 converts the sampled signal (i.e., the signal at time T s) is converted into a second n-bit digital representation output at the output terminal 106. At this point, the other ADC units 102 are disabled. After Ts seconds, the controller adjusts the input switch 112 and the output switch 114 to enable the third ADC unit 102 (ADC3) to sample the input analog signal (i.e., at time Ts*2) and generate a corresponding digital output at the output terminal 106. At this point, the other ADC units 102 are disabled. This process continues until the last ADC unit 102 (ADCN) has sampled the input analog signal and generated a corresponding digital output at the output terminal 106. The next sampling cycle begins with the first ADC unit 102 (ADC1) sampling the input signal, and the process continues.
[0068] When the digital outputs at the output 106 are available sequentially in the same order as the sampling operations, the digital values are reassembled into a digital signal having a sampling rate f s Therefore, although the individual ADC unit 102 only uses f s / N, but the overall operation of the TIADC device 100 is equivalent to f s Simply put, the input analog signal is divided into different ADC time slices, which are processed separately by individual ADC units 102 and then recombined at the output to form a high data rate digital representation of the input analog signal.
[0069] Because, in any real-world TI ADC device 100 , each individual ADC cell 102 will exhibit slightly different gain and offset errors, those errors will be passed on to the output of the TI ADC device 100 .
[0070] To illustrate, Figure 2 Several graphs are shown depicting errors that may be introduced by ADC conversions performed by different ADC units 102 within a TI ADC device 100, where the ADC units 102 have different gain and offset errors. Graph 202 depicts ADC conversions performed by a first ADC unit 102 (e.g., ADC1) in a TI ADC device 100, while graph 204 depicts ADC conversions performed by a second ADC unit 102 (e.g., ADC2) in a TI ADC device 100. Graph 206 depicts outputs generated by the TI ADC device 100 (e.g., the outputs of combined ADC1 and ADC2). In each of graphs 202, 204, and 206, the horizontal axis represents time, while the vertical axis represents signal magnitudes, either digital or analog.
[0071] An analog signal 208 is depicted in each of the graphs 202, 204, and 206, where the analog signal 208 represents an analog input (A) to the TI ADC device 100. The various vertical bars depicted in each of the graphs 202, 204, and 206 represent digital values produced by the first ADC unit 102, the second ADC unit 102, and the TI ADC device 100, respectively. In an ideal ADC, the vertical bars (or digital values) would have magnitudes that exactly match the analog signal.
[0072] As shown by the graph 202, the first ADC unit 102 is characterized by gain and offset errors that cause the first ADC unit 102 to tend to produce an output digital value 210 that represents an analog value that is slightly lower than the actual value of the analog signal 208. Therefore, the first ADC unit 102 is generally characterized by a negative offset error, but the device may additionally exhibit other non-linear errors, as described herein. Similarly, the second ADC unit 102 is characterized by gain and offset errors that cause the second ADC unit 102 to produce an output digital value 212 that represents an analog value that is slightly higher than the actual value of the analog signal 208. Therefore, the second ADC unit 102 is generally characterized by a positive offset error, but the device may additionally exhibit other non-linear errors, as described herein.
[0073] As depicted by the digital value 214 output by the TI ADC device 100 representing a time-interleaved combination of the digital values 210 and 212 generated by the first and second ADC units 102 , both positive and negative offset errors generated by the first and second ADC units 102 are included.
[0074] In addition to the offset errors produced by the individual ADC units, the time-interleaved digital output of the TI ADC device 100 may additionally include additional spurious content, referred to as "interleaving spurs." Interleaving spurious artifacts are typically time-domain errors that may be introduced by analog impairments in the communication channel of the TI ADC device 100, which are modulated with the sliced conversion signals produced by each ADC unit 102 due to the interleaving process and ultimately appear in the final digitized output.
[0075] In addition to these types of errors and noise, the performance and accuracy of the individual ADC units 102 may be affected by process-voltage-temperature (PVT) factors that produce error offsets between the ADC units 102. Among these types of variations, process variations are responsible for deviations in the semiconductor component manufacturing process of the components of the ADC units 102. Process variations may be attributed to changes in manufacturing conditions such as temperature, pressure, and dopant concentrations. The various voltages utilized within the various ICs that make up the TI ADC device 100 (e.g., supply voltages, reference voltages, bias conditions on the device) may vary from established design values during daily operation and also vary over their lifetime, which may affect the operation of the circuit components of the ADC units 102 in the TI ADC device 100. Similarly, temperature fluctuations may affect the operating parameters of the various circuit components while the device is operating.
[0076] In order to compensate for these and other errors that may not be consistent across all ADC units 102 in a TI ADC device 100, the present disclosure provides an ADC implementation in which a trained neural network "knows" what the analog-to-digital conversion errors of each individual ADC unit 102 are, and also "knows" when to compensate for those errors. As mentioned herein, an analog-to-digital conversion error may be any type of distortion caused by the circuitry within the ADC unit 102 that causes the digital output signal of the TI ADC device 100 to not be a true or "ideal" digital conversion of the input analog signal. Such analog-to-digital conversion errors may be caused by nonlinearities within the ADC circuitry or any other imperfections or processing parameters (which may be in the form of noise, distortion, harmonics, etc.) that may produce such ADC errors. The TI ADC device 100 may be any type of time-interleaved ADC implemented within any type of device or circuitry that utilizes an ADC. The neural network described herein may be implemented with any type of hardware and / or software (as will be further disclosed herein), for example, within an appropriate set of digital circuitry suitable for implementing the neural network. The illustration of the neural network within the accompanying drawings is not limited to the scope of the embodiments of the present disclosure. In the various drawings, "(A)" indicates that the signal is an analog signal, and "(D)" indicates that the signal is a digital signal.
[0077] Figure 3A is a block diagram depicting a TI ADC device 100 (eg, Figure 1 The components of the system 300 include a TI ADC device 100 and a neural network 302 configured to provide active correction of the output signal of the TI ADC device 100 to compensate for the errors generated by the individual ADC units 102 therein. The TI ADC device 100 is connected to a controller 311. The controller 311 is configured to send control signals to the TI ADC device 100 to adjust the output signal of the TI ADC device 100 at f sThe individual ADC units 102 of the TI ADC device 100 are sequentially enabled at a frequency of f, thereby enabling the TI ADC device 100 to generate a sampling rate f at the terminal 304. s The time-interleaved output of the N individual ADC units 102 is only f s In an embodiment, the neural network 302 may have a relatively small number of neurons and interconnections, thereby enabling the implementation of the neural network 302 using low-cost hardware solutions. In one example, the neural network 302 may have less than 3 layers with less than 10 neurons in each layer (e.g., a 2×5 neural network with 2 layers and 5 neurons in each layer). Figure 3A System 300 is shown in a configuration that implements training of a neural network 302 based on an input training signal.
[0078] System 300 is configured to receive an input analog signal at input terminal 306, take samples of the input analog signal, and generate an output digital signal at terminal 304. As described below, neural network 302 is configured to apply an appropriate correction scheme to generate an output signal that can be used to modify the output digital signal at terminal 304 to generate a corrected digital signal output at node 308. In some embodiments, this may involve neural network 302 applying the correction scheme and outputting the corrected signal itself, such as Figure 3A However, in other embodiments (and as Figure 4 ), the neural network 302 may be configured to generate an error signal using a correction scheme, which may be applied to the output of the TI ADC device 100 (eg, via subtraction) to correct the output of the TI ADC device 100.
[0079] The neural network 302 is trained using a training signal representing a known analog input signal and a corresponding "ideal" digital signal representing an accurate conversion of the known analog input signal into a digital signal stream. Specifically, the known analog training signal is supplied at input 306. At the same time, the training digital signal is supplied to digital training signal input 309. During training of the neural network 302, the TI ADC device 100 processes the known analog input signal to produce a corresponding digital output signal. The output of the TI ADC device 100 is fed through the neural network 302, and then the output of the neural network 302 (i.e., the "corrected" TI ADC device 100 output signal) is compared to the digital training signal received at input 309. If the TI ADC device 100 is operating without errors, and / or the neural network 302 is accurately trained to correct all errors, the two signals will be equal and the error will be zero. However, because the TI ADC device 100 introduces some errors into the output signal of the TI ADC device 100 that are not accurately corrected by the neural network 302 when the neural network 302 is trained, this error (e.g., determined by determining the difference between the output of the neural network 302 and the digital training signal) is fed back to the neural network 302, thereby enabling the neural network 302 to develop a more accurate error correction scheme.
[0080] In various embodiments, the training digital signal may be filtered by a digital filter 310 that may be configured to filter out harmonics and excess noise in the training digital stream. The output of the filter 310 thus represents an "ideal" digital signal that would typically be produced by a reference ADC that produces less conversion error than the TI ADC device 100. In some embodiments, while training the neural network 302, a delay element may be positioned between the output 304 of the TI ADC device 100 and the neural network 302 to compensate for any delays (e.g., propagation delays) inherent in the operation of the filter 310. While in other embodiments, if the filter 310 has a larger word width, an optional quantizer 312 may be introduced after the filter 310 to reduce the number of bits in the output signal of the filter 310. The quantizer 312 may be any device configured to perform the desired quantization (e.g., truncation, rounding, scaling, establishing the highest or lowest number of bits, etc.). According to alternative embodiments of the present disclosure, one or more quantizers may be used throughout the system to limit the word width in the digital data stream.
[0081] According to certain embodiments of the present disclosure, neural network 302 is trained by an error signal ('error'), which is the difference between the output signal from neural network 302 and the ideal signal produced by filter 310 and / or optional quantizer 312 .
[0082] To accomplish this, a training digital signal (e.g., filtered by filter 310 and processed by optional quantizer 312) is subtracted from a corrected output signal output by neural network 302 by comparator 314. The error signal resulting from this comparison is passed through a cost function used to train neural network 302 (e.g., using a Levenberg-Marquardt algorithm, a back-propagation (gradient descent) algorithm, or the like). Back-propagation calculates the gradient in the weight space of a feed-forward neural network with respect to the loss function. Back-propagation is used to adapt the computational parameters (e.g., weights and / or biases) of the neurons of neural network 302 (e.g., following the method of steepest descent, also known as gradient descent). However, any suitable cost function and training algorithm may be used to train a neural network.
[0083] This training phase of the neural network 302 may be performed continuously or iteratively until the corrected output signal of the neural network 302 at node 308 has a minimal error (or at least a desired smaller set of errors) compared to the training digital signal.
[0084] In some embodiments, the neural network 302 may be trained using a predetermined training signal directly retrieved from a memory storage location provided to both the TI ADC device 100 and the neural network. In this case, the training signal does not need to be quantized or filtered. Thus, the neural network 302 may be trained using locally available training information (e.g., "on-chip" training data) rather than a training signal received from an external source.
[0085] To further optimize this training method, in system 300, additional inputs are provided to enable the neural network 302 to identify which particular ADC unit 102 is responsible for producing the current digital value output by the TI ADC device 100 at the output node 304. This additional information enables the neural network 302 to be trained to implement a correction scheme customized for each individual ADC unit 102. Because different ADC units 102 within a TI ADC device 100 may present unique error distributions, as described above, this additional information enables the neural network 302 to establish a different correction scheme for each ADC unit 102.
[0086] To implement this functionality, the system 300 includes an ADC tag generator 316 configured to send an identifier value to the neural network 302, wherein the identifier identifies a particular ADC unit 102 in the TI ADC device 100 that is currently enabled and responsible for generating a digital value output by the TI ADC device 100 at the output node 304. Typically, the identifier value output by the ADC tag generator 316 will be a digital value that identifies the ADC unit 102. In a TI ADC device 100 having six ADC units 102, for example, the identifier values output by the ADC tag generator 316 may be, for example, binary values 000, 001, 010, 011, 100, and 101. The identifier value is processed by the neural network 302 along with the input digital data stream generated by the TI ADC device 100 to enable the neural network 302 to learn a customized error correction scheme specific to the particular ADC unit 102 that is enabled.
[0087] To provide synchronous operation, the ADC flag generator 316 receives a signal having a frequency f s The clock signal input of TI ADC device 100 enables ADC tag generator 316 to coordinate the sending of the ADC tag identifier with the time of enabling the identified ADC unit 102 within TI ADC device 100.
[0088] Specifically, when the controller 311 initiates the operation of the TI ADC device 100 (e.g., by enabling the operation of the first ADC unit 102 of the TI ADC device 100), the controller 311 is configured to send a synchronization signal to the ADC tag generator 316 to notify the ADC tag generator 316 that the TI ADC device 100 has started operating and the first ADC unit 102 has been enabled. At this time, the ADC tag generator 316 is configured to output an ADC unit identifier output signal, which identifies the currently enabled first ADC unit 102. Next, the ADC tag generator 316 is configured to use the clock signal to generate an ADC unit identifier output signal, which identifies which specific ADC unit 102 is currently enabled within the TI ADC device 100.
[0089] Specifically, when the TI ADC device 100 has N individual ADC units 102, the ADC flag generator 316 is configured to generate an output indicating that the first ADC unit 102 has been enabled after receiving a synchronization signal from the controller 311 (indicating that the operation of the TI ADC device 100 has been initiated). When the clock signal switches for the first time after receiving the synchronization signal (at which time the controller 311 enables the second ADC unit 102 of the TI ADC device 100), the ADC flag generator 316 is configured to generate an output indicating that the second ADC unit 102 has been enabled. When the clock signal switches for the second time after receiving the synchronization signal (at which time the controller 311 enables the third ADC unit 102 of the TI ADC device 100), the ADC flag generator 316 is configured to generate an output indicating that the third ADC unit 102 has been enabled. The ADC tag generator 316 continues to operate in this manner, wherein the ADC tag generator 316 sequentially identifies the ADC unit 102 with each clock signal toggle until the ADC unit 102N is enabled, at which time the ADC tag generator 316 returns to identifying the first ADC unit 102 at the next clock signal toggle. The process then continues, with the ADC units 102 within the TI ADC device 100 being enabled sequentially, and the ADC tag generator 316 synchronously outputting the identifier of the currently enabled ADC unit 102.
[0090] Thus, in system 300, ADC tag generator 316 generates N unique tag identifier values and outputs those values in a loop that is consistent with the loop used to iterate through the various ADC units 102 of TI ADC device 100 and at a sampling frequency equal to f s The sampling rate is coordinated with the algorithm.
[0091] Thus, each digital output sample from the TI ADC device 100 that is provided to the neural network 302 is tagged with the identifier of the originating ADC unit 102 for that data sample. This in turn informs the neural network 302 which ADC unit 102 is responsible for producing which output data value. If the neural network knows the ADC unit 102 that produced each particular output value, the neural network 302 thereby gains the ability to compensate for the unique errors produced by that particular ADC unit 102.
[0092] According to certain embodiments of the present disclosure, one or more PVT sensors 318 may optionally be incorporated into the system 300 to provide PVT parameters to the neural network 302 so that the neural network 302 can further modify the output of the TI ADC device 100 to correct for PVT variations. Thus, when the neural network 302 is trained, its cost function will take into account the PVT impact of errors caused by such processed inputs to be used to adjust weights and / or biases within the nodes of the neural network 302.
[0093] According to various embodiments of the present disclosure, any type of process-related information, such as various technology parameters (e.g., electron mobility, transistor parameters (e.g., Vt, fT, β, doping), resistor parameters (e.g., nominal resistor value, voltage dependency), or capacitor parameters (e.g., nominal capacitor value, voltage dependency) may be added as PVT input to the neural network 302. As long as there is a way to measure the process-related parameters, the output of this measurement may be provided as PVT input to the neural network 302.
[0094] Despite Figure 3A The neural network 302 trained according to the training protocol is described above. Figure 3B The system 300 is depicted in a production or active configuration for incorporation into an end-use application. Figure 3B , the neural network 302 has already been trained, and therefore, does not need to receive a training digital input signal previously used to train the neural network 302. Therefore, the signal path from the digital training signal input 309 through the filter 310 and the optional quantizer 312 is removed. Similarly, no error signal feedback is required, so the comparator 314 and its connections can be removed.
[0095] During normal operation, an analog input signal is applied to an input 306 of the TI ADC device 100, and the TI ADC device 100 generates a digital output signal at an output 304, which is subsequently modified by the trained neural network 302 to generate a compensated or calibrated corrected output signal at a node 308, which can be utilized by other circuit systems or components coupled to the output node 308. In order to enable the neural network 302 to apply an appropriate error correction scheme to correct the output of the TI ADC device 100, the ADC flag generator 316 informs the neural network 302 of the particular ADC unit 102 responsible for generating the output being corrected by the neural network 302. Similarly, one or more PVT sensors 318 provide PVT inputs to the neural network 302, thereby enabling the neural network 302 to apply a particular correction to compensate for the PVT condition of the ADC unit 102. Upon receiving the ADC unit 102 identifier value, the neural network 302 is trained to apply an error correction scheme customized for that particular ADC unit 102 to correct the digital output signal output by the TI ADC device 100 and received by the neural network 302.
[0096] In system 300, ADC signature generator 316 may be implemented by any suitable circuit system or component configured to receive a signal having a frequency f s The TI ADC device 100 receives as input a clock signal of the ADC unit 102 and, in response, generates an output signal including an identifier value that can be used to identify a particular ADC unit 102 of the TI ADC device 100 .
[0097] ADC signature generator 316 is typically a digital circuit configured to associate data samples with specific ADC units of the TI ADC device to enable the neural network to correct the output while taking into account the individual error distribution of each ADC unit. In some cases, ADC signature generator 316 may be implemented as a binary counter configured to count the total number N of ADC units 102 before being reset.
[0098] As discussed above, implementation Figure 3A and 3B , such that the neural network 302 is trained to process the output signal of the TI ADC device 100 to produce a correction signal output by the neural network 302 as the corrected output signal at the node 308. However, in other embodiments, the neural network 302 may be configured to generate an error signal or a correction signal using a correction scheme, which may be applied to the output of the TI ADC device 100 (e.g., via subtraction) to correct the output of the TI ADC device 100.
[0099] Figure 4 Depicts a system 400 including a neural network 402 configured to generate an error correction signal configured to correct ADC unit specific errors within a TI ADC device. The system 400 is configured to generate an error correction signal configured to correct ADC unit specific errors within a TI ADC device. Figure 3A and 3B 4 and is configured in a manner similar to the system 300 described above. Thus, the system 400 has a Figure 3A and 3B During operation, the neural network 402 analyzes the output of the TI ADC device 100, and in conjunction with the ADC unit 102 identifier information received from the ADC tag generator 316, the neural network 402 is configured to calculate an error signal representing the difference between the actual output and the desired or ideal output of the TI ADC device 100. The neural network 402 is configured to output the error signal to the summing node 404.
[0100] At the same time, the output of the TI ADC device 100 is provided to the summing node 404 through an optional delay element 406. The optional delay element 406 is any appropriately configured circuit component that is configured to delay the transmission of the output signal of the TI ADC device 100 to the node 404 by a time delay that is equal to the time delay of the signal passing through the neural network 402. In this way, the original output of the TI ADC device 100 and the error signal generated by the neural network 402 arrive at the summing node 404 substantially simultaneously. The summing node 404 is then configured to subtract the error signal from the original output signal of the TI ADC device 100 to produce a corrected output signal.
[0101] Figure 5A and 5B is a schematic diagram depicting an example implementation of the ADC signature generator 316. Figure 5A In the example, the ADC flag generator 500 is implemented as a modulo-N counter ( Figure 5A 4, where N is the number of ADC units 102 in the TI ADC device 100. In the modulo 4 counter configuration, the ADC flag generator 500 includes two flip-flops: flip-flop 502 and flip-flop 504. The clock input (clk) of flip-flop 502 receives a clock signal having a frequency f. s The clock signal is used as input. The data input terminal (D) and the inverting output terminal of the flip-flop 502 The non-inverting output terminal (Q) of flip-flop 502 is connected to the clock input terminal (clk) of flip-flop 504. The data input terminal (D) and the inverting output terminal (Q) of flip-flop 504 are connected to each other and represent the least significant bit (LSB) of the counter. are connected to each other and represent the most significant bit (MSB) output of the 2-bit counter of the ADC flag generator 500.
[0102] As the clock signal input to the ADC signature generator 500 switches, the output of the ADC signature generator 500 will count through binary values from a minimum value of zero up to a maximum binary value of 11 (i.e., a value of 3) or N-1, where N is the number of ADC units 102 in the TI ADC device 100. The LSB and MSB values are supplied to the neural network 302 as ADC unit identifiers, enabling the neural network 302 to identify the particular ADC unit 102 that generated the data processed by the neural network 302.
[0103] exist Figure 5B In , ADC flag generator 550 is implemented as an n-bit shift register. Figure 5B In the example depicted in , n is equal to 4. ADC flag generator 550 is composed of four flip-flops 552 , 554 , 556 , and 558 .
[0104] The clock input (clk) of each flip-flop 552, 554, 556, and 558 receives a clock having a frequency f s The clock signal of flip-flop 556 is used as input. The data input terminal (D) of flip-flop 552 is connected to the non-inverting output terminal (Q) of flip-flop 558. The non-inverting output terminal (Q) of flip-flop 552 is connected to the data input terminal (D) of flip-flop 554. The non-inverting output terminal (Q) of flip-flop 554 is connected to the data input terminal (D) of flip-flop 556. The non-inverting output terminal (Q) of flip-flop 556 is connected to the data input terminal (D) of flip-flop 558.
[0105] When the clock signal switches, a high output value (eg, a binary value of '1') is output at terminal Q0 and transitions to Q 1、 The '1' value returns to being output at terminal Q0 after another clock toggle, and the process continues.
[0106] The output values Q0 , Q1 , Q2 , and Q3 are supplied to the neural network 302 as ADC unit 102 identifiers, enabling the neural network 302 to identify the particular ADC unit 102 that generated the data processed by the neural network 302 .
[0107] As described herein, the present disclosure provides an improved TI ADC system in which an error correction system (e.g., a neural network) is configured to perform corrections in the TI ADC system to correct ADC conversions occurring in individual ADC units of the TI ADC system. Specifically, the error correction system is configured to receive input indicating which of several ADC units are enabled (and actively sending digital values to the output of the TI ADC system), and then apply an error correction scheme that may be unique to that particular ADC unit. In this way, improved error correction may be achieved.
[0108] To illustrate, Fig. 6A is a graph depicting the frequency domain output of a Fast Fourier Transform (FFT) of the output signal of a conventional TI ADC device applied to an analog input signal. The horizontal axis represents frequency, while the vertical axis represents magnitude. Errors resulting from time interleaving, which may be referred to as 'IL spurs' (interleaving spurs), occur at a fixed frequency (for a sinusoidal input) and are Fig. 6A Generally speaking, a TI ADC with N ADC units has a peak value of f OS =(k / N)f s These offset spurs are generated at f G =(k / N)f s ±f in The corresponding gain spur is found at .
[0109] For example, if a particular TI ADC device has N=6 ADC units, then the offset spurs will be at f=0, f=f s / 6, f = f s / 3, f = 2f s / 3(aliased to f s / 3), f=5f s / 6(aliased to f s / 6). Fig. 6A In FIG. 5 , the spectrum of a TI ADC with an interleaving factor of 6 is shown. Spurious content associated with time interleaving can be found highlighted at the characteristic frequencies discussed above. In this example, fs = 47 MHz and fin = 559 kHz.
[0110] Figure 6B3. The same output is shown with corrections applied by an error correction system (e.g., neural network 302) configured in accordance with the present disclosure. Specifically, the error correction system has been trained using ADC unit identifiers generated by an ADC tag generator (e.g., ADC tag generator 316). This enables the error correction system to apply error corrections that are customized for the specific errors (gain, offset, or other) generated by each individual ADC unit. Thus, time interleaving spurs in the output of the TI ADC device can be corrected. Figure 6B As seen in , the magnitude of those IL spurs has been greatly reduced, indicating a significant reduction in the error in the corrected TI ADC device output.
[0111] As described herein, various embodiments of the present disclosure implement a machine learning system configured to compensate for ADC errors in a specific ADC unit of a TI ADC device. Such a machine learning system may be implemented in hardware and / or software, for example, in one or more data processing systems. However, the functionality described herein is not limited to implementation into any specific hardware / software platform.
[0112] Figure 7 is a flow chart of an example method 700 for correcting the output of a TI ADC device by applying an ADC unit specific correction. The method 700 may be implemented by any suitable error correction system (e.g., Figure 3A , 3B The error correction system is implemented by a neural network 302 of TI ADC device 100 and 4, the error correction system is connected to the output of the TI ADC device and is configured to receive an input specifying which particular ADC unit within the TI ADC device is enabled. At step 710, the error correction system receives a first digital output signal from the TI ADC device (e.g., TI ADC device 100). As described herein, the TI ADC device includes a plurality of individual ADC units (e.g., ADC unit 102) configured to measure the magnitude of an input signal and output a digital value corresponding to the magnitude of the input signal.
[0113] At step 720, the error correction system receives a first identifier from the ADC tag generator. The first identifier identifies a specific ADC unit from a plurality of ADC units that is enabled and currently generating a digital output signal for the entire TI ADC device. After receiving the first identifier information, at step 730, the error correction system modifies the first digital output signal by applying a first error correction scheme associated with the first ADC to compensate for a first analog-to-digital conversion error occurring within the first ADC to generate a first corrected digital output signal.
[0114] After the ADC unit has completed sampling its input signal, the ADC unit is disabled and another ADC unit of the TI ADC device is enabled. At this point, the error correction system receives a signal identifying the second ADC unit and begins applying an error correction scheme specific to that ADC unit. This process continues until all N ADC units contained within the TI ADC device have been sequentially enabled (with the neural network applying the appropriate error correction to each individual ADC unit). The process then repeats with the first ADC unit of the plurality of ADC units enabled.
[0115] As will be appreciated by those skilled in the art, aspects of the present disclosure may be embodied as systems, processes, methods, and / or program products. Thus, aspects of the present disclosure (e.g., machine learning systems) may take the form of fully hardware embodiments, fully software embodiments (including firmware, resident software, microcode, etc.), or embodiments combining software and hardware aspects (which may be generally referred to herein as "circuits," "circuitry," "modules," or "systems"). Furthermore, aspects of the present disclosure may take the form of program products embodied in one or more computer-readable storage media having computer-readable program code embodied thereon. (However, any combination of one or more computer-readable media may be utilized. The computer-readable medium may be a computer-readable signal medium or a computer-readable storage medium.)
[0116] A computer-readable storage medium may be, for example, but not limited to, an electronic, magnetic, optical, electromagnetic, infrared, biological, atomic, or semiconductor system, device, controller, or apparatus, or any suitable combination of the foregoing, wherein the computer-readable storage medium itself is not a transient signal. More specific examples of computer-readable storage media (a non-exhaustive list) may include the following: an electrical connector having one or more wires, a portable computer disk, a hard disk, a random access memory ("RAM"), a read-only memory ("ROM"), an erasable programmable read-only memory ("EPROM" or flash memory), an optical fiber, a portable compact disc read-only memory ("CD-ROM"), an optical storage device, a magnetic storage device, or any suitable combination of the foregoing. In the context of this document, a computer-readable storage medium may be any tangible medium that may contain or store a program for use by or in conjunction with an instruction execution system, device, controller, or apparatus. Program code embodied in a computer-readable signal medium may be transmitted using any suitable medium, including but not limited to wireless, wire, fiber optic cable, RF, etc., or any suitable combination of the foregoing.
[0117] A computer readable signal medium may include a propagated data signal in which a computer readable program code is embodied, for example, in baseband or as part of a carrier wave. Such a propagated signal may take any of a variety of forms, including but not limited to electromagnetic, optical, or any suitable combination thereof. A computer readable signal medium may be any computer readable medium that is not a computer readable storage medium and that can communicate, propagate, or transport a program for use by or in conjunction with an instruction execution system, device, controller, or apparatus.
[0118] The block diagrams in the figures illustrate the architecture, functionality, and operation of possible implementations of circuit systems, systems, methods, processes, and program products according to various embodiments of the present disclosure. In this regard, certain blocks in the block diagrams (e.g., neural network 302) may represent portions of modules, segments, or codes that include one or more executable program instructions for implementing specified logical functions. It should also be noted that in some embodiments, the functions annotated in the boxes may not occur in the order annotated in the figures. For example, depending on the functions involved, two boxes shown in succession may actually be executed substantially simultaneously, or the boxes may sometimes be executed in reverse order.
[0119] Modules implemented in software for various types of processors to execute may, for example, include one or more physical or logical blocks of computer instructions, which may, for example, be organized as objects, programs, or functions. However, the executable files of the identified modules do not have to be physically located together, but may include different instructions stored in different locations, which include the modules when logically combined together and realize the stated purpose of the modules. In fact, the module of the executable code may be a single instruction, or many instructions, and may even be distributed on several different code segments among different programs, and across several memory devices. Similarly, operational data (e.g., a knowledge base of adapted weights and / or biases described herein) may be identified and shown in the module herein, and may be embodied and organized in any suitable form in any suitable type of data structure. The operational data may be collected as a single data set, or may be distributed in different locations, including being distributed on different storage devices. Data may provide electronic signals on a system or network.
[0120] These program instructions may be provided to a general purpose computer, one or more processors and / or controllers of a special purpose computer or other programmable data processing devices (e.g., a controller) to produce a machine, so that the instructions executed by the processor of the computer or other programmable data processing devices create a circuit system or method for implementing the functions / actions specified in one or more blocks of the block diagram.
[0121] It should also be noted that each block of the block diagram, and combinations of blocks in the block diagram, may be implemented by a dedicated hardware-based system (e.g., which may include one or more graphics processing units) or a combination of dedicated hardware and computer instructions that performs the specified functions or actions. For example, a module (e.g., neural network 302) may be implemented as a hardware circuit that includes custom VLSI circuits or gate arrays, off-the-shelf semiconductors such as logic chips, transistors, controllers, or other discrete components. A module (e.g., neural network 302) may also be implemented in a programmable hardware device such as a field programmable gate array, programmable array logic, programmable logic device, dedicated IC, microcontroller, system on chip, general purpose processor, microprocessor, etc.
[0122] Computer program code, i.e., instructions, for performing operations of various aspects of the present disclosure may be written in any combination of one or more programming languages, including object-oriented programming languages such as Java, Smalltalk, Python, C++, conventional procedural programming languages such as the "C" programming language or similar programming languages, or any of the machine learning software disclosed herein.
[0123] These program instructions may also be stored in a computer-readable storage medium, which may direct a computer system, other programmable data processing device, controller or other apparatus to function in a specific manner so that the instructions stored in the computer-readable medium produce an article of manufacture, which includes instructions for implementing the functions / actions specified in one or more blocks of the block diagram.
[0124] The program instructions may also be loaded onto a computer, other programmable data processing device, controller or other apparatus to cause a series of operating steps to be executed on the computer, other programmable device or other apparatus to produce a computer-implemented process, so that the instructions executed on the computer, other programmable device or other apparatus provide a process for implementing the functions / actions specified in one or more blocks of the block diagram.
[0125] The preceding detailed description is merely illustrative in nature and is not intended to limit the embodiments of the subject matter or the application and uses of such embodiments.
[0126] As used herein, the word "exemplary" means "serving as an example, instance, or illustration." Any embodiment described herein as exemplary is not necessarily to be construed as preferred or advantageous over other embodiments. Furthermore, there is no intention to be bound by any expressed or implied theory presented in the preceding technical field, background, or detailed description.
[0127] The connecting lines shown in each figure contained herein are intended to represent exemplary functional relationships and / or physical couplings between various elements. It should be noted that many alternative or additional functional relationships or physical connections may exist in the embodiments of the subject matter. In addition, certain terms may also be used herein only for reference purposes, so these terms are not intended to be restrictive, and unless the context clearly indicates, the terms "first", "second" and other such numerical terms referring to structures do not imply order or sequence.
[0128] As used herein, "node" means any internal or external reference point, connection point, junction, signal line, conductive element, etc., at which a given signal, logic level, voltage, data pattern, current or quantity exists. In addition, two or more nodes may be implemented by one physical element (and two or more signals may be multiplexed, modulated or otherwise distinguished even if received or output at a common node).
[0129] The foregoing description refers to elements or nodes or features being "connected" or "coupled" together. As used herein, unless otherwise expressly stated, "connected" means that one element is directly connected to (or is in direct communication with) another element, and not necessarily mechanically. Likewise, unless otherwise expressly stated, "coupled" means that one element is directly or indirectly connected to (or is in direct communication with, electrically or otherwise) another element, and not necessarily mechanically. Thus, while the schematic diagrams shown in the figures depict one exemplary arrangement of elements, additional intervening elements, devices, features, or components may be present in embodiments of the depicted subject matter.
[0130] Although at least one exemplary embodiment has been presented in the foregoing specific embodiments, it should be understood that there are a large number of variations. It should also be understood that one or more exemplary embodiments described herein are not intended to limit the scope, applicability, or configuration of the claimed subject matter in any way. In fact, the foregoing specific embodiments will provide a convenient guide for implementing the described embodiment or embodiments to those skilled in the art. It should be understood that various changes may be made to the function and arrangement of elements without departing from the scope defined by the claims, and the scope defined by the claims includes known equivalents and foreseeable equivalents at the time of filing this patent application.
Claims
1. A system, characterized in that: include: A time-interleaved analog-to-digital converter (ADC) configured to sample at a rate f s Operation, the time-interleaved ADC comprises: An input terminal configured to receive an analog input signal, an output terminal configured to output a digital signal, wherein the digital signal is a digital approximation of the analog input signal, and A plurality of ADC units, wherein each ADC unit of the plurality of ADC units is configured to: measuring an analog value of the analog input signal, converting the analog value into a digital value, and sending the digital value to the output of the time-interleaved ADC; A controller configured to: The receiving device has s determines the frequency of the clock signal, and enabling ADC units in the plurality of ADC units sequentially at a rate determined by the clock signal; An ADC signature generator configured to: receiving the clock signal, and generating an ADC unit identifier output signal using the clock signal, the ADC unit identifier output signal identifying a currently enabled ADC unit among the plurality of ADC units; and A machine learning system configured to: receiving the digital signal from the output of the time-interleaved ADC, receiving the ADC unit identifier signal from the ADC tag generator, and The digital signal is modified to produce a corrected digital signal by applying an error correction scheme associated with the ADC unit identified by the ADC unit identifier signal to compensate for analog-to-digital conversion errors occurring within the ADC unit, wherein the machine learning system determines the error correction scheme using the ADC unit identifier signal.
2. The system according to claim 1, characterized in that Also included is a sensor configured to provide a parameter to the machine learning system, wherein the output of the ADC unit varies with the value of the parameter.
3. The system according to claim 1, characterized in that The machine learning system includes a neural network having a plurality of neurons arranged into a network having fewer than three layers and fewer than ten neurons in each layer.
4. A system, characterized in that: include: A time-interleaved analog-to-digital converter (ADC) comprising a plurality of ADC cells connected in parallel between an input and an output of the time-interleaved ADC; an ADC flag generator configured to output a first identifier of a first ADC unit among the plurality of ADCs when the first ADC unit is enabled; and An error correction system configured to: receiving a first digital signal from the output of the time-interleaved ADC, receiving the first identifier of the first ADC unit from the ADC signature generator, modifying the first digital signal to produce a first corrected digital signal by applying a first error correction scheme associated with the first ADC unit to compensate for a first analog-to-digital conversion error occurring within the first ADC unit, and The first corrected digital signal is outputted.
5. The system according to claim 4, characterized in that: The ADC flag generator is configured to output a second identifier of a second ADC unit among the plurality of ADCs when the second ADC unit is enabled; and The error correction system is configured to: receiving a second digital signal from the output of the time-interleaved ADC, receiving the second identifier of the second ADC unit from the ADC signature generator, modifying the second digital signal to produce a second corrected digital signal by applying a second error correction scheme associated with the second ADC unit to compensate for a second analog-to-digital conversion error occurring within the second ADC unit, and The second corrected digital signal is output.
6. The system according to claim 4, characterized in that The time-interleaved ADC comprises a controller configured to: The receiving device has s determines the frequency of the clock signal, and ADC units in the plurality of ADC units are enabled sequentially at a rate determined by the clock signal.
7. The system according to claim 6, characterized in that The ADC signature generator is configured to: receiving the clock signal; and An ADC unit identifier output signal is generated using the clock signal to sequentially identify the ADC units among the plurality of ADC units when the ADC units are enabled, and the ADC unit identifier output signal is sent to the error correction system.
8. A method, characterized in that include: receiving a first digital output signal from a time-interleaved analog-to-digital converter (ADC), wherein the time-interleaved ADC comprises a plurality of ADC cells; receiving a first identifier from an ADC tag generator, wherein the first identifier identifies a first ADC unit that is currently enabled among the plurality of ADC units; as well as The first digital output signal is modified by applying a first error correction scheme associated with the first ADC to compensate for a first analog-to-digital conversion error occurring within the first ADC to produce a first corrected digital output signal.
9. The method according to claim 8, characterized in that Also includes: receiving a second digital output signal from the time-interleaved ADC; receiving a second identifier from the ADC tag generator, wherein the second identifier identifies a second ADC unit among the plurality of ADC units that generates the second digital output signal; as well as The second digital output signal is modified by applying a second error correction scheme associated with the second ADC to compensate for a second analog-to-digital conversion error occurring within the second ADC to produce a second corrected digital output signal.
10. The method according to claim 8, characterized in that The time-interleaved ADC has a sampling rate f s , and the method further comprises: The receiving device has s a clock signal of a certain frequency; and An ADC unit identifier output signal is generated using the clock signal, the ADC unit identifier output signal identifying a currently enabled ADC unit among the plurality of ADC units.
