Radio measurement calibration method and system based on machine learning

Abstract

The invention relates to the technical field of machine learning, and discloses a radio measurement calibration method and system based on machine learning, and the method comprises the steps: carrying out the parameter measurement of a radio signal source, and obtaining an original measurement data set; based on the original measurement data set, extracting a frequency precision feature, a power flatness feature, a phase noise feature, a harmonic ratio feature and an environment change feature to obtain a feature data set; dividing the feature data set into a training data subset, a verification data subset and a test data subset, and constructing a first neural network model containing a double-branch neural network based on the training data subset; and inputting the verification data subset into the first neural network model for model parameter optimization to obtain a second neural network model, thereby improving the calculation precision of calibration parameters and the accuracy of environment compensation.

Description

Technical Field

[0001] The present application relates to the field of machine learning technology, and in particular to a radio metrology calibration method and system based on machine learning. Background Art

[0002] With the rapid development of radio measurement technology, the calibration accuracy and stability requirements of signal sources as key measurement equipment are constantly increasing. Traditional signal source calibration methods mainly rely on manual experience and fixed calibration parameters, which are difficult to deal with measurement errors and drift problems under complex environmental conditions. Especially in high-frequency band measurements, factors such as ambient temperature and humidity have a more significant impact on the measurement results, and traditional calibration methods cannot effectively compensate for the impact of these environmental factors.

[0003] Existing calibration technologies generally have problems such as long calibration cycles, frequent manual intervention, and low calibration efficiency. At the same time, due to the lack of in-depth mining and utilization of historical calibration data, it is impossible to establish a mapping relationship between environmental factors and measurement errors, which makes it difficult to ensure the reliability and stability of the calibration results. In addition, existing calibration methods lack adaptive optimization capabilities and cannot dynamically adjust calibration parameters according to equipment usage status and environmental changes. In large-scale measurement application scenarios, the performance of the signal source will drift over time, requiring frequent calibration and maintenance. However, traditional calibration methods cannot achieve online real-time calibration, and lack the ability to predict and analyze long-term performance change trends of equipment, which seriously restricts the use efficiency of measurement equipment and the reliability of measurement results. Summary of the invention

[0004] The present application provides a radio metrology calibration method and system based on machine learning, thereby improving the calculation accuracy of calibration parameters and the accuracy of environmental compensation.

[0005] The first aspect of the present application provides a radio metrology calibration method based on machine learning, and the radio metrology calibration method based on machine learning includes: Perform parameter measurements on the radio signal source to obtain an original measurement data set; Based on the original measurement data set, frequency accuracy features, power flatness features, phase noise features, harmonic ratio features and environmental change features are extracted to obtain a feature data set; Dividing the feature data set into a training data subset, a validation data subset, and a test data subset, and constructing a first neural network model including a two-branch neural network based on the training data subset; The verification data subset is input into the first neural network model to optimize model parameters to obtain a second neural network model.

[0006] A second aspect of the present application provides a radio metrology calibration system based on machine learning, the radio metrology calibration system based on machine learning comprising: A measurement module, used to measure parameters of a radio signal source to obtain an original measurement data set; An extraction module, used to extract frequency accuracy features, power flatness features, phase noise features, harmonic ratio features and environmental change features based on the original measurement data set to obtain a feature data set; A construction module, used for dividing the feature data set into a training data subset, a verification data subset and a test data subset, and constructing a first neural network model including a two-branch neural network based on the training data subset; The optimization module is used to input the verification data subset into the first neural network model to optimize the model parameters to obtain a second neural network model.

[0007] Compared with the prior art, the present application has the following beneficial effects: by constructing a dual-branch neural network structure, parameter calibration and environmental compensation are modeled separately, which effectively reduces the complexity of the model, and at the same time improves the calculation accuracy of the calibration parameters and the accuracy of environmental compensation. The network structure design using the self-attention mechanism and residual connection enhances the model's feature extraction ability for measurement data at different frequency points and different power points, and at the same time improves the convergence performance of the model. The LSTM network structure is introduced to model environmental parameters, and combined with the temporal attention mechanism, the dynamic impact of environmental parameter changes on the measurement results is accurately captured, and precise control of environmental compensation is achieved. A loss function and optimization strategy based on physical constraints are designed to ensure that the calibration results meet the physical characteristics requirements of the signal source and avoid unreasonable calibration parameters output by the model. Through the incremental learning method and the hierarchical parameter update strategy, the online optimization of the model is achieved, the model's memory ability for learned features is maintained, and the model's adaptability to new data is improved. BRIEF DESCRIPTION OF THE DRAWINGS

[0008] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative labor.

[0009] The structures, proportions, sizes, etc. illustrated in the drawings of this specification are only used to match the contents disclosed in the specification so as to facilitate understanding and reading by persons familiar with this technology. They are not used to limit the conditions under which the present invention can be implemented, and therefore have no substantive technical significance. Any structural modification, change in proportion or adjustment of size, without affecting the effects and purposes that can be achieved by the present invention, should still fall within the scope of the technical contents disclosed by the present invention.

[0010] Figure 1 It is a flow chart of a radio metrology calibration method based on machine learning provided by an embodiment of the present invention; Figure 2 It is a schematic block diagram of the structure of a radio metrology calibration system based on machine learning provided in an embodiment of the present invention. DETAILED DESCRIPTION

[0011] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0012] The flowcharts shown in the accompanying drawings are only examples and do not necessarily include all the contents and operations / steps, nor must they be executed in the order described. For example, some operations / steps may also be decomposed, combined or partially merged, so the actual execution order may change according to actual conditions.

[0013] It should also be understood that the terms used in this application specification are only for the purpose of describing specific embodiments and are not intended to limit the application. As used in this application specification and the appended claims, unless the context clearly indicates otherwise, the singular forms "a", "an" and "the" are intended to include plural forms.

[0014] It should be further understood that the term "and / or" used in the specification and appended claims of this application refers to any combination and all possible combinations of one or more of the associated listed items, and includes these combinations. Figure 1 , an embodiment of the radio metrology calibration method based on machine learning in the embodiment of the present application includes: Step 100: Measure parameters of a radio signal source to obtain an original measurement data set; It is understandable that the execution subject of the present application may be a radio metrology calibration system based on machine learning, or a terminal or a server, which is not specifically limited here. The present application embodiment is described by taking a server as the execution subject as an example.

[0015] Specifically, a series of settings are made for the radio signal source, including selecting multiple discrete frequency points to ensure that the measured data can fully cover the frequency response range of the signal source. For discrete frequency points, multiple repeated frequency measurements are required for each frequency point to reduce measurement errors and enhance the stability and reliability of the data. While completing the frequency measurement, the output power of the signal source is scanned to obtain signal characteristic data at different power levels. During the power scanning process, each power point is repeatedly measured to ensure the accuracy of the measurement data and reduce errors caused by instantaneous fluctuations or environmental changes. At the same time, a phase noise analyzer is used to measure the phase noise of the signal source, and repeated measurements are performed for the preset frequency deviation points to obtain the phase noise characteristic data of the signal source. The harmonic characteristics of the signal source are measured. Harmonic measurement can reveal the harmonic distribution of the signal source in different frequency ranges, thereby helping to evaluate its signal quality. Parameters such as ambient temperature, humidity, and atmospheric pressure are recorded synchronously to construct an environmental parameter data set. The acquisition of environmental data relies on high-precision temperature and humidity sensors and air pressure sensors, which monitor the changes in environmental parameters in real time and record them in the measurement database for analysis and compensation in subsequent data processing and model training stages. All measurement data are time synchronized using a unified timestamp format. Frequency measurement data, power measurement data, phase noise measurement data, harmonic measurement data, and environmental parameter data are aligned according to the same time reference to form a complete, time-synchronized raw measurement data set. The implementation of time synchronization relies on a high-precision time synchronization protocol or a high-stability clock source, such as a GPS synchronized clock or a high-precision local time reference. By time aligning all data, it is ensured that during data analysis and model training, all types of measurement data can correctly correspond to the same time window, thereby improving the timeliness and accuracy of the data.

[0016] Step 200: extract frequency accuracy features, power flatness features, phase noise features, harmonic ratio features, and environmental change features based on the original measurement data set to obtain a feature data set; Specifically, the original measurement data is cleaned and preprocessed to ensure the quality and reliability of the data. Since the radio measurement process is affected by equipment errors, environmental interference or instantaneous fluctuations, the 3σ criterion is used to detect anomalies in the original measurement data. By calculating the mean and standard deviation of the data, the data that exceeds the range of three times the standard deviation is marked as anomalies to obtain the anomaly marking results. After removing the outliers, the remaining non-abnormal data is normalized to eliminate the differences in the dimensions of different measurement parameters, so that different features can be learned on the same scale to obtain normalized measurement data. Frequency features are extracted from the normalized measurement data to characterize the frequency stability and accuracy of the radio signal source. The frequency accuracy is calculated, that is, the deviation between the measured frequency and the standard frequency. At the same time, the frequency stability is calculated to evaluate the stability of the signal source on different time scales, and the frequency resolution is analyzed to determine the measurement system's perception of small frequency changes. The long-term drift trend analysis is performed on each frequency measurement point to examine the changes in the signal frequency on a longer time scale, and the short-term jitter characteristics are extracted to reflect the stability changes of the signal in a short time. Through the above calculations, complete frequency feature data is obtained. At the same time, the phase noise data in the normalized measurement data is analyzed, the noise values ​​at multiple frequency deviation points are extracted, and the phase noise characteristic data is constructed. The phase noise values ​​are calculated at multiple frequency deviation points (such as 1Hz, 10Hz, 100Hz, 1kHz, etc.) to fully characterize the phase noise characteristics of the signal source. The harmonic data in the normalized measurement data is processed, and the harmonic levels of different harmonics and their ratios to the fundamental wave are calculated to obtain harmonic characteristic data. The environmental parameters in the normalized measurement data are analyzed in time and frequency using wavelet transform to reveal the pattern of environmental parameters changing over time. The temperature change rate, humidity change rate and their correlation coefficients with the measurement parameters are calculated to quantify the impact of environmental factors on the stability of the radio signal. Through this step, the key environmental factors that can affect the measurement accuracy are extracted and used as part of the characteristic data to enhance the environmental adaptability and robustness of the model. The frequency characteristic data, phase noise characteristic data, harmonic characteristic data and environmental characteristic data are combined to form a characteristic data set.

[0017] Step 300: Divide the feature data set into a training data subset, a verification data subset, and a test data subset, and construct a first neural network model including a two-branch neural network based on the training data subset; It should be noted that the feature data set is divided to ensure that the training, verification and testing of the model can cover the diversity of data and have good generalization ability. The feature data set is divided into training data subset, verification data subset and test data subset in a ratio of 7:2:1, where the training data is used to learn the model parameters, the verification data is used to adjust the hyperparameters and prevent overfitting, and the test data is used for the final model performance evaluation. After the division is completed, in order to enhance the robustness of the model and avoid the deviation caused by the data order, the training data subset is randomly shuffled to ensure that the model does not depend on the input order of the data during the training process, avoid the emergence of local optimal solutions, and obtain the preprocessed training data. When constructing the neural network model, since radio metrology calibration involves multiple physical property data, a dual-branch neural network architecture is adopted, in which the first branch is used to process the calibration parameters of the radio signal itself, and the second branch is used to process the compensation coefficient of environmental factors. In the first branch, an eight-layer fully connected neural network, namely the parameter calibration network, is designed. The first three layers contain 1024, 512, and 256 neurons, respectively, for feature extraction and nonlinear mapping. These three layers use the ReLU activation function to effectively avoid the gradient vanishing problem and improve the learning ability of the model. The self-attention mechanism is introduced in the fourth and fifth layers, in which the dimension of the weight matrix is ​​set to 256×256, thereby enhancing the model's ability to capture long-range dependencies, enabling it to more accurately learn the relationship between different frequency points, power levels, and other calibration parameters. Between these two self-attention layers, a residual connection is set to ensure the effective propagation of the gradient and improve the training stability of the model. In the sixth to eighth layers, 256, 128, and 64 neurons are set, respectively, for feature compression and parameter calibration output generation. The last layer uses a linear activation function to ensure that the calibration parameters output by the model are optimized and applied in a continuous numerical space. At the same time, in the second branch, a three-layer LSTM network, namely the environmental compensation network, is constructed to process environmental feature data and calculate environmental compensation coefficients. Since environmental factors (such as temperature, humidity, and atmospheric pressure) are time-dependent, LSTM (Long Short-Term Memory) is a suitable structural choice. Each layer of LSTM contains 128 hidden units, which effectively captures long-term dependencies in time series. Between LSTM units, dropout layers are added to prevent overfitting and ensure that the model still has good generalization ability when facing complex environmental data. In order to enable the model to pay attention to the characteristics of environmental data at different time points and give higher weights to time windows with greater measurement impact, a temporal attention mechanism is introduced in the output part of LSTM. Through the adaptive weighting method, the environmental feature data with greater impact on measurement accuracy is emphasized, and the environmental compensation output is obtained. The calibration parameter output is fused with the environmental compensation output to obtain the final model output.By using a weighted summation operation and setting a set of weight parameters, the outputs of the two branches are linearly weighted, so that environmental compensation can reasonably correct the calibration parameters. In order to ensure the dynamic adaptability of the fusion process, these weight coefficients are adaptively updated through back propagation to ensure that the model is continuously optimized during the training process, so that the influence of environmental factors is reasonably compensated without causing excessive interference to the final calibration results. After the above modeling process, a complete first neural network model is obtained, whose dual-branch structure can effectively consider the physical measurement characteristics of radio signals and external environmental factors at the same time, and achieve high-precision metrology calibration through adaptive optimization.

[0018] Step 400: Input the verification data subset into the first neural network model to optimize the model parameters to obtain a second neural network model.

[0019] Specifically, the validation data subset is input into the first neural network model, and the relative frequency deviation of the frequency measurement data is calculated, that is, the ratio difference between the measured frequency and the reference frequency, to measure the accuracy of the model in frequency calibration. The absolute deviation of the power measurement data is calculated, that is, the direct difference between the measured power and the standard power, to evaluate the accuracy of power calibration. The root mean square error of each frequency deviation point is calculated for the phase noise measurement data, and the mean of the square error between the model prediction value and the true value at different frequency deviation points is statistically calculated to quantify the performance of the model in phase noise compensation, and obtain the error values ​​of each parameter, including frequency error, power error, phase noise error and harmonic error. A loss function is constructed based on the output results of the first neural network model to clarify the optimization direction and goal. Since radio signal measurement involves multiple important physical characteristics, different weight coefficients are assigned to different error terms when constructing the loss function to reflect their impact on the final measurement accuracy. The frequency error term is assigned a first weight coefficient to ensure that the model can achieve high accuracy in frequency calibration. At the same time, the power error term is assigned a second weight coefficient to ensure the accuracy of power calibration. Since phase noise has a great influence on the stability of radio signals, a third weight coefficient is assigned to the phase noise error term to enhance the model's ability to calibrate phase noise, and a fourth weight coefficient is assigned to the harmonic error term to ensure that the model can effectively calibrate the harmonic components and make the measurement results more accurate. Through the weighted loss function, the model pays attention to the importance of different error terms during the optimization process and ensures that the optimization target meets the actual needs of radio measurement. Physical constraints are added on the basis of the weighted loss function to construct a constrained loss function that is more in line with the measurement standard. These physical constraints include: ensuring that the frequency error does not exceed the inherent resolution of the measurement equipment within a certain range to prevent meaningless small deviations from interfering with the optimization results; limiting the upper and lower bounds of the power error to ensure that the calibrated power output meets the dynamic range of the measurement equipment; constraining the optimization range of phase noise to ensure that the optimization process does not affect the essential characteristics of the signal; and setting reasonable constraints on the harmonic error to prevent over-calibration from causing excessive harmonic suppression, thereby affecting the normal transmission characteristics of the signal. By introducing these physical constraints, the optimization process can better meet the needs of the real measurement system, so that the final calibration model has a higher physical credibility. During the optimization process, the Adam optimization algorithm is used. The Adam algorithm combines the advantages of momentum gradient descent and adaptive learning rate, so that the model can converge quickly during training and avoid falling into the local optimal solution. Based on the constrained loss function and the error values ​​of each parameter, the Adam optimization algorithm calculates the gradient and updates the model weights, so that the model can gradually reduce the error and improve the calibration accuracy. During the optimization process, the first-order moment estimation and second-order moment estimation of the Adam algorithm can ensure the adaptive adjustment of the learning rate, making the optimization more stable, avoiding unstable oscillations caused by too high a learning rate, and preventing the convergence speed from being too slow due to too low a learning rate.As the optimization progresses, the model parameters are constantly adjusted, so the first neural network model is iteratively trained. Through multiple rounds of optimization, the loss function is continuously reduced, and the errors of each parameter gradually converge to a smaller range. In each iterative training, the model is updated according to the optimized parameters, and a new error calculation is performed using the validation data set to ensure that the optimization direction is correct. When the loss function converges to the preset threshold range, that is, when each error term meets the requirements of the radio measurement standard, the trained model parameters are obtained. Based on the trained model parameters, the first neural network model is finally updated to form the second neural network model.

[0020] After the model is deployed, the neural network model is dynamically updated based on the signal source performance data in actual use, so that it can adapt to the long-term changing measurement environment and maintain a high calibration accuracy. In actual use, the performance of the signal source will change over time, so long-term stability data, environmental adaptability data and reliability data are continuously monitored and collected to ensure the effectiveness of the model. When the cumulative amount of new data exceeds 10% of the initial training data subset, the system will trigger a model update signal to determine whether the model needs to be adjusted and optimized to ensure the long-term effectiveness and adaptability of the radio measurement calibration model. After the model update is triggered, in order to ensure the stability of the neural network and make full use of the existing training results, the model optimization is carried out by fixing some parameters and updating some parameters. In the second neural network model, the parameters of the first five layers of the parameter calibration network will be fixed, including 1024, 512, and 256 neurons in layers 1-3, and the self-attention layer parameters of 256×256 dimensions in layers 4-5. These layers are mainly responsible for low-level feature extraction and global feature attention calculation. They have learned the basic characteristic patterns of signal source calibration, so they do not need to be adjusted frequently to ensure the stability of the basic structure of the model while avoiding overfitting or forgetting historical data. After fixing these parameters, a fixed parameter layer is obtained. While ensuring the stability of the fixed parameter layer, the last three layers of the parameter calibration network are updated to enhance the model's adaptability to new data. The 256, 128, and 64 neuron parameters of layers 6-8 will be retrained so that the model can better learn new signal source characteristics and improve its generalization ability under different measurement environments. At the same time, in order to ensure the stability of the environmental compensation mechanism, the three-layer LSTM network parameters of the environmental compensation network remain unchanged, because the LSTM part has learned the long-term impact of environmental factors on radio signal measurement and will not deviate significantly with changes in short-term measurement data. By fixing some parameters and updating some parameters, the adaptability of the model is effectively improved, while retaining the learning results of the original network, and an updateable parameter layer is obtained to adapt to the latest data distribution. After the model is updated, its calibration accuracy is verified, a subset of test data is input into the updated neural network model, and the calibration error and measurement deviation are calculated to evaluate the calibration accuracy of the new model. By comparing the performance with the old model, it is determined whether the updated neural network model has improved the accuracy of radio metering and ensured that it is applicable to the latest signal source data distribution. The calibration accuracy evaluation results are used for model fine-tuning to optimize the application effect of the model in the actual measurement environment. On this basis, in order to optimize the performance of the model and prevent the forgetting of existing knowledge due to the addition of new data, the incremental learning method is used to fine-tune the updateable parameter layer. Incremental learning enables the model to adapt to the new data without affecting the learned knowledge, and ensures that the model does not completely forget the previous measurement features due to changes in the training data.During the fine-tuning process, based on the calibration accuracy evaluation results, the last three layers of the parameter calibration network are adjusted, and the parameters of the model are gradually optimized through the back-propagation algorithm, while keeping the feature extraction capability of the fixed parameter layer unchanged. Thus, while continuously learning new data, the measurement features of old data are retained, allowing the neural network to operate stably for a long time and adapt to the ever-changing radio measurement environment. After the fine-tuning is completed, the fixed parameter layer and the adjusted updateable parameter layer are combined to form the final target calibration model.

[0021] The frequency measurement data in the test data subset is subjected to Fourier transform, and the signal is converted from the time domain to the frequency domain through the transformation, so as to more effectively analyze its frequency characteristics. In the frequency domain space, each frequency component is extracted, and the power spectrum density of the frequency component is calculated to quantify the stability of the radio signal. In this way, the frequency stability evaluation data is obtained. Based on the output of the parameter calibration network of the first neural network model, the frequency stability evaluation data is self-attention weighted to enhance the model's perception of key frequency features. The weight distribution of different frequency components is calculated on the 256×256 dimensional weight matrix, so that the model can reasonably adjust the contribution of the calibration parameters according to the degree of influence of different frequency components on the measurement error. In this way, the frequency accuracy score is obtained and the frequency calibration evaluation parameters are formed to quantify the optimization effect of the model on the frequency characteristics and ensure that it can still maintain high-precision calibration capabilities during long-term operation. At the same time, the phase noise measurement data in the test data subset is analyzed to evaluate the model's compensation capability for phase noise. The noise root mean square value is calculated for multiple different frequency offset points to quantify the phase jitter of the signal source under different frequency offset conditions. Based on the prediction results of the second neural network model, the phase noise compensation coefficient is calculated to correct the phase noise deviation caused by environmental or equipment changes during the measurement process. Phase noise evaluation parameters are obtained through these calculations to measure the optimization effect of the model on phase noise and ensure that the calibration results can effectively reduce the phase jitter of the radio signal. In order to evaluate the adaptability of the model to environmental changes, the output results of the three-layer LSTM unit of the environmental compensation network are analyzed with the environmental parameter data in the test data subset for time series correlation to calculate the influence of environmental factors on the measurement accuracy of the radio signal. In this process, environmental variables such as the temperature change rate and the humidity change rate are calculated, and their influence coefficients on the calibration results are analyzed. Through this step, environmental compensation evaluation parameters are obtained to measure the adaptability of the model under different environmental conditions and ensure that it can still maintain high measurement accuracy in a complex external environment. Frequency calibration evaluation parameters, phase noise evaluation parameters and environmental compensation evaluation parameters are adaptively weighted to obtain an initial evaluation score. Based on the learning results of the neural network model, the weights of each parameter in the calibration accuracy evaluation are adaptively adjusted so that the final evaluation result can fully consider the comprehensive effects of multiple factors such as frequency characteristics, phase noise and environmental influences. Based on the output of the fusion layer of the dual-branch neural network model, residual compensation is performed on the initial evaluation score, that is, the evaluation score is dynamically adjusted using the output error of the fusion layer to ensure that it is closer to the distribution of the actual measurement data. Through the residual compensation mechanism, the compensated evaluation score is obtained, and the physical constraint verification is performed on it to ensure that the evaluation result conforms to the physical laws and measurement standards of radio measurement, and finally the calibration accuracy evaluation result is obtained.

[0022] In the embodiment of the present application, by constructing a dual-branch neural network structure, parameter calibration and environmental compensation are modeled separately, which effectively reduces the complexity of the model, and at the same time improves the calculation accuracy of the calibration parameters and the accuracy of environmental compensation. The network structure design using the self-attention mechanism and residual connection enhances the model's feature extraction ability for measurement data at different frequency points and different power points, while improving the convergence performance of the model. The LSTM network structure is introduced to model environmental parameters, and combined with the temporal attention mechanism, the dynamic impact of environmental parameter changes on the measurement results is accurately captured, and precise control of environmental compensation is achieved. A loss function and optimization strategy based on physical constraints are designed to ensure that the calibration results meet the physical property requirements of the signal source and avoid unreasonable calibration parameters output by the model. Through the incremental learning method and the hierarchical parameter update strategy, the online optimization of the model is achieved, the model's memory ability for learned features is maintained, and the model's adaptability to new data is improved.

[0023] In a specific embodiment, the process of executing step 100 may specifically include the following steps: A plurality of discrete frequency points are set for the radio signal source, and frequency measurement is repeated for each discrete frequency point to obtain frequency measurement data, and the output power of the signal source is scanned, and each power point is repeatedly measured to obtain power measurement data; Repeatedly measure the signal source at a preset frequency deviation point based on a phase noise analyzer to obtain phase noise measurement data, and perform harmonic measurement on the signal source to obtain harmonic measurement data; Record the environmental temperature, humidity and atmospheric pressure parameters to obtain environmental parameter data; Frequency measurement data, power measurement data, phase noise measurement data, harmonic measurement data and environmental parameter data are time synchronized based on a unified timestamp format to obtain an original measurement data set.

[0024] Specifically, multiple discrete frequency points are set for the radio signal source to ensure that the collected data can fully cover the working range of the signal source. These discrete frequency points are selected as equally spaced frequency points or specific key frequency points according to the experimental requirements. For example, for a signal source working in the range of 1 GHz to 10 GHz, a discrete frequency point set is selected. ), each of which represents a selected discrete frequency point. For example, if the interval is 100 MHz, then 1 GHz, GHz GHz. For each discrete frequency point, multiple repeated frequency measurements are performed to reduce measurement errors and improve data reliability. ,conduct Repeat the measurement to get a set of measurement data ( ),in Indicates The average measured value is then calculated and the measurement standard deviation , the formula is as follows: ; ; in, represents the frequency mean of the measurement, Reflects the dispersion of the measured value, that is, the measurement stability. At the same time, the output power of the signal source is scanned to analyze the output characteristics at different power levels. Set the output power range to , and in this range with a step value Perform a power scan to get a series of power points: ; in, Indicates Power measurement points, is the total number of measured power points. , repeated measurements are performed to obtain a set of measurement values And calculate its mean and standard deviation: ; ; in, represents the measured power mean, Reflects the uncertainty of power measurement. In order to analyze the phase noise of the signal, a phase noise analyzer is used to measure the signal source at different preset frequency deviation points. ( ), such as 1 Hz, 10 Hz, 100 Hz, 1 kHz, etc., and measure the phase noise power spectral density at these frequency offset points , which is defined as follows: ; in, Indicates frequency deviation The phase noise power spectrum density at . This data reflects the phase jitter of the radio signal at different frequency offsets. In order to evaluate the quality of the signal source, the harmonics are measured. Set the fundamental frequency , and measure its high-order harmonic components , calculate the harmonic ratio: ; in, It is The power of the harmonics, is the fundamental power. This indicates that the nonlinear distortion of the signal source is small and the signal quality is high. At the same time, in order to study the impact of environmental factors on the measurement results, the ambient temperature, humidity and atmospheric pressure are recorded, namely: ; in, represents the temperature changing with time, Indicates humidity, Indicates air pressure. Calculates the rate of temperature change and humidity change rate : ; Environmental data helps with subsequent machine learning calibration model training. Use a unified timestamp format Time synchronization of all measurement data. For example, frequency measurement data , power measurement data , Phase noise measurement data , Harmonic measurement data And environmental parameter data Recording is done on the same time basis to facilitate subsequent data fusion and feature extraction, ultimately obtaining a complete set of original measurement data.

[0025] In a specific embodiment, the process of executing step 200 may specifically include the following steps: Based on the 3σ criterion, anomaly detection is performed on the original measurement data set to obtain anomaly value marking results, and non-abnormal data in the original measurement data set is normalized to obtain normalized measurement data; Extract frequency characteristics from normalized measurement data, calculate frequency accuracy, frequency stability and frequency resolution, analyze long-term drift trend and short-term jitter characteristics for each frequency measurement point, and obtain frequency characteristic data; Extracting multiple frequency deviation point noise values ​​from the phase noise data in the normalized measurement data to obtain phase noise characteristic data, and calculating harmonic levels and ratios thereof from the harmonic data in the normalized measurement data to obtain harmonic characteristic data; The environmental parameters in the normalized measurement data are analyzed in time and frequency through wavelet transform, and the correlation coefficients between the temperature change rate, humidity change rate and the measurement parameters are calculated to obtain the environmental characteristic data. The frequency characteristic data, phase noise characteristic data, harmonic characteristic data and environmental characteristic data are combined into a characteristic data set.

[0026] Specifically, statistical analysis is performed on the original measurement data set to calculate the mean and standard deviation of each measurement parameter. Assume that the set of observation values ​​of a certain measurement parameter is ,in Representative The measured value is For the total number of measurement samples, the mean is calculated as: ; in represents the mean of the parameter. At the same time, the standard deviation is calculated by the following formula: ; in Represents the standard deviation of the parameter. According to the 3σ criterion, if a measurement value satisfy: ; The measured value is considered an outlier and is marked in the outlier labeling result. All non-outlier data will be retained and normalized. The normalization method uses minimum-maximum normalization: ; in and They are the minimum and maximum values ​​in the data set, respectively, and the normalized data It is limited to the interval [0, 1], so that the numerical scales of different physical quantities are consistent, improving the stability of neural network training. Frequency feature extraction is performed on the normalized measurement data to calculate frequency accuracy, frequency stability and frequency resolution. For the frequency measurement data set , defining the frequency accuracy for: ; in, is the measured average frequency value: ; and is the base reference frequency. The frequency stability is calculated using the Allan variance: ; in, represents the sampling time interval, Reflects the signal on the time scale The frequency resolution is defined as: ; in, It reflects the ability of the measuring device to distinguish small frequency changes. For each measured frequency point, the long-term drift trend is calculated, that is, based on linear regression solution: ; in, represents the drift rate and is measured by least squares fitting. The short-term jitter characteristic extracts the spectral components of the frequency jitter through fast Fourier transform, thereby quantifying the impact of high-frequency noise components. For phase noise feature extraction, the noise values ​​of multiple frequency offset points in the normalized measurement data are calculated. Phase noise is expressed as The unit is defined as: ; in, Indicates the offset frequency The phase noise at Represents the phase noise power spectral density. The phase noise data at different frequency offset points will be recorded and used for subsequent model training. For harmonic feature extraction, the power of different harmonic frequency components is calculated, and the harmonic level is calculated: ; in, It is The power of the harmonics, is the fundamental power. To quantify the relative energy ratio of the harmonics to the fundamental, the total harmonic distortion is calculated: ; This indicator is used to evaluate the degree of nonlinear distortion of radio signals. In terms of environmental feature extraction, wavelet transform is used to perform time-frequency analysis on parameters such as temperature, humidity, and atmospheric pressure. is a sequence of temperature measurements, then the wavelet transform is defined as: ; in, is the scale parameter, is the time offset, is the wavelet basis function. Through wavelet transform, the temperature change rate is calculated and humidity change rate : ; Perform correlation analysis between environmental parameter data and radio measurement data. is the measurement parameter, is an environmental variable, then the correlation coefficient The calculation is as follows: ; like If the value is larger, it means that the environmental parameter has a greater impact on the measured data. The frequency characteristic data, phase noise characteristic data, harmonic characteristic data and environmental characteristic data are combined to form a complete characteristic data set.

[0027] In a specific embodiment, the process of executing step 300 may specifically include the following steps: The feature data set is divided into a training data subset, a validation data subset, and a test data subset in a ratio of 7:2:1, and the training data subset is randomly shuffled to obtain preprocessed training data; An eight-layer fully connected network structure of the first branch was constructed, in which the number of neurons in the 1st to 3rd layers were 1024, 512, and 256, respectively, the 4th to 5th layers adopted a self-attention mechanism, the weight matrix dimension was 256×256, and the number of neurons in the 6th to 8th layers were 256, 128, and 64, respectively, to obtain a parameter calibration network; A three-layer LSTM network structure of the second branch is constructed, each layer contains 128 hidden units, the input is environmental feature data, and the output is the environmental compensation coefficient, thus obtaining an environmental compensation network; The ReLU activation function is used in the 1st to 3rd layers of the parameter calibration network, a residual connection is set between the 4th and 5th self-attention layers, and a linear activation function is used in the last layer to obtain the calibration parameter output; A dropout layer is set between the three-layer LSTM units of the environmental compensation network, and the environmental features at different times are weighted through the temporal attention mechanism to obtain the environmental compensation output; The calibration parameter output and the environmental compensation output are fused based on a weighted sum operation, and the weight coefficients are adaptively updated through back propagation to obtain a first neural network model.

[0028] Specifically, the feature data set is divided into a training data subset, a validation data subset, and a test data subset in a ratio of 7:2:1. The training data subset is used for model parameter optimization, the validation data subset is used to adjust hyperparameters and prevent overfitting, and the test data subset is used to evaluate the generalization performance of the final model. Assume that the feature data set is ,in Representative Input features, represents the corresponding calibration label, then the training data subset Contains 70% of the data, validation data subset Contains 20% of the data, a test data subset Contains 10% of the data, namely: ; In order to prevent the data from having sequential dependencies during the training process, the training data subsets are randomly shuffled to ensure that the model will not have learning bias due to the order of data input. The data sequence is shuffled to make it randomly distributed, and the preprocessed training data is obtained. The first neural network model is constructed, in which the first branch uses an eight-layer fully connected neural network, that is, a parameter calibration network. The first three layers of the network perform feature extraction and dimensionality reduction, in which the first layer contains 1024 neurons, the second layer contains 512 neurons, and the third layer contains 256 neurons. Its input is feature data , mapped to the latent feature space through linear transformation and nonlinear activation function: ; ; ; in, They are the hidden layer outputs of the first three layers, are the weight matrices, They are bias terms, and ReLU as a nonlinear activation function helps improve the model's expressiveness and prevents gradient disappearance. In the fourth and fifth layers, the self-attention mechanism is introduced to calculate the weighted relationship matrix between input features and enhance the model's ability to focus on key features. (output from the third layer), define the query matrix , Key matrix , Value matrix : ; in, is a trainable parameter matrix, and the attention weight matrix is ​​calculated: ; in, is the dimension of the key matrix, and the weighted output is finally calculated: ; A residual connection is set between the fourth and fifth layers to ensure gradient flow and improve the training stability of the model. The sixth to eighth layers are used for further feature extraction, containing 256, 128, and 64 neurons respectively, and finally the calibration parameter output is obtained: ; The last layer uses a linear activation function to ensure that the output can be adaptively adjusted within a continuous numerical range. At the same time, the second branch constructs a three-layer LSTM network, namely the environmental compensation network, which is used to process environmental feature data. LSTM is a neural network suitable for time series data. Assume that the environmental feature input is ,in Representative The environment data of time steps, the hidden state of LSTM is updated as follows: ; ; ; ; ; in, are the activation values ​​of the forget gate, input gate, and output gate, respectively. is the cell state, is the hidden state. Each LSTM layer contains 128 hidden units. Dropout layers are added between layers to prevent overfitting. Finally, the environmental features at different times are weighted through the temporal attention mechanism: ; ; In the final fusion stage, a weighted summation operation is used to output the calibration parameters. With ambient compensation output To perform the fusion: ; in, are trainable weight parameters that are adaptively updated through back-propagation to minimize the loss function: ; Finally, the first neural network model is obtained, whose dual-branch structure can comprehensively consider the calibration parameters of the radio signal and the environmental compensation factors, thereby improving the measurement accuracy and adapting to different measurement environments.

[0029] Among them, the eight-layer fully connected network structure of the first branch is constructed, including: performing collaborative domain alignment on the preprocessed training data, calculating the first-order moment and second-order moment statistical characteristics of the data, and obtaining the source domain feature distribution; calculating the marginal probability density of the source domain feature distribution based on the kernel density estimation method to obtain the marginal distribution matrix; performing conditional probability calculation on the marginal distribution matrix, using KL divergence to measure the distribution difference between different domains, and obtaining the conditional distribution matrix; inputting the marginal distribution matrix and the conditional distribution matrix into the joint optimizer, updating the network parameters by minimizing the distribution difference, and obtaining the domain alignment parameters; and performing the optimization on the 4th to 5th layers of the eight-layer fully connected network. Some shrinkage layers are added, and the Jacobian matrix of the hidden layer activation values ​​to the input is calculated to obtain the feature sensitivity matrix; the Frobenius norm is calculated based on the feature sensitivity matrix, and it is added to the loss function as a regularization term to obtain the shrinkage constraint term; according to the domain alignment parameters and the shrinkage constraint term, the weights and biases of the eight-layer fully connected network are optimized in stages, some shrinkage layer parameters are fixed in the pre-training stage, and all parameters are unfrozen and updated in the fine-tuning stage to obtain the optimized parameter configuration; the optimized parameter configuration is applied to the eight-layer fully connected network, and the network is reconstructed to obtain a parameter calibration network with domain migration capability.

[0030] In a specific embodiment, the process of executing step 400 may specifically include the following steps: Input the verification data subset into the first neural network model, calculate the relative frequency deviation of the frequency measurement data, calculate the absolute deviation of the power measurement data, and calculate the root mean square error of each frequency deviation point of the phase noise measurement data to obtain the error value of each parameter; A loss function is constructed according to the output result of the first neural network model, a first weight coefficient is assigned to the frequency error term, a second weight coefficient is assigned to the power error term, a third weight coefficient is assigned to the phase noise error term, and a fourth weight coefficient is assigned to the harmonic error term, so as to obtain a weighted loss function; Add physical constraints to the weighted loss function to obtain a constrained loss function; Based on the constrained loss function and the error values ​​of each parameter, the Adam optimization algorithm is used to optimize the parameters and obtain the optimized parameter configuration; The first neural network model is iteratively trained to obtain trained model parameters, and the first neural network model is updated based on the trained model parameters to obtain a second neural network model.

[0031] Specifically, the validation data subset is input into the first neural network model to evaluate its actual error in radio metrology calibration. The frequency measurement data is analyzed and the relative frequency deviation is calculated to quantify the model's error in frequency calibration. Assuming the true reference frequency is , the frequency value predicted by the neural network is , then the relative frequency deviation Defined as: ; This value is used to measure the deviation ratio between the predicted frequency and the true frequency. If the value is too large, it means that the model has significant errors in frequency calibration and needs to be optimized. At the same time, the power measurement data is analyzed and the absolute deviation is calculated to measure the power calibration capability of the model. Assume that the real power value is , the prediction power of the neural network is , then the absolute deviation The calculation is as follows: ; This indicator reflects the absolute error between the predicted power and the actual power. This shows that the power prediction of the model is more accurate, while the larger This means that the power compensation capability of the calibration network needs to be improved. In terms of phase noise measurement, the root mean square error is calculated for multiple frequency offset points to measure the phase noise compensation accuracy of the model. , if the predicted phase noise of the neural network is The actual measurement value is , then the root mean square error is defined as follows: ; This error reflects the model's ability to compensate for phase noise at different frequency offsets. A lower RMSE value means a better calibration effect. After obtaining the error values ​​of each parameter, a loss function is constructed to guide the optimization process of the neural network. Since different error terms have different importance in the calibration process, corresponding weights are assigned to different error terms. Assume that the weight of the frequency error term is , the weight of the power error term is , the weight of the phase noise error term is , the weight of the harmonic error term is , then the weighted loss function The definition is as follows: ; in, Indicates Harmonic power, is the actual measured value, Used to measure the model's ability to optimize harmonic errors. After constructing a weighted loss function, in order to ensure that the optimization process conforms to the physical laws of radio measurement, physical constraints are added to construct a constrained loss function. For example, in order to ensure that the calibration result does not exceed the capabilities of the measurement equipment, a frequency error constraint is added: ; in is the maximum relative frequency error allowed by the device. The power error also needs to be subject to physical constraints: ; in is the maximum allowable error in power measurement. Phase noise compensation is also limited to an acceptable range: ; The constrained loss function is expressed as: , ; in, To constrain the penalty coefficient, if the error exceeds the allowable range, the loss function value will increase significantly to force the optimization process to meet the physical constraints. During the optimization process, the Adam optimization algorithm is used to update the neural network parameters. Adam combines the advantages of momentum gradient descent and adaptive learning rate adjustment to update the parameters in each iteration. : ; ; ; in, is the first-order momentum estimate, is the second-order momentum estimate, is the learning rate, is a stable term. The Adam optimization algorithm can effectively accelerate convergence and adapt to optimization problems with different gradient change rates. Through multiple iterative training, the loss function of the neural network gradually converges, and finally the trained model parameters are obtained. The optimized parameters are used to update the first neural network model so that it can perform radio metrology calibration more accurately, and the second neural network model is obtained.

[0032] In a specific embodiment, executing the radio metrology calibration method based on machine learning further includes the following steps: Based on the signal source performance data in actual use, which includes long-term stability data, environmental adaptability data and reliability data, when the cumulative amount of new data exceeds 10% of the training data subset, a model update trigger signal is obtained; The parameters of the first five layers of the parameter calibration network in the second neural network model are fixed, including the 1024, 512, and 256 neurons in the 1st to 3rd layers and the self-attention layer parameters of the 4th to 5th layers with a dimension of 256×256, to obtain a fixed parameter layer; Update the parameters of the last three layers of the parameter calibration network, including the 256, 128, and 64 neuron parameters of the 6th to 8th layers, while keeping the three-layer LSTM network parameters of the environmental compensation network unchanged to obtain an updateable parameter layer; Inputting the test data subset into the updated neural network model to perform calibration accuracy evaluation, and obtaining a calibration accuracy evaluation result; The updateable parameter layer is adjusted based on the calibration accuracy evaluation result, and the parameters are fine-tuned using the incremental learning method to maintain the learned feature extraction capability in the fixed parameter layer, thereby obtaining the adjusted neural network model; The fixed parameter layer and the adjusted updateable parameter layer are combined to obtain the target calibration model.

[0033] Specifically, the neural network model is optimized based on the signal source performance data in actual use, and the long-term stability and environmental adaptability of radio metrology calibration are ensured. The long-term stability data, environmental adaptability data, and reliability data of the signal source are continuously monitored so that the model can be updated when the data distribution changes. The initial size of the training data subset is set to , when the amount of new data satisfy: ; That is, when the cumulative amount of new data exceeds 10% of the training data subset, the model update signal is triggered to ensure that the model can continue to adapt to the new measurement environment and improve the calibration accuracy. After the model update is triggered, the parameter calibration network in the second neural network model is partially updated. In order to avoid affecting the original feature extraction capability, the parameters of the first five layers of the parameter calibration network are fixed, including 1024, 512, and 256 neurons in layers 1-3, and 256×256 dimensional self-attention layer parameters in layers 4-5. These layers are used for basic feature extraction and global attention calculations. Their parameters have learned effective feature mappings on large-scale training data, so there is no need to adjust them frequently. Suppose the parameters of the original model are , where the parameter sets of the first five layers are: ; During the model update process, these parameters are frozen, namely: ; Ensure that these parameters are not updated with gradients during training. At the same time, the model still needs to adapt to the new data distribution, so the parameters of the last three layers of the parameter calibration network are updated, that is, the 256, 128, and 64 neuron parameters of the 6th to 8th layers, which are responsible for the final calibration parameter prediction. Suppose the parameter set of the last three layers is: ; During the update process, these parameters participate in gradient calculation and weight adjustment to adapt to the new measurement data distribution. At the same time, in order to keep the compensation ability of environmental factors unaffected, the three-layer LSTM network parameters of the environmental compensation network remain unchanged, namely: ; in, Represents the parameter set of the LSTM network. After completing the parameter update, evaluate the calibration accuracy of the model, input the test data subset into the updated neural network model, and calculate the calibration error. For frequency measurement data, calculate the relative frequency deviation: ; in, is the prediction frequency of the neural network, is the reference measurement frequency. For power measurement data, calculate the absolute deviation: ; in, is the power value predicted by the model, and is the true power measurement. For phase noise, calculate the RMS error at multiple frequency offset points: ; in, is the predicted phase noise, is the actual phase noise measurement value. The error calculation result is used as the basis for calibration accuracy evaluation. After obtaining the calibration accuracy evaluation result, the updateable parameter layer is adjusted and the parameters are fine-tuned using the incremental learning method to ensure that the model can adapt to new data without forgetting the existing knowledge. Incremental learning optimizes: ; in, is the learning rate, is the loss function, is the gradient of the updateable parameters. In order to prevent the model from forgetting old data, the incremental learning process uses regularization methods, such as: ; in, is the parameter before updating, is the regularization coefficient, which is used to maintain the consistency of the new and old models. The fixed parameter layer and the adjusted updateable parameter layer are combined to form a new target calibration model: ; This target calibration model not only inherits the effective feature extraction capability of previous training, but also adapts to the new data distribution through incremental learning, so that the model can still maintain high measurement accuracy and environmental adaptability during long-term use, and can further optimize its calibration capability through a continuous data update mechanism in the future.

[0034] In a specific embodiment, the execution step inputs the test data subset into the updated neural network model for calibration accuracy evaluation, and the process of obtaining the calibration accuracy evaluation result may specifically include the following steps: Perform Fourier transform on the frequency measurement data in the test data subset, extract the frequency components in the frequency domain space, calculate the power spectrum density for each frequency component, and obtain the frequency stability evaluation data; Based on the parameter calibration network output of the first neural network model, self-attention weighting is performed on the frequency stability evaluation data, and the frequency accuracy score is calculated through a 256×256 dimensional weight matrix to obtain the frequency calibration evaluation parameters; Calculating the noise root mean square values ​​of multiple frequency deviation points for the phase noise measurement data in the test data subset, and calculating the phase noise compensation coefficient according to the prediction result of the second neural network model to obtain the phase noise evaluation parameter; The output results of the three-layer LSTM unit of the environmental compensation network are analyzed with the environmental parameter data in the test data subset in a time series correlation manner, and the influence coefficients of the temperature change rate and humidity change rate on the calibration results are calculated to obtain the environmental compensation evaluation parameters. Adaptively weighting the frequency calibration evaluation parameter, the phase noise evaluation parameter, and the environmental compensation evaluation parameter to obtain an initial evaluation score; Based on the fusion layer output of the dual-branch neural network model, residual compensation is performed on the initial evaluation score to obtain the compensated evaluation score, and physical constraint verification is performed on the compensated evaluation score to obtain the calibration accuracy evaluation result.

[0035] Specifically, the frequency measurement data in the test data subset is Fourier transformed, and the frequency components are extracted in the frequency domain space to calculate the power spectrum density and obtain the frequency stability evaluation data. To obtain the time domain measurement data, assume that the frequency measurement signal is ,in As a time variable, the signal is transformed into the frequency domain by discrete Fourier transform (DFT): ; in, Indicates discrete frequency components, is the signal length, is the sampling interval. Calculate the power spectral density (PSD) of each frequency component: ; in, Representative The power density of each frequency component is used to quantify the energy distribution of the signal at different frequency points. , evaluate the stability of the frequency measurement data, and obtain the frequency stability evaluation data. Based on the parameter calibration network output of the first neural network model, the frequency stability evaluation data is self-attention weighted to calculate the frequency accuracy score. Suppose the parameter calibration output of the neural network is ,in is the feature dimension, calculate the query matrix , Key matrix Sum Value Matrix : ; in, is the trainable parameter matrix, and then calculates the attention weight matrix: ; in, is the dimension of the key matrix, and the weighted frequency accuracy score is finally calculated: ; in, represents the frequency calibration evaluation parameter, which is used to quantify the optimization effect of the neural network on frequency stability. At the same time, the noise root mean square value of multiple frequency deviation points is calculated for the phase noise measurement data in the test data subset, and the phase noise compensation coefficient is calculated according to the prediction result of the second neural network model to obtain the phase noise evaluation parameter. Assume that the phase noise measurement data is ,in is the frequency deviation point set: ; in, is the phase noise value predicted by the neural network model, and the phase noise compensation coefficient is calculated by the neural network: ; The phase noise evaluation parameters are expressed as: ; in, This reflects the optimization effect of the calibration model on phase noise. In order to quantify the impact of environmental factors on measurement accuracy, the output results of the three-layer LSTM unit of the environmental compensation network are analyzed with the environmental parameter data in the test data subset for time series correlation, and the influence coefficients of the temperature change rate and humidity change rate on the calibration results are calculated to obtain the environmental compensation evaluation parameters. Assume that the environmental characteristic data is , the LSTM is calculated as follows: ; in, is the hidden state of the LSTM unit, and the influence coefficient of temperature and humidity on the calibration result is calculated using time series correlation analysis: ; ; in, and Reflects the effect of temperature and humidity on the measured parameters The influence of is thus formed into environmental compensation evaluation parameters. The frequency calibration evaluation parameters, phase noise evaluation parameters and environmental compensation evaluation parameters are adaptively weighted to obtain the initial evaluation score. Suppose the weighting parameter is ,but: ; in, Represents the initial evaluation score. After obtaining the initial evaluation score, the residual compensation is performed on the initial evaluation score based on the fusion layer output of the dual-branch neural network model, and the compensated evaluation score is finally calculated: ; in, is the residual compensation weight, is the residual term calculated by the neural network. Physical constraints are verified on the compensated evaluation score to ensure that the error meets the radio measurement standard: ; in, is the allowable measurement error threshold to ensure that the final calibration accuracy evaluation result meets the accuracy requirements of the measuring equipment.

[0036] The above describes the radio metrology calibration method based on machine learning in the embodiment of the present application. The following describes the radio metrology calibration system 10 based on machine learning in the embodiment of the present application. Figure 2 In the embodiment of the present application, an embodiment of the radio metrology calibration system 10 based on machine learning includes: The measurement module 11 is used to measure the parameters of the radio signal source to obtain an original measurement data set; An extraction module 12 is used to extract frequency accuracy features, power flatness features, phase noise features, harmonic ratio features and environmental change features based on the original measurement data set to obtain a feature data set; A construction module 13 is used to divide the feature data set into a training data subset, a verification data subset and a test data subset, and to construct a first neural network model including a two-branch neural network based on the training data subset; The optimization module 14 is used to input the verification data subset into the first neural network model to optimize the model parameters to obtain the second neural network model.

[0037] Through the cooperation of the above components, the parameter calibration and environmental compensation are modeled separately by constructing a dual-branch neural network structure, which effectively reduces the complexity of the model and improves the calculation accuracy of the calibration parameters and the accuracy of environmental compensation. The network structure design using the self-attention mechanism and residual connection enhances the model's ability to extract features from measurement data at different frequency points and power points, while improving the convergence performance of the model. The LSTM network structure is introduced to model environmental parameters, and combined with the temporal attention mechanism, the dynamic impact of environmental parameter changes on the measurement results is accurately captured, and precise control of environmental compensation is achieved. The loss function and optimization strategy based on physical constraints are designed to ensure that the calibration results meet the physical characteristics requirements of the signal source and avoid unreasonable calibration parameters output by the model. Through the incremental learning method and hierarchical parameter update strategy, the online optimization of the model is achieved, the model's memory ability for learned features is maintained, and the model's adaptability to new data is improved.

[0038] Those skilled in the art can clearly understand that, for the convenience and brevity of description, the specific working processes of the above-described systems, systems and units can refer to the corresponding processes in the aforementioned method embodiments and will not be repeated here.

[0039] If the integrated unit is implemented in the form of a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present application is essentially or the part that contributes to the prior art or all or part of the technical solution can be embodied in the form of a software product. The computer software product is stored in a storage medium, including several instructions to enable an electronic device (which can be a personal computer, server, or network device, etc.) to perform all or part of the steps of the method described in each embodiment of the present application. The aforementioned storage medium includes: U disk, mobile hard disk, read-only memory (ROM), random access memory (RAM), disk or optical disk, etc., various media that can store program codes.

[0040] As described above, the above embodiments are only used to illustrate the technical solutions of the present application, rather than to limit them. Although the present application has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the aforementioned embodiments, or make equivalent replacements for some of the technical features therein. However, these modifications or replacements do not deviate the essence of the corresponding technical solutions from the spirit and scope of the technical solutions of the embodiments of the present application.

Claims

1. A radio metrology calibration method based on machine learning, characterized in that: The method comprises: Perform parameter measurements on the radio signal source to obtain an original measurement data set; Based on the original measurement data set, frequency accuracy features, power flatness features, phase noise features, harmonic ratio features and environmental change features are extracted to obtain a feature data set; Dividing the feature data set into a training data subset, a validation data subset, and a test data subset, and constructing a first neural network model including a two-branch neural network based on the training data subset; The verification data subset is input into the first neural network model to optimize model parameters to obtain a second neural network model.

2. The radio metrology calibration method based on machine learning according to claim 1, characterized in that: The parameter measurement of the radio signal source is performed to obtain an original measurement data set, including: A plurality of discrete frequency points are set for the radio signal source, and frequency measurement is repeated for each discrete frequency point to obtain frequency measurement data, and the output power of the signal source is scanned, and each power point is repeatedly measured to obtain power measurement data; Repeatedly measure the signal source at a preset frequency deviation point based on a phase noise analyzer to obtain phase noise measurement data, and perform harmonic measurement on the signal source to obtain harmonic measurement data; Record the environmental temperature, humidity and atmospheric pressure parameters to obtain environmental parameter data; The frequency measurement data, the power measurement data, the phase noise measurement data, the harmonic measurement data and the environmental parameter data are time synchronized based on a unified timestamp format to obtain an original measurement data set.

3. The radio metrology calibration method based on machine learning according to claim 2, characterized in that: The extracting frequency accuracy features, power flatness features, phase noise features, harmonic ratio features and environmental change features based on the original measurement data set to obtain a feature data set includes: Performing anomaly detection on the original measurement data set based on the 3σ criterion to obtain an outlier labeling result, and performing normalization processing on non-abnormal data in the original measurement data set to obtain normalized measurement data; Extracting frequency characteristics from the normalized measurement data, calculating frequency accuracy, frequency stability and frequency resolution, analyzing long-term drift trends and short-term jitter characteristics for each frequency measurement point, and obtaining frequency characteristic data; Extracting multiple frequency deviation point noise values ​​from the phase noise data in the normalized measurement data to obtain phase noise characteristic data, and calculating harmonic levels and ratios thereof from the harmonic data in the normalized measurement data to obtain harmonic characteristic data; The environmental parameters in the normalized measurement data are subjected to time-frequency analysis by wavelet transform, the correlation coefficients between the temperature change rate, the humidity change rate and the measurement parameters are calculated to obtain environmental characteristic data, and the frequency characteristic data, the phase noise characteristic data, the harmonic characteristic data and the environmental characteristic data are combined into a characteristic data set.

4. The radio metrology calibration method based on machine learning according to claim 3 is characterized in that: The step of dividing the feature data set into a training data subset, a verification data subset, and a test data subset, and constructing a first neural network model including a two-branch neural network based on the training data subset includes: Dividing the feature data set into a training data subset, a verification data subset, and a test data subset in a ratio of 7:2:1, and randomly shuffling the training data subset to obtain preprocessed training data; An eight-layer fully connected network structure of the first branch was constructed, in which the number of neurons in the 1st to 3rd layers were 1024, 512, and 256, respectively, the 4th to 5th layers adopted a self-attention mechanism, the weight matrix dimension was 256×256, and the number of neurons in the 6th to 8th layers were 256, 128, and 64, respectively, to obtain a parameter calibration network; A three-layer LSTM network structure of the second branch is constructed, each layer contains 128 hidden units, the input is environmental feature data, and the output is the environmental compensation coefficient, thus obtaining an environmental compensation network; The ReLU activation function is used in the 1st to 3rd layers of the parameter calibration network, a residual connection is set between the 4th and 5th self-attention layers, and a linear activation function is used in the last layer to obtain a calibration parameter output; A dropout layer is set between the three-layer LSTM units of the environmental compensation network, and environmental features at different times are weighted through a temporal attention mechanism to obtain an environmental compensation output; The calibration parameter output and the environmental compensation output are fused based on a weighted sum operation, and the weight coefficients are adaptively updated through back propagation to obtain a first neural network model.

5. The radio metrology calibration method based on machine learning according to claim 4, characterized in that: The step of inputting the verification data subset into the first neural network model to optimize model parameters to obtain a second neural network model comprises: Inputting the verification data subset into the first neural network model, calculating the relative frequency deviation of the frequency measurement data, calculating the absolute deviation of the power measurement data, and calculating the root mean square error of each frequency deviation point of the phase noise measurement data to obtain the error value of each parameter; Constructing a loss function according to the output result of the first neural network model, assigning a first weight coefficient to the frequency error term, assigning a second weight coefficient to the power error term, assigning a third weight coefficient to the phase noise error term, and assigning a fourth weight coefficient to the harmonic error term, to obtain a weighted loss function; Adding a physical constraint term to the weighted loss function to obtain a constrained loss function; Based on the constrained loss function and the error values ​​of each parameter, the Adam optimization algorithm is used to perform parameter optimization to obtain an optimized parameter configuration; The first neural network model is iteratively trained to obtain trained model parameters, and the first neural network model is updated based on the trained model parameters to obtain a second neural network model.

6. The radio metrology calibration method based on machine learning according to claim 5, characterized in that: The radio metrology calibration method based on machine learning also includes: Based on the signal source performance data in actual use, the signal source performance data includes long-term stability data, environmental adaptability data and reliability data, when the cumulative amount of new data exceeds 10% of the training data subset, a model update trigger signal is obtained; Fixing the parameters of the first five layers of the parameter calibration network in the second neural network model, including 1024, 512, and 256 neurons in the 1st to 3rd layers and the self-attention layer parameters of the 4th to 5th layers with a dimension of 256×256, to obtain a fixed parameter layer; The parameters of the last three layers of the parameter calibration network are updated, including the 256, 128, and 64 neuron parameters of the 6th to 8th layers, while the parameters of the three-layer LSTM network of the environmental compensation network are kept unchanged to obtain an updateable parameter layer; Inputting the test data subset into the updated neural network model to perform calibration accuracy evaluation to obtain a calibration accuracy evaluation result; The updateable parameter layer is adjusted based on the calibration accuracy evaluation result, and the parameters are fine-tuned using an incremental learning method to maintain the learned feature extraction capability in the fixed parameter layer, thereby obtaining an adjusted neural network model; The fixed parameter layer and the adjusted updateable parameter layer are combined to obtain a target calibration model.

7. The radio metrology calibration method based on machine learning according to claim 6, characterized in that: The step of inputting the test data subset into the updated neural network model to perform calibration accuracy evaluation to obtain a calibration accuracy evaluation result includes: Performing Fourier transformation on the frequency measurement data in the test data subset, extracting frequency components in the frequency domain space, calculating the power spectrum density for each frequency component, and obtaining frequency stability evaluation data; Based on the parameter calibration network output of the first neural network model, self-attention weighting is performed on the frequency stability evaluation data, and a frequency accuracy score is calculated through a weight matrix of 256×256 dimensions to obtain a frequency calibration evaluation parameter; Calculating the noise root mean square values ​​of multiple frequency deviation points for the phase noise measurement data in the test data subset, and calculating the phase noise compensation coefficient according to the prediction result of the second neural network model to obtain the phase noise evaluation parameter; Perform time series correlation analysis on the output results of the three-layer LSTM unit of the environmental compensation network and the environmental parameter data in the test data subset, calculate the influence coefficients of the temperature change rate and the humidity change rate on the calibration result, and obtain the environmental compensation evaluation parameters; Adaptively weighting the frequency calibration evaluation parameter, the phase noise evaluation parameter, and the environmental compensation evaluation parameter to obtain an initial evaluation score; Based on the fusion layer output of the dual-branch neural network model, residual compensation is performed on the initial evaluation score to obtain a compensated evaluation score, and physical constraint verification is performed on the compensated evaluation score to obtain a calibration accuracy evaluation result.

8. A radio metrology calibration system based on machine learning, characterized in that: For executing the radio metrology calibration method based on machine learning according to any one of claims 1 to 7, the system comprises: A measurement module, used to measure parameters of a radio signal source to obtain an original measurement data set; An extraction module, used to extract frequency accuracy features, power flatness features, phase noise features, harmonic ratio features and environmental change features based on the original measurement data set to obtain a feature data set; A construction module, used for dividing the feature data set into a training data subset, a verification data subset and a test data subset, and constructing a first neural network model including a two-branch neural network based on the training data subset; The optimization module is used to input the verification data subset into the first neural network model to optimize the model parameters to obtain a second neural network model.

Images