An EVM evaluation method and system based on a convolutional neural network
By constructing a feature tensor and using a convolutional neural network for multi-round error pattern recognition and EVM correction, the instability problem of existing EVM evaluation methods in multi-error coupling scenarios is solved, and stable and reliable EVM evaluation under high-order modulation conditions is achieved.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- CHENGDU SIPAI TECH CO LTD
- Filing Date
- 2026-05-08
- Publication Date
- 2026-07-21
Smart Images
Figure CN122437787A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of wireless communication signal processing technology, and more specifically to an EVM evaluation method and system based on convolutional neural networks. Background Technology
[0002] In wireless communication systems, the quality of modulated signals is typically evaluated using the Error Vector Magnitude (EVM). EVM reflects the deviation of received symbols from the ideal constellation point and is a crucial indicator for measuring modulation accuracy, link performance, and equipment consistency. In practical engineering, EVM is widely used in transmitter commissioning, receiver performance evaluation, and system-level optimization, especially in high-order modulation scenarios (such as 64QAM and 256QAM), where higher requirements are placed on the accuracy and stability of EVM.
[0003] Existing EVM evaluation methods are mostly based on direct differential calculations between the received symbol and the reference constellation points, obtaining the overall EVM value by statistically analyzing the error over a time window. These methods are simple to implement and can obtain relatively reliable evaluation results when the channel is stable and the noise distribution is close to ideal. However, in practical applications, the received signal is often affected by a combination of non-ideal factors, such as quantization error, gain mismatch, phase noise, and residual frequency offset. These factors exhibit different distribution characteristics on the constellation diagram and overlap, making it difficult to distinguish the source of error.
[0004] Furthermore, traditional EVM evaluation methods are prone to significant fluctuations in short time windows or dynamic scenarios. Due to the limited number of statistical samples, the results of a single calculation are highly sensitive to local outliers, leading to insufficient stability of the evaluated values. Simultaneously, existing methods typically output only a single EVM value, lacking the ability to characterize the error structure and failing to provide a valid basis for subsequent link optimization or parameter adjustment. This problem is even more pronounced under conditions of high-order modulation and low-bit-width receiver links, where the EVM results may deviate from the true level and fail to reflect the dominant error types.
[0005] In recent years, although some studies have attempted to introduce machine learning methods to evaluate signal quality, most of them have remained at the level of directly using constellation diagrams for feature extraction and regression prediction. They have not specifically modeled the error distribution characteristics, nor have they had a constraint mechanism for the stability of the results in the multi-round evaluation process, making it difficult to balance evaluation accuracy and stability.
[0006] Therefore, a new EVM evaluation method is urgently needed to address the problems of unstable evaluation results, difficulty in distinguishing error sources, and lack of convergence determination mechanism in existing EVM evaluation methods under multi-error coupling scenarios. Summary of the Invention
[0007] The purpose of this invention is to provide an EVM evaluation method and system based on convolutional neural networks, so as to at least solve the problems of unstable evaluation results, difficulty in characterizing error structure, and lack of convergence determination mechanism in existing EVM evaluation methods under multiple error coupling conditions.
[0008] To achieve the above objectives, a first aspect of the present invention provides an EVM evaluation method based on a convolutional neural network. The method includes: acquiring a complex baseband symbol sequence and a corresponding reference constellation symbol sequence output by a wireless communication receiver to determine an error vector sequence, and performing compensation processing on the complex baseband symbol sequence to obtain a corrected symbol sequence; constructing a feature tensor based on the corrected symbol sequence and the error vector sequence, and performing distribution adjustment processing on the feature tensor to obtain a balanced feature tensor; inputting the balanced feature tensor into a convolutional neural network model, performing multi-round error pattern recognition and EVM correction processing to obtain a corresponding corrected EVM sequence; calculating the change between adjacent corrected EVMs based on the corrected EVM sequence, and outputting the corresponding corrected EVM as an EVM evaluation result when the change satisfies a stability criterion.
[0009] Optionally, obtaining the complex baseband symbol sequence and the corresponding reference constellation symbol sequence output by the wireless communication receiver to determine the error vector sequence includes: segmenting the complex baseband signal output by the wireless communication receiver according to the symbol period, and selecting sampling points as corresponding symbol values within each symbol period to construct a complex baseband symbol sequence; mapping each complex baseband symbol to the constellation point with the smallest distance according to the decision boundary of the modulation constellation based on the modulation scheme corresponding to the complex baseband symbol sequence to determine the reference constellation symbol sequence; and performing a symbol-by-symbol complex difference operation based on the complex baseband symbol sequence and the reference constellation symbol sequence to determine the error vector sequence.
[0010] Optionally, compensation processing is performed on the complex baseband symbol sequence to obtain a corrected symbol sequence, including: calculating the corresponding amplitude for each complex baseband symbol in the complex baseband symbol sequence, and performing amplitude segmentation statistical processing on the complex baseband symbol sequence based on the amplitude to obtain the symbol distribution characteristics within each amplitude interval; determining the distribution change trend corresponding to each amplitude interval based on the symbol distribution characteristics within each amplitude interval, and calculating the corresponding compensation offset based on the distribution change trend; and performing amplitude direction correction processing on each complex baseband symbol in the complex baseband symbol sequence based on the compensation offset to obtain the corrected symbol sequence.
[0011] Optionally, constructing a feature tensor based on the correction symbol sequence and the error vector sequence includes: performing two-dimensional coordinate mapping processing on each complex symbol in the correction symbol sequence to construct a constellation distribution grid based on the in-phase and orthogonal components of each complex symbol, and counting the number of symbols in each grid cell to obtain a density matrix for characterizing the constellation distribution; extracting amplitude and phase angle information from each error vector in the error vector sequence, and performing partitioned statistical processing on the error vector based on the amplitude and phase angle information to obtain a distribution matrix for characterizing the error structure; and performing channel combination processing based on the density matrix and the distribution matrix to obtain the corresponding feature tensor.
[0012] Optionally, channel combination processing is performed based on the density matrix and the distribution matrix to obtain the corresponding feature tensor, including: performing normalization processing on the density matrix to obtain a first feature channel, used to characterize the relative density of the constellation distribution; performing normalization processing on the amplitude statistics in the distribution matrix to obtain a second feature channel, used to characterize the error amplitude distribution; performing angle partitioning encoding processing on the direction statistics in the distribution matrix to obtain a third feature channel, used to characterize the error direction distribution; and performing concatenation processing based on the first feature channel, the second feature channel, and the third feature channel according to the channel dimension to obtain the corresponding feature tensor.
[0013] Optionally, a distribution adjustment process is performed on the feature tensor to obtain a balanced feature tensor, including: calculating the distribution function of each feature region based on the density information corresponding to the constellation distribution in the feature tensor, and determining the corresponding cumulative distribution relationship based on the distribution function; constructing a feature mapping relationship based on the cumulative distribution relationship, and performing coordinate remapping processing on each feature region in the feature tensor according to the feature mapping relationship to adjust the distribution density of each feature region; and constructing a balanced feature tensor based on the remapped feature data.
[0014] Optionally, the balanced feature tensor is input into a convolutional neural network model, and multiple rounds of error pattern recognition and EVM correction processing are performed to obtain a corresponding corrected EVM sequence. This includes: inputting the balanced feature tensor as initial input data into the convolutional neural network model, performing error pattern recognition processing to obtain the corresponding dominant error pattern, and correcting the baseline EVM based on the dominant error pattern to obtain a first corrected EVM; performing an update process based on the balanced feature tensor to obtain an updated balanced feature tensor, and re-inputting the updated balanced feature tensor into the convolutional neural network model to perform error pattern recognition and EVM correction processing to obtain a next corrected EVM; repeatedly performing the balanced feature tensor update process and the error pattern recognition and EVM correction process to obtain multiple corrected EVMs, and arranging each corrected EVM according to the processing order to construct a corrected EVM sequence.
[0015] Optionally, an update process is performed based on the balanced feature tensor to obtain an updated balanced feature tensor, including: dividing the balanced feature tensor into multiple feature sub-regions according to spatial partitioning rules, and calculating anomaly indicators based on the error distribution corresponding to each feature sub-region to obtain the anomaly degree of each feature sub-region; sorting the feature sub-regions according to the anomaly indicators of each feature sub-region, and selecting the top n feature sub-regions in terms of anomaly indicators, or selecting feature sub-regions in which the anomaly indicators are greater than a threshold; determining the corresponding symbol data set based on the selected feature sub-regions, and performing a supplementary sampling process on the symbol data set to obtain supplementary symbol data; and performing an update process on the balanced feature tensor based on the supplementary symbol data to obtain an updated balanced feature tensor.
[0016] Optionally, the calculation of the change between adjacent modified EVMs based on the modified EVM sequence, and the output of the corresponding modified EVM as the EVM evaluation result when the change satisfies the stability criterion, includes: calculating the difference between two adjacent modified EVMs according to the processing order based on the modified EVM sequence, and performing absolute value processing on the difference to obtain the corresponding change sequence; determining m consecutive change quantities based on the change sequence, where m is an integer greater than or equal to 2, and determining whether the m consecutive change quantities are all less than a threshold; when the m consecutive change quantities are all less than the threshold, determining that the modified EVM sequence satisfies the stability criterion, and selecting the last modified EVM corresponding to the m consecutive change quantities as the EVM evaluation result.
[0017] A second aspect of the present invention provides an EVM evaluation system based on a convolutional neural network. The system includes: a data acquisition unit, configured to acquire a complex baseband symbol sequence and a corresponding reference constellation symbol sequence output by a wireless communication receiver to determine an error vector sequence, and perform compensation processing on the complex baseband symbol sequence to obtain a corrected symbol sequence; a processing unit, configured to construct a feature tensor based on the corrected symbol sequence and the error vector sequence, and perform distribution adjustment processing on the feature tensor to obtain a balanced feature tensor; an identification unit, configured to input the balanced feature tensor into a convolutional neural network model, perform multi-round error pattern recognition and EVM correction processing to obtain a corresponding corrected EVM sequence; and a result output unit, configured to calculate the change between adjacent corrected EVMs based on the corrected EVM sequence, and output the corresponding corrected EVM as an EVM evaluation result when the change satisfies a stability criterion.
[0018] Through the above technical solution, this invention compensates for the complex baseband symbol sequence and constructs an error vector sequence by combining it with a reference constellation symbol sequence. This weakens the impact of quantization errors and uneven amplitude distribution on subsequent processing, thereby improving the consistency of input data. Furthermore, it constructs a feature tensor based on the correction symbol sequence and the error vector sequence, and performs distribution adjustment processing to ensure a balanced representation of the constellation distribution and error structure in the feature space, which is beneficial for improving the effectiveness of feature representation. On this basis, a convolutional neural network is introduced to perform multi-round error pattern recognition and EVM correction processing, resulting in a continuous corrected EVM sequence, thus avoiding random fluctuations caused by single evaluations. Finally, by determining the change amount of the corrected EVM sequence, convergence constraints on the evaluation results are achieved, enabling the output of stable and reliable EVM evaluation results under multiple error coupling conditions.
[0019] Other features and advantages of the embodiments of the present invention will be described in detail in the following detailed description section. Attached Figure Description
[0020] The accompanying drawings are provided to further illustrate embodiments of the present invention and form part of the specification. They are used together with the following detailed description to explain the embodiments of the present invention, but do not constitute a limitation thereof. In the drawings: Figure 1 This is a flowchart of the steps of an EVM evaluation method based on a convolutional neural network provided in one embodiment of the present invention; Figure 2 This is a schematic diagram of the structure of a convolutional neural network model provided in one embodiment of the present invention; Figure 3 This is a system architecture diagram of an EVM evaluation system based on a convolutional neural network provided in one embodiment of the present invention; Figure 4 This is an internal structural diagram of a computer device provided in one embodiment of the present invention. Detailed Implementation
[0021] The specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings. It should be understood that the specific embodiments described herein are for illustration and explanation only and are not intended to limit the present invention.
[0022] like Figure 1 As shown, embodiments of the present invention provide an EVM evaluation method based on a convolutional neural network, the method comprising: Step S1: Obtain the complex baseband symbol sequence and the corresponding reference constellation symbol sequence output by the wireless communication receiver to determine the error vector sequence, and perform compensation processing on the complex baseband symbol sequence to obtain the corrected symbol sequence.
[0023] Specifically, obtaining the complex baseband symbol sequence and the corresponding reference constellation symbol sequence output by the wireless communication receiver to determine the error vector sequence includes: segmenting the complex baseband signal output by the wireless communication receiver according to the symbol period, and selecting sampling points as corresponding symbol values within each symbol period to construct the complex baseband symbol sequence; mapping each complex baseband symbol to the constellation point with the smallest distance according to the decision boundary of the modulation constellation based on the modulation scheme corresponding to the complex baseband symbol sequence to determine the reference constellation symbol sequence; and performing a symbol-by-symbol complex difference operation based on the complex baseband symbol sequence and the reference constellation symbol sequence to determine the error vector sequence.
[0024] Further, compensation processing is performed on the complex baseband symbol sequence to obtain a corrected symbol sequence, including: calculating the corresponding amplitude for each complex baseband symbol in the complex baseband symbol sequence, and performing amplitude segmentation statistical processing on the complex baseband symbol sequence based on the amplitude to obtain the symbol distribution characteristics within each amplitude interval; determining the distribution change trend corresponding to each amplitude interval based on the symbol distribution characteristics within each amplitude interval, and calculating the corresponding compensation offset based on the distribution change trend; and performing amplitude direction correction processing on each complex baseband symbol in the complex baseband symbol sequence based on the compensation offset to obtain the corrected symbol sequence.
[0025] In this embodiment of the invention, the starting point for EVM evaluation does not directly focus on error calculation, but rather on the structured reconstruction of the receiver's output signal. After completing down-conversion and basic demodulation, the wireless communication receiver outputs a complex baseband signal, which is continuously distributed along the time axis, but its effective information is actually carried in units of symbols. Therefore, it is necessary to convert the continuous signal into a symbol-level expression for subsequent constellation mapping and error analysis.
[0026] In practical implementation, the complex baseband signal can be segmented according to the system's symbol period. The symbol period can be derived from system configuration parameters or provided by the synchronization module. Within each symbol period, a representative sampling point is selected as the complex value of that symbol. This sampling point can be a single point at the symbol center or a weighted average within a certain range, thus achieving a balance between noise immunity and response speed. Through this process, the original continuous signal is regularized into a discrete complex baseband symbol sequence.
[0027] After obtaining the complex baseband symbol sequence, it is necessary to establish its correspondence with the ideal modulation symbol. This correspondence is determined by the constellation structure corresponding to the modulation scheme. Specifically, for each complex baseband symbol, based on its position in the complex plane, the distance to all possible constellation points in the modulation constellation is calculated, and the constellation point with the smallest distance is selected as its corresponding reference constellation symbol. This process is essentially a minimum distance decision, and its result constitutes the reference constellation symbol sequence. Through this mapping relationship, a one-to-one correspondence is established between the actual received symbols and the ideal symbols, providing a basis for error calculation.
[0028] After establishing the correspondence between symbols and reference constellation points, the error vector can be calculated symbol by symbol. The error vector reflects the offset of the received symbol relative to the ideal constellation points, and its direction and amplitude together characterize the type and degree of signal distortion. By performing a differential operation on the entire symbol sequence, a complete error vector sequence can be obtained.
[0029] In practical engineering environments, complex baseband symbol sequences are often affected by quantization errors, gain nonlinearity, and dynamic range limitations, resulting in a non-uniform distribution of symbols across different amplitude ranges. This non-uniformity directly impacts the stability of subsequent feature construction; therefore, compensation processing of the symbol sequence is necessary before error analysis. This compensation is based on structural corrections made according to the symbol distribution characteristics.
[0030] Specifically, the amplitude of complex baseband symbols is calculated, and the symbols are segmented and statistically analyzed according to amplitude intervals. By statistically analyzing the distribution of the number of symbols within each amplitude interval, the distribution characteristics of the symbols in the amplitude space can be obtained. Furthermore, by analyzing the distribution differences between adjacent amplitude intervals, the changing trend of the symbol distribution can be obtained. This trend reflects the degree of concentration or sparsity of symbols in certain amplitude intervals, thus indirectly reflecting the quantization compression or expansion effect.
[0031] Based on the aforementioned distribution trends, the corresponding compensation offset can be calculated. The direction of the compensation offset is adjusted along the symbol amplitude direction, and its magnitude is related to the degree of distribution change. For densely symbolized regions, an expansion offset is appropriately introduced to alleviate the compression effect; for sparsely symbolized regions, a contraction offset is used to improve distribution consistency. By applying this amplitude direction correction to all symbols, a corrected symbol sequence can be obtained. This sequence is statistically closer to the ideal distribution, thus providing a more stable input for subsequent feature construction.
[0032] Step S2: Construct a feature tensor based on the correction symbol sequence and the error vector sequence, and perform distribution adjustment processing on the feature tensor to obtain a balanced feature tensor.
[0033] Specifically, constructing a feature tensor based on the correction symbol sequence and the error vector sequence includes: performing two-dimensional coordinate mapping processing on each complex symbol in the correction symbol sequence to construct a constellation distribution grid based on the in-phase and orthogonal components of each complex symbol, and counting the number of symbols in each grid cell to obtain a density matrix for characterizing the constellation distribution; extracting amplitude and phase angle information from each error vector in the error vector sequence, and performing partitioned statistical processing on the error vector based on the amplitude and phase angle information to obtain a distribution matrix for characterizing the error structure; and performing channel combination processing based on the density matrix and the distribution matrix to obtain the corresponding feature tensor.
[0034] Furthermore, based on the density matrix and the distribution matrix, channel combination processing is performed to obtain the corresponding feature tensor, including: performing normalization processing on the density matrix to obtain a first feature channel, used to characterize the relative density of the constellation distribution; performing normalization processing on the amplitude statistics in the distribution matrix to obtain a second feature channel, used to characterize the error amplitude distribution; performing angle partitioning encoding processing on the direction statistics in the distribution matrix to obtain a third feature channel, used to characterize the error direction distribution; and performing concatenation processing on the first feature channel, the second feature channel, and the third feature channel according to the channel dimension to obtain the corresponding feature tensor.
[0035] Specifically, performing distribution adjustment processing on the feature tensor to obtain a balanced feature tensor includes: calculating the distribution function of each feature region based on the density information corresponding to the constellation distribution in the feature tensor, and determining the corresponding cumulative distribution relationship based on the distribution function; constructing a feature mapping relationship based on the cumulative distribution relationship, and performing coordinate remapping processing on each feature region in the feature tensor according to the feature mapping relationship to adjust the distribution density of each feature region; and constructing a balanced feature tensor based on the remapped feature data.
[0036] In this embodiment of the invention, the correction symbol sequence is essentially distributed in the complex plane, and its statistical form is a constellation diagram in the traditional sense. However, directly processing it using point cloud form is difficult to unify the input structure, and local density differences will be amplified by the model. Therefore, mapping complex symbols from continuous space to a discrete grid is a more stable representation.
[0037] In practical implementation, a two-dimensional mesh can be constructed on the complex plane using in-phase and quadrature components as coordinate axes. The mesh resolution can be set according to the modulation order and signal-to-noise ratio; for example, a higher resolution is used in 256QAM scenarios to preserve details. Each complex symbol falls into its corresponding mesh cell according to its coordinates, and the number of symbols in each cell is counted to form a density matrix. This matrix reflects the degree of symbol clustering in space, and its distribution pattern can directly reflect the constellation distortion.
[0038] Error vector sequences provide another type of information. Unlike the position of the symbol itself, error vectors emphasize the offset direction and amplitude structure. After decomposing the error vector into amplitude and phase angle, it can be partitioned and statistically analyzed in polar coordinate space. For example, the amplitude can be divided into several intervals, and the phase angle into equal-angle sectors, with the number of errors or the cumulative amplitude counted within each sector. The resulting distribution matrix is no longer a simple set of errors, but an expression of error patterns with structural meaning.
[0039] The two types of matrices correspond to signal distribution and error structure, respectively. Using either type alone is incomplete; only by combining them can a comprehensive description of the signal state be formed. During combination, they are not directly superimposed but organized in a channel format, aligning different types of information spatially and separating them semantically. The density matrix, after normalization, serves as the first channel, its value reflecting relative density rather than absolute counts. The error amplitude distribution matrix, after normalization, serves as the second channel, used to express the error intensity distribution. The error direction distribution forms the third channel through angular partitioning encoding. Its encoding method can be a histogram of equally divided angular intervals or a periodic encoding method; the specific form is not limited.
[0040] The three types of channels maintain a consistent spatial dimension, and are concatenated along the channel dimension to form a feature tensor. At this point, the feature tensor already possesses a structure similar to an image, but each channel has a clear physical meaning, rather than simple grayscale information. This construction method allows convolution operations to learn the correlation patterns between different features while preserving spatial relationships.
[0041] Based on the aforementioned feature tensor, further distribution adjustment is needed. This is because the distribution density of constellation points is not uniform across different regions; inner-layer points are often denser, while outer-layer points are relatively sparse. This non-uniformity manifests as a data distribution bias in the feature space, which can easily cause the model to become overly sensitive to high-density regions during training or inference. To alleviate this problem, the feature tensor needs to be redistributed.
[0042] The core of distribution adjustment lies in establishing a mapping relationship from the original distribution to the equilibrium distribution. The probability distribution of each grid cell can be obtained through the statistical density matrix, and the cumulative distribution relationship can be calculated accordingly. The cumulative distribution function has monotonicity and can serve as a basis for coordinate transformation. Mapping the original coordinates to the cumulative probability space can achieve an effect similar to histogram balancing, resulting in a more uniform representation of different regions in the new space.
[0043] This mapping applies not only to the density channel but also simultaneously to the error correlation channel, thus maintaining spatial consistency across channels. After remapping, the distribution of each feature region tends to be more balanced, and the sample weights obtained by the model in different regions are more consistent, which helps to improve the overall discrimination stability.
[0044] It should be noted that the grid division method, the number of amplitude intervals, the granularity of angle partitions, and the calculation method of the distribution function can all be adjusted according to the specific application scenario. For example, in a low signal-to-noise ratio environment, the grid resolution can be appropriately reduced to enhance robustness. In a high-order modulation environment, the angle partitions can be refined to improve the ability to represent the error direction. The above adjustments are all optional implementations of this embodiment.
[0045] In the specific calculation process, let the corrected symbol sequence obtained after the aforementioned compensation processing be: ; in, Indicates the first One correction complex number symbol; Indicates the first The in-phase component of a corrected complex symbol; Indicates the first Orthogonal components of a complex number sign correction; Represents the imaginary unit; This indicates the total number of symbols in the current processing window; Indicates the symbol number.
[0046] To map the correction symbol sequence to a regular two-dimensional grid space, discrete grids can be established on the in-phase axis and the orthogonal axis, respectively. Let the lower bound of the in-phase axis be... The grid spacing of the phase axis is The lower bound of the orthogonal axis is: The grid spacing of the orthogonal axes is Then the first The grid coordinates corresponding to each correction complex number symbol can be represented as: ; in, Indicates the first The grid index corresponding to each correction complex number symbol in the in-phase axis direction; Indicates the first The grid index corresponding to the corrected complex number sign in the orthogonal axis direction; Indicates the starting value for mesh generation on the same phase axis; Indicates the starting value for orthogonal axis mesh generation; Indicates the interval between the centers of adjacent grids on the same phase axis; This indicates the interval between the centers of adjacent grids on the orthogonal axis.
[0047] After obtaining the grid positions corresponding to all the corrected complex symbols, the number of symbols in each grid cell can be counted to obtain the density matrix: ; in, Indicates grid coordinates as The number of symbols in the grid cells; Represents the grid index variable along the phase axis; Represents the grid index variable along the orthogonal axis; This represents an indicator function, which takes the value 1 when the condition inside the parentheses is true, and takes the value 0 otherwise. Indicates the first The coordinates of the corrected complex number symbols are: Grid cells.
[0048] To avoid the absolute number of symbols affecting subsequent processing, the density matrix can be normalized to obtain the first feature channel: ; in, Indicates grid coordinates as The value of the first feature channel at that location is used to characterize the relative symbol density at that position; This represents the total number of symbols in all grid cells, which in this embodiment is equal to the total number of symbols in the current processing window. .
[0049] For the error vector sequence, let the reference constellation symbol sequence be... Then the first The error vector corresponding to each symbol can be represented as: ; in, Indicates the first One error vector; Indicates the relationship with the first The reference constellation symbol corresponding to each correction complex number symbol; Indicates the first A correction complex number symbol.
[0050] The magnitude and phase angle of the error vector can be written as: ; ; in, Indicates the first The magnitude of each error vector; Represents the complex error vector The modulus; Indicates the first Phase angle of each error vector; Represents the complex error vector Argument in the complex plane.
[0051] To construct the error structure distribution matrix, the error vector magnitude can be divided into... The error vector phase angle is divided into several amplitude intervals. Let the ... Each amplitude range is denoted as , No. The angle intervals are denoted as The error distribution matrix can then be expressed as: ; in, Indicates that the amplitude belongs to the first The interval and the phase angle belongs to the th interval. The number of error vectors in each interval; Indicates the amplitude range index. ; Indicates the angle range index. ; Indicates the first One error amplitude range; Indicates the first One error phase angle interval; Indicates the first The magnitude of the error vector falls within the first... A range; Indicates the first The phase angle of the error vector is entered at the first error vector. A range.
[0052] After normalizing the error distribution matrix, we can obtain the characteristic channels used to characterize the error magnitude distribution and error direction distribution. The second characteristic channel can be written as: ; in, Indicates the amplitude range index as Angle interval index is The value of the second feature channel at that location; This represents the total number of error vectors in all error partitions.
[0053] In one alternative implementation, to highlight the directional distribution characteristics, a third feature channel can be formed based on the proportion of each angle interval: ; in, Indicates the first The proportion of directional distribution corresponding to each angle interval; Indicates the phase corner entering the first The total number of error vectors in each interval; This represents the total number of error vectors. If it is necessary to maintain the same two-dimensional spatial dimensions as the first and second feature channels, then... The matrix is copied and expanded along the amplitude range to form a two-dimensional directional encoding matrix.
[0054] After obtaining the first, second, and third feature channels, they can be concatenated along the channel dimension to form a feature tensor: ; in, Represents the characteristic tensor; This indicates channel concatenation operations; The first feature channel represents the relative density of constellation distribution; This represents the second characteristic channel corresponding to the error amplitude distribution; This represents the third feature channel corresponding to the error direction distribution.
[0055] To perform distribution adjustment on the feature tensor, it is necessary to first calculate the distribution function based on the constellation distribution density information within it. Let the constellation distribution region be divided into... The nth feature region, then the nth The probability distribution of each feature region can be expressed as: ; in, Indicates the first The distribution function values corresponding to each feature region; This represents the total number of characteristic regions along the same phase axis. This represents the total number of feature regions along the orthogonal axis. Because... It has been normalized, therefore, in actual implementation, Can be directly taken as The above formula is used to explain its probabilistic meaning.
[0056] Based on this, the cumulative distribution relationship is constructed: ; in, This indicates the accumulation from the origin region to the nth... The cumulative distribution value corresponding to each feature region; and These are all intermediate summation index variables.
[0057] Based on the cumulative distribution relationship, a mapping relationship can be established between the original feature region coordinates and the remapped coordinates. Let the remapped coordinates be... Then it can be expressed as: ; ; in, Indicates the coordinate index of the in-phase axis after remapping; Indicates the coordinate index of the orthogonal axis direction after remapping; This represents a one-dimensional cumulative distribution function along the phase axis. This represents a one-dimensional cumulative distribution function along the orthogonal axes; and These represent the maximum index values in the two directions, respectively.
[0058] After obtaining the mapping relationship, the original feature tensor located in Location feature data remapped to Position, to obtain the equilibrium feature tensor: ; in, Represents the equilibrium feature tensor in the remapped coordinates Passing the exam The values at each channel; Represents the original feature tensor in the original coordinates Passing the exam The values at each channel; This represents the channel index, which typically corresponds to any one of the first, second, and third feature channels.
[0059] Step S3: Input the balanced feature tensor into the convolutional neural network model and perform multi-round error pattern recognition and EVM correction processing to obtain the corresponding corrected EVM sequence.
[0060] Specifically, the balanced feature tensor is used as initial input data to the convolutional neural network model, error pattern recognition processing is performed to obtain the corresponding dominant error pattern, and the baseline EVM is corrected based on the dominant error pattern to obtain the first corrected EVM; an update process is performed based on the balanced feature tensor to obtain the updated balanced feature tensor, and the updated balanced feature tensor is re-inputted into the convolutional neural network model to perform error pattern recognition and EVM correction processing to obtain the next corrected EVM; the balanced feature tensor update processing and error pattern recognition and EVM correction processing are repeatedly performed to obtain multiple corrected EVMs, and the corrected EVMs are arranged according to the processing order to construct a corrected EVM sequence.
[0061] Further, an update process is performed based on the balanced feature tensor to obtain an updated balanced feature tensor, including: dividing the balanced feature tensor into multiple feature sub-regions according to spatial partitioning rules, and calculating anomaly indicators based on the error distribution corresponding to each feature sub-region to obtain the anomaly degree of each feature sub-region; sorting the feature sub-regions according to the anomaly indicators of each feature sub-region, and selecting the top n feature sub-regions with the highest anomaly indicators, or selecting feature sub-regions with anomaly indicators greater than a threshold; determining the corresponding symbol data set based on the selected feature sub-regions, and performing a supplementary sampling process on the symbol data set to obtain supplementary symbol data; and performing an update process on the balanced feature tensor based on the supplementary symbol data to obtain an updated balanced feature tensor.
[0062] In this embodiment of the invention, the balanced feature tensor is fed as the initial input into a convolutional neural network model. This model receives spatial data with a multi-channel structure, where different channels carry information such as constellation density, error magnitude, and error direction. The network extracts local spatial features through convolution operations and then combines them between layers to form an overall feature representation. During this process, the spatial structures corresponding to different error patterns exhibit distinguishable statistical characteristics. For example, magnitude compression typically manifests as radial contraction, while phase noise manifests as angular diffusion. The convolutional neural network can capture these differences without explicit modeling and output the corresponding error pattern discrimination results.
[0063] After obtaining the error pattern recognition result, the EVM is not directly output. Instead, the baseline EVM is corrected based on the pattern. The baseline EVM can be understood as an initial value directly calculated based on the error vector sequence, without distinguishing the source of the error. Different error patterns have different characteristics in their impact on the EVM; therefore, by introducing pattern information, the baseline EVM can be adjusted accordingly. This adjustment result is denoted as the first corrected EVM.
[0064] Since the input data still contains local anomalies, the results of a single iteration may be affected by local feature biases. Therefore, an iterative mechanism is needed. In subsequent steps, the input feature tensor is updated, causing changes in the input for the next round, thereby gradually eliminating the impact of local anomalies on the overall evaluation.
[0065] A balanced feature tensor can be viewed as spatially divided into several feature sub-regions, each corresponding to a portion of the sign distribution and error structure. Within these regions, the error distribution is often uneven, with some regions exhibiting concentrated large deviations. By quantizing these regions, the parts requiring focused correction can be identified.
[0066] The calculation of anomaly indicators relies on error distribution information. Specifically, it considers both the dispersion of error amplitude and the diffusion of error direction. For example, within a certain characteristic sub-region, if the error amplitude fluctuates significantly and the direction distribution is relatively dispersed, it indicates strong instability in that region. By calculating and ranking the anomaly indicators for all regions, the relative distribution of anomaly severity can be obtained.
[0067] After identifying the outlier regions, the process reverts to the original symbolic level. Each feature sub-region corresponds to a set of symbolic data, which constitute the statistical basis of that region. By performing supplementary sampling on these symbolic data, the sample size of that region can be increased, allowing it to occupy a more reasonable weight when reconstructing the feature tensor. Supplementary sampling methods can include repeated sampling, neighborhood interpolation, or generating new samples based on perturbations, and the specific implementation method can be selected according to system resources and real-time requirements.
[0068] After obtaining the resampled symbolic data, the feature construction and distribution adjustment process needs to be repeated to obtain an updated balanced feature tensor. This updated tensor is then fed back into the convolutional neural network to repeatedly perform error pattern recognition and EVM correction processing, resulting in the next corrected EVM. As iterations proceed, the input data gradually becomes more balanced, the error pattern recognition results gradually stabilize, and the variation of the corrected EVM gradually decreases.
[0069] Through multiple iterations, a set of corrected EVM values arranged in chronological order can be obtained. Constructing these values into a sequence reflects the convergence process of the evaluation results. This sequence is not only used to output the final result but also provides a basis for stability determination.
[0070] In one specific implementation, let the equilibrium feature tensor be: ; in, Represents the initial equilibrium feature tensor; and These represent the spatial dimensions of the characteristic tensor along the in-phase axis and the orthogonal axis, respectively; Indicates the number of channels.
[0071] The convolutional neural network model is denoted as a mapping function. ,in, Indicates model parameters. In the... In the round of iteration, the input is The output includes error mode and corrected EVM: ; in, Indicates the first Error patterns obtained from wheel recognition; Indicates the first Wheel correction EVM.
[0072] Anomalies are calculated at the feature sub-region level. Let the feature tensor be divided into... Sub-region, the first The abnormal indicators corresponding to each region can be represented as: ; in, Indicates the first Abnormal indicators of each characteristic sub-region; This represents the set of error magnitudes within the region; This represents the set of error phase angles within this region; Indicates variance operation; and This is a weighting coefficient used to balance the effects of amplitude and direction dispersion.
[0073] Sort by abnormal indicators and select the top Given a region, obtain the region index set. For sets For each region in the dataset, the corresponding set of symbolic data is denoted as: ; Perform a complement sampling process on this set to obtain an expanded set. ,in, This indicates the supplementary sampling operator.
[0074] Replace or expand the symbol set of all updated regions to the overall symbol set, and reconstruct the feature tensor: ; in, This represents the construction function from the symbol set to the feature tensor.
[0075] Through the above update mechanism, a series of modified EVMs can be obtained: ; This sequence is the modified EVM sequence.
[0076] It should be noted that the calculation method of outlier indicators, the supplementary sampling strategy, and the feature reconstruction method can all be adjusted according to specific application requirements. The key to this implementation method lies in using feature-driven region selection and data augmentation to gradually stabilize the input features during the iteration process, thereby ensuring the convergence and consistency of the EVM evaluation results.
[0077] In one specific implementation, such as Figure 2 The EVM evaluation method based on convolutional neural networks first obtains a balanced feature tensor as input data. This balanced feature tensor consists of a constellation density channel, an error amplitude channel, and an error direction channel, which respectively characterize the spatial distribution of the signal in the complex plane, the error amplitude distribution, and the error direction structure. The input splicing layer performs splicing processing on each channel to form a unified input feature tensor with dimensions H×W×3.
[0078] The input feature tensor is fed into the convolutional feature encoding backbone. The convolutional feature encoding backbone includes multiple convolutional blocks and a downsampling structure. By extracting local spatial features layer by layer, it achieves joint encoding of the constellation distribution structure and the error distribution structure. In this embodiment, the convolutional blocks are 3×3 and combined with batch normalization and nonlinear activation functions to enhance feature representation capabilities; downsampling reduces the spatial dimension and expands the receptive field, thereby obtaining more stable global features. After global average pooling, a shared feature vector is output. This is used for subsequent multitasking.
[0079] In the error pattern recognition path, the shared feature vector The input is a fully connected layer with feature mapping, and the dominant error pattern is output through a softmax classification layer. The error pattern is used to characterize the main error types in the current signal, including but not limited to quantization-dominant error, amplitude imbalance error, phase spread error, or mixed error types. This error pattern is used as conditional information for subsequent EVM correction calculations.
[0080] In the conditional EVM correction path, the baseline EVM calculated based on the error vector sequence is denoted as... and combine it with the shared feature vector and dominant error mode The input is then subjected to joint encoding and fed into the correction function layer to obtain the corresponding correction factor. Furthermore, the correction factor is combined with the baseline EVM to obtain the modified EVM: ; This method allows for targeted correction of the EVM based on different error structures, enabling the EVM results to reflect the true error level.
[0081] In multi-round iterative processing, the corrected EVM obtained in the current round and the corresponding feature information are used to update the balanced feature tensor. The update method may include filtering out error anomaly regions and performing supplementary sampling on the corresponding symbolic data, thereby reconstructing the balanced feature tensor. The updated feature tensor is then input into the convolutional neural network again, and the error pattern recognition and EVM correction processing are repeated to obtain the corrected EVM for the next round. Through multiple iterations, a corrected EVM sequence was obtained. In the output stage, the modified EVM sequence is subjected to change analysis. When the change between adjacent modified EVMs meets the preset stability criterion, the corresponding modified EVM is selected as the final EVM evaluation result.
[0082] It should be noted that the use of a convolutional neural network to process the feature tensor in this embodiment is necessary. Since the input features contain spatial information about constellation distribution and error structure, error patterns often manifest as distribution changes or directional shifts within local regions. Traditional statistical methods are insufficient to effectively characterize such spatial structural features. Convolutional neural networks, through local receptive fields and weight sharing mechanisms, can extract spatially correlated features and distinguish different error patterns, thus providing a reliable basis for subsequent EVM correction.
[0083] Furthermore, the specific structure, parameter configuration, and feature processing methods of each of the above modules can be adjusted according to actual application requirements. For example, the number of convolutional layers, the number of channels, the form of the activation function, and the specific expression of the correction function are not limited to the structure shown in this embodiment. The multi-round iterative update strategy can also be implemented using different anomaly screening and data reconstruction methods. Without departing from the overall technical concept of this application, any equivalent substitutions or modifications to the above technical features should fall within the protection scope of this application.
[0084] Step S4: Calculate the change between adjacent modified EVMs based on the modified EVM sequence, and output the corresponding modified EVM as the EVM evaluation result when the change satisfies the stability criterion.
[0085] Specifically, based on the modified EVM sequence, the difference between two adjacent modified EVMs is calculated according to the processing order, and the absolute value processing is performed on the difference to obtain the corresponding change sequence; based on the change sequence, m consecutive changes are determined, where m is an integer greater than or equal to 2, and it is determined whether the m consecutive changes are all less than a threshold; when the m consecutive changes are all less than the threshold, the modified EVM sequence is determined to meet the stability criterion, and the last modified EVM corresponding to the m consecutive changes is selected as the EVM evaluation result.
[0086] In this embodiment of the invention, the modified EVM sequence essentially reflects the convergence trajectory of the EVM during the iteration process. As iterations proceed, if the input features tend to stabilize, the change between two adjacent modified EVMs should gradually decrease. Based on this characteristic, it is possible to determine whether the sequence has entered the stable region by analyzing the amount of change between adjacent results. Compared to directly setting a fixed number of iterations, this method is more suitable for dynamic convergence processes under different signal conditions.
[0087] In practice, the corrected EVM sequence is arranged in the order of its generation. Each element in the sequence corresponds to the corrected EVM value output in one iteration. For two adjacent results, the difference can be calculated to characterize the degree of change in the current iteration step size. Since EVM itself is a non-negative scalar, the difference may have positive or negative variations; therefore, taking the absolute value of the difference can uniformly describe the magnitude of the change.
[0088] After obtaining the sequence of changes, a continuous interval can be found within this sequence, where the changes are all within a relatively small range. This continuous interval can be understood as a marker that the system has entered a stable state. When the changes in several consecutive iterations are all below a certain threshold, it can be considered that further iterations have limited impact on the result, and the sequence has essentially converged. In this case, continuing iterations is not very meaningful, and the modified EVM corresponding to the end of the current stable interval is directly selected as the final evaluation result.
[0089] It should be noted that the number of consecutive iterations can be adjusted according to actual application requirements. For example, in scenarios with large signal fluctuations, the required number of consecutive iterations can be appropriately increased to avoid misjudgments. In scenarios with high real-time requirements, the number of consecutive iterations can be appropriately reduced to decrease computational overhead. Similarly, the threshold selection can also be related to the system's EVM level; for example, it can be set as a certain percentage of the baseline EVM, rather than being limited to a fixed constant.
[0090] Furthermore, the calculation form of the change is not limited to absolute difference. In some implementations, normalized difference or relative rate of change can be used to accommodate differences in numerical ranges under different power or modulation conditions. This implementation uses the absolute difference form for explanation, the core of which is to determine the convergence state by quantifying the trend of change between adjacent results.
[0091] In one specific embodiment, the corrected EVM sequence obtained through multiple iterations is as follows: ; in, Indicates the first The corrected EVM value output by the round of iterations; For iterative index, , This indicates the maximum number of iterations or the number of iterations currently executed.
[0092] The difference between two consecutive corrected EVM values is defined as: ; in, Indicates the first Wheel and the first The difference between the wheels.
[0093] Perform absolute value processing on the above differences to obtain the sequence of changes: ; in, Indicates the first The change corresponding to the wheel is used to characterize the magnitude of change between two adjacent corrections to the EVM.
[0094] In the sequence of changes, select one of lengths... Continuous subsequences: ; in, Let be the number of continuously changing quantities, satisfying ; Represents the position index in the current sequence, and .
[0095] The stability criterion can be expressed as: ; in, This represents the threshold value for the amount of change, used to limit the upper limit of the change range; This means that all indexes within the specified range are true.
[0096] When the above conditions are met, it is considered to be in the interval The internally corrected EVM change is small enough that the system enters a stable state. At this point, the corresponding final evaluation result is selected as: ,middle, This indicates the output EVM evaluation results; This represents the corrected EVM value corresponding to the last iteration within the stable interval.
[0097] Through the above mechanism, the results of multiple iterations can be transformed into a final output with clear convergence criteria. This method can adaptively determine the output timing under different signal conditions, without relying on a fixed number of iterations, thereby improving the consistency and reliability of the evaluation results. It is important to understand the length of the continuous interval. and threshold All values can be set according to specific system requirements, and their range does not constitute a limitation of this implementation method.
[0098] In other implementations, during multiple iterations, not only are abnormal regions filtered based on error magnitude, but the concentration of error directions within each feature sub-region is further calculated to obtain the corresponding directional consistency index. For regions with low directional consistency, it is assumed that their internal error structure exhibits multi-directional superposition characteristics, which easily interferes with convolutional feature extraction. Therefore, when updating the balanced feature tensor, directional normalization processing is performed on the symbol data corresponding to such regions to make the error directions tend to concentrate in the local region.
[0099] For regions with high directional consistency, the original distribution is maintained, and only the amplitude dimension is updated through supplementary sampling. This approach simultaneously constrains the error amplitude and directional structure during the feature update stage, resulting in stronger spatial consistency in the feature tensors input to the convolutional neural network, which improves the stability of error pattern recognition. This implementation provides a supplementary processing method for feature perturbation caused by the discrete distribution of error directions, and exhibits better adaptability in complex error coupling scenarios.
[0100] like Figure 3As shown, this invention provides an EVM evaluation system based on a convolutional neural network. The system includes: a data acquisition unit, configured to acquire a complex baseband symbol sequence and a corresponding reference constellation symbol sequence output by a wireless communication receiver to determine an error vector sequence, and perform compensation processing on the complex baseband symbol sequence to obtain a corrected symbol sequence; a processing unit, configured to construct a feature tensor based on the corrected symbol sequence and the error vector sequence, and perform distribution adjustment processing on the feature tensor to obtain a balanced feature tensor; an identification unit, configured to input the balanced feature tensor into a convolutional neural network model, perform multi-round error pattern recognition and EVM correction processing to obtain a corresponding corrected EVM sequence; and a result output unit, configured to calculate the change between adjacent corrected EVMs based on the corrected EVM sequence, and output the corresponding corrected EVM as the EVM evaluation result when the change satisfies the stability criterion.
[0101] In one embodiment, a computer device is provided, which may be a server, and its internal structure diagram may be as follows: Figure 4 As shown, the computer device includes a processor A01, a network interface A02, memory (not shown), and a database (not shown) connected via a system bus. The processor A01 provides computing and control capabilities. The memory includes internal memory A03 and a non-volatile storage medium A04. The non-volatile storage medium A04 stores an operating system B01, a computer program B02, and a database (not shown). The internal memory A03 provides an environment for the operation of the operating system B01 and the computer program B02 stored in the non-volatile storage medium A04. The network interface A02 is used for communication with external terminals via a network connection. When the computer program B02 is executed by the processor A01, it implements an EVM evaluation method based on a convolutional neural network.
[0102] Those skilled in the art will understand that all or part of the steps in the methods of the above embodiments can be implemented by a program instructing related hardware. This program is stored in a storage medium and includes several instructions to cause a microcontroller, chip, or processor to execute all or part of the steps of the methods described in the various embodiments of the present invention. The aforementioned storage medium includes various media capable of storing program code, such as a USB flash drive, a portable hard drive, a read-only memory (ROM), a random access memory (RAM), a magnetic disk, or an optical disk.
[0103] The optional embodiments of the present invention have been described in detail above with reference to the accompanying drawings. However, the embodiments of the present invention are not limited to the specific details described above. Within the scope of the technical concept of the embodiments of the present invention, various simple modifications can be made to the technical solutions of the embodiments of the present invention, and these simple modifications all fall within the protection scope of the embodiments of the present invention. It should also be noted that the various specific technical features described in the above specific embodiments can be combined in any suitable manner without contradiction. To avoid unnecessary repetition, the embodiments of the present invention will not further describe the various possible combinations.
[0104] Furthermore, various different embodiments of the present invention can be combined in any way, as long as they do not violate the spirit of the embodiments of the present invention, they should also be regarded as the content disclosed by the embodiments of the present invention.
Claims
1. An EVM evaluation method based on convolutional neural networks, characterized in that, The method includes: The complex baseband symbol sequence and the corresponding reference constellation symbol sequence output by the wireless communication receiver are obtained to determine the error vector sequence, and the complex baseband symbol sequence is compensated to obtain the correction symbol sequence. A feature tensor is constructed based on the correction symbol sequence and the error vector sequence, and a distribution adjustment process is performed on the feature tensor to obtain a balanced feature tensor; The balanced feature tensor is input into the convolutional neural network model, and multiple rounds of error pattern recognition and EVM correction processing are performed to obtain the corresponding corrected EVM sequence. The change between adjacent modified EVMs is calculated based on the modified EVM sequence, and when the change satisfies the stability criterion, the corresponding modified EVM is output as the EVM evaluation result.
2. The EVM evaluation method based on convolutional neural networks according to claim 1, characterized in that, To determine the error vector sequence, the complex baseband symbol sequence output by the wireless communication receiver and the corresponding reference constellation symbol sequence are obtained, including: The complex baseband signal output by the wireless communication receiver is segmented according to the symbol period, and sampling points are selected in each symbol period as the corresponding symbol values to construct a complex baseband symbol sequence. Based on the modulation scheme corresponding to the complex baseband symbol sequence, each complex baseband symbol is mapped to the constellation point with the smallest distance according to the decision boundary of the modulation constellation, so as to determine the reference constellation symbol sequence; Perform symbol-by-symbol complex difference operations based on the complex baseband symbol sequence and the reference constellation symbol sequence to determine the error vector sequence.
3. The EVM evaluation method based on convolutional neural networks according to claim 2, characterized in that, The complex baseband symbol sequence is subjected to compensation processing to obtain a corrected symbol sequence, including: The amplitude of each complex baseband symbol in the complex baseband symbol sequence is calculated, and the amplitude segmentation statistical processing is performed on the complex baseband symbol sequence based on the amplitude to obtain the symbol distribution characteristics in each amplitude interval; Based on the symbol distribution characteristics within each amplitude interval, the distribution change trend corresponding to each amplitude interval is determined, and the corresponding compensation offset is calculated according to the distribution change trend. Based on the compensation offset, the amplitude direction of each complex baseband symbol in the complex baseband symbol sequence is corrected to obtain the corrected symbol sequence.
4. The EVM evaluation method based on convolutional neural networks according to claim 1, characterized in that, Constructing a feature tensor based on the correction symbol sequence and the error vector sequence includes: Two-dimensional coordinate mapping is performed on each complex symbol in the correction symbol sequence to construct a constellation distribution grid based on the in-phase and orthogonal components of each complex symbol, and the number of symbols in each grid cell is counted to obtain a density matrix for characterizing the constellation distribution. The amplitude and phase angle information of each error vector in the error vector sequence are extracted, and the error vectors are subjected to partition statistical processing based on the amplitude and phase angle information to obtain a distribution matrix for characterizing the error structure; Channel combination processing is performed based on the density matrix and the distribution matrix to obtain the corresponding feature tensor.
5. The EVM evaluation method based on convolutional neural networks according to claim 4, characterized in that, Based on the density matrix and the distribution matrix, channel combination processing is performed to obtain the corresponding feature tensor, including: The density matrix is normalized to obtain a first feature channel, which is used to characterize the relative density of the constellation distribution; The amplitude statistics in the distribution matrix are normalized to obtain a second feature channel, which is used to characterize the error amplitude distribution. An angle partitioning encoding process is performed on the directional statistical results in the distribution matrix to obtain a third feature channel, which is used to characterize the error direction distribution; Based on the first feature channel, the second feature channel, and the third feature channel, a concatenation process is performed according to the channel dimension to obtain the corresponding feature tensor.
6. The EVM evaluation method based on convolutional neural networks according to claim 5, characterized in that, Perform distribution adjustment processing on the feature tensor to obtain a balanced feature tensor, including: The distribution function of each feature region is calculated based on the density information corresponding to the constellation distribution in the feature tensor, and the corresponding cumulative distribution relationship is determined according to the distribution function. Based on the cumulative distribution relationship, a feature mapping relationship is constructed, and coordinate remapping processing is performed on each feature region in the feature tensor according to the feature mapping relationship to adjust the distribution density of each feature region; Construct a balanced feature tensor based on the remapped feature data.
7. The EVM evaluation method based on convolutional neural networks according to claim 1, characterized in that, The balanced feature tensor is input into a convolutional neural network model, and multiple rounds of error pattern recognition and EVM correction processing are performed to obtain the corresponding corrected EVM sequence, including: The balanced feature tensor is used as the initial input data to input the convolutional neural network model, and error pattern recognition processing is performed to obtain the corresponding dominant error pattern. The baseline EVM is then corrected based on the dominant error pattern to obtain the first corrected EVM. Based on the balanced feature tensor, an update process is performed to obtain an updated balanced feature tensor. The updated balanced feature tensor is then re-input into the convolutional neural network model to perform error pattern recognition and EVM correction processing to obtain the next corrected EVM. The process of updating the balanced feature tensor and the error pattern recognition and EVM correction are repeatedly performed to obtain multiple corrected EVMs. The corrected EVMs are then arranged in the order of processing to construct a corrected EVM sequence.
8. The EVM evaluation method based on convolutional neural networks according to claim 7, characterized in that, An update process is performed based on the balanced feature tensor to obtain an updated balanced feature tensor, including: The balanced feature tensor is divided into multiple feature sub-regions according to spatial partitioning rules, and anomaly indexes are calculated based on the error distribution corresponding to each feature sub-region to obtain the degree of anomaly of each feature sub-region. The feature sub-regions are sorted according to the abnormal indicators of each feature sub-region, and the top n feature sub-regions with the highest abnormal indicators are selected, or the feature sub-regions with abnormal indicators greater than the threshold are selected. Based on the selected feature sub-region, a corresponding symbol data set is determined, and the symbol data set is subjected to a supplementary sampling process to obtain the supplementary symbol data; The balanced feature tensor is updated based on the sampled symbolic data to obtain the updated balanced feature tensor.
9. The EVM evaluation method based on convolutional neural networks according to claim 8, characterized in that, Based on the modified EVM sequence, the change between adjacent modified EVMs is calculated, and when the change satisfies the stability criterion, the corresponding modified EVM is output as the EVM evaluation result, including: Based on the modified EVM sequence, the difference between two adjacent modified EVMs is calculated according to the processing order, and the absolute value processing is performed on the difference to obtain the corresponding change sequence. Based on the change sequence, determine m consecutive changes, where m is an integer greater than or equal to 2, and determine whether all m consecutive changes are less than a threshold. When all m consecutive changes are less than the threshold, the modified EVM sequence is determined to meet the stability criterion, and the modified EVM corresponding to the last of the m consecutive changes is selected as the EVM evaluation result.
10. An EVM evaluation system based on a convolutional neural network, characterized in that, The system includes: The data acquisition unit is used to acquire the complex baseband symbol sequence and the corresponding reference constellation symbol sequence output by the wireless communication receiver to determine the error vector sequence, and to perform compensation processing on the complex baseband symbol sequence to obtain the correction symbol sequence; The processing unit is configured to construct a feature tensor based on the correction symbol sequence and the error vector sequence, and perform distribution adjustment processing on the feature tensor to obtain a balanced feature tensor; The recognition unit is used to input the balanced feature tensor into the convolutional neural network model and perform multi-round error pattern recognition and EVM correction processing to obtain the corresponding corrected EVM sequence. The result output unit is used to calculate the change between adjacent modified EVMs based on the modified EVM sequence, and output the corresponding modified EVM as the EVM evaluation result when the change satisfies the stability criterion.