Radio frequency coaxial cable line loss degree prediction method based on deep learning

Abstract

This invention discloses a deep learning-based method for predicting the line loss of radio frequency (RF) coaxial cables, comprising the following steps: setting the test frequency band range for the RF coaxial cable and dividing it into frequencies; collecting line loss-related test data; preprocessing the line loss test dataset; performing frequency-based mapping on standardized line loss feature data based on the sub-frequency band set; inputting the frequency band sub-feature dataset and the frequency band test quality index set into the frequency band feature encoding module of the frequency-division evidence deep regression network; inputting the frequency band feature vector set into the frequency band evidence regression module to model the line loss; generating a global line loss prediction result based on the frequency band evidence parameter set and the frequency band test quality index set; and generating a predictive detection result based on the global line loss prediction result and the frequency band evidence parameter set. This invention uses a frequency-division evidence deep regression network to achieve RF coaxial cable line loss prediction.

Description

Technical Field

[0001] This invention relates to the fields of radio frequency testing and intelligent signal processing technology, and in particular to a method for predicting the degree of line loss of radio frequency coaxial cables based on deep learning. Background Technology

[0002] Radio frequency (RF) coaxial cables, as a crucial carrier for RF signal transmission, are widely used in communication systems, test and measurement systems, and RF equipment connection scenarios. Their line loss directly impacts signal transmission quality and system performance stability. Currently, RF coaxial cable line loss detection typically relies on testing equipment such as vector network analyzers. Insertion loss parameters are obtained through frequency sweep testing, and these parameters are then used to evaluate cable performance. This method primarily relies on static test results and is mainly used for factory testing or periodic inspections, making it difficult to predict and analyze line loss trends.

[0003] As equipment complexity and application frequency bands continue to increase, some existing technologies attempt to introduce data modeling or simple regression methods to analyze line loss data. However, most methods only model the overall data across the entire frequency band, failing to fully consider the differences in line loss characteristics across different frequency bands, and lacking a systematic characterization of test data quality and prediction uncertainty. Furthermore, existing methods typically output a single predicted value directly, failing to provide information on prediction reliability, making it difficult to support predictive testing and health status assessment applications, and exhibiting certain limitations in long-term operational monitoring and fault prediction of RF coaxial cables.

[0004] Therefore, how to provide a deep learning-based method for predicting the line loss of radio frequency coaxial cables is a problem that urgently needs to be solved by those skilled in the art. Summary of the Invention

[0005] One objective of this invention is to propose a deep learning-based method for predicting the line loss of radio frequency (RF) coaxial cables. This invention comprehensively utilizes RF parameter testing technology, frequency band division modeling methods, and evidence-based deep regression networks to model and predict the line loss characteristics of RF coaxial cables at different frequency bands, achieving predictive detection of line loss levels. This invention characterizes the differences in line loss across different frequency bands through frequency division processing, introduces evidence parameters to quantify prediction uncertainty, and combines test quality indicators to weightedly fuse the prediction results for each frequency band. It can output predicted line loss values ​​and corresponding prediction reliability information, possessing advantages such as high reliability, strong frequency band adaptability, and applicability to health status assessment.

[0006] The method for predicting the line loss of radio frequency coaxial cables based on deep learning according to embodiments of the present invention includes the following steps: Define the test frequency range for the RF coaxial cable, and divide the test frequency range into sub-frequency bands based on the frequency sensitivity criterion to generate a set of sub-frequency bands; The radio frequency parameters of the coaxial cable under test are tested within the test frequency band, and line loss related test data are collected to generate a line loss test dataset. Preprocess the line loss test dataset to generate standardized line loss feature data; Based on the sub-band set, the standardized line loss feature data is divided into frequencies and mapped to construct a frequency band sub-feature dataset, and a set of frequency band test quality indicators is extracted. The frequency band sub-feature dataset and the frequency band test quality index set are input into the frequency band feature encoding module of the frequency division evidence deep regression network to generate a set of frequency band feature vectors; The set of frequency band feature vectors is input into the frequency band evidence regression module of the frequency division evidence deep regression network to model the degree of line loss and output the line loss sub-prediction results and the set of frequency band evidence parameters. In the frequency division evidence deep regression network, a global line loss prediction result is generated based on the frequency band evidence parameter set and the frequency band test quality index set. Based on the global line loss prediction results and the frequency band evidence parameter set, predictive detection results are generated.

[0007] Optionally, the step of setting the test frequency band range of the RF coaxial cable and dividing the test frequency band range into sub-band sets based on frequency sensitivity criteria specifically includes: Set the test frequency range for the line loss of the RF coaxial cable. Within the test frequency range, perform line loss related tests on the RF coaxial cable under test to generate initial line loss curve data. Based on the initial line loss curve data, a frequency sensitivity criterion is constructed and defined as a frequency sensitivity index. Based on the preset frequency sensitivity threshold and combined with the frequency sensitivity criteria, the frequency sensitivity indicators corresponding to each frequency point within the test frequency band are compared one by one to select the frequency sensitive interval. Using the boundary frequencies of the frequency-sensitive interval and the endpoint frequencies of the test frequency band as frequency division nodes, the test frequency band is divided based on the frequency sensitivity criterion to generate multiple continuous sub-frequency bands. All sub-bands are combined in ascending order of frequency to generate a sub-band set.

[0008] Optionally, the step of performing RF parameter tests on the RF coaxial cable under test within the test frequency band, collecting line loss-related test data, and generating a line loss test dataset specifically includes: Within the test frequency band, RF parameters of the RF coaxial cable under test are tested, and the insertion loss parameters of the RF coaxial cable are collected at each test frequency. Based on the insertion loss parameter, determine the line loss value of the RF coaxial cable at each test frequency; The line loss values ​​corresponding to each test frequency within the test frequency band are sorted in order from low to high according to the test frequency to construct a line loss test data sequence. Based on the sub-band set, the line loss test data sequence is mapped to the frequency range corresponding to each sub-band, and the line loss test data belonging to each sub-band is collected to form a line loss test dataset.

[0009] Optionally, the preprocessing includes frequency alignment, resampling, noise suppression, and amplitude normalization.

[0010] Optionally, the step of performing frequency-based mapping on standardized line loss feature data based on sub-band sets to construct frequency band sub-feature datasets and extracting frequency band test quality indicator sets specifically includes: Obtain standardized line loss characteristic data and read the aligned frequency sequence corresponding to each standardized line loss characteristic data; Based on the sub-band set, each aligned frequency in the aligned frequency sequence is mapped to construct a frequency band sub-feature dataset; For each frequency band sub-feature dataset, calculate the frequency band line loss dispersion index and the frequency band fluctuation index; Based on the frequency band line loss dispersion index and the frequency band fluctuation index, a set of frequency band test quality indices corresponding to each sub-frequency band is constructed.

[0011] Optionally, the step of inputting the frequency band sub-feature dataset and the frequency band test quality index set into the frequency band feature encoding module of the frequency-division evidence deep regression network to generate the frequency band feature vector set specifically includes: Obtain the frequency band sub-feature dataset and the frequency band test quality index set. For each sub-frequency band, construct the frequency band input sequence in ascending order of frequency. For the frequency band input sequence, a frequency position encoding vector is constructed for each aligned frequency point and combined with the corresponding normalized line loss value to generate a frequency band point-level embedding sequence; The quality indicators of the frequency band test are vectorized to obtain the quality vector, and the quality modulation parameters are generated based on the quality vector to obtain the quality gate coefficient vector and the quality bias vector. The frequency band point-level embedding sequence is input into the multi-scale frequency domain coding unit of the frequency band feature coding module of the frequency division evidence deep regression network to generate a multi-scale coded feature sequence. Based on the quality gating coefficient vector and the quality bias vector, the multi-scale encoded feature sequence is subjected to quality gating modulation processing to generate a quality modulated feature sequence. Attention convergence processing is performed on the quality modulation feature sequence to generate frequency band feature vectors. The frequency band feature vectors corresponding to all sub-frequency bands are then aggregated to generate a frequency band feature vector set.

[0012] Optionally, the step of inputting the set of frequency band feature vectors into the frequency band evidence regression module of the frequency-division evidence deep regression network to model the degree of line loss and output the line loss sub-prediction results and the set of frequency band evidence parameters specifically includes: Obtain the set of frequency band feature vectors, input the frequency band feature vectors into the line loss regression head in the frequency band evidence regression module of the frequency sub-evidence deep regression network, and generate line loss sub-prediction results; The frequency band feature vector is input into the evidence parameter regression head of the frequency band evidence regression module of the frequency sub-frequency evidence deep regression network to generate the frequency band evidence parameter vector, and the frequency band evidence parameter vectors corresponding to all sub-frequency bands are collected to construct the frequency band evidence parameter vector set. The frequency band evidence parameter vector is nonnegated to obtain a nonnegative evidence parameter vector. Based on the nonnegative evidence parameter vector, a set of frequency band evidence parameters is constructed. The line loss sub-prediction results corresponding to all sub-frequency bands are aggregated to generate a set of line loss sub-prediction results.

[0013] Optionally, the step of generating a global line loss prediction result based on the frequency band evidence parameter set and the frequency band test quality index set in the frequency division evidence deep regression network specifically includes: Obtain the set of sub-prediction results for line loss, the set of frequency band evidence parameters, and the set of frequency band test quality indicators; For each sub-band, calculate the sub-band evidence strength scalar based on the band evidence parameter vector; The quality strength scalar is calculated based on the frequency band test quality index, and the quality strength scalar is used as the quality convergence value of the frequency band test quality index. Based on the unnormalized evidence weights of the evidence strength scalar and the quality strength scalar, the unnormalized evidence weights are normalized to obtain the evidence weights. Based on the preset threshold for changes in evidence weights of adjacent sub-bands, change constraint processing is performed on the evidence weights corresponding to adjacent sub-bands; Based on the preset line loss consistency constraint condition with frequency, the line loss sub-prediction results are subjected to consistency constraint processing. Based on the evidence weights of each sub-band after the constraint processing of the change in evidence weights of adjacent sub-bands, the sub-prediction results of line loss of each sub-band after the constraint processing of line loss with frequency consistency are subjected to weighted fusion processing to generate the global line loss degree prediction result.

[0014] Optionally, the generation of predictive detection results based on the global line loss prediction results and the frequency band evidence parameter set specifically includes: Based on the set of frequency band evidence parameters, the corresponding prediction confidence index is calculated for each sub-frequency band to obtain the confidence value. Based on a preset confidence threshold, the confidence values ​​corresponding to each sub-band are compared. When the confidence value is lower than the confidence threshold, the sub-band is marked as a low-confidence sub-band and a low-confidence sub-band identifier is generated. The global line loss prediction results, confidence values, and low-confidence sub-band identifiers are integrated to construct a line loss prediction result set; The predicted line loss results are output as predictive test results for RF coaxial cables.

[0015] The beneficial effects of this invention are: This invention systematically processes line loss data collected from RF coaxial cables within the test frequency band, expanding the traditional static-based line loss detection method into a data-modeling-based line loss prediction method. By performing frequency alignment, resampling, noise suppression, and normalization on the line loss test data, and combining this with a sub-band division mechanism, this invention can stably acquire structured line loss characteristic data under wide frequency band conditions. This provides a consistent input foundation for subsequent modeling, thereby improving the adaptability of the line loss prediction process to changes in test conditions.

[0016] Building upon this foundation, this invention introduces a frequency-division evidence deep regression network as the core modeling structure. By performing frequency-division encoding and independent regression modeling on the line loss characteristics of different sub-bands, it achieves targeted prediction of the line loss degree in each sub-band. While outputting the line loss sub-prediction results, the network simultaneously generates corresponding frequency band evidence parameters, ensuring that the prediction results not only include numerical values ​​of the line loss degree but also a quantitative representation of prediction uncertainty. This enhances the line loss prediction model's ability to characterize frequency band differences and data uncertainty in complex spectral environments.

[0017] Furthermore, this invention weights and fuses the prediction results of each sub-band based on frequency band evidence parameters and frequency band test quality indicators. During the fusion process, it introduces constraints on the weight changes of adjacent sub-bands and constraints on the consistency of line loss with frequency, ensuring the continuity and consistency of the global line loss prediction results across the frequency dimension. By outputting a prediction result set that includes predicted line loss values, prediction confidence information, and low-confidence sub-band identifiers, this invention provides more complete detection results for predictive testing and health status assessment of RF coaxial cables, improving the usability of line loss prediction results in engineering applications. Attached Figure Description

[0018] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings: Figure 1 This is a flowchart of the deep learning-based method for predicting the line loss of radio frequency coaxial cables proposed in this invention. Figure 2 This is a schematic diagram of the overall structure of the frequency division evidence deep regression network in the deep learning-based method for predicting the line loss of radio frequency coaxial cables proposed in this invention. Figure 3 This is a schematic diagram of the frequency band feature encoding module in the frequency division evidence deep regression network of the deep learning-based RF coaxial cable loss prediction method proposed in this invention. Detailed Implementation

[0019] The present invention will now be described in further detail with reference to the accompanying drawings. These drawings are simplified schematic diagrams, illustrating only the basic structure of the invention, and therefore only show the components relevant to the invention.

[0020] refer to Figures 1-3 A deep learning-based method for predicting the line loss of RF coaxial cables includes the following steps: Define the test frequency range for the RF coaxial cable, and divide the test frequency range into sub-frequency bands based on the frequency sensitivity criterion to generate a set of sub-frequency bands; The radio frequency parameters of the coaxial cable under test are tested within the test frequency band, and line loss related test data are collected to generate a line loss test dataset. Preprocess the line loss test dataset to generate standardized line loss feature data; Based on the sub-band set, the standardized line loss feature data is divided into frequencies and mapped to construct a frequency band sub-feature dataset, and a set of frequency band test quality indicators is extracted. The frequency band sub-feature dataset and the frequency band test quality index set are input into the frequency band feature encoding module of the frequency division evidence deep regression network to generate a set of frequency band feature vectors; The set of frequency band feature vectors is input into the frequency band evidence regression module of the frequency division evidence deep regression network to model the degree of line loss and output the line loss sub-prediction results and the set of frequency band evidence parameters. In the frequency division evidence deep regression network, a global line loss prediction result is generated based on the frequency band evidence parameter set and the frequency band test quality index set. Based on the global line loss prediction results and the frequency band evidence parameter set, predictive detection results are generated.

[0021] In this embodiment, the step of setting the test frequency band range of the RF coaxial cable and dividing the test frequency band range into sub-band sets based on frequency sensitivity criteria specifically includes: The test frequency range for the line loss of the radio frequency coaxial cable is defined by the minimum test frequency and the maximum test frequency. The minimum test frequency is used to characterize the lower limit of the frequency covered by the line loss test of the radio frequency coaxial cable, and the maximum test frequency is used to characterize the upper limit of the frequency covered by the line loss test of the radio frequency coaxial cable. The continuous frequency interval between the minimum test frequency and the maximum test frequency is used as the test frequency range of the radio frequency coaxial cable. Based on the initial line loss curve data, a frequency sensitivity criterion is constructed to characterize the degree of line loss change with frequency. Specifically, the rate of change of line loss with frequency is used as the frequency sensitivity index. The rate of change of line loss with frequency is obtained by calculating the ratio of the change in line loss at adjacent frequency points to the change in the corresponding frequency, and the absolute value of the rate of change is taken as the frequency sensitivity index. The frequency sensitivity index is used to characterize the sensitivity of line loss to frequency changes. The greater the rate of change of line loss, the higher the corresponding frequency sensitivity. Based on the preset frequency sensitivity threshold and combined with the frequency sensitivity criteria, the frequency sensitivity index corresponding to each frequency point within the test frequency band is compared one by one, and the continuous frequency point intervals whose frequency sensitivity index reaches or exceeds the frequency sensitivity threshold are selected, and the frequency range formed by the continuous frequency points is determined as the frequency sensitive interval. Using the start and end frequencies of the frequency-sensitive interval as the frequency division boundary, and the lowest and highest test frequencies of the test frequency band as the frequency division endpoints, under the constraint of the frequency sensitivity criterion, the test frequency band is continuously divided in order from low to high frequency, so that several frequency intervals are formed between adjacent frequency division boundaries, and each frequency interval is determined as a corresponding sub-frequency band, thereby generating multiple continuous sub-frequency bands. All sub-bands are combined in ascending order of frequency to generate a sub-band set.

[0022] In this embodiment, the step of performing RF parameter testing on the RF coaxial cable under test within the test frequency band, collecting line loss-related test data, and generating a line loss test dataset specifically includes: Within the test frequency band, the radio frequency parameters of the radio frequency coaxial cable under test are tested. The radio frequency excitation signal is applied to the radio frequency coaxial cable through the radio frequency test instrument, and the frequency of the radio frequency excitation signal is changed point by point according to the preset frequency sampling step size to perform radio frequency parameter testing in the form of frequency sweep. At each test frequency, the corresponding radio frequency transmission parameters of the radio frequency coaxial cable are collected. The radio frequency transmission parameters include the insertion loss parameters used to characterize the signal transmission attenuation characteristics, and a correspondence is established between each test frequency and the corresponding radio frequency transmission parameters. Based on the insertion loss parameters collected at each test frequency, the line loss values ​​of the RF coaxial cable at each test frequency are determined according to the correspondence between insertion loss and line loss. Each line loss value is used to characterize the signal transmission attenuation of the RF coaxial cable under the corresponding test frequency conditions, and a one-to-one correspondence is established between the line loss values ​​corresponding to each test frequency and its test frequency. The line loss values ​​corresponding to each test frequency within the test frequency band are sorted in order from low to high according to the test frequency. The line loss values ​​are then connected sequentially using the sequential correspondence between the test frequency and the corresponding line loss value as an index to construct a line loss test data sequence that characterizes the relationship between the line loss of the RF coaxial cable and the frequency. Based on the sub-band set, the line loss test data sequence is mapped to the frequency range corresponding to each sub-band, and the line loss test data belonging to each sub-band is collected to form the line loss test dataset of the corresponding sub-band set.

[0023] In this embodiment, the preprocessing of the line loss test dataset to generate standardized line loss feature data specifically includes: Obtain the line loss test dataset corresponding to the sub-band set, and read the test frequency information corresponding to each data point and the line loss value corresponding to each test frequency from the line loss test dataset. The test frequency is used to characterize the frequency position of the RF coaxial cable when the line loss test is performed, and the line loss value is used to characterize the signal transmission attenuation degree of the RF coaxial cable under the corresponding test frequency condition, and establish the correspondence between the test frequency and the line loss value. Within the test frequency band, an alignment frequency sequence for frequency alignment is generated according to a preset frequency alignment grid. The frequency alignment grid is used to limit the frequency interval between adjacent frequency points in the alignment frequency sequence. The alignment frequency sequence takes the lowest test frequency in the test frequency band as the starting frequency and the highest test frequency in the test frequency band as the ending frequency. Multiple frequency points are arranged sequentially between the starting frequency and the ending frequency according to the frequency alignment grid, thereby forming a unified frequency reference sequence for subsequent frequency alignment processing. The line loss values ​​corresponding to each test frequency in the line loss test dataset are mapped to the alignment frequency sequence according to the correspondence between the test frequency and each frequency point in the alignment frequency sequence, thereby generating an alignment line loss sequence that corresponds one-to-one with the alignment frequency sequence. When a frequency point in the aligned frequency sequence is not completely consistent with the actual test frequency in the line loss test dataset, the line loss value corresponding to the frequency point is determined by interpolation based on the line loss value corresponding to the known test frequencies adjacent to the frequency point, so as to ensure the continuity and integrity of the aligned line loss sequence throughout the entire test frequency band. Resampling is performed on the alignment loss sequence. Specifically, according to the preset sampling step size, the frequency points of the alignment frequency sequence are reselected. When the interval between adjacent frequency points in the alignment frequency sequence is less than the sampling step size, the alignment loss sequence is downsampled. When the interval between adjacent frequency points in the aligned frequency sequence is greater than the sampling step size, the aligned line loss sequence is upsampled, and the corresponding line loss value is determined at the newly added frequency point, thereby generating a resampled line loss sequence that corresponds one-to-one with the resampled frequency points, so that the resampled line loss sequence has a uniform frequency resolution in the entire test frequency band. Noise suppression processing is performed on the resampled line loss sequence. Specifically, a sliding window is set in the resampled line loss sequence, and the sliding window moves point by point in the frequency order. Multiple line loss values ​​contained in each window are smoothed and calculated. The smoothed calculation result is used to replace the line loss value corresponding to the center frequency point of the window, thereby weakening the sudden change in line loss value caused by test noise, instantaneous fluctuations or occasional anomalies, and generating a continuous and smooth denoised line loss sequence in the entire test frequency band. The denoised line loss sequence is subjected to amplitude normalization processing, specifically: the minimum and maximum line loss values ​​among all line loss values ​​in the denoised line loss sequence are obtained, and the minimum and maximum line loss values ​​are used as normalization benchmarks. Each line loss value in the denoised line loss sequence is linearly scaled, so that each line loss value, after deducting the minimum line loss value, is divided by the difference between the maximum line loss value and the minimum line loss value, thereby converting the denoised line loss sequence into a normalized line loss sequence with a uniform numerical range.

[0024] In this embodiment, the step of performing frequency-based mapping on standardized line loss feature data based on sub-band sets, constructing frequency band sub-feature datasets, and extracting frequency band test quality indicator sets specifically includes: Acquire standardized line loss characteristic data and synchronously read the aligned frequency sequence corresponding to each standardized line loss characteristic data. Each frequency point in the aligned frequency sequence is used to characterize the specific frequency position corresponding to the standardized line loss characteristic data within the test frequency band. Read the normalized line loss value corresponding to each aligned frequency point. The normalized line loss value is used to characterize the line loss of the radio frequency coaxial cable after amplitude normalization under the corresponding aligned frequency condition, thereby establishing the correspondence between the aligned frequency point and the normalized line loss value. Based on the sub-band set, each aligned frequency in the aligned frequency sequence is mapped according to its sub-band interval. The aligned frequencies falling within the same sub-band interval and their corresponding normalized line loss values ​​are collected to construct the frequency band sub-feature datasets corresponding to each sub-band. Each frequency band sub-feature dataset contains all aligned frequencies and their corresponding normalized line loss values ​​within the sub-band. For each frequency band sub-feature dataset, a frequency band line loss dispersion index is calculated to characterize the stability of the line loss test. The frequency band line loss dispersion index is used to reflect the degree of dispersion of each normalized line loss value within the same sub-frequency band relative to the overall level of the sub-frequency band. Specifically, firstly, all normalized line loss values ​​contained in the frequency band sub-feature dataset are statistically analyzed to obtain the average normalized line loss value corresponding to the sub-frequency band. Then, the deviation between each normalized line loss value and the average normalized line loss value is calculated. Based on the overall distribution of the deviation, the frequency band line loss dispersion index corresponding to the sub-frequency band is generated to characterize the stability level of the line loss test data within the sub-frequency band. For each frequency band sub-feature dataset, a frequency band fluctuation index is calculated to characterize the stability of line loss with frequency variation. Specifically, the normalized line loss values ​​corresponding to adjacent aligned frequency points in the frequency band sub-feature dataset are compared one by one in order of alignment frequency from low to high. The difference between two adjacent normalized line loss values ​​is calculated, and the difference is used as a characterization of the magnitude of line loss variation with frequency. Within a sub-band, the differences between all adjacent frequency points are aggregated to generate a frequency band fluctuation index for the corresponding sub-band. The frequency band fluctuation index is used to reflect the stability of line loss within the sub-band as frequency changes. Based on the frequency band line loss dispersion index and frequency band fluctuation index corresponding to each sub-frequency band, the reliability of the line loss test data of each sub-frequency band is comprehensively characterized. Specifically, the frequency band line loss dispersion index and frequency band fluctuation index obtained within the same sub-frequency band are jointly processed to form a test quality evaluation result that reflects the overall stability and consistency of the sub-frequency band line loss test data. The test quality evaluation results corresponding to each sub-frequency band are then collected to construct a set of frequency band test quality indicators that correspond one-to-one with the set of sub-frequency bands. The set of frequency band test quality indicators is associated with the corresponding frequency band sub-feature dataset to characterize the reliability of line loss test data for each sub-frequency band.

[0025] In this embodiment, the step of inputting the frequency band sub-feature dataset and the frequency band test quality index set into the frequency band feature encoding module of the frequency division evidence deep regression network to generate the frequency band feature vector set specifically includes: Obtain the frequency band sub-feature dataset and the frequency band test quality index set, and associate them according to the order of the sub-frequency bands. For each sub-frequency band, extract the frequency band sub-feature dataset corresponding to the sub-frequency band to characterize the normalized line loss feature distribution within the sub-frequency band range. At the same time, extract the frequency band test quality index corresponding to the frequency band sub-feature dataset to characterize the reliability of the sub-frequency band line loss test data. Establish the correspondence between the frequency band sub-feature dataset and the frequency band test quality index on a sub-frequency band basis. For each sub-band sub-feature dataset, the aligned frequency points and their corresponding normalized line loss values ​​within the sub-band range are arranged in order from low to high frequency. The aligned frequency points and normalized line loss values ​​are then combined to construct the frequency input sequence corresponding to the sub-band. The frequency band input sequence is used to characterize the sequential relationship of line loss characteristics within a sub-band as frequency changes, and to record the number of aligned frequency points contained within the sub-band as an input scale parameter. For each aligned frequency point in the frequency band input sequence, a frequency position encoding vector corresponding to the aligned frequency point is constructed based on the relative frequency position of the aligned frequency point in the corresponding sub-frequency band. The frequency position encoding vector is combined with the normalized line loss value corresponding to the aligned frequency point to form a point-level feature representation that characterizes the line loss features and frequency position information of the aligned frequency point. According to the order of frequency from low to high, the point-level feature representations corresponding to all aligned frequency points in the sub-band are collected to generate the frequency band point-level embedding sequence corresponding to the sub-band. The frequency band point-level embedding sequence is used to characterize the distribution and location information of line loss features in the frequency dimension in the sub-band. The frequency band test quality index is vectorized to generate a quality vector that corresponds one-to-one with the corresponding sub-frequency band. The quality vector is then input into the first quality mapping unit and the second quality mapping unit in the frequency band feature encoding module of the frequency division evidence deep regression network. The first quality mapping unit and the second quality mapping unit each include at least one layer of linear transformation structure and a nonlinear transformation structure connected to the linear transformation structure. The first quality mapping unit performs linear and nonlinear transformations on the quality vector and outputs a quality gating coefficient vector consistent with the encoded feature dimension. The second quality mapping unit performs linear and nonlinear transformations on the quality vector and outputs a quality bias vector consistent with the encoded feature dimension. The quality gating coefficient vector and the quality bias vector are associated with the corresponding sub-bands and used as control parameters for quality gating modulation of the sub-band coding features. The frequency band point-level embedding sequence is input into the multi-scale frequency domain coding unit in the frequency band feature coding module of the frequency division evidence deep regression network. At least two parallel one-dimensional convolutional branches are set in the multi-scale frequency domain coding unit, and each one-dimensional convolutional branch corresponds to a different convolutional kernel size. Parallel convolutional coding is performed on the frequency band point-level embedding sequence through each one-dimensional convolutional branch to extract the feature information of the frequency band point-level embedding sequence at different frequency scales. Subsequently, feature fusion processing is performed on the encoding results output by each one-dimensional convolution branch. The encoding features obtained from multiple convolution branches are aggregated to generate a multi-scale encoding feature sequence for characterizing multi-scale frequency domain information of sub-band line loss features. Based on the quality gate coefficient vector and the quality bias vector, quality gate modulation processing is performed on the multi-scale coded feature sequence. Specifically, according to the arrangement order of each feature dimension in the multi-scale coded feature sequence, the quality gate coefficient vector is used to perform dimension-wise scaling processing on the coded features of the corresponding feature dimensions. After the scaling processing is completed, the quality bias vector is superimposed on the corresponding coded features dimension by dimension, thereby performing amplitude modulation and bias modulation on the multi-scale coded feature sequence. By using quality-gated modulation processing, the multi-scale coded feature sequence is combined with frequency band test quality information to form a quality modulation feature sequence while maintaining the original frequency domain structure. The quality modulation feature sequence is then used as the input result for feature convergence processing. Attention convergence processing is performed on the quality modulation feature sequence, specifically: based on the feature information corresponding to each aligned frequency point in the quality modulation feature sequence, attention weights corresponding to each aligned frequency point are calculated to form an attention weight sequence, wherein each attention weight is used to characterize the importance of the corresponding aligned frequency point feature in the current sub-band line loss characterization; Subsequently, in order of alignment frequency from low to high, the attention weight sequence is used to weight each feature in the quality modulation feature sequence, and the weighted features are then aggregated. The quality modulation feature sequence is then weighted and summed in the frequency dimension to generate a frequency band feature vector that is used to characterize the line loss features of the sub-band as a whole. The frequency band feature vectors obtained from each sub-band are collected in a unified manner according to the order of the sub-bands in the sub-band set. The frequency band feature vectors corresponding to each sub-band are sequentially included in the same set, thus forming a frequency band feature vector set containing the frequency band feature vectors of all sub-bands. The set of frequency band feature vectors is composed of frequency band feature vectors corresponding to each sub-frequency band, and maintains an independent correspondence between the feature vectors of each frequency band.

[0026] In this embodiment, the step of inputting the set of frequency band feature vectors into the frequency band evidence regression module of the frequency-division evidence deep regression network, performing line loss level modeling, and outputting the line loss sub-prediction result and the set of frequency band evidence parameters specifically includes: Obtain the set of frequency band feature vectors, and divide the set of frequency band feature vectors by index according to the order of the sub-frequency bands in the set of sub-frequency bands, and determine the frequency band feature vector corresponding to each sub-frequency band; In this process, the frequency band feature vector corresponding to the first sub-band is denoted as the first frequency band feature vector, the frequency band feature vector corresponding to the second sub-band is denoted as the second frequency band feature vector, and so on, until the frequency band feature vector corresponding to the last sub-band is determined, thereby establishing a one-to-one correspondence between sub-bands and frequency band feature vectors. The frequency band feature vector corresponding to each sub-frequency band is input into the line loss regression unit in the frequency band evidence regression module of the frequency division evidence deep regression network, and regression mapping processing is performed on the frequency band feature vector. The regression mapping process includes performing at least one layer of linear transformation on the frequency band feature vector to generate intermediate features for line loss regression, performing nonlinear transformation on the intermediate features for line loss regression, and obtaining the line loss sub-prediction results for the corresponding sub-frequency band. The line loss sub-prediction results are used to characterize the predicted line loss of radio frequency coaxial cables within the sub-frequency band range, and maintain a one-to-one correspondence with the input frequency band feature vector, thereby forming a set of line loss sub-prediction results corresponding to each sub-frequency band. The frequency band feature vector corresponding to each sub-frequency band is input into the evidence parameter regression unit in the frequency band evidence regression module of the frequency division evidence deep regression network, and the evidence parameter regression processing is performed on the frequency band feature vector. The evidence parameter regression processing includes performing at least one layer of vector mapping transformation on the frequency band feature vector to generate an intermediate evidence vector for characterizing the uncertainty of line loss prediction. The intermediate evidence vector has a dimension greater than one and is used to simultaneously characterize multiple evidence-related parameters. A non-negative constraint transformation is performed on the intermediate evidence vector to make the values ​​of each parameter component in the intermediate evidence vector non-negative, thereby obtaining the frequency band evidence parameter vector of the corresponding sub-frequency band. The frequency band evidence parameter vector is used to characterize the prediction confidence information and uncertainty information corresponding to the sub-band line loss sub-prediction results, and establishes a one-to-one correspondence with the line loss sub-prediction results of the corresponding sub-band, thereby forming a set of frequency band evidence parameter vectors corresponding to each line loss sub-prediction result. For each sub-band, the frequency band evidence parameter vector is non-negatively processed. Specifically, for each parameter component in the frequency band evidence parameter vector, exponential mapping or soft addition mapping is used to convert the original parameter components into non-negative values, thereby eliminating the semantic conflicts caused by negative parameter values. By performing nonnegation processing, each parameter component in the obtained evidence parameter vector satisfies the nonnegation constraint requirement, forming a nonnegative evidence parameter vector that corresponds one-to-one with the corresponding sub-frequency band. Based on the nonnegative evidence parameter vector obtained after nonnegation processing, a corresponding set of frequency band evidence parameters is constructed for each sub-frequency band. Specifically, each parameter component in the nonnegative evidence parameter vector is used as an evidence parameter to characterize the uncertainty of the sub-frequency band line loss prediction result. The evidence parameters are organized according to the preset evidence parameter structure to form a set of frequency band evidence parameters that corresponds one-to-one with the corresponding sub-frequency band. The set of frequency band evidence parameters is used to describe the prediction reliability level and uncertainty of the sub-frequency band line loss sub-prediction results, and to establish a correspondence with the corresponding line loss sub-prediction results; The line loss sub-prediction results obtained from each sub-band are collected in a unified manner according to the order of the sub-band in the sub-band set. The line loss sub-prediction results corresponding to each sub-band are then included in the same result set in sequence, thus forming a line loss prediction result set that includes the line loss sub-prediction results of all sub-bands.

[0027] In this embodiment, the step of generating a global line loss prediction result based on the frequency band evidence parameter set and the frequency band test quality index set in the frequency division evidence deep regression network specifically includes: Obtain the set of sub-prediction results for line loss, the set of frequency band evidence parameters, and the set of frequency band test quality indicators; Specifically, for each sub-frequency band, the line loss sub-prediction results, the frequency band evidence parameters, and the frequency band test quality indicators corresponding to the sub-frequency band are extracted respectively. The correspondence between the line loss sub-prediction results, the frequency band evidence parameters, and the frequency band test quality indicators is established to form a multi-element associated data structure indexed by the sub-frequency band. For each sub-band, the frequency band evidence parameters corresponding to the sub-band are aggregated. Specifically, the evidence parameter components contained in the frequency band evidence parameter vector corresponding to the sub-band are read, and the evidence parameter components are accumulated one by one to obtain the evidence strength scalar used to characterize the overall evidence strength of the sub-band. The evidence strength scalar is used to reflect the degree of evidence support for the sub-band line loss sub-prediction results; For each sub-band, the corresponding band test quality index is read, and the quality evaluation components contained in the band test quality index are aggregated. Specifically, the various quality evaluation components used to characterize the stability and consistency of the sub-band line loss test are uniformly synthesized to obtain a quality intensity scalar used to characterize the overall test quality level of the sub-band. The quality intensity scalar is used to reflect the reliability of the sub-band line loss test data and serves as the quality aggregation result of the sub-band frequency band test quality indicators. For each sub-band, the evidence strength scalar and the quality strength scalar corresponding to the sub-band are jointly calculated. Specifically, the evidence strength scalar is used as the quantitative result reflecting the degree of support of the predicted evidence in the sub-band, and the quality strength scalar is used as the quantitative result reflecting the reliability of the test data in the sub-band. The two are multiplied to obtain the unnormalized evidence weight corresponding to the sub-band. The unnormalized evidence weight is obtained by multiplying the evidence strength scalar and the quality strength scalar, and is used to simultaneously and comprehensively characterize the evidence sufficiency and test quality level of the sub-band line loss sub-prediction results. Normalization processing is performed on the unnormalized evidence weights corresponding to all sub-bands. Specifically, the unnormalized evidence weights corresponding to each sub-band are used as numerators, and the sum of the unnormalized evidence weights corresponding to all sub-bands is used as a unified normalization benchmark. The unnormalized evidence weights of each sub-band are scaled proportionally to obtain the evidence weights corresponding to each sub-band. Based on the preset threshold for changes in the evidence weights of adjacent sub-bands, change constraint processing is performed on the evidence weights corresponding to adjacent sub-bands. Specifically, the evidence weights of two adjacent sub-bands are compared sequentially according to the order of the sub-bands in the sub-band set. When the difference between the evidence weight of the latter sub-band and the evidence weight of the former sub-band exceeds the threshold for changes in the evidence weights of adjacent sub-bands, the evidence weight of the latter sub-band is adjusted so that the difference is limited to the range allowed by the change threshold. By applying change constraints, the variation range between the evidence weights of any adjacent sub-bands is ensured to not exceed the threshold for the variation of evidence weights of adjacent sub-bands. The evidence weights of each sub-band after the change constraint processing are used as the constraint evidence weights for the weighted fusion of line loss sub-prediction results. Based on the preset line loss consistency constraint condition with frequency, the line loss sub-prediction results corresponding to each sub-frequency band are subjected to consistency constraint processing. Specifically, according to the frequency order of the sub-frequency band in the sub-frequency band set, the line loss sub-prediction results corresponding to adjacent sub-frequency bands are compared one by one. When the line loss sub-prediction result corresponding to the later sub-frequency band is less than the line loss sub-prediction result corresponding to the earlier sub-frequency band, the line loss sub-prediction result of the later sub-frequency band is adjusted so that it is not less than the line loss sub-prediction result corresponding to the earlier sub-frequency band. Through consistency constraint processing, the line loss sub-prediction results corresponding to each sub-band maintain a non-decreasing relationship in the frequency dimension, thereby forming a line loss sub-prediction result sequence that meets the consistency requirement of line loss changing with frequency. The line loss sub-prediction results after consistency constraint processing are used as the constrained line loss sub-prediction results adopted for weighted fusion processing. Based on the evidence weights of each sub-band after the constraint processing of the change in evidence weights of adjacent sub-bands, a weighted fusion processing is performed on the sub-prediction results of line loss of each sub-band after the constraint processing of line loss with frequency consistency. Specifically, the constraint evidence weights corresponding to each sub-band are multiplied with the constraint line loss sub-prediction results corresponding to the sub-band to obtain the weighted line loss prediction value corresponding to the sub-band. Subsequently, the weighted line loss prediction values ​​corresponding to all sub-bands are summed to generate a global line loss prediction result; The global line loss prediction result is jointly determined by the line loss sub-prediction results of each sub-band under the influence of their corresponding evidence weights, and is used to characterize the overall line loss prediction value of the RF coaxial cable in the entire test frequency band.

[0028] In this embodiment, generating predictive detection results based on the global line loss prediction results and the frequency band evidence parameter set specifically includes: Based on the set of frequency band evidence parameters, the corresponding prediction credibility index is calculated for each sub-frequency band. Specifically, the frequency band evidence parameters corresponding to the sub-frequency band are aggregated to obtain a credibility value that characterizes the credibility of the sub-frequency band line loss prediction result, and a one-to-one correspondence is established between the credibility value and the corresponding sub-frequency band. Based on a preset confidence threshold, the confidence values ​​corresponding to each sub-band are compared. When the confidence value corresponding to a certain sub-band is lower than the confidence threshold, the sub-band is marked as a low confidence sub-band, and a low confidence sub-band identifier corresponding to each sub-band is generated. The global line loss prediction results, the confidence values ​​of each sub-band, and the low confidence sub-band identifiers are integrated to construct a line loss prediction result set. The line loss prediction result set includes at least the overall line loss prediction value, the prediction confidence information of each sub-band, and the low confidence sub-band identifier information. The line loss prediction result set is output as the predictive test result of the RF coaxial cable, which is used to characterize the line loss prediction status and corresponding prediction confidence distribution of the RF coaxial cable within the test frequency band.

[0029] Example 1: To verify the feasibility of this invention in practice, it was applied to a predictive testing scenario for radio frequency coaxial cables in a communication equipment system. In this system, the radio frequency coaxial cable operates in a wide-band signal transmission state for extended periods. With increasing usage time, aging of the internal conductors, changes in shielding performance, and fluctuations in connection status easily lead to a gradual increase in line loss. However, traditional line loss detection methods based on single tests cannot reflect the trend of line loss changes in a timely manner, failing to meet the needs of predictive maintenance and health status assessment.

[0030] In this application scenario, the test frequency band range covering the actual working requirements of the system is first set for the RF coaxial cable under test. This test frequency band range is then divided into multiple consecutive sub-bands based on frequency sensitivity criteria. Subsequently, the RF coaxial cable is subjected to frequency sweep testing within the test frequency band range using RF parameter testing equipment. Insertion loss RF parameters are collected, and the collected data is converted into line loss test data at the corresponding frequency points, forming a line loss test dataset. To ensure the comparability of data under different test conditions, frequency alignment, resampling, noise suppression, and amplitude normalization are performed on the line loss test dataset to generate standardized line loss characteristic data.

[0031] In the data modeling phase, standardized line loss feature data are frequency-mapping based on the aforementioned sub-band set, constructing multiple frequency band sub-feature datasets. The dispersion and fluctuation levels of line loss data within each sub-band are calculated to form a set of frequency band test quality indicators. Subsequently, the frequency band sub-feature datasets and the set of frequency band test quality indicators are input into a frequency-division evidence deep regression network. The frequency band feature encoding module encodes the line loss features of each sub-band, and a quality gating mechanism modulates the frequency band features at different quality levels, generating a set of frequency band feature vectors. Based on this, the frequency band evidence regression module models the line loss degree of each frequency band feature vector, outputting the line loss sub-prediction results and frequency band evidence parameters corresponding to each sub-band.

[0032] In the prediction result fusion stage, based on the frequency band evidence parameters and frequency band test quality indicators corresponding to each sub-band, the evidence weight of each line loss sub-prediction result is determined. Under the constraints of evidence weight changes between adjacent sub-bands and line loss consistency with frequency, the line loss sub-prediction results are weighted and fused to generate a global line loss prediction result. Simultaneously, the prediction confidence of each sub-band is calculated based on the frequency band evidence parameters, and sub-bands with low prediction confidence are identified. Finally, a predictive detection result containing the overall line loss prediction value, prediction confidence information, and low-confidence sub-band identification is generated.

[0033] To verify the beneficial effects of this invention, the predicted line loss results were compared and analyzed with multiple measured line loss results. The results show that this invention can effectively reflect the reliability differences in prediction results across different sub-frequency bands while maintaining high prediction accuracy, avoiding significant interference from low-quality test data on the overall prediction results. Under multiple continuous test conditions, the global line loss prediction results output by this invention maintain good continuity in the frequency dimension, with the prediction error stably controlled within a reasonable range. The fluctuation of the prediction results is significantly less than that of traditional single-model regression methods, providing reliable data support for predictive testing and health status assessment of RF coaxial cables.

[0034] Table 1. Comparative Data on Line Loss Prediction of RF Coaxial Cables Based on Frequency Division Evidence Deep Regression Network

[0035] As can be seen from the table above, this invention significantly improves upon traditional holistic regression prediction methods in terms of accuracy, stability, and reliability in predicting the line loss of RF coaxial cables. Firstly, regarding the prediction accuracy for each sub-band, the prediction errors of this invention in the low-frequency, second-low-frequency, and mid-frequency sub-bands are controlled at 0.03dB, 0.03dB, and 0.04dB, respectively, all significantly lower than the prediction deviations produced by traditional holistic regression methods in the same frequency band. This indicates that by independently characterizing the line loss characteristics of different frequency bands through frequency-specific modeling, the prediction deviation caused by differences in frequency band characteristics can be effectively reduced, improving the consistency between the predicted line loss and the measured data.

[0036] In the mid-to-high frequency and high-frequency sub-bands, the prediction error of traditional global regression prediction methods increases significantly, especially in the high-frequency sub-bands, where the predicted values ​​deviate considerably from the measured line loss. This reflects the insufficient adaptability of traditional methods to nonlinear changes in line loss and test fluctuations in the high-frequency band. In contrast, under the same high-frequency sub-band conditions, although the prediction error increases, it remains within a relatively controllable range. Furthermore, by using a prediction confidence mechanism, this sub-band is identified as a low-confidence sub-band, thereby reducing its impact on the overall prediction results during subsequent full-band fusion. This result demonstrates that this invention can identify the source of prediction uncertainty under unstable high-frequency test conditions, avoiding excessive interference from a single high-error frequency band on the global prediction.

[0037] From the full-band fusion results, the error between the global line loss prediction value generated by this invention and the measured average line loss is 0.07 dB, which is significantly smaller than the prediction error level of the traditional overall regression method. Furthermore, the fused prediction results maintain good continuity in the frequency dimension. Combined with the prediction confidence distribution shown in the table, it can be seen that this invention uses the frequency band evidence parameters output by the frequency-division evidence deep regression network and the frequency band test quality indicators to jointly participate in weight determination, allowing high-confidence sub-bands to occupy a more reasonable weight proportion in the global prediction, thereby improving the stability and reliability of the overall prediction results. These results fully demonstrate that this invention, in the application of RF coaxial cable line loss prediction, can balance prediction accuracy and prediction confidence, providing more reliable data support for predictive detection and health status assessment.

[0038] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A method for predicting the line loss of radio frequency coaxial cables based on deep learning, characterized in that, Includes the following steps: Define the test frequency range for the RF coaxial cable, and divide the test frequency range into sub-frequency bands based on the frequency sensitivity criterion to generate a set of sub-frequency bands; The radio frequency parameters of the coaxial cable under test are tested within the test frequency band, and line loss related test data are collected to generate a line loss test dataset. Preprocess the line loss test dataset to generate standardized line loss feature data; Based on the sub-band set, the standardized line loss feature data is divided into frequencies and mapped to construct a frequency band sub-feature dataset, and a set of frequency band test quality indicators is extracted. The frequency band sub-feature dataset and the frequency band test quality index set are input into the frequency band feature encoding module of the frequency division evidence deep regression network to generate a set of frequency band feature vectors; The set of frequency band feature vectors is input into the frequency band evidence regression module of the frequency division evidence deep regression network to model the degree of line loss and output the line loss sub-prediction results and the set of frequency band evidence parameters. In the frequency division evidence deep regression network, a global line loss prediction result is generated based on the frequency band evidence parameter set and the frequency band test quality index set. Based on the global line loss prediction results and the frequency band evidence parameter set, predictive detection results are generated.

2. The method for predicting the line loss of radio frequency coaxial cables based on deep learning according to claim 1, characterized in that, The process of setting the test frequency range for the RF coaxial cable and dividing the test frequency range into sub-frequency bands based on frequency sensitivity criteria to generate a set of sub-frequency bands specifically includes: Set the test frequency range for the line loss of the RF coaxial cable. Within the test frequency range, perform line loss related tests on the RF coaxial cable under test to generate initial line loss curve data. Based on the initial line loss curve data, a frequency sensitivity criterion is constructed and defined as a frequency sensitivity index. Based on the preset frequency sensitivity threshold and combined with the frequency sensitivity criteria, the frequency sensitivity indicators corresponding to each frequency point within the test frequency band are compared one by one to select the frequency sensitive interval. Using the boundary frequencies of the frequency-sensitive interval and the endpoint frequencies of the test frequency band as frequency division nodes, the test frequency band is divided based on the frequency sensitivity criterion to generate multiple continuous sub-frequency bands. All sub-bands are combined in ascending order of frequency to generate a sub-band set.

3. The method for predicting the line loss of radio frequency coaxial cables based on deep learning according to claim 1, characterized in that, The process of performing RF parameter tests on the coaxial cable under test within the test frequency band, collecting line loss-related test data, and generating a line loss test dataset specifically includes: Within the test frequency band, RF parameters of the RF coaxial cable under test are tested, and the insertion loss parameters of the RF coaxial cable are collected at each test frequency. Based on the insertion loss parameter, determine the line loss value of the RF coaxial cable at each test frequency; The line loss values ​​corresponding to each test frequency within the test frequency band are sorted in order from low to high according to the test frequency to construct a line loss test data sequence. Based on the sub-band set, the line loss test data sequence is mapped to the frequency range corresponding to each sub-band, and the line loss test data belonging to each sub-band is collected to form a line loss test dataset.

4. The method for predicting the line loss of radio frequency coaxial cables based on deep learning according to claim 1, characterized in that, The preprocessing includes frequency alignment, resampling, noise suppression, and amplitude normalization.

5. The method for predicting the line loss of radio frequency coaxial cables based on deep learning according to claim 1, characterized in that, The process of performing frequency-based mapping on standardized line loss feature data based on sub-band sets, constructing frequency band sub-feature datasets, and extracting frequency band test quality indicator sets specifically includes: Obtain standardized line loss characteristic data and read the aligned frequency sequence corresponding to each standardized line loss characteristic data; Based on the sub-band set, each aligned frequency in the aligned frequency sequence is mapped to construct a frequency band sub-feature dataset; For each frequency band sub-feature dataset, calculate the frequency band line loss dispersion index and the frequency band fluctuation index; Based on the frequency band line loss dispersion index and the frequency band fluctuation index, a set of frequency band test quality indices corresponding to each sub-frequency band is constructed.

6. The method for predicting the line loss of radio frequency coaxial cables based on deep learning according to claim 1, characterized in that, The step of inputting the frequency band sub-feature dataset and the frequency band test quality index set into the frequency band feature encoding module of the frequency division evidence deep regression network to generate the frequency band feature vector set specifically includes: Obtain the frequency band sub-feature dataset and the frequency band test quality index set. For each sub-frequency band, construct the frequency band input sequence in ascending order of frequency. For the frequency band input sequence, a frequency position encoding vector is constructed for each aligned frequency point and combined with the corresponding normalized line loss value to generate a frequency band point-level embedding sequence; The quality indicators of the frequency band test are vectorized to obtain the quality vector, and the quality modulation parameters are generated based on the quality vector to obtain the quality gate coefficient vector and the quality bias vector. The frequency band point-level embedding sequence is input into the multi-scale frequency domain coding unit of the frequency band feature coding module of the frequency division evidence deep regression network to generate a multi-scale coded feature sequence. Based on the quality gating coefficient vector and the quality bias vector, the multi-scale encoded feature sequence is subjected to quality gating modulation processing to generate a quality modulated feature sequence. Attention convergence processing is performed on the quality modulation feature sequence to generate frequency band feature vectors. The frequency band feature vectors corresponding to all sub-frequency bands are then aggregated to generate a frequency band feature vector set.

7. The method for predicting the loss level of radio frequency coaxial cables based on deep learning according to claim 1, characterized in that, The step of inputting the set of frequency band feature vectors into the frequency band evidence regression module of the frequency-division evidence deep regression network to model the degree of line loss and output the line loss sub-prediction results and the set of frequency band evidence parameters specifically includes: Obtain the set of frequency band feature vectors, input the frequency band feature vectors into the line loss regression head in the frequency band evidence regression module of the frequency sub-evidence deep regression network, and generate line loss sub-prediction results; The frequency band feature vector is input into the evidence parameter regression head of the frequency band evidence regression module of the frequency sub-frequency evidence deep regression network to generate the frequency band evidence parameter vector, and the frequency band evidence parameter vectors corresponding to all sub-frequency bands are collected to construct the frequency band evidence parameter vector set. The frequency band evidence parameter vector is nonnegated to obtain a nonnegative evidence parameter vector. Based on the nonnegative evidence parameter vector, a set of frequency band evidence parameters is constructed. The line loss sub-prediction results corresponding to all sub-frequency bands are aggregated to generate a set of line loss sub-prediction results.

8. The method for predicting the line loss of radio frequency coaxial cables based on deep learning according to claim 1, characterized in that, The generation of global line loss prediction results in the frequency-division evidence deep regression network, based on the frequency band evidence parameter set and the frequency band test quality index set, specifically includes: Obtain the set of sub-prediction results for line loss, the set of frequency band evidence parameters, and the set of frequency band test quality indicators; For each sub-band, calculate the sub-band evidence strength scalar based on the band evidence parameter vector; The quality strength scalar is calculated based on the frequency band test quality index, and the quality strength scalar is used as the quality convergence value of the frequency band test quality index. Based on the unnormalized evidence weights of the evidence strength scalar and the quality strength scalar, the unnormalized evidence weights are normalized to obtain the evidence weights. Based on the preset threshold for changes in evidence weights of adjacent sub-bands, change constraint processing is performed on the evidence weights corresponding to adjacent sub-bands; Based on the preset line loss consistency constraint condition with frequency, the line loss sub-prediction results are subjected to consistency constraint processing. Based on the evidence weights of each sub-band after the constraint processing of the change in evidence weights of adjacent sub-bands, the sub-prediction results of line loss of each sub-band after the constraint processing of line loss with frequency consistency are subjected to weighted fusion processing to generate the global line loss degree prediction result.

9. The method for predicting the line loss of radio frequency coaxial cables based on deep learning according to claim 1, characterized in that, The generation of predictive detection results based on the global line loss prediction results and frequency band evidence parameter set specifically includes: Based on the set of frequency band evidence parameters, the corresponding prediction confidence index is calculated for each sub-frequency band to obtain the confidence value. Based on a preset confidence threshold, the confidence values ​​corresponding to each sub-band are compared. When the confidence value is lower than the confidence threshold, the sub-band is marked as a low-confidence sub-band and a low-confidence sub-band identifier is generated. The global line loss prediction results, confidence values, and low-confidence sub-band identifiers are integrated to construct a line loss prediction result set; The predicted line loss results are output as predictive test results for RF coaxial cables.

Images