Digital pre-distortion system, device and medium
Patent Information
- Authority / Receiving Office
- WO · WO
- Patent Type
- Applications
- Current Assignee / Owner
- Filing Date
- 2025-02-10
- Publication Date
- 2026-08-13
Smart Images

Figure CN2025076669_13082026_PF_FP_ABST
Abstract
Description
DIGITAL PRE-DISTORTION SYSTEM, DEVICE AND MEDIUMBACKGROUND1. Field of the Disclosure
[0001] The present disclosure relates to the field of communication technology, and more specifically, to a digital pre-distortion system and method used in connection with 4th Generation, 5th Generation and 6th Generation (4G / 5G / 6G) wireless telecommunications systems. 2. Description of Related Art
[0002] Power amplifiers (PA) are one of the most critical components in wireless communication systems, significantly impacting signal coverage, signal quality, and system energy consumption. To address the nonlinear characteristics of PAs, especially broadband PAs (with instantaneous bandwidth IBW ≥ 200 MHz) , linearization techniques such as digital predistortion (DPD) are applied at the digital intermediate frequency stage. DPD aims to correct for PA nonlinearities by applying an inverse distortion to the input signal, thereby improving output signal linearity.
[0003] Broadband PAs exhibit complex nonlinear behaviors across multiple dimensions including occupied bandwidth (oBW) , component carriers (cc) , signal power levels, operating temperature, and equipment aging. Existing DPD technologies often rely on pre-calculated and stored lookup tables of model parameters tailored to various combinations of these dimensions and their respective granularities. For instance, consider a scenario where IBW = 450 MHz with 3 component carriers; cc1 has an oBW of 100 MHz, cc2 has an oBW of 60 MHz, and cc3 has an oBW of 40 MHz; signal power levels within each oBW vary from -29 dBFS to -12 dBFS in 0.5 dB increments; PA operating temperatures are considered at high and low extremes; equipment aging conditions are evaluated at 3 years, 6 years, and 9 years of usage.
[0004] In this example, the total number of condition combinations is 3 × 35 × 2 × 3 = 630, implying the need to pre-store 630 parameter configuration lookup tables. Additional conditions or finer granularity would require even more lookup tables, leading to substantial storage resource requirements in Field Programmable Gate Arrays (FPGAs) or application-specific integrated circuits (ASICs) . This approach of trading space for performance is inefficient and lacks true generalization and adaptability.
[0005] Accordingly, there is a need for a digital pre-distortion system that overcomes, alleviates, and / or mitigates one or more of the aforementioned and other deleterious effects of prior art pre-distortion systems used in wireless telecommunication systems.SUMMARY
[0006] Accordingly, what is needed is a digital pre-distortion system that consumes significantly fewer storage resources compared to traditional digital predistortion systems.
[0007] It is further desired to provide a digital pre-distortion system that achieves excellent generalization performance under various power amplifier operating conditions.
[0008] It is still further desired to provide a digital pre-distortion system that leverages an artificial neural network and can be used with 4th / 5th / 6th Generation (4G / 5G / 6G) telecommunications systems.
[0009] In one configuration, an artificial neural network digital pre-distortion (AI-DPD) system is provided. The AI-DPD includes: 1) a Preprocessor; 2) a Deep Feature Extractor; 3) a Shallow Feature Extractor; 4) a Feature Enhancer; and 5) a Predistortion Filter.
[0010] 1) The Preprocessor is used to address the nonlinear characteristics of Radio Frequency (RF) PAs by converting a complex signal vector into an equal-length real vector. The Preprocessor will also perform transformation processing on the operating conditions indication (OCI) . The transformed values are then combined with the real vector to form the final real signal vector.
[0011] 2) The Deep Feature Extractor uses a deep neural network (DNN) to employ data-driven learning to extract deep-level features correlated with the nonlinear characteristics of the PA from the real signal vector to output a deep feature vector of equal length to the complex signal vector.
[0012] 3) The Shallow Feature Extractor uses a neural network with a single hidden layer and employs data-driven learning methods to extract shallow-level features that are highly correlated with the nonlinear characteristics of the PA from the real signal vector to output a shallow feature vector of equal length to the complex signal vector.
[0013] 4) The Feature Enhancer performs a weighted summation of the deep feature vector and the shallow feature vector to output a key feature vector of equal length to the complex signal vector.
[0014] 5) The Predistortion Filter uses the key feature vector as filter coefficients to filter the complex signal vector to perform linearization pre-correction of the PA’s nonlinear distortion and output a scalar pre-correction signal.
[0015] The AI-DPD system utilizes one or more processors and includes a memory. The memory is configured to store one or more programs. When the one or more programs are executed by the one or more processors, they cause the one or more processors to implement the AI-DPD system, which includes the various steps and features described above.
[0016] It is contemplated that the AI-DPD system can be used in connection with a telecommunications system including 4G and 5G wireless telecommunications systems.
[0017] For this application, the following terms and definitions shall apply:
[0018] The term “data” as used herein means any indicia, signals, marks, symbols, domains, symbol sets, representations, and any other physical form or forms representing information, whether permanent or temporary, whether visible, audible, acoustic, electric, magnetic, electromagnetic or otherwise manifested. The term “data” as used to represent predetermined information in one physical form shall be deemed to encompass any and all representations of the same predetermined information in a different physical form or forms.
[0019] The term “network” as used herein includes both networks and internetworks of all kinds, including the Internet, and is not limited to any particular type of network or inter-network.
[0020] The term "user equipment" as used herein encompasses any suitable type of wireless user device, such as mobile phones, portable data processing devices, portable web browsers, or in-vehicle mobile stations.
[0021] The terms “first” and “second” are used to distinguish one element, set, data, object or thing from another, and are not used to designate relative position or arrangement in time.
[0022] The terms “coupled” , “coupled to” , “coupled with” , “connected” , “connected to” , and “connected with” as used herein each means a relationship between or among two or more devices, apparatus, files, programs, applications, media, components, networks, systems, subsystems, and / or means, constituting any one or more of (a) a connection, whether direct or through one or more other devices, apparatus, files, programs, applications, media, components, networks, systems, subsystems, or means, (b) a communications relationship, whether direct or through one or more other devices, apparatus, files, programs, applications, media, components, networks, systems, subsystems, or means, and / or (c) a functional relationship in which the operation of any one or more devices, apparatus, files, programs, applications, media, components, networks, systems, subsystems, or means depends, in whole or in part, on the operation of any one or more others thereof.
[0023] The term "automatic" and variations thereof, as used herein, refers to any process or operation done without material human input when the process or operation is performed. However, a process or operation can be automatic, even though performance of the process or operation uses material or immaterial human input, if the input is received before performance of the process or operation. Human input is deemed to be material if such input influences how the process or operation will be performed. Human input that consents to the performance of the process or operation is not deemed to be "material. "
[0024] In one configuration, a digital predistortion system is provided comprising: a preprocessor configured to receive a complex signal vector that includes: historical and current values of a transmit signal after digital up-conversion, and power amplifier operating condition indictors (OCIs) . The digital predistortion system is provided such that the preprocessor converts the complex signal vector into a real number signal vector having a length equal to the complex signal vector and performing transformation processing on the OCIs and combines the transformed OCIs with the real vector to form the final real signal vector. The digital predistortion system further comprises a Deep Feature Extractor (DFE) configured to extract features related to nonlinearity characteristics of a power amplifier (PA) from the real number signal vector. The digital predistortion system is provided such that the DFE outputs a deep feature vector having a length equal to the complex signal vector. The digital predistortion system still further comprises a Shallow Feature Extractor (SFE) configured to extract features related to the nonlinearity characteristics of the PA from the real number signal vector. The digital predistortion system is provided such that the SFE outputs a shallow feature vector having a length equal to the complex signal vector. The digital predistortion system also comprises a Feature Enhancer (FE) configured to overlay the deep feature vector and the shallow feature vector by performing a weighted summation of the deep feature vector and the shallow feature vector to generate and output a key feature vector having a length equal to the complex signal vector. Finally, the digital predistortion system comprises a Predistortion Filter (PF) receiving the key feature vector and filtering the complex signal vector by linearization pre-correction of the PA’s nonlinear distortion to generate a scalar pre-correction signal, where the PF outputs the scalar pre-correction signal.
[0025] In another configuration, the digital predistortion system is provided as an artificial neural network digital pre-distortion (AI-DPD) system.
[0026] The above-described and other features and advantages of the present disclosure will be appreciated and understood by those skilled in the art from the following detailed description, drawings, and appended claims. DESCRIPTION OF THE DRAWINGS
[0027] FIG. 1 is a schematic diagram illustrating the deployment of the AI-DPD provided by an embodiment of the present application in the RRU unit of a 4G wireless communication system or the AAU unit of a 5G / 6G system.
[0028] FIG. 2 is a block diagram illustrating the AI-DPD used in connection with a 4G wireless communication system or a 5G / 6G system.
[0029] FIG. 3 is a block diagram illustrating an implementation structure of the AI-DPD according to FIG. 2.
[0030] FIG. 4 is a comparison chart illustrating nonlinear distortion characteristics of a narrowband PA and a broadband PA.
[0031] FIG. 5 is a block diagram illustrating the implementation structure of the Preprocessor according to FIG. 3.
[0032] FIG. 6 is a block diagram illustrating the implementation structure of the Deep Feature Extractor according to FIG. 3.
[0033] FIG. 7 is a block diagram illustrating the implementation structure of the Shallow Feature Extractor according to FIG. 3.
[0034] FIG. 8 is a block diagram illustrating the implementation structure of the Feature Enhancer according to FIG. 3.
[0035] FIG. 9 is a block diagram illustrating the implementation structure of the Pre-distortion Filter according to FIG. 3.
[0036] FIG. 10 illustrates two architectural diagrams for training the AI-DPD according to FIG. 3.
[0037] FIG. 11 is a chart showing test performance results of the AI-DPD according to FIG. 10.
[0038] FIG. 12 is a block diagram illustrating the structure of the AI-DPD device according to FIG. 10.DETAILED DESCRIPTION
[0039] In the following text, various embodiments of the present disclosure will be described with reference to the accompanying drawings. The following description is provided with reference to the accompanying drawings, which illustrate specific configurations of the invention.
[0040] In one configuration, an artificial neural network digital pre-distortion system is (AI-DPD) is provided. The AI-DPD can be configured in a radio remote unit (RRU) 50 of a 4G wireless communication system or, for example, the antenna array unit (AAU) 60 of a 5G / 6G system.
[0041] FIG. 1 diagrammatically illustrates the location of an RRU 50 in a 4G system or an AAU 60 in a 5G / 6G system within their respective systems according to the prior art.
[0042] FIG. 2 is a block diagram illustrating the AI-DPD 100 within the RRU 50 of a 4G system or the AAU 60 of a 5G / 6G system.
[0043] FIG. 3 is a block diagram illustrating one implementation structure of AI-DPD 100. In this configuration, the AI-DPD 100 includes a Preprocessor 110, a Deep Feature Extractor 120, a Shallow Feature Extractor 130, a Feature Enhancer 140, and a Predistortion Filter 150.
[0044] In FIG. 3, the main input to the Preprocessor comprises a complex signal vector formed by the historical and current values of the transmitted signal after digital up-conversion. For example, let the current transmitted complex signal value be: where xI (n) and xQ (n) represent the in-phase component and quadrature component of the current signal, respectively. The historical signal values before the current moment are: {x (n-1) , …, x (n-M) } . Here M is a positive integer constant that reflects the strength of the memory effect specific to the PA. Then, the complex signal vector is: [x (n-M) , …, x (n-1) , x (n) ] Nonlinear characteristics of Radio Frequency (RF) PAs are primarily caused by their memory effects. Memory effect refers to the fact that the current output value of the PA depends not only on its current input value, but also on its historical input values, making it a typical memory system and causal system. The Preprocessor 110 converts the complex signal vector: [x (n-M) , …, x (n-1) , x (n) ] into an equal-length real vector: [s (n-M) , …, s (n-1) , s (n) ] .
[0045] Another input to the Preprocessor 110 is the operating conditions indication (OCI) of PAs. It is a real scalar signal, or a real signal vector used to indicate the primary operational conditions of the power amplifier at the current time. The Preprocessor 110 will perform certain transformation processing on the OCI, the transformed values are then combined with the real vector to form the final real signal vector. For example, suppose the transformed value of the OCI input is a scalar signal: {oci} , then the real signal vector is: [s (n-M) , …, s (n-1) , s (n) , oci]
[0046] As can be seen in FIG. 3, the Preprocessor 110 is connected to the Deep Feature Extractor 120 via a fully connected layer. The Deep Feature Extractor 120, through a deep neural network (DNN) , employs data-driven learning or training methods to extract deep-level features that are highly correlated with the nonlinear characteristics of the power amplifier from the real signal vector. It then outputs a complex deep feature vector of equal length to the aforementioned complex signal vector:
[0047] The Preprocessor 110 is also connected to the Shallow Feature Extractor 130 via another fully connected layer. The Shallow Feature Extractor 130 uses a neural network with a single hidden layer and similarly employs data-driven learning or training methods to extract shallow-level features that are highly correlated with the nonlinear characteristics of the power amplifier from the real signal vector. The Shallow Feature Extractor 130 then outputs a complex shallow feature vector of equal length to the complex signal vector:
[0048] In FIG. 3, both the Deep Feature Extractor 120 and the Shallow Feature Extractor 130 are connected to the Feature Enhancer 140 via dot connections.
[0049] The Feature Enhancer 140 reinforces and highlights key features by performing a weighted summation of the deep feature vector and the shallow feature vector. It then outputs a complex key feature vector of equal length to the original complex signal vector:
[0050] The Predistortion Filter 150 uses the key feature vector as filter coefficients to filter the complex signal vector. This process accomplishes the linearization pre-correction of the power amplifier's nonlinear distortion and outputs a scalar pre-correction signal: y (n) .
[0051] Another configuration further clarifies the power amplifier operating condition indicators (OCI) input to the Preprocessor 110, as well as how the Preprocessor 110 converts the complex signal vector into a real vector.
[0052] The occupied bandwidth (oBW) is a positive integer that is measured in MHz. It indicates the actual occupied bandwidth of one or more component carriers (cc) within the instantaneous bandwidth (IBW) of the RF PA. For example: IBW=450 (MHz) , cc1=3.5 (GHz) , oBW1=100 (MHz) ; cc2=3.85 (GHz) , oBW2=100 (MHz) , oBW=oBW1+oBW2=200 (MHz) .
[0053] The number of component carriers (Ncc) is also a positive integer but is dimensionless, indicating the number of cc within the IBW of the RF PA. In the above example, NCC=2.
[0054] The operating temperature (OT) is a non-zero integer, with units in degrees Celsius (℃) . It indicates the average internal temperature of the PA device during operation. For example, the typical operating temperature range for a gallium nitride (GaN) power amplifier is -40℃ to +150℃. If the current temperature exactly matches the lower limit of the tolerable temperature, then: OT=-40 (℃) . When the operating temperature is 0℃, OT is set to the smallest positive integer, i.e., OT=1 (℃) .
[0055] The operational age (OA) is also a positive integer, with units measured in years. It indicates the number of years the current PA device has been in operation. If the operational age is less than one year, it is set to one year. For example: OA=6 years.
[0056] FIG. 4 is a comparative illustration of the nonlinear characteristics of two PAs in different application scenarios: on the left, a narrowband PA with IBW of 100 MHz, one cc (namely NCC=1) , and oBW of 40 MHz; on the right, a wideband PA with IBW of 450 MHz, two cc (namely NCC=2) , and oBW of 200 MHz. From FIG. 4, it is evident that the nonlinear characteristics of the two power amplifiers differ significantly.
[0057] Therefore, to enable the AI-DPD to have broad adaptability across multiple conditional dimensions, it is necessary to include OCI as input neurons, allowing the weights of the connections between these input neurons and the first hidden layer of the neural network (NN) to be incorporated into the training or learning process of the NN. For example, if the AI-DPD needs to generalize and adapt to variations in oBW, then OCI=oBW can be set, meaning OCI is a scalar represented by a positive integer; if the AI-DPD needs to generalize and adapt to four conditions -occupied bandwidth (oBW) , number of component carriers (NCC) , operating temperature (OT) , and operational age (OA) -then OCI= [oBW, NCC, OT, OA] can be set, meaning OCI is a vector composed of non-zero integers.
[0058] FIG. 5 illustrates one implementation structure of Preprocessor 110. It computes the sum of squares of the in-phase component (I) and the quadrature component (Q) for each element of the complex signal vector, as shown below: Or compute their square root, as shown below: Where x (n-j) representing the jth historical transmit signal prior to the current transmit signal at moment n after Digital Up-Conversion (DUC) . When j=0, it indicates the current signal value at moment n. real (·) and imag (·) respectively denote extracting the in-phase component (I) and the quadrature component (Q) from a complex signal. represents the current output value of the jth neuron in the input layer (i.e., layer 0) of the AI-DPD. Through the aforementioned computational processing, the complex signal vector: [x (n-M) , …, x (n-1) , x (n) ] is converted into a real vector:
[0059] In still another configuration specific processing methods of the Preprocessor 110 on the OCI is discussed including Encoding and / or Scaling.
[0060] The process of Encoding as used herein refers to encoding the OCI according to specific rules to reduce the bit-width overhead of the numerical values.
[0061] Taking OCI=oBW as an example, suppose the possible scenarios for occupied bandwidth oBW include: 5, 10, 15, 20, 25, 30, or 35 MHz, seven in total. A binary encoding using a 3-bit width is applied to these scenarios, yielding corresponding encoded values (ev) as indicated below:
[0062] Above, the ev>0 to ensure that, for the first hidden layer of the mentioned AI-DPD, it constitutes a valid input. It will be understood that the OCI may be a scalar or a vector; therefore, ev will also be a scalar or a vector.
[0063] The process of Scaling as used herein refers to multiplying the OCI or its encoded value (ev) after Encoding by a specific scaling factor (sf) , so that the magnitude of the product matches the size of the initialization parameter values of the NN. It will be understood that the OCI may be a scalar or a vector; therefore, the sf will also be a scalar or a vector.
[0064] The two processes described here, Encoding and Scaling, can be applied individually (i.e., utilizing only one) , or they can be applied in sequence with Encoding performed first followed by Scaling.
[0065] A detailed explanation of the specific technical implementations of the Deep Feature Extractor 120, Shallow Feature Extractor 130, Feature Enhancer 140, and Predistortion Filter 150 will now be discussed.
[0066] The Deep Feature Extractor 120 (DFE) employs a deep learning network, leveraging its robust feature learning capability to extract deep features from the input signal that are highly correlated with the nonlinear characteristics of the PA.
[0067] In contrast, the Shallow Feature Extractor 130 (SFE) uses a single-hidden-layer neural network to extract shallow features that contribute to the PA's nonlinear characteristics but with lower computational complexity.
[0068] This combination of outputs from both the DFE 120 and the SFE 130 performed in the Feature Enhancer 140 ensures comprehensive and accurate feature extraction while incorporating the structural advantages of residual networks, including main path and shortcut connection in parallel. Such a structure can effectively mitigate the vanishing or exploding gradient phenomena that commonly occurs during the training of neural networks. In other words, this connection structure not only enhances the accuracy and comprehensiveness of feature extraction, but also improves the stability and efficiency of the NN training process.
[0069] Unlike in classic residual networks where the main path and shortcut outputs are directly added, the feature enhancer performs a self-learning weighted combination of the deep and shallow features. This process strengthens and highlights the key features that characterize the nonlinear characteristics of the PA. The Predistortion Filter 150 then uses these key feature vectors to filter the complex signal vector, thereby achieving pre-correction of the signal distortion caused by the nonlinearities of the PA.
[0070] FIG. 6 illustrates one possible implementation structure of the Deep Feature Extractor 120. It includes at least one real fully connected layer with a real-valued weight parameter matrix R (k) and a real-valued bias parameter vector A (k) . The parameter definitions and signal processing procedures are described as follows: x (k+1) =σ (R (k+1) ·x (k) +A (k+1) ) , (k=0, …, K-2) where, K represents the number of hidden layers of the Deep Feature Extractor 120. Nk is the number of neurons in the kth hidden layer. N0=M+1+m, Here, m is the number of elements in the OCI, 1≤m≤4; NK=M+1. σ (·) is the activation function for the neurons can be set to, for example, but is not limited to, the following functions: where, α can be any real number, for example but not limited to: 2-d (where d is any positive integer) or -1. When α=-1, the Leaky_ReLU (·) function will be equivalent to the absolute value function: abs (·) .
[0071] In the implementation structure of the Deep Feature Extractor 120, after one or more real fully connected layers as described above, there is also a complex fully connected layer cascaded with a complex-valued weight parameter matrix C (K) . The parameter definitions and signal processing procedures are as follows:
[0072] FIG. 7 illustrates one possible implementation structure of the Shallow Feature Extractor 130. It comprises a complex fully connected layer with a complex-valued weight parameter matrix C and a complex-valued bias parameter vector B. The parameter definitions and signal processing procedures are as follows:
[0073] FIG. 8 illustrates one possible implementation structure of the Feature Enhancer 140. It processes the real and imaginary components of the received deep feature vectors and shallow feature vectors separately by applying weighted handling, resulting in the real and imaginary components of the key feature vector, as described below: where, the weighting parameters wu (u=1, …, M+1) are automatically obtained through the neural network training process. Their initial values are random numbers drawn from a uniform distribution in the interval [0, 1] .
[0074] FIG. 9 illustrates one possible implementation structure of the Predistortion Filter 150. It performs an element-wise multiplication of the complex signal vector with the key feature vector, and then separately sums the in-phase and quadrature components of the resulting complex signal vector to obtain the in-phase and quadrature components of the scalar pre-corrected signal, as described below:
[0075] In another configuration, a description of how to train the AI-DPD using both direct and indirect training architectures is discussed. It also details the construction of a loss function based on Mean Squared Error (MSE) and / or Adjacent Channel Leakage Ratio (ACLR) .
[0076] Referring now to FIG. 10, direct learning architecture is illustrated where the flow order of signal X is first through the AI-DPD to obtain signal Y, and then through the PA to obtain signal Z. This is consistent with the actual signal flow order. The specific training process is as follows:
[0077] Step 1. Forward Inference (FI) refers to the process where signal X sequentially passes through the AI-DPD and PA, resulting in the overall output signal Z.
[0078] Step 2. Loss Calculation (LC) refers to the process of using the overall input signal X as the expected output or label and computing the error and corresponding loss function between the actual output Z and this label.
[0079] Step 3. Backward Propagation (BP) refers to the process of calculating the gradients of the loss function with respect to the parameters of each layer of the AI-DPD in a reverse, layer-by-layer manner.
[0080] Step 4. Parameters Update (PU) refers to the process of updating the parameters of each layer of the AI-DPD based on the gradient values calculated in Step 3 above.
[0081] The training steps are carried out sequentially and repeatedly until the performance objectives or other training termination conditions are met.
[0082] In training based on the indirect training architecture in FIG. 10, the process is divided into major iterations and minor iterations. During the first major iteration and each minor iteration, the signal flow order is first through the PA and then through the AI-DPD, which is opposite to the actual signal flow order. The specific training procedure is as follows:
[0083] During the first major iteration, the AI-DPD of the main path processes the signal in a straight-through manner, meaning Y = X. Then, a preset number of minor iterations begins, where signal Y enters the PA and is subsequently fed back through the feedback loop to the AI-DPD of the auxiliary path, resulting in output X_DPD. The error and corresponding loss function are calculated between this output X_DPD and signal Y. Following this, backward propagation (BP) for gradient calculation and PU are performed sequentially, exactly as in Steps 3 and 4 of the direct training procedure described above. This minor iteration loop continues until the preset number of minor iterations is reached, at which point it terminates.
[0084] Following this, the second major iteration begins. At this point, the parameters of the AI-DPD of the main path are copied from the current parameter values of the auxiliary path AI-DPD, and it no longer processes signal X in a straight-through manner, resulting in signal Y. Then, the process follows the same minor iteration procedure as described for the first major iteration, including the handling and processing flow for each minor iteration, until the preset number of minor iterations is reached, at which point it terminates.
[0085] The process for subsequent major iterations is the same as that of the second major iteration, continuing until the preset total number of major iterations is reached, at which point the training ends.
[0086] In the direct training architecture, the loss function is constructed as follows: Loss_direct=α·MSE_direct+β·LACLR_direct+γ·HACLR_direct where, refers to the spectral density estimate of the signal Z; f0 refers to the center frequency of the main band; LACLR and HACLR refer to the leakage ratios of the first adjacent channels to the left (lower frequency) and right (higher frequency) of the main band, respectively. α, β, γ are positive integers that satisfy: α+β+γ=1. When setting: α=1; β=γ=0, it represents the loss function constructed solely based on mean squared error (MSE) ; when setting: α=0; β+γ=1, it represents the loss function constructed solely based on adjacent channel leakage ratio (ACLR) .
[0087] In the indirect training architecture, the loss function is constructed as follows: Loss_indirect=α·MSE_indirect+β·LACLR_indirect+γ·HACLR_indirect where, refers to the spectral density estimate of signal XDPD.
[0088] FIG. 11 illustrates the test performance under following scenario: IBW=280MHz, 2cc, cc1=3870MHz, oBW1=100MHz , cc2=3940MHz, oBW2=100MHz ; Sumitomo SG40N90T-H GaN-HEMT PA. Using the direct training architecture, the AI-DPD was trained with input and output samples of the PA at different power levels. After training, the AI-DPD parameters were fixed, and then entirely new input data at various power levels were used for testing to evaluate the performance metrics. Compared to a traditional DPD, the AI-DPD requires only one set of parameters to accommodate various power levels, greatly reducing the consumption of storage resources.
[0089] FIG. 12 is a structural diagram of a digital predistortion device provided by an embodiment of this application. As shown in FIG. 12, the device provided includes: a processor 1210 and a memory 1220. The number of processors 1210 in the device can be one or more processors, where FIG. 12 illustrates an example with one processor 1210. Similarly, the number of memories 1220 can comprise any number, where FIG. 12 also illustrates an example with one memory 1220. The processor 1210 and memory 1220 within the device can be connected through a bus or other means, where FIG. 12 provides an example where they are connected via a bus.
[0090] The memory 1220, serving as a computer-readable storage medium, can be configured to store software programs, computer-executable programs, and modules, such as the program instructions / modules corresponding to the digital predistortion device described in any embodiment of this application. The memory 1220 can include a program storage area and a data storage area, where the program storage area can store operating systems, application programs required for at least one function, and so on; the data storage area can store data created based on the use of the device. Additionally, the memory 1220 can encompass high-speed random-access memory and may also include non-volatile memory, such as at least one magnetic disk storage device, flash memory device, or other non-volatile solid-state storage devices. In some instances, the memory 1220 can further include remote storage relative to the processor 1210, which can be connected to the device via a network. Examples of such networks include but are not limited to the Internet, intranets, local area networks, mobile communication networks, and combinations thereof.
[0091] The device provided herein can be configured to execute the implementation methods of the digital predistortion system as described in any of the embodiments above, thereby achieving corresponding functionalities and effects.
[0092] In another configuration, a storage medium (e.g., memory 1220) contains computer-executable instructions. When these instructions are executed by a computer processor (e.g., processor 1210) , they perform the implementation method of a digital predistortion system.
[0093] In general, various embodiments of this application can be implemented in hardware or dedicated circuits, software, logic, or any combination thereof. For example, some aspects can be implemented in hardware, while other aspects can be implemented in firmware or software that can be executed by a controller, microprocessor, or other computing device, although the application is not limited to these implementations.
[0094] The embodiments of this application can be realized by the data processor of a mobile device executing computer program instructions, for instance, within a processor entity, either through hardware, software, or a combination of both. The computer program instructions can consist of assembly instructions, instruction set architecture (ISA) instructions, machine instructions, machine-dependent instructions, microcode, firmware instructions, state-setting data, or source code or object code written in one or more programming languages, including an object-oriented programming language and conventional procedural programming languages.
[0095] The block diagrams in the accompanying drawings of this application can represent program steps, or they may represent interconnected logical circuits, modules, and functions, or a combination of program steps with logical circuits, modules, and functions. The computer programs can be stored on memory. The memory can have any type suitable for the local technical environment and can be implemented using any appropriate data storage technology, including but not limited to: Read-Only Memory (ROM) , Random Access Memory (RAM) , optical storage devices and systems (such as Digital Video Disc (DVD) or Compact Disk (CD) ) , etc. The computer-readable medium can include non-transitory storage media. The data processor can be of any type suitable for the local technical environment, including but not limited to: general-purpose computers, special-purpose computers, microprocessors, digital signal processors (DSPs) , application-specific integrated circuits (ASICs) , field-programmable gate arrays (FPGAs) , and processors based on multi-core processor architectures.
[0096] While the present disclosure has been described with reference to one or more exemplary embodiments, it will be understood by those skilled in the art that various changes can be made and equivalents can be substituted for elements thereof without departing from the scope of the present disclosure. In addition, many modifications can be made to adapt a particular situation or material to the teachings of the disclosure without departing from the scope thereof. Therefore, it is intended that the present disclosure should not be limited to the particular embodiment (s) disclosed as the best mode contemplated, but that the disclosure will include all embodiments falling within the scope of the appended claims.
Claims
1.A digital predistortion system, comprising:a preprocessor configured to receive a complex signal vector that includes:historical and current values of a transmit signal after digital up-conversion, andpower amplifier operating condition indictors (OCIs) ;wherein the preprocessor converts the complex signal vector into a real number signal vector having a length equal to the complex signal vector and performing transformation processing on the OCIs;wherein the preprocessor combines the transformed OCIs with the real number signal vector to form the final real number signal vector;a Deep Feature Extractor (DFE) configured to extract features related to nonlinearity characteristics of a power amplifier (PA) from the final real number signal vector;wherein the DFE generates and outputs a deep feature vector having a length equal to the complex signal vector;a Shallow Feature Extractor (SFE) configured to extract features related to the nonlinearity characteristics of the PA from the final real number signal vector;wherein the SFE generates and outputs a shallow feature vector having a length equal to the complex signal vector;a Feature Enhancer (FE) configured to overlay the deep feature vector and the shallow feature vector by performing a weighted summation of the deep feature vector and the shallow feature vector to generate and output a key feature vector having a length equal to the complex signal vector; anda Predistortion Filter (PF) receiving the key feature vector and filtering the complex signal vector by linearization pre-correction of the PA’s nonlinear distortion to generate a scalar pre-correction signal;wherein the PF outputs the scalar pre-correction signal.2.The digital predistortion system according to claim 1, wherein the power amplifier OCIs are selected from the group consisting of: occupied bandwidth, number of component carriers, operating temperature, operational age, and any combinations thereof.3.The digital predistortion system according to claim 1, wherein the processing of the power amplifier OCIs comprises: encoding and / or scaling.4.The digital predistortion system according to claim 1, wherein the Deep Feature Extractor and the Shallow Feature Extractor implement feature extraction based on a deep neural network and a single hidden layer neural network, respectively.5.The digital predistortion system according to claim 1, wherein the Preprocessor calculates the square sum or the square root of each complex signal's in-phase and quadrature components in the complex signal vector element by element to generate the real number signal vector.6.The digital predistortion system according to claim 1, wherein the Deep Feature Extractor comprises at least one real-valued fully connected layer with real weight parameter matrices and real bias vectors cascaded with a complex-valued fully connected layer with complex weight parameter matrices.7.The digital predistortion system according to claim 1, wherein the Shallow Feature Extractor comprises a complex-valued fully connected layer with complex weight parameter matrices and complex bias vectors.8.The digital predistortion system according to claim 1, wherein the Feature Enhancer performs weighted processing on real and imaginary parts of the received deep feature vector and shallow feature vector separately, obtaining real and imaginary parts of the key feature vector.9.The digital predistortion system according to claim 1, wherein the Predistortion Filter performs point-by-point multiplication between the complex signal vector and the key feature vector, and then sums the in-phase and quadrature components of the multiplied complex signal vector separately, yielding the in-phase and quadrature components of the scalar pre-corrected signal.10.The digital predistortion system according to claim 1, wherein the system is trained using either direct training architecture or indirect training architecture, and a loss function used for training is constructed based on mean square error or adjacent channel leakage ratio or combinations thereof.11.The digital predistortion system according to claim 1, further comprising:one or more processors; anda memory accessible by the one or more processors and configured to store one or more programs;wherein when the one or more programs are executed by the one or more processors, the one or more processors to implement the functions of the Preprocessor, the DFE, the SFE, the Feature Enhancer, and the Predistortion Filter.12.The digital predistortion system according to claim 11, wherein the memory has saved thereon only one set of parameters for use by the system, and wherein the single set of parameters is used to accommodate different power levels of the PA.