5g base station phase compensation method and system based on network configuration information

Abstract

The present application relates to the technical field of information transmission, and in particular to a 5G base station phase compensation method and system based on network configuration information, comprising: a base station running a first prediction model and an online learning model in parallel; the first prediction model uses a working temperature sequence and network configuration information to predict a first phase error; the online learning model uses the temperature and network configuration information at the current time to obtain a second phase error; the two kinds of phase errors are fused using a weight coefficient to obtain a phase compensation error and perform phase compensation; the weight coefficient is updated based on the difference between the first phase error and the phase compensation error at a reference time; and the weight coefficient update amplitude is adjusted using an actual phase error determined by an SRS signal. The present application realizes real-time and continuous compensation of the phase error of a 5G base station, and improves the beamforming precision and the stability of the communication system.

Description

Technical Field

[0001] This invention relates to the field of information transmission technology, and specifically to a 5G base station phase compensation method and system based on network configuration information. Background Technology

[0002] In 5G Massive MIMO communication systems, base stations achieve beamforming and multi-user MIMO technology by precisely controlling the phase of the transmitted signals of each channel of the antenna array, thereby accurately directing signal energy to the target user and simultaneously serving multiple users. To ensure beam pointing accuracy and signal orthogonality between users, it is necessary to compensate and calibrate the phase errors between each radio frequency channel in real time. Existing phase compensation techniques mainly adopt periodic offline calibration methods, which require interrupting normal service communication and transmitting a specific calibration reference signal during a dedicated quiet time slot, and calculating the phase error by measuring the response of each channel at the receiver. However, this method directly leads to a decline in user experience and network capacity loss, consuming valuable time and frequency resources and reducing spectral efficiency. In addition, phase errors drift in real time with temperature changes, device aging, and dynamic adjustments to network configuration. Existing calibration methods cannot be performed frequently due to high resource overhead, and compensation can usually only be performed at startup or periodically, making it difficult to achieve continuous, real-time tracking compensation. This leads to a gradual deterioration of system phase consistency during the compensation period, a decrease in beam performance, an increase in interference between users, and a serious impact on the capacity and stability of the communication system. Summary of the Invention

[0003] To address the aforementioned issues, this invention provides a 5G base station phase compensation method and system based on network configuration information.

[0004] The 5G base station phase compensation method and system based on network configuration information of the present invention adopts the following technical solution: One embodiment of the present invention provides a 5G base station phase compensation method based on network configuration information, the method comprising the following steps: The base station runs a first prediction model and an online learning model in parallel. The first prediction model uses the base station's operating temperature sequence before any time and the network configuration information at any time to predict the first phase error at several times after any time, wherein the next time after any time is marked as the reference time. The online learning model uses the temperature and network configuration information at any time to obtain the second phase error at any time. The first prediction model is delayed after the online learning model. The second phase error at the current moment and the most recently obtained first phase error at the current moment are fused using weighting coefficients to obtain the phase compensation error at the current moment; phase compensation is then performed using the phase compensation error at the current moment. After the current time, when the current time is marked as a reference time, the weighting coefficient is updated based on the difference between the most recently obtained first phase error at the current time and the phase compensation error at the current time. The initial value of the weighting coefficient is a preset value. The phase error determined by the base station through the transmission of SRS signals at the agreed time is recorded as the actual phase error. The difference between the actual phase error and the phase compensation error at the agreed time is used to adjust the update amplitude when updating the weight coefficients.

[0005] Preferably, the phase compensation error at the current moment is positively correlated with the most recently obtained first phase error at the current moment, the second phase error at the current moment, and the weighting coefficient, respectively.

[0006] Preferably, the specific steps for updating the weighting coefficients based on the difference between the most recently obtained first phase error at the current time and the phase compensation error at the current time are as follows: The difference between the first phase error most recently obtained at the current time when the current time is marked as the reference time and the phase compensation error at the current time is denoted as the first difference; The difference between the value of the weight coefficient before the update and the initial value of the weight coefficient is recorded as the basic update amount. The basic update amount is corrected using the first difference to obtain the target update amount. The target update amount is summed with the initial value of the weight coefficient to obtain the updated weight coefficient. The target update amount is positively correlated with the first difference.

[0007] Preferably, the target update amount R = R0 × (1 + r), where R0 represents the basic update amount, r represents the update coefficient, the update coefficient is equal to the product of the update amplitude and the direction coefficient, where the direction coefficient is positively correlated with the first difference, and the initial value of the update amplitude is a preset value.

[0008] Preferably, the specific steps for adjusting the update magnitude of the weight coefficient update by utilizing the difference between the actual phase error and the phase compensation error at the agreed time are as follows: The absolute value of the difference between the actual phase error and the phase compensation error at the agreed time is calculated, and the adjusted update amplitude is positively correlated with the absolute value of the difference; the agreed time refers to the time point after each fixed period.

[0009] Preferably, the specific steps for performing phase compensation using the phase compensation error at the current moment are as follows: The phase compensation error ΔΦ is converted into a complex compensation factor: exp(-j×ΔΦ), where exp() represents an exponential function with the natural constant as the base and j represents the imaginary unit; the compensation factor is multiplied onto the precoding weight of each channel to complete the phase compensation.

[0010] Preferably, the network configuration information includes downtilt angle, azimuth angle, carrier frequency, and bandwidth. Preferably, the specific formula for the phase compensation error at the current moment is as follows: The phase compensation error at the current time t0 is E(t0) = w1 × E1(t0) + w2 × E2(t0), where w1 represents the weighting coefficient, w2 represents the preset reference coefficient, E1(t0) represents the first phase error most recently obtained at the current time t0, and E2(t0) represents the second phase error at the current time t0.

[0011] Preferably, the online learning model adopts a recursive least squares model.

[0012] Another embodiment of the present invention provides a 5G base station phase compensation system based on network configuration information. The system includes a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor runs the computer program, it implements all the steps of the 5G base station phase compensation method based on network configuration information described above.

[0013] The beneficial effects of the technical solution of the present invention are: This invention achieves real-time and continuous compensation for phase errors in 5G base stations by running a first prediction model and an online learning model in parallel and fusing the phase error prediction results of the two models using weight coefficients. This effectively solves the service interruption problem caused by periodic offline calibration in existing technologies. The first prediction model predicts the phase error at several future moments based on the operating temperature sequence and network configuration information, while the online learning model obtains a real-time phase error estimate based on the current temperature and network configuration information. The parallel operation of the two models balances prediction accuracy and real-time requirements. The phase compensation error is obtained by fusing the two phase error prediction results through weight coefficients, which utilizes the high-precision prediction potential of the first prediction model and combines the real-time output capability of the online learning model, achieving continuity and accuracy in phase compensation. An adaptive optimization mechanism is established by updating the weight coefficients based on the difference at the reference time, enabling the fusion strategy to be dynamically adjusted according to high-confidence supervision signals, thus improving long-term compensation accuracy. The difference between the actual phase error determined by the SRS signal and the phase compensation error is used to adjust the update amplitude of the weight coefficients, introducing a second layer of closed-loop supervision to ensure the stability and convergence efficiency of the entire adaptive optimization process. While ensuring phase compensation accuracy, this invention achieves real-time tracking and compensation for dynamic phase errors, thereby improving beamforming accuracy and communication system stability. Attached Figure Description

[0014] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0015] Figure 1 This is a flowchart illustrating the steps of a 5G base station phase compensation method based on network configuration information, provided in one embodiment of the present invention. Detailed Implementation

[0016] To further illustrate the technical means and effects adopted by the present invention to achieve its intended purpose, the following, in conjunction with the accompanying drawings and preferred embodiments, details the specific implementation, structure, features, and effects of the 5G base station phase compensation method and system based on network configuration information proposed in this invention. In the following description, different "one embodiment" or "another embodiment" do not necessarily refer to the same embodiment. Furthermore, specific features, structures, or characteristics in one or more embodiments can be combined in any suitable form.

[0017] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this invention pertains.

[0018] The following description, in conjunction with the accompanying drawings, details the specific scheme of the 5G base station phase compensation method and system based on network configuration information provided by this invention.

[0019] Please see Figure 1 The diagram illustrates a flowchart of a 5G base station phase compensation method based on network configuration information according to an embodiment of the present invention. The method includes the following steps: Step S101: The base station runs the first prediction model and the online learning model in parallel; the first prediction model uses the base station's operating temperature sequence before any time and the network configuration information at any time to predict the first phase error at several times after any time, wherein the next time after any time is marked as the reference time; the online learning model uses the temperature and network configuration information at any time to obtain the second phase error at any time; the first prediction model is delayed after the online learning model.

[0020] The online learning model boasts high computation speed and real-time output, but its simplicity, relying solely on single-moment information, makes it difficult to accurately depict the complex relationship between phase error and historical temperature changes, resulting in insufficient accuracy. The first prediction model, while structurally complex and capable of handling working temperature sequences with high potential prediction accuracy, suffers from significant computational latency and poor real-time performance. Parallel execution allows this embodiment to leverage the advantages of both models simultaneously, providing a technical foundation for fusing the two prediction results in subsequent steps to balance real-time performance and accuracy.

[0021] This embodiment further describes the input data of the first prediction model. The operating temperature sequence, as time-series data, reflects the dynamic accumulation process of phase error caused by factors such as temperature drift and device aging in base station hardware (e.g., power amplifier). The network configuration information (such as downtilt angle and carrier frequency) directly determines the external commands for the beamforming network phase state. Compared to inputting only single-moment data, inputting an operating temperature sequence containing historical information allows the model to learn and predict the dynamic law and long-term trend of phase error changes with temperature. This provides richer input information for solving the technical problem of dynamic phase error tracking and lays the foundation for obtaining more accurate long-term prediction results (first phase error).

[0022] This embodiment further describes the output data of the first prediction model. The output of the first phase error at several future moments represents an estimate of the channel phase deviation over a future period. This output characteristic enables the model to have forward-looking predictive capabilities; even if the first prediction model itself has computational delays, its output predictions can still cover multiple future compensation moments. This provides data reserves to address the technical problem that traditional periodic calibration cannot achieve continuous, real-time compensation, enabling this embodiment to provide relatively accurate phase error references for multiple future moments after the first prediction model outputs its results.

[0023] The reference time is the next time step corresponding to the operating temperature sequence input into the first prediction model. Since the first phase error at this time is predicted based on its own previous (no delay) operating temperature sequence, its prediction result is theoretically closer to the true phase error and has higher reliability. Introducing the reference time provides a key time anchor and judgment basis for the subsequent monitoring signal weight coefficient update mechanism.

[0024] It should be further clarified that reference time is not only a time in the future that can be marked as a reference time. Rather, a reference time is only marked when the following condition is met: the first phase error obtained at that time is obtained by processing the working temperature sequence before the previous time through the first prediction model.

[0025] For example, if at any time t1, the previous time is t2, and the working temperature sequence within a preset time period before t2 (including t2) is denoted as T(t2), if the first phase error at time t1 (that is, the next time after t2) is obtained by inputting T(t2) and the network configuration information C(t2) of t2 into the first prediction model, then time t1 (that is, the next time after t2) is marked as the reference time.

[0026] Furthermore, this embodiment describes the input data of the online learning model. Using only temperature and network configuration information at a single moment (e.g., the current moment), it captures the instantaneous state of the phase error under the current environment and configuration. This simplified input method significantly reduces the computational complexity of the model, enabling it to achieve extremely low processing latency. Compared to the first prediction model, the online learning model sacrifices some completeness of input information in exchange for real-time output capability, ensuring a usable phase error estimate (second phase error) is available at any given time, thus providing the possibility for real-time compensation.

[0027] The second phase error is a rapid estimate of the phase deviation at a single moment based on a simplified model. Although its absolute accuracy may be affected by model simplification, its output is highly timely. This output serves as an immediate data source for real-time compensation, resolving the problem of being unable to perform any phase compensation during the delay period caused by the delay of the first prediction model, thus ensuring the continuity of phase compensation operations.

[0028] The delay in the first prediction model refers to the computational overhead caused by its complexity and the need to process historical sequence data. This embodiment does not attempt to eliminate this delay, but rather adapts to and utilizes this characteristic through a parallel architecture and multi-step prediction.

[0029] Step S102: Use weighting coefficients to fuse the second phase error at the current time and the most recently obtained first phase error at the current time to obtain the phase compensation error at the current time; use the phase compensation error at the current time to perform phase compensation.

[0030] Assuming the current time is time t0 (the following analysis and description will all take time t0 as the anchor point), after the working temperature sequence of the preset time period (including time t0) before time t0 and the network configuration information at time t0 are input into the first prediction model, due to the delay characteristics of the first prediction model, it will not immediately output the first phase error at several times after time t0 in a short period of time.

[0031] It is easy to see that the first phase error at the current moment (i.e., time t0) is obtained based on the operating temperature sequence and network configuration information at certain moments before time t0. For example, the first phase error at time t0 can be obtained based on the operating temperature sequence for a preset time period before time t0-6, or it can also be obtained based on the operating temperature sequence for a preset time period before time t0-5. Here, t0-5 and t0-6 represent the 5th and 6th moments before time t0.

[0032] Therefore, the current moment (i.e., time t0) may correspond to multiple first phase errors. This embodiment only considers the most recent or latest first phase error obtained at the current moment (i.e., time t0) (that is, the first phase error obtained later at the same moment covers the first phase error obtained earlier). For example, the most recent first phase error is obtained based on the working temperature sequence of a preset time period before time t0-5. Then it can be seen that the first phase error at time t0 uses the working temperature sequence of the preset time period before t0-5. Therefore, it can be said that the most recent first phase error obtained at the current moment has an information delay.

[0033] The second phase error at the current moment has high real-time performance but limited accuracy, while the most recently obtained first phase error (predicted based on an earlier operating temperature sequence) has higher accuracy potential but suffers from information delay. Directly using either result is insufficient. The fusion operation of weighting coefficients aims to combine the effective information of the two predictions to obtain a phase error estimate that strikes a balance between real-time performance (relying on the online learning model) and accuracy (incorporating information from the first prediction model). This partially solves the problem in step S101 of how to provide a timely and relatively accurate error estimate for the current moment due to the respective shortcomings of the two models.

[0034] The weighting coefficients adjust the relative proportions of the real-time contribution (second phase error) and the historical accuracy contribution (first phase error) in the final compensation error. The value of the weighting coefficients determines the bias of the fusion result.

[0035] The fused output is the phase compensation error used for actual compensation. It represents the best estimate of the antenna channel phase deviation at the current moment, given by comprehensive evaluation in this embodiment. As a direct input to the compensation operation, its accuracy directly determines the beamforming accuracy and the effectiveness of inter-user interference suppression.

[0036] Using the calculated phase compensation error to pre-correct the phase of the transmitted signal in the radio frequency channel can cancel or reduce the actual phase error between channels, thereby ensuring accurate beam pointing and orthogonality of multi-user signals. This avoids problems such as beam performance degradation and capacity loss caused by phase error and is the final step in improving the performance of 5G Massive MIMO systems.

[0037] Step S103: After the current time, when the current time is marked as a reference time, update the weighting coefficients based on the difference between the most recently obtained first phase error at the current time and the phase compensation error at the current time.

[0038] It should be noted that as time progresses, the timeline passes through the aforementioned time t0. When it reaches a certain moment after time t0, the aforementioned time t0 is marked as the reference time. At this point, this step describes a specific weight coefficient update trigger condition. According to step S101, the first phase error corresponding to the reference time is predicted based on the no-delay operating temperature sequence, thus having higher reliability and serving as a reliable monitoring signal. The technical effect of this trigger condition is that the weight coefficient update is only initiated when a high-reliability monitoring signal can be obtained, thereby ensuring the effectiveness and relevance of the update process. After the weight coefficients are updated, the updated weight coefficients are used when the fusion operation is performed again at subsequent moments (specifically referring to the aforementioned current time, i.e., moments after time t0). This step solves the problem of when to optimize and adjust the weight coefficients, avoiding blind updates when the prediction results are unreliable.

[0039] The difference reflects the deviation between the high-confidence predicted value and the historical fusion output value. This difference quantifies the gap between the compensation error caused by the historical fusion weights at the reference time and the more accurate predicted value. The larger the difference, the greater the error of the historical fusion strategy; the larger the difference, the smaller the error of the historical fusion strategy. Based on this difference, the weight coefficients are updated to adjust in a direction that can reduce the fusion error at similar times in the future (reference time). This realizes a closed-loop feedback mechanism, which continuously optimizes the fusion strategy and adaptively improves the long-term accuracy of the fusion result in step S102.

[0040] Step S104: The phase error determined by the base station by sending SRS signals at the agreed time is recorded as the actual phase error. The difference between the actual phase error and the phase compensation error at the agreed time is used to adjust the update amplitude when updating the weight coefficient.

[0041] This step introduces the most reliable external supervisory data source from physical layer measurements. The actual phase error determined by the SRS signal represents the near-true channel phase error obtained by the base station through actual wireless measurements at a specific moment. Although this method consumes a small amount of service resources and is discontinuous, its results have the highest reliability. Using this as the absolute benchmark for calibrating the entire prediction and fusion process provides the highest level of accuracy anchor for this embodiment.

[0042] Specifically, the difference between the actual phase error and the phase compensation error describes the final output error of the entire prediction fusion process (including dual-model prediction and weighted fusion). It comprehensively reflects the overall deviation under the combined effect of multiple factors such as model prediction accuracy and the effectiveness of fusion weights.

[0043] Furthermore, based on this difference index with the highest confidence level, the intensity (update magnitude) of the weight coefficient update process in step S103 is adjusted. For example, when the actual difference is large, it indicates a serious deviation, requiring a larger (more aggressive) adjustment of the weight coefficients to achieve rapid convergence; when the actual difference is small, it indicates good performance, requiring a smaller (more conservative) fine-tuning of the weight coefficients to avoid oscillations. The technical effect of this adjustment operation is to introduce a second layer of closed-loop supervision, supervising the first-layer update mechanism, thereby solving the problem of over-update or under-update that may occur in the internal update of step S103, ensuring that the entire adaptive optimization process smoothly and efficiently approaches the global optimum.

[0044] In summary, all the above steps, using time t0 (i.e., the current time mentioned above) as the anchor point, describe a method for fusing the first prediction model and the online learning model into a phase compensation error based on weight coefficients and performing phase compensation. The method for updating the weight coefficients under specific triggering conditions is also described. At each time point after time t0, the updated weight coefficients are used to fuse the phase compensation error and perform phase compensation, with continuous triggering of weight coefficient updates. This embodiment, overall, achieves real-time tracking and compensation of dynamic phase errors while ensuring phase compensation accuracy.

[0045] As an example, the first prediction model uses the base station's operating temperature sequence before any time and the network configuration information at any time to predict the first phase error at several times after any time. The specific methods include: In this embodiment, each moment is defined as 1 millisecond (one subframe length). In other examples, each moment can be defined as 5 milliseconds (5 subframe lengths) to reduce computational burden.

[0046] For any time t, obtain the operating temperature sequence T(t) within a preset time period before time t, and simultaneously obtain the network configuration information C(t) at time t. The network configuration information includes parameters such as downtilt angle, azimuth angle, carrier frequency, and bandwidth. Concatenate the operating temperature sequence T(t) and the network configuration information C(t) as the input vector, and input it into the trained first prediction model. The first prediction model outputs the first phase error sequence E=[E1(t+1), E1(t+2), ..., E1(t+M)] for the next M times, where E1(t+M) represents the predicted first phase error value at time t+M (i.e., the M times after time t).

[0047] In some embodiments, all temperature data and network configuration information refer to data after removing dimensions and reducing orders of magnitude. As an example, methods for removing dimensions and orders of magnitude include: dividing the acquired temperature by 27°C, dividing the downtilt angle and azimuth angle by 0.5π, dividing the carrier frequency by 3GHz (taking FR1 as an example), and dividing the bandwidth by 25MHz, thereby achieving dimension removal and order-of-magnitude reduction. In other examples, network configuration information may also include other data, such as the lifespan of the RF module; this embodiment will not elaborate on network configuration information. The method for removing dimensions and reducing orders of magnitude for lifespan is: dividing the lifespan by 3 years (or the maximum lifespan of the RF module).

[0048] It should be noted that the above embodiment mainly focuses on removing dimensions and reducing the order of magnitude, but it is not limited to reducing the order of magnitude to the range [0, 1]. Other methods can be used to remove dimensions and reduce the order of magnitude in other examples, and this example does not impose any limitations.

[0049] In this embodiment, a temperature sensor (such as an NTC thermistor sensor) is integrated into the radio frequency channel to collect temperature data at a frequency of 10Hz. The preset time period refers to 2 seconds before time t (inclusive). All temperatures collected within the preset time period constitute the operating temperature sequence. Furthermore, this embodiment uses M=10 as an example for description.

[0050] As an example, the training method for the first prediction model is as follows: A large number of radio frequency (RF) devices (or 5G base stations) are collected. Each RF device is tested in an anechoic chamber. The test includes: operating the RF device under simulated ambient temperature and different preset network configuration information; having the RF device emit a test signal; and recording the phase difference between the receiver signal and the test signal as the measurement phase error. During the test, the operating temperature sequence within a preset time period before each moment and the network configuration information at each moment are used as samples, and the measurement phase error at several moments after each moment is used as a label. Through multiple tests on a large number of radio frequency devices, all the samples and labels obtained were used as a dataset. The first prediction model was trained using the dataset. The mean squared error loss function was used during training, and the ratio of the training set to the test set was 7:3.

[0051] As an example, the first prediction model uses a Long Short-Term Memory (LSTM) network, with Adam as the optimizer during training, a learning rate of 0.02, 6 training batches, and a maximum training iteration count of 10. 4 .

[0052] It should be noted that in this embodiment, the operating temperature sequence and network configuration information corresponding to a certain moment are input into the trained first prediction model. After the first prediction model outputs the first phase error at several moments, the operating temperature sequence and network configuration information corresponding to the next moment are input into the trained first prediction model. After the first prediction model outputs the first phase error at several moments, the operating temperature sequence and network configuration information corresponding to the second moment after that moment are input into the trained first prediction model, and so on, so that the first prediction model continuously outputs the first phase error.

[0053] It should be further explained that, due to the lag characteristic of the first prediction model, if the n0th first phase error output by the first prediction model occurs at or before the current time, it indicates that the lag characteristic of the first prediction model may cause the first prediction model to fail to reliably cover the current time for several time points. Therefore, the operating temperature sequence and network configuration information corresponding to the current time are directly input into the trained first prediction model to ensure that the first phase errors output by the first prediction model for several time points can continuously cover the current time. The preferred value range for n0 is [M / 2, M]. This example uses n0=7 as an example.

[0054] In specific cases, such as when the delay characteristic of the first prediction model is too large or the computing power is insufficient, there may be situations where the output of the first prediction model cannot cover the current time at certain moments (i.e., no first phase error can be obtained at certain moments). In this case, the phase compensation error is directly set to equal the second phase error. In this case, this embodiment degenerates into using only the online learning model for phase compensation.

[0055] As an example, the online learning model employs a recursive least squares (RLS) model. The input to this model is the temperature and network configuration information at each time step, and the output is the second phase error at each time step. Its working principle is as follows: The least squares RLS model is initialized, and then the second phase error is output in real time using this model. Whenever a predetermined time step arrives, the model's parameters are updated using the actual phase error at that predetermined time step.

[0056] Since the Recursive Least Squares (RLS) model is a well-known technique, the method for updating its parameters is also well-known, and will not be described in detail in this embodiment. The initialization of the RLS model can utilize the dataset obtained above, and the process is also well-known.

[0057] As an example, starting from base station startup, at fixed intervals (e.g., every 0.5 hours) designated as agreed-upon times, the base station schedules one or more User Equipments (UEs) to transmit Sounding Reference Signals (SRS) on their uplink. SRS is a predefined signal in the 5G standard, the contents of which are fully known to the base station. The base station's antenna array receives this uplink SRS signal through each radio frequency channel. The receiver in each channel performs down-conversion, analog-to-digital conversion, and other processing on the signal to obtain the baseband digital signal. For the SRS signal received on each channel, the base station compares it with the known original SRS sequence using standard channel estimation algorithms such as correlation or least squares to obtain the phase offset of each channel, i.e., the actual phase error.

[0058] This process is well-known and will not be described in detail in this embodiment.

[0059] As a preferred example, the second phase error at the current moment and the most recently obtained first phase error at the current moment are fused using weighting coefficients to obtain the phase compensation error at the current moment, including the following formula: The phase compensation error at the current moment, i.e., time t0, is E(t0) = w1 × E1(t0) + w2 × E2(t0). Here, E1(t0) represents the most recently obtained first phase error at the current moment, E2(t0) represents the second phase error at the current moment, and the weighting coefficient w1 is used to adjust the relative proportions of the first and second phase errors in the fusion result. w2 is a preset reference coefficient, which is a constant. As an example, w2 = 0.4, and the initial value of the weighting coefficient w1 is set to 0.7. It should be noted that since the actual phase error can be larger or smaller than E1(t0) or E2(t0), and may also fall between E1(t0) and E2(t0), this embodiment does not require the sum of w1 and w2 to be equal to 1. This ensures that the phase compensation error will not be limited to the first and second phase errors during subsequent supervised update adjustments, allowing the phase compensation error sufficient room for value exploration.

[0060] As a preferred example, phase compensation is performed using the phase compensation error at the current moment, including: The calculated phase compensation error ΔΦ (in radians) is converted into a complex compensation factor: exp(-j×ΔΦ), where exp() represents an exponential function with the natural constant as the base, and j represents the imaginary unit. This compensation factor is then multiplied by the original precoding weights of each channel to complete the phase compensation. The specific process is well-known and will not be elaborated upon in this example.

[0061] As a preferred example, updating the weighting coefficients based on the difference between the most recently obtained first phase error at the current time and the phase compensation error at the current time includes the following method: The difference between the first phase error F1 most recently obtained at the current time (i.e., time t0) and the phase compensation error F2 at the current time (i.e., time t0) is denoted as the first difference f; Note that F1 at this point is not the same value as the first phase error mentioned in step S102. This is because the first phase error mentioned in step S102 refers to the most recent first phase error obtained at time t0 (i.e., the current time mentioned in step S102) when the timeline reaches time t0 as time progresses. In this example, F1 refers to the most recent first phase error obtained at time t0 (i.e., the current time mentioned in step S102) when the timeline reaches time t0 after t0 (let's say time t1), and time t0 is precisely marked as the reference time.

[0062] In summary, the first phase error mentioned in step S102 refers to the first phase error corresponding to time t0 when the timeline reaches t0, while F1 refers to the first phase error corresponding to time t0 when the timeline reaches t1. Due to the time difference effect caused by the delay characteristics of the first prediction model, F1 and the first phase error mentioned in step S102 are not the same value.

[0063] The difference between the weight coefficient's value before the update and its initial value is recorded as the basic update amount, representing the change in the weight coefficient before the update. The basic update amount is corrected using the first difference f to obtain the target update amount. The target update amount is summed with the initial value of the weight coefficient to obtain the updated weight coefficient. The target update amount is positively correlated with the first difference f. A larger first difference f indicates that the phase compensation error F2 obtained at the current time (i.e., time t0) is less likely to approach the more accurate and larger value of F1. In this case, a larger target update amount is needed so that when the updated weight coefficients are fused again, the fused phase compensation error can change towards a larger value. Conversely, a smaller first difference f indicates that the phase compensation error F2 obtained at the current time (i.e., time t0) is less likely to approach the more accurate and smaller value of F1. In this case, a smaller target update amount is needed so that when the updated weight coefficients are fused again, the fused phase compensation error can change towards a smaller value.

[0064] As an example, the target update amount is obtained by correcting the base update amount using the first difference f, including the following methods: The target update amount R = R0 × (1 + r), where R0 represents the base update amount and r represents the update coefficient. The update coefficient r consists of two parts: the update amplitude and the direction coefficient. The update coefficient r is equal to the product of the update amplitude and the direction coefficient. The direction coefficient is positively correlated with the first difference f, and its main function is to control whether the weight coefficient is adjusted to be larger or smaller. The update amplitude is mainly used to control the magnitude of the weight coefficient update.

[0065] In one example, the direction coefficient x = f / (|F1| + 0.05π). It should be noted that in this embodiment, the phase, i.e., all phase error quantities involved, are expressed in radians. Adding 0.05π to the denominator is to avoid the denominator being zero. Where x is less than 0, it indicates that the control weight coefficient is adjusted to a smaller value; x is greater than 0, it indicates that the control weight coefficient is adjusted to a larger value. The initial value of the update amplitude is set to 1.0.

[0066] Specifically, when the target update amount is less than -0.4 or greater than 0.7, the target update amount is set to -0.4 or 0.7 to avoid significant adjustments to the weighting coefficients.

[0067] As a preferred example, the update magnitude of the weight coefficients is adjusted by utilizing the difference between the actual phase error and the phase compensation error at the agreed time. The steps include: Calculate the absolute value of the difference between the actual phase error and the phase compensation error at the agreed time. The update amplitude is positively correlated with the absolute value of the difference. The larger the absolute value of the difference, the more serious the deviation, and the more aggressively the weight coefficient needs to be adjusted to achieve rapid convergence. The smaller the absolute value of the difference, the better the work is, and the weight coefficient should be fine-tuned with a smaller amplitude to avoid oscillation.

[0068] As an example, the formula for calculating the update magnitude includes: The adjusted update amplitude H = H0 × (1 + q / (|q0| + 0.05π)), where H0 represents the initial value of the update amplitude, q0 represents the actual phase error, and q represents the absolute value of the difference. The reason for using (q0 + 0.05π) as the denominator is to avoid the denominator being equal to 0.

[0069] Another embodiment of the present invention provides a 5G base station phase compensation system based on network configuration information. The system includes several (e.g., 32) radio frequency channels (32T32R) in the 5G base station, and a compensation module for performing phase compensation for all radio frequency channels. The compensation module includes a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor runs the computer program, it implements all the steps of all the above embodiments. In addition, the module is equipped with different GPU devices, which are used to run the first prediction model and the online learning model corresponding to each radio frequency channel, respectively. The first prediction model and the online learning model corresponding to the same radio frequency channel are respectively in independent running resources (independent processes).

[0070] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A 5G base station phase compensation method based on network configuration information, characterized in that, The method includes the following steps: The base station runs a first prediction model and an online learning model in parallel. The first prediction model uses the base station's operating temperature sequence before any time and the network configuration information at any time to predict the first phase error at several times after any time, wherein the next time after any time is marked as the reference time. The online learning model uses the temperature and network configuration information at any time to obtain the second phase error at any time. The first prediction model is delayed after the online learning model. The second phase error at the current moment and the most recently obtained first phase error at the current moment are fused using weighting coefficients to obtain the phase compensation error at the current moment; phase compensation is then performed using the phase compensation error at the current moment. After the current time, when the current time is marked as a reference time, the weighting coefficient is updated based on the difference between the most recently obtained first phase error at the current time and the phase compensation error at the current time. The initial value of the weighting coefficient is a preset value. The phase error determined by the base station through the transmission of SRS signals at the agreed time is recorded as the actual phase error. The difference between the actual phase error and the phase compensation error at the agreed time is used to adjust the update amplitude when updating the weight coefficients.

2. The 5G base station phase compensation method based on network configuration information according to claim 1, characterized in that, The phase compensation error at the current moment is positively correlated with the most recently obtained first phase error at the current moment, the second phase error at the current moment, and the weighting coefficient.

3. The 5G base station phase compensation method based on network configuration information according to claim 1, characterized in that, The specific steps for updating the weighting coefficients based on the difference between the most recently obtained first phase error at the current time and the phase compensation error at the current time are as follows: The difference between the first phase error most recently obtained at the current time when the current time is marked as the reference time and the phase compensation error at the current time is denoted as the first difference; The difference between the weight coefficient before the update and the initial value of the weight coefficient is recorded as the basic update amount. The basic update amount is corrected using the first difference to obtain the target update amount. The target update amount is summed with the initial value of the weight coefficient to obtain the updated weight coefficient. The target update volume is positively correlated with the first difference.

4. The 5G base station phase compensation method based on network configuration information according to claim 3, characterized in that, The target update amount R = R0 × (1 + r), where R0 represents the basic update amount, r represents the update coefficient, the update coefficient is equal to the product of the update amplitude and the direction coefficient, the direction coefficient is positively correlated with the first difference, and the initial value of the update amplitude is a preset value.

5. The 5G base station phase compensation method based on network configuration information according to claim 4, characterized in that, The method of adjusting the update amplitude of the weight coefficient update by utilizing the difference between the actual phase error and the phase compensation error at the agreed time includes the following specific steps: calculating the absolute value of the difference between the actual phase error and the phase compensation error at the agreed time, and the adjusted update amplitude is positively correlated with the absolute value of the difference; the agreed time refers to the time point after each fixed period.

6. The 5G base station phase compensation method based on network configuration information according to claim 1, characterized in that, The specific steps for performing phase compensation using the phase compensation error at the current moment are as follows: The phase compensation error ΔΦ is converted into a complex compensation factor: exp(-j×ΔΦ), where exp() represents an exponential function with the natural constant as the base and j represents the imaginary unit; the compensation factor is multiplied onto the precoding weight of each channel to complete the phase compensation.

7. The 5G base station phase compensation method based on network configuration information according to claim 1, characterized in that, The network configuration information includes downtilt angle, azimuth angle, carrier frequency, and bandwidth.

8. The 5G base station phase compensation method based on network configuration information according to claim 2, characterized in that, The specific formula for the phase compensation error at the current moment is as follows: The phase compensation error at the current time t0 is E(t0) = w1 × E1(t0) + w2 × E2(t0), where w1 represents the weighting coefficient, w2 represents the preset reference coefficient, E1(t0) represents the first phase error most recently obtained at the current time t0, and E2(t0) represents the second phase error at the current time t0.

9. The 5G base station phase compensation method based on network configuration information according to claim 1, characterized in that, The online learning model adopts a recursive least squares model.

10. A 5G base station phase compensation system based on network configuration information, the system comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor runs the computer program, it implements all the steps of the 5G base station phase compensation method based on network configuration information as described in any one of claims 1 to 9.

Images