A bayesian optimization-based wideband impedance spectroscopy pattern recognition method and system

Abstract

A broadband impedance spectrum pattern recognition method and system based on Bayesian optimization includes the following steps: acquiring and processing the broadband impedance spectrum of a cable to generate a training dataset; performing smoothing filtering and normalization on the acquired data, extracting features using a frequency band division strategy and combining them into a multi-dimensional feature vector, and calculating a comprehensive evaluation index; determining the optimal multi-dimensional feature vector with the comprehensive evaluation index as the objective; inputting the multi-dimensional feature vector into a predefined fault classification model to optimize the model's hyperparameters; identifying the fault type from the impedance spectrum data of the cable under test; obtaining impedance spectrum samples from the impedance spectrum data of the cable under test, calculating the joint diagnostic reliability, and correcting the model accordingly. This invention adaptively adjusts the broadband impedance spectrum feature extraction parameters through Bayesian optimization and combines a fault classification model with a feedback update mechanism, thereby improving the stability and representativeness of the cable impedance spectrum features.

Description

Technical Field

[0001] This invention relates to the field of cable fault diagnosis technology, and in particular to a broadband impedance spectrum pattern recognition method and system based on Bayesian optimization. Background Technology

[0002] Broadband impedance spectroscopy is an emerging method that has been increasingly applied in cable fault diagnosis and condition assessment in recent years. As a non-destructive testing technique based on frequency domain analysis, broadband impedance spectroscopy can reveal the internal structure, defect characteristics, and aging state of a cable by measuring the characteristics of its input impedance as a function of frequency. Compared with traditional fault diagnosis methods, broadband impedance spectroscopy has significant advantages in capturing complex cable characteristics and is currently a hot research topic.

[0003] Analysis shows that broadband impedance spectroscopy, by acquiring complex impedance data of cables and combining it with spectral feature extraction and pattern analysis, enables the identification and location of local defects (bubbles, scratches, dirt, etc.) in cables. Impedance spectral data not only reflects the dielectric properties of the cable but also contains important information about the microstructure of the insulation material and the distribution of defects, offering advantages such as high diagnostic accuracy and wide applicability. In recent years, with the development of signal processing technology, broadband impedance spectroscopy has gradually moved from laboratory research to practical engineering applications.

[0004] Currently, the process of broadband impedance spectroscopy analysis mainly includes the following steps: measuring the input impedance of the cable over a wide frequency range using a dedicated impedance analyzer to obtain curves showing the amplitude and phase changes with frequency; processing and extracting features from the acquired impedance spectrum to extract key information related to the cable's condition, such as frequency feature points and amplitude changes; and determining the location and type of cable faults by combining specific fault models or diagnostic algorithms. Existing technical document 1 (CN119293625A) discloses a cable condition assessment method based on Bayesian optimization and XGBoost. This method uses a Bayesian optimization algorithm to adjust the hyperparameters of a machine learning model (XGBoost) to improve the model's prediction accuracy for test data. This scheme achieves cable health status assessment by modeling statistical parameters such as runtime and fault occurrence, but its optimization process is limited to the model parameter level and does not involve the physical analysis and feature extraction of cable electrical characteristic data.

[0005] However, despite the broad application potential of broadband impedance spectroscopy, existing technologies still have some problems and limitations, including but not limited to: 1. When processing broadband impedance spectra, fixed frequency band divisions, fixed filtering parameters, or feature extraction methods set manually based on experience are typically used. This makes it difficult to automatically adjust parameters for different cable structures, operating environments, and noise levels, resulting in insufficient representativeness of the extracted features and affecting the accuracy of fault identification. 2. In actual measurement processes, impedance spectrum data are inevitably affected by random noise, measurement errors, and environmental interference. Existing technologies typically lack a comprehensive evaluation mechanism for feature stability and signal-to-noise ratio, leading to significant fluctuations in extracted features under different noise conditions. 3. Most existing diagnostic systems are static models, making it difficult to dynamically correct model parameters based on new measurement samples and environmental changes during actual operation. When data distribution drift or new fault modes occur, the system struggles to adjust in a timely manner, leading to a decrease in diagnostic reliability. Summary of the Invention

[0006] To address the shortcomings of existing technologies, this invention provides a broadband impedance spectrum pattern recognition method and system based on Bayesian optimization. By introducing Bayesian optimization technology, it effectively extracts key features from broadband impedance spectrum data, adaptively optimizes the parameters of the pattern recognition model, and improves the accuracy and stability of fault identification. This method is applicable to diverse cable designs and complex operating environments. The method includes the following steps: acquiring and processing the broadband impedance spectrum of the cable to generate a training dataset; performing smoothing filtering and normalization on the acquired data, extracting features using a frequency band division strategy and combining them into a multi-dimensional feature vector, and calculating a comprehensive evaluation index; determining the optimal multi-dimensional feature vector based on the comprehensive evaluation index; inputting the multi-dimensional feature vector into a predefined fault classification model to optimize the model's hyperparameters; identifying the fault type from the impedance spectrum data of the cable under test; obtaining impedance spectrum samples from the impedance spectrum data of the cable under test, calculating the joint diagnostic reliability, and correcting the model accordingly. This invention adaptively adjusts the broadband impedance spectrum feature extraction parameters through Bayesian optimization and combines a fault classification model with a feedback update mechanism, thereby improving the stability and representativeness of the cable impedance spectrum features.

[0007] The present invention adopts the following technical solution: A first aspect of the present invention provides a broadband impedance spectrum pattern recognition method based on Bayesian optimization, comprising: S1. Collect and process the broadband impedance spectrum of the cable to generate a training dataset; S2. Perform smoothing filtering and normalization on each sample in the training dataset, extract features using a frequency band partitioning strategy, and combine them into a multi-dimensional feature vector; use the frequency band partitioning strategy and the filtering window size as the data-side parameter set, initialize multiple sets of data-side parameter sets, and calculate the physical consistency index, signal-to-noise ratio index, and feature stability index. Take the logarithm of each index and weight them to obtain the exponent of a predefined exponential function, thus obtaining a comprehensive evaluation index; using the comprehensive evaluation index as the objective, optimize the data-side parameter set using Bayesian optimization to determine the optimal number. S3. Input the multidimensional feature vectors under the set of side parameters into a predefined fault classification model and train the model. Based on the trained model, identify the impedance spectrum data of the cable under test and determine the fault type. S4. Based on the impedance spectrum data of the cable under test, generate impedance spectrum samples and calculate the corresponding comprehensive evaluation index, confidence interval of the comprehensive evaluation index, and drift health. Based on the fault classification model, output the class probability of the samples and calculate the prediction entropy certainty. Based on the comprehensive evaluation index, drift health, and prediction entropy certainty of the samples, calculate the joint diagnosis credibility and correct the model parameters.

[0008] Preferably, in S1, the reactance component of the imaginary part and the complex impedance spectrum in the measured broadband impedance spectrum data are used as the original impedance spectrum data, and each original impedance spectrum data and the actual cable state corresponding to the measurement are used as the truth label to generate a training dataset.

[0009] Preferably, the process of extracting features using a specific frequency band division strategy in S2 is as follows: Within a wide-band measurement range, a preset number of frequency bands are used to divide the frequency bands and obtain the frequency band boundary parameters. The filtered and normalized impedance spectrum data is then divided into sub-bands according to the frequency band boundary parameters. For any sub-band, the frequency corresponding to the maximum impedance amplitude within the corresponding frequency band is taken as the peak position of the impedance amplitude. Take the logarithm of the frequency to obtain the logarithmic frequency; after selecting a predetermined number of discrete sampling points in each sub-band, use least squares linear regression to fit the phase-frequency relationship, and use the slope as the phase change rate. For each sub-band, the difference between the impedance amplitude of the next frequency and the current frequency within the sub-band is taken as the amplitude difference at the current frequency. If the amplitude difference at the previous frequency is greater than zero and the amplitude difference at the current frequency is not greater than zero, and the impedance amplitude at the current frequency is greater than a predefined peak significance threshold, the current frequency is taken as a local peak candidate point. When the product of the reactance components at the current frequency and the next frequency is less than zero, a candidate frequency is obtained by linear interpolation. Based on the impedance amplitude at the adjacent frequencies of the candidate frequency, the impedance amplitude of the corresponding candidate frequency is obtained by linear interpolation. The local peak candidate points and candidate frequencies are taken as candidate resonance points to construct a candidate resonance point set. When there are multiple candidate resonance points in the candidate resonance point set, the one with the largest amplitude is taken as the main resonance point frequency of the frequency band. The peak positions, phase change rates, and main resonance frequencies extracted from each frequency band are sequentially concatenated to form a fixed-dimensional multidimensional feature vector.

[0010] Preferably, the process of calculating the physical consistency index in S2 is as follows: Acquire filtered impedance spectrum data and peak positions, phase change rates, and main resonance frequencies of different sub-bands using the data-side parameter set. For each sub-band, calculate the corresponding reference reactance components based on the resistance components at the peak positions and main resonance frequencies using the Cramer-Kronig transform. Determine the phase of each frequency based on the resistance components in the sub-band, and use least squares linear regression to fit the phase-frequency relationship to determine the reference phase change rate. For any sub-band, the reactance component at the peak position and main resonance frequency is subtracted from the reference reactance component at the corresponding frequency, and then divided by the mean of the reference reactance components at the corresponding frequency in all sub-bands, and then squared to obtain the reactance error at the corresponding frequency. The difference between the phase change rate of the corresponding sub-band and the reference phase change rate is taken as the numerator, and the mean of the reference phase change rates in all sub-bands is taken as the denominator. The square of the ratio of the numerator to the denominator is taken as the phase change error. The average of the reactance error and phase change error at the peak position and main resonance frequency of all sub-bands is taken, the square root is taken, and then normalized. The normalized value is subtracted from 1 to obtain the physical consistency index.

[0011] Preferably, the process of calculating the signal-to-noise ratio in S2 is as follows: For each sampling frequency, the impedance amplitude of the original impedance spectrum data and the filtered impedance spectrum data at the corresponding frequency is subtracted to obtain the noise amplitude at the corresponding frequency. The average of the squared impedance amplitudes of the original impedance spectrum data at all sampling frequencies is taken as the numerator, and the average of the squared noise amplitudes at all sampling frequencies is taken as the denominator. The logarithm of the ratio of the numerator to the denominator is multiplied by a predefined signal-to-noise ratio coefficient and then normalized. The signal-to-noise ratio index is obtained by subtracting the normalized value from 1.

[0012] Preferably, the process of calculating the characteristic stability index in S2 is as follows: Multiple different noises were added to the original impedance spectrum data, and the peak position, phase change rate, and main resonance frequency with different noise disturbances were obtained based on the data side parameter set. For each sub-band, the mean and standard deviation of the peak position, phase change rate, and main resonance frequency of the corresponding band under all noise disturbances were calculated. The ratio of the standard deviation to the mean was used as the coefficient of variation, and the mean of the coefficients of variation of the peak position, phase change rate, and main resonance frequency was subtracted from 1 to obtain the characteristic stability index.

[0013] Preferably, the process of adjusting the data-side parameter set in S2 is as follows: All initial data-side parameter sets and corresponding comprehensive evaluation indicators are used as observation data to form an observation dataset; the maximum value of the comprehensive evaluation indicator in the observation dataset is determined as the optimal comprehensive evaluation indicator. In each round of optimization iteration, based on the observation dataset at the current iteration number, the distribution of the comprehensive evaluation index is fitted by Gaussian process regression; the expected improvement acquisition function is adopted to determine the set of data-side parameters when the comprehensive evaluation index is maximized within the predefined parameter space; Take the comprehensive evaluation index corresponding to the data-side parameter set under the current iteration number, and add the data-side parameter set and the corresponding comprehensive evaluation index as new observation samples to the observation dataset; when the maximum iteration number is reached or the difference between the optimal comprehensive evaluation index under adjacent iteration numbers is less than a predetermined difference threshold, output the data-side parameter set with the maximum comprehensive evaluation index in the current observation dataset as the optimal data-side parameter set.

[0014] Preferably, the process of calculating the confidence interval of the comprehensive evaluation index and the drift health in S4 is as follows: Collect the most recent impedance spectrum samples using a rolling window of length W, and calculate the comprehensive evaluation index of the corresponding impedance spectrum samples under the optimal data side parameter set; The comprehensive evaluation index of the impedance spectrum samples is resampled to obtain the confidence interval of the comprehensive evaluation index; The maximum mean difference between the impedance spectrum samples and the training samples is calculated and normalized. The drift health is obtained by subtracting the normalized value from 1.

[0015] Preferably, the process of calculating the confidence level of the combined diagnosis in S4 is as follows: Based on the trained fault classification model, the class probability of the output impedance spectrum sample under different fault categories is calculated. For all fault categories, the logarithm of the class probability under different fault categories is multiplied by the corresponding class probability and then averaged. The average value is then subtracted from 1 to obtain the prediction entropy and certainty. The comprehensive evaluation index is normalized, and the normalized comprehensive evaluation index, prediction entropy certainty, and drift health are multiplied by a predefined exponential coefficient to obtain the joint diagnostic credibility of the corresponding impedance spectrum sample.

[0016] Preferably, the process of correcting the model parameters in S4 is as follows: When the confidence level of the joint diagnosis is not greater than a predetermined confidence threshold, a feedback sample is constructed. When the upper limit of the confidence level of the comprehensive evaluation index of the impedance spectrum samples is less than the predetermined confidence threshold, the data-side parameter set is corrected based on the updated training dataset; the hyperparameters of the fault classification model are updated based on the updated training dataset until the joint diagnostic confidence of the impedance spectrum samples is greater than the predetermined confidence threshold, at which point the correction stops.

[0017] Preferably, the process of constructing feedback samples is as follows: Obtain the fault category label and actual spatial location of the impedance spectrum sample; use the fault category label and the corresponding original cable impedance spectrum data as feedback samples; calculate the classification error between the predicted fault category and the fault category label of the feedback sample; use the difference between the predicted spatial location and the actual spatial location as the numerator and the difference between the upper and lower limits of the spatial location as the denominator to calculate the positioning error. Subtract the joint diagnosis confidence from 1 to get the sample unreliability; subtract the drift health from 1 to get the sample drift; and weight the sample unreliability, sample drift, classification error, and localization error of the feedback sample and limit them to a predetermined range to obtain the sample selection weight. When the sample selection weight is greater than the predetermined selection threshold, the corresponding feedback sample is added to the training dataset.

[0018] A second aspect of the present invention provides a broadband impedance spectrum pattern recognition system based on Bayesian optimization, employing a broadband impedance spectrum pattern recognition method based on Bayesian optimization, comprising: The data acquisition module collects and processes the broadband impedance spectrum of the cable to generate a training dataset. The feature extraction module performs smoothing filtering and normalization on each sample in the training dataset, extracts features using a frequency band partitioning strategy, and combines them into a multi-dimensional feature vector. The frequency band partitioning strategy and the filtering window size are used as the data-side parameter set to initialize multiple sets of data-side parameters. Physical consistency, signal-to-noise ratio, and feature stability indices are calculated. The logarithms of each index are taken and weighted to obtain the exponent of a predefined exponential function, yielding a comprehensive evaluation index. Using the comprehensive evaluation index as the objective, the data-side parameter set is optimized using Bayesian methods to determine the multi-dimensional feature vector under the optimal data-side parameter set. The fault category identification module inputs multi-dimensional feature vectors into a predefined fault classification model to train the model; based on the trained model, it identifies the fault type from the impedance spectrum data of the cable under test. The fault classification model correction module generates impedance spectrum samples based on the impedance spectrum data of the cable under test and calculates the corresponding comprehensive evaluation index, confidence interval of the comprehensive evaluation index, and drift health. Based on the fault classification model, it outputs the class probability of the samples and calculates the prediction entropy certainty. Based on the comprehensive evaluation index, drift health, and prediction entropy certainty of the samples, it calculates the joint diagnosis credibility and corrects the model parameters.

[0019] A third aspect of the present invention provides a computer device including a memory and a processor, the memory storing a computer program, wherein the processor executes the computer program to implement the steps of a Bayesian-optimized broadband impedance spectrum pattern recognition method.

[0020] A fourth aspect of the present invention provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps of a Bayesian-optimized broadband impedance spectrum pattern recognition method.

[0021] Compared with the prior art, the beneficial effects of the present invention include at least the following: 1. This invention constructs a data-side parameter set by combining frequency band division strategies and filter window sizes, and dynamically adjusts this parameter set using Bayesian optimization. A comprehensive evaluation index is used as the optimization target to achieve adaptive optimization of the impedance spectrum feature extraction process. Simultaneously, physical consistency, signal-to-noise ratio, and feature stability indices are introduced into the comprehensive evaluation index to provide multi-dimensional evaluation of impedance spectrum features. This makes the extracted peak positions, phase change rates, and main resonance frequencies more consistent with the physical characteristics of the cable, thereby improving the accuracy and representativeness of feature extraction.

[0022] 2. This invention constructs a multi-dimensional feature vector based on broadband impedance spectrum data, and inputs the Bayesian-optimized features into a fault classification model for training, enabling the model to more effectively distinguish different fault types. Simultaneously, by extracting and analyzing the peak positions, amplitude differences, and resonance points in the impedance spectrum, the feature differences corresponding to impedance discontinuities can be amplified, thereby enhancing the contrast between defect signals and normal region signals, improving the accuracy of cable fault type identification, and increasing the precision of fault spatial location.

[0023] 3. This invention introduces comprehensive evaluation indicators—confidence interval, drift health, and prediction entropy certainty—in the diagnostic phase. A joint diagnostic credibility is constructed using these indicators to comprehensively evaluate the reliability of the identification results. When the joint diagnostic credibility falls below a preset threshold, feedback samples are constructed and the training dataset is updated. This dynamically corrects the data-side parameter set and model hyperparameters, forming a closed-loop feedback mechanism that maintains stable diagnostic performance under different cable types, operating environments, and conditions with noise interference or data distribution changes, thereby improving the reliability and robustness of fault identification results. Attached Figure Description

[0024] Figure 1 This is a flowchart of a broadband impedance spectrum pattern recognition method based on Bayesian optimization provided by the present invention. Detailed Implementation

[0025] To make the objectives, technical solutions, and advantages of this invention clearer, the technical solutions of this invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of this invention. The described embodiments are merely some embodiments of this invention, and not all embodiments. Based on the spirit of this invention, all other embodiments obtained by those skilled in the art without creative effort are within the protection scope of this invention.

[0026] Example 1 Embodiment 1 of the present invention provides a broadband impedance spectrum pattern recognition method based on Bayesian optimization, see below. Figure 1 This includes the following steps: Step 1: Collect and process the broadband impedance spectrum of the cable to generate a training dataset.

[0027] Preferably, a wideband impedance analyzer, such as, but not limited to, an Agilent E4990A, is used to measure the cable and obtain curves showing the amplitude and phase of its input impedance as a function of frequency. Preferably, but not limitingly, the measurement frequency range is from 10 Hz to 10 MHz to ensure coverage of the main frequency bands of the cable's characteristic response.

[0028] The measured broadband impedance spectrum data is stored in complex form and recorded as follows: ,in, For resistance components, The reactance component; the reactance component of the imaginary part in the measured impedance spectrum data. The complex impedance spectrum is used as the raw impedance spectrum data, and the raw impedance spectrum data of all cables at the measurement frequency are regarded as a multi-dimensional input vector. Here, n represents the number of samples, i.e., the number of cables collected. Each raw data vector and its corresponding actual cable condition (e.g., normal, bubble, scratch, etc.) is used as a ground truth label to generate a training dataset. This training dataset contains complete data of raw electrical information across the entire frequency band and its corresponding ground truth labels, providing a foundation for subsequent Bayesian optimization through iterative backtracking.

[0029] Step 2: Perform smoothing filtering and normalization on the raw impedance spectrum data of each sample in the training dataset. Extract features using a frequency band partitioning strategy and combine them into a multidimensional feature vector. Use the frequency band partitioning strategy and the filter window size as the data-side parameter set to initialize multiple sets of data-side parameter sets. Based on the multidimensional feature vectors under the data-side parameter sets, calculate the physical consistency index, signal-to-noise ratio index, and feature stability index respectively. Take the logarithm of each index and weight them to obtain the exponent of an exponential function with the natural constant as the base, thus obtaining the comprehensive evaluation index. Using the comprehensive evaluation index under the data-side parameter set as the objective, dynamically adjust the data-side parameter set through Bayesian optimization to determine the multidimensional feature vector under the optimal data-side parameter set.

[0030] As a preferred implementation method, the process of extracting features using a specific frequency band division strategy is as follows: Based on multiple predefined frequency bands, the collected wide-range raw data is divided into impedance spectrum data of the corresponding frequency bands; Wideband measurement range The frequency bands are divided based on a preset number of frequency bands K (K=3 in this embodiment), and the frequency band boundary parameters are expressed as follows: ,in And satisfy To ensure that the dimension of the subsequent eigenvectors remains constant; the filtered and normalized impedance spectrum data Z(f) (and its amplitude (|Z(f)|) and phase By boundary The frequency bands were divided into sub-bands. ; For any sub-band, the frequency corresponding to the maximum impedance amplitude within that band is taken as the peak position of the impedance amplitude; specifically expressed as: ; In the formula, This represents the peak position of the impedance amplitude in the k-th sub-band; Let Z(f) be the amplitude of the impedance at frequency f; Phase expansion to eliminate Jump effect, i.e., take Considering scale consistency across a wide frequency range, taking the logarithm of the frequency is... The logarithmic frequency is obtained; after selecting a predetermined number of discrete sampling points within each sub-band, the phase-frequency is fitted using least-squares linear regression. The relationship, where the slope is used as the rate of phase change, is specifically expressed as: ; in, The phase change rate in the kth sub-band; This represents the predetermined number of samples in the k-th sub-band. Let be the logarithmic frequency of the i-th discrete sampling point; The phase of the i-th discrete sampling point; The average logarithmic frequency; Average phase; For each sub-band, the impedance amplitude of the next frequency within the sub-band is subtracted from the current frequency. This serves as the amplitude difference at the current frequency; when the amplitude difference at the previous frequency is greater than zero and the amplitude difference at the current frequency is not greater than zero, i.e. and If the impedance amplitude at the current frequency is greater than the predefined peak significance threshold, the current frequency is used as a local peak candidate point to suppress noise spurious peaks. When the product of the reactance components at the current frequency and the next frequency is less than zero, that is... At that time, it is considered that in the current frequency and the next frequency range The memory is at the zero crossover point, and linear interpolation is used. Candidate frequencies are obtained; based on the impedance magnitudes at adjacent frequencies of the candidate frequencies, linear interpolation is performed. This yields the impedance amplitude at the corresponding candidate frequency; Local peak candidate points and candidate frequencies are used as candidate resonance points to construct a set of candidate resonance points. When the set of candidate resonance points When multiple candidate resonance points exist, to ensure a fixed number of output features for each frequency band, it is preferable to select only the one with the largest amplitude as the main resonance frequency for that frequency band. If no candidate resonance point exists in the frequency band, the main resonance point frequency will be determined according to the preset default rules. The values ​​are assigned and processed uniformly in the subsequent process; in this embodiment, the peak position of the impedance amplitude is taken as the main resonant frequency. Extracting each frequency band The features are sequentially concatenated to form a fixed-dimensional multi-dimensional feature vector, which is then used as input to the classification model. During the Bayesian optimization iteration process, only the following adjustments are made: The inner boundary parameters and filtering window parameters are optimized using classification accuracy or equivalent evaluation index as the objective function, thereby achieving adaptive optimization of the feature extraction process while keeping the feature dimension constant.

[0031] In this embodiment, based on the typical characteristics of the cable impedance spectrum, the wideband range is divided into several specific frequency bands; the principle of frequency band division is to ensure that each frequency band covers a region in the impedance spectrum where a certain type of characteristic changes significantly. For example, the low-frequency band reflects the overall trend of the cable's reactance, the mid-frequency band is the region where the impedance spectrum shows inflection points or rapid phase changes, and the high-frequency band often exhibits resonance peaks or subtle fluctuations. When the measurement frequency range is 10 Hz to 10 MHz, this range is divided into three sub-bands: 10 Hz to 1 kHz (low-frequency band), 1 kHz to 100 kHz (mid-frequency band), and 100 kHz to 10 MHz (high-frequency band). Specifically, b0 = 10 Hz, b3 = 10 MHz, and [b0, b1] is designated as the low-frequency band, [b1, b2] as the mid-frequency band, and [b2, b3] as the high-frequency band. The initial values ​​of b1 and b2 can be set to 1 kHz and 100 kHz, respectively. Parameters such as b1, b2, and the filter window size are used as variables to be optimized in Bayesian optimization, while the number of frequency bands K remains constant. This division ensures that the impedance spectrum behavior within each frequency band exhibits a relatively consistent pattern, facilitating the extraction of corresponding features.

[0032] As a preferred implementation method, the process of calculating the comprehensive evaluation index is as follows: The impedance spectrum data after filtering under the data measurement parameter set, as well as the peak position, phase change rate, and main resonance frequency of different sub-bands, are obtained. For each sub-band, the corresponding reference reactance component is calculated based on the resistance component of the peak position and the main resonance frequency through Kramer-Kronig (KK) transform. Based on the resistance component of each frequency in the sub-band, the phase of the corresponding frequency is determined, and the phase-frequency relationship is fitted by least squares linear regression, with the slope used as the reference phase change rate. For any sub-band, the reactance component at the peak position and main resonance frequency is subtracted from the reference reactance component at the corresponding frequency, then divided by the mean of the reference reactance components at the corresponding frequency in all sub-bands, and squared to obtain the reactance error at the corresponding frequency. The difference between the phase change rate of the corresponding sub-band and the reference phase change rate is used as the numerator, and the mean of the reference phase change rates in all sub-bands is used as the denominator. The square of the ratio of the numerator to the denominator is used as the phase change error. The average of the reactance error and phase change error at the peak position and main resonance frequency of all sub-bands is taken, the square root is normalized, and 1 is subtracted from the normalized value to obtain the physical consistency index, specifically expressed as: ; In the formula, As a physical consistency indicator; The normalization function can be a maximum-minimum normalization function, a sigmoid function, etc. , These are the resistance component and the reference reactance component at the peak position of sub-frequency band k, respectively; This is the average of the reference reactance components at the peak positions across all sub-bands; , These are the resistance component and the reference reactance component at the main resonant frequency under sub-band k, respectively. This is the average of the reference reactance components at the main resonant frequency in all sub-bands; The reference phase change rate at sub-frequency band k; This is the average of the reference phase change rates across all sub-bands; For each sampling frequency, the difference between the impedance amplitude of the original impedance spectrum data and the filtered impedance spectrum data at the corresponding frequency is taken as the noise amplitude at that frequency. The average of the squared impedance amplitudes of the original impedance spectrum data at all sampling frequencies is used as the numerator, and the average of the squared noise amplitudes at all sampling frequencies is used as the denominator. The logarithm of the ratio of the numerator to the denominator is multiplied by a predefined signal-to-noise ratio coefficient and then normalized. Subtracting the normalized value from 1 yields the signal-to-noise ratio (SNR) index. The specific formula is as follows: ; In the formula, Signal-to-noise ratio; It is a logarithmic function with base 10; For the filtered impedance spectrum data at frequency The impedance amplitude below; For frequency The noise amplitude is denoted by n; n is the total number of sampling frequencies. Multiple different noises were added to the original impedance spectrum data, and based on the data-side parameter set, the peak position, phase change rate, and main resonance frequency with different noise perturbations were obtained. For each sub-band, the mean and standard deviation of the peak position, phase change rate, and main resonance frequency under all noise perturbations were calculated. The ratio of the standard deviation to the mean was used as the coefficient of variation, and the mean of the coefficients of variation of the peak position, phase change rate, and main resonance frequency was subtracted from 1 to obtain the characteristic stability index. The specific formula is expressed as follows: ; In the formula, As a characteristic stability index; , Let K be the standard deviation and mean of the peak position of frequency band k under all noise disturbances; , Let be the standard deviation and mean of the phase change rate of frequency band k under all noise disturbances; , Let K be the standard deviation and mean of the main resonance frequency of frequency band k under all noise disturbances; In this embodiment, each perturbation adds small-amplitude random noise that follows a normal distribution to the real and imaginary parts of the original impedance data. By adding multiple perturbations, the original impedance data under multiple perturbations is obtained.

[0033] The physical consistency index, signal-to-noise ratio, and characteristic stability index under the data measurement parameter set are weighted by taking the logarithm and then using the weighted sum as the exponent of an exponential function with the natural constant as the base, to obtain the comprehensive index; specifically expressed as: ; In the formula, A comprehensive index based on a set of data parameters; It is an exponential function with the natural constant as its base; For the parameter set of data, the index m is used; These represent the physical consistency index, signal-to-noise ratio, and feature stability index, respectively. A logarithmic function with the natural constant as its base; To prevent singular constants; The adjustment coefficient for index m, with values ​​all within the interval (0,1) and .

[0034] As a preferred implementation method, the process of adjusting the data-side parameter set is as follows: The number of frequency bands K is preset to be a fixed constant, and the inner boundary of the frequency band and the size of the filter window are defined as the parameter vector to be optimized, generating a set of data-side parameters; specifically, it is expressed as follows: ; in, For the set of parameters on the data side; For the k-th frequency band, and satisfying w is the size of the smoothing filter window; To avoid confusing the performance of the classification model output with the quality of the input data itself, the parameter set involved in this method is divided into two categories: data-side parameter set and model-side parameter set. Used to characterize impedance spectrum preprocessing and feature extraction processes, including but not limited to frequency band boundary parameters. (in and the size of the smoothing filter window w etc.; Model-side parameter set Used to characterize the composition and training hyperparameters of a classification model, including but not limited to network structure, learning rate, regularization coefficient, batch size, optimizer, and loss function.

[0035] Initialize multiple sets of data-side parameters, obtain the multi-dimensional feature vector under each set of data-side parameters, and calculate the corresponding comprehensive evaluation index; use all initial sets of data-side parameters and corresponding comprehensive evaluation indices as observation data to form an observation dataset; determine the maximum value of the comprehensive evaluation index in the observation dataset as the optimal comprehensive evaluation index. In each optimization iteration, based on the observation dataset at the current iteration number, the distribution of the comprehensive evaluation index is fitted using Gaussian process regression; the expected improvement acquisition function is used to determine the set of data-side parameters that maximize the comprehensive evaluation index within a predefined parameter space; the specific formula is expressed as: ; In the formula, This is the set of data-side parameters for the current iteration number. For the set of parameters on the data side; For data-side parameter set The following comprehensive evaluation indicators; The optimal comprehensive evaluation index; The expected function is defined in this embodiment as the range of values ​​for the parameters in the data-side parameter set. Obtain the comprehensive evaluation index corresponding to the data-side parameter set under the current iteration number, and add the data-side parameter set and the corresponding comprehensive evaluation index as new observation samples to the observation dataset; when the maximum iteration number is reached or the difference between the optimal comprehensive evaluation index under adjacent iteration numbers is less than a predetermined difference threshold, output the data-side parameter set when the comprehensive evaluation index in the current observation dataset is the largest, as the optimal data-side parameter set.

[0036] As an alternative implementation method, optimization algorithms such as particle swarm optimization, ant colony optimization, and genetic algorithm can be used to obtain the optimal set of data-side parameters by using the comprehensive evaluation index as the fitness function.

[0037] Step 3: Input the multidimensional feature vectors of all training samples into the predefined fault classification model, and optimize the model hyperparameters with the fault classification accuracy of the model as the target; based on the trained fault classification model, identify the impedance spectrum data of the cable under test and determine the fault type.

[0038] Preferably, but not limitingly, step 3 specifically includes: Step 3.1: Use a machine learning model to classify the feature vectors extracted in Step 2.

[0039] Preferably, but not restrictively, a neural network is used as the fault mode classification model, with a multidimensional feature vector as input.

[0040] This embodiment automatically adjusts the hyperparameters of the classification model through Bayesian optimization. Preferably, the hyperparameters include the kernel function type, regularization coefficient, and learning rate among the collected parameters, thereby optimizing its classification performance.

[0041] Step 3.2: Construct a Bayesian optimization model, including: establishing the objective function, modeling and predicting the objective function, selecting the next set of parameter combinations and updating the objective function model in each round of optimization.

[0042] Preferably, but not restrictively, the iterative process of the Bayesian optimization terminates when any of the following stopping conditions are met: 1) Maximum number of iterations condition: the search stops when the number of iterations reaches a preset upper limit; 2) Performance convergence condition: the failure mode classification accuracy is used as the optimization objective function value. = ( When continuous Optimal classification accuracy obtained from rounds of iteration + If the improvement is insufficient, convergence is determined and the process stops, i.e., it satisfies | + + ∣< (in The iteration terminates when the preset precision threshold is reached. Preferably, but not limitingly, step 3.2 specifically includes: Establish an objective function with the goal of optimizing the fault mode classification accuracy. Construct a Bayesian optimization objective function; the objective function has the following form:

[0043] Where Objective is the objective function; To optimize parameters; Let it be the expected function; It is a function of random variables; This is an impedance spectrum feature dataset.

[0044] Modeling and predicting the objective function includes constructing a Bayesian optimization model and optimizing the hyperparameters of the kernel function. Preferably, Gaussian Process Regression is used as a surrogate model for the optimizer.

[0045] Using fault classification accuracy as the objective function, a Bayesian optimization model is constructed using a Gaussian process: ; in, Represent a Gaussian process; Let be the mean function of a Gaussian process; This is a kernel function used to characterize the correlation between the objective function values ​​of different hyperparameter combinations; A set of hyperparameters; This is another set of candidate hyperparameters.

[0046] Furthermore, for the kernel function obtained in step 2, the hyperparameters of the kernel function are optimized by maximizing the marginal likelihood: ; In the formula, It is a logarithmic function; The hyperparameter sample set is used for evaluation. θ i For the first i The combination of hyperparameters of the classification model selected in the next iteration. K It is the covariance matrix; for K The determinant of; For kernel hyperparameters, at least the length scale l and signal variance are included. .

[0047] Those skilled in the art will readily recognize that optimizing the hyperparameters of the kernel function allows the Gaussian process to better fit the observed data, thereby improving the efficiency of Bayesian optimization and enhancing the overall accuracy of pattern recognition and model performance. By dynamically adjusting the feature extraction frequency band segmentation strategy and the filter window size through Bayesian optimization, the representativeness and stability of the input features are improved, which in turn affects the effect of the kernel function. These two aspects complement each other, jointly optimizing the entire feature extraction and classification process.

[0048] As a preferred implementation, in each round of optimization, the next set of parameter combinations is selected using the acquisition function, and the objective function model is updated; the specific process is as follows: Preferably, but not restrictively, the next set of parameter combinations is selected based on the Expected Improvement (EI) criterion, and the objective function model is updated accordingly.

[0049]

[0050] in, and These are the mean and variance predicted by the Gaussian process, which are obtained by optimizing the hyperparameters of the kernel function by maximizing the marginal likelihood and outputting the posterior mean and variance for any point. To explore the parameters, a value of 0.01 is used in this embodiment; The value of x is approximated in the positive direction; The cumulative distribution function of the standard normal distribution; The cumulative density function of the standard normal distribution; These are the standardized normal variables.

[0051] In one preferred implementation, during actual pattern recognition and diagnosis, the test data consists of on-site data of unknown fault types. The model analyzes this data to predict the defect type and determines the corresponding fault location based on the fault type and the corresponding raw impedance spectrum data. During the model training or validation phase, the test data consists of data labeled with known faults, used to evaluate the model's accuracy and optimization performance.

[0052] The prediction results are classified into patterns. Preferably, the fault types are: bubble defects, scratch defects, and dirt defects, etc., and the defect types and locations of bubbles, scratches and dirt are identified. The recognition accuracy is gradually improved by dynamically updating the hyperparameters of the classification model.

[0053] Step 4: Based on a predefined length, obtain impedance spectrum samples from the impedance spectrum data of the cable under test, and calculate the comprehensive evaluation index, confidence interval of the comprehensive evaluation index, and drift health of the corresponding samples; based on the trained fault classification model, output the class probability of the impedance spectrum samples and calculate the corresponding prediction entropy certainty; based on the comprehensive evaluation index, drift health, and prediction entropy certainty of the impedance spectrum samples, calculate the joint diagnosis credibility; when it is not greater than the predetermined credibility threshold, construct feedback samples to correct the model parameters.

[0054] Preferably, but not limitingly, during runtime, the length is... W scrolling window Collect the most recent impedance spectrum samples and calculate the comprehensive evaluation index of the corresponding impedance spectrum samples under the optimal data side parameter set; specifically expressed as: ; in, For scrolling windows Comprehensive evaluation index of impedance spectrum samples; This is the optimal set of data-side parameters; Preferably, the comprehensive evaluation index for impedance spectrum samples Bootstrap resampling Next, the confidence interval of the comprehensive evaluation index was obtained. , used to characterize the uncertainty of online estimation (where (This is a preset constant, preferably 1000).

[0055] Calculate and normalize the maximum mean difference between the impedance spectrum samples and the training samples. Subtract the normalized value from 1 to obtain the drift health score. .

[0056] Preferably, but not restrictively, based on the trained fault classification model, the class probabilities of the output impedance spectrum samples under different fault categories are calculated; for all fault categories, the logarithm of the class probabilities under different fault categories is taken, multiplied by the corresponding class probabilities, and then averaged. The prediction entropy certainty is obtained by subtracting the average from 1; the specific formula is as follows: ; In the formula, Let be the prediction entropy certainty of the t-th impedance spectrum sample; This represents the total number of fault categories; Let be the class probability of the t-th impedance spectrum sample under fault category c; To prevent singular constants; The comprehensive evaluation index is normalized, and then multiplied by the normalized comprehensive evaluation index, prediction entropy certainty, and drift health using a predefined exponential coefficient to obtain the joint diagnostic confidence of the corresponding impedance spectrum sample; specifically expressed as: ; in, The reliability of the joint diagnosis for the t-th impedance spectrum sample; This is a truncation function, used here to restrict the normalized comprehensive evaluation index to the interval (0,1); and These are the predefined lower and upper limits of the comprehensive evaluation indicators; , , These are the index coefficients for comprehensive evaluation indicators, predictive entropy certainty, and drift health, respectively. When the confidence level of the joint diagnosis of impedance spectrum samples is not greater than a predetermined confidence threshold, i.e. At that time, the corresponding classification result of the impedance spectrum sample is marked as low confidence output and the feedback acquisition process is triggered to obtain the actual cable status and generate feedback samples with truth labels.

[0057] As a preferred implementation method, the process of constructing feedback samples is as follows: In delay Then, the actual cable condition, i.e., the fault category label, is obtained from the impedance spectrum sample. and actual spatial location Based on the fault category label and the corresponding original cable impedance spectrum data as feedback samples; calculate the classification error between the predicted fault category and the fault category label of the feedback sample, where 1 indicates an incorrect prediction and 0 indicates a correct one; use the difference between the predicted spatial location and the actual spatial location as the numerator and the difference between the upper and lower limits of the spatial location as the denominator to calculate the positioning error; specifically expressed as: ; ; In the formula, This represents classification error; To predict the type of failure; This is an indicator function; the value inside the parentheses is 1 if the condition is met, and 0 otherwise. This is the positioning error; For the predicted spatial location; , The upper and lower limits of the predefined spatial locations; Subtracting the joint diagnostic confidence from 1 yields the sample unreliability; subtracting the drift health from 1 yields the sample drift; and the sample selection weight is obtained by weighting the sample unreliability, sample drift, classification error, and localization error of the feedback sample and limiting it within a predetermined range; specifically expressed as: ; In the formula, Weights are selected for the samples to be used in the subsequent rolling update process for sample selection and weighted training. The larger the weight, the greater the difference between the sample and the training samples and the greater the model prediction error. The corresponding feedback sample should be included in the training dataset for retraining. , , , These are adjustment coefficients for sample unreliability, sample drift, classification error, and localization error, respectively. When the sample selection weight is greater than the predetermined selection threshold, the corresponding feedback sample is added to the training dataset, and the training dataset is updated.

[0058] As a preferred implementation method, the process of correcting the model parameters is as follows: When the upper confidence limit of the comprehensive evaluation index of the impedance spectrum sample is less than the predetermined confidence threshold When the drift is significant (in this embodiment, it means that the sample drift of the corresponding sample is greater than a predetermined drift threshold), the data-side parameter set is corrected based on the updated training dataset, that is, step 2 is triggered to maximize the value. The Bayesian optimization process yields a new set of data-side parameters. Based on the updated training dataset, the hyperparameters of the fault classification model are updated, which triggers step 3; the correction stops when the joint diagnostic confidence of the impedance spectrum samples is greater than the predetermined confidence threshold.

[0059] Example 2 Embodiment 2 of the present invention provides a broadband impedance spectrum pattern recognition system based on Bayesian optimization, which uses a broadband impedance spectrum pattern recognition method based on Bayesian optimization, including: The data acquisition module collects and processes the broadband impedance spectrum of the cable to generate a training dataset. The feature extraction module performs smoothing filtering and normalization on each sample in the training dataset, extracts features using a frequency band partitioning strategy, and combines them into a multi-dimensional feature vector. The frequency band partitioning strategy and the filtering window size are used as the data-side parameter set to initialize multiple sets of data-side parameters. Physical consistency, signal-to-noise ratio, and feature stability indices are calculated. The logarithms of each index are taken and weighted to obtain the exponent of a predefined exponential function, yielding a comprehensive evaluation index. Using the comprehensive evaluation index as the objective, the data-side parameter set is optimized using Bayesian methods to determine the multi-dimensional feature vector under the optimal data-side parameter set. The fault category identification module inputs multi-dimensional feature vectors into a predefined fault classification model to train the model; based on the trained model, it identifies the fault type from the impedance spectrum data of the cable under test. The fault classification model correction module generates impedance spectrum samples based on the impedance spectrum data of the cable under test and calculates the corresponding comprehensive evaluation index, confidence interval of the comprehensive evaluation index, and drift health. Based on the fault classification model, it outputs the class probability of the samples and calculates the prediction entropy certainty. Based on the comprehensive evaluation index, drift health, and prediction entropy certainty of the samples, it calculates the joint diagnosis credibility and corrects the model parameters.

[0060] This disclosure can be a system, method, and / or computer program product. A computer program product may include a computer-readable storage medium having computer-readable program instructions loaded thereon for causing a processor to implement various aspects of this disclosure.

[0061] Computer-readable storage media can be tangible devices capable of holding and storing instructions for use by an instruction execution device. Computer-readable storage media can be, for example—but not limited to—electrical storage devices, magnetic storage devices, optical storage devices, electromagnetic storage devices, semiconductor storage devices, or any suitable combination of the foregoing. More specific examples (a non-exhaustive list) of computer-readable storage media include: portable computer disks, hard disks, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM or flash memory), static random access memory (SRAM), portable compact disc read-only memory (CD-ROM), digital multifunction disc (DVD), memory sticks, floppy disks, mechanical encoding devices, such as punch cards or recessed protrusions storing instructions thereon, and any suitable combination of the foregoing. The computer-readable storage media used herein are not to be construed as transient signals themselves, such as radio waves or other freely propagating electromagnetic waves, electromagnetic waves propagating through waveguides or other transmission media (e.g., light pulses through fiber optic cables), or electrical signals transmitted through wires.

[0062] The computer-readable program instructions described herein can be downloaded from computer-readable storage media to various computing / processing devices, or downloaded via a network, such as the Internet, local area network, wide area network, and / or wireless network, to an external computer or external storage device. The network may include copper transmission cables, fiber optic transmission, wireless transmission, routers, firewalls, switches, gateway computers, and / or edge servers. A network adapter card or network interface in each computing / processing device receives the computer-readable program instructions from the network and forwards them to the computer-readable storage media in the respective computing / processing device.

[0063] Computer program instructions used to perform the operations of this disclosure may be assembly instructions, instruction set architecture (ISA) instructions, machine instructions, machine-dependent instructions, microcode, firmware instructions, status setting data, or source code or object code written in any combination of one or more programming languages, including object-oriented programming languages ​​such as Smalltalk, C++, etc., and conventional procedural programming languages ​​such as the "C" language or similar programming languages. The computer-readable program instructions may execute entirely on the user's computer, partially on the user's computer, as a standalone software package, partially on the user's computer and partially on a remote computer, or entirely on a remote computer or server. In cases involving a remote computer, the remote computer may be connected to the user's computer via any type of network—including a local area network (LAN) or a wide area network (WAN)—or may be connected to an external computer (e.g., via the Internet using an Internet service provider). In some embodiments, electronic circuitry, such as programmable logic circuitry, field-programmable gate arrays (FPGAs), or programmable logic arrays (PLAs), is personalized by utilizing the status information of the computer-readable program instructions to implement various aspects of this disclosure.

[0064] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and not to limit it. Although the present invention has been described in detail with reference to the above embodiments, those skilled in the art should understand that modifications or equivalent substitutions can still be made to the specific implementation of the present invention. Any modifications or equivalent substitutions that do not depart from the spirit and scope of the present invention should be covered within the protection scope of the claims of the present invention.

Claims

1. A broadband impedance spectrum pattern recognition method based on Bayesian optimization, characterized in that, include: S1. Collect and process the broadband impedance spectrum of the cable to generate a training dataset; S2. Perform smoothing filtering and normalization on each sample in the training dataset. Extract features using a frequency band division strategy and combine them into a multi-dimensional feature vector. Use the frequency band division strategy and the filter window size as the data-side parameter set, initialize multiple sets of data-side parameter sets, and calculate the physical consistency index, signal-to-noise ratio, and feature stability index. Take the logarithm of each index and weight it as the exponent of a predefined exponential function to obtain a comprehensive evaluation index. Using the comprehensive evaluation index as the target, optimize the data-side parameter set through Bayesian optimization to determine the multi-dimensional feature vector under the optimal data-side parameter set. S3. Input the multi-dimensional feature vector into a predefined fault classification model and train the model. Based on the trained model, identify the fault type using the impedance spectrum data of the cable under test. S4. Based on the impedance spectrum data of the cable under test, generate impedance spectrum samples and calculate the corresponding comprehensive evaluation index, confidence interval of the comprehensive evaluation index, and drift health. Based on the fault classification model, the class probability of the output sample is calculated, and the prediction entropy and certainty are determined. Based on the comprehensive evaluation indicators of the samples, drift health, and predictive entropy certainty, the credibility of joint diagnosis is calculated and the model parameters are corrected.

2. A broadband impedance spectrum pattern recognition method based on Bayesian optimization as described in claim 1, characterized in that: In S1, the reactance component of the imaginary part and the complex impedance spectrum in the measured broadband impedance spectrum data are used as the original impedance spectrum data. Each original impedance spectrum data and the actual cable state corresponding to the measurement are used as the true value label to generate a training dataset.

3. A broadband impedance spectrum pattern recognition method based on Bayesian optimization as described in claim 1, characterized in that: The process of extracting features in S2 using a specific frequency band division strategy is as follows: Within a wide-band measurement range, a preset number of frequency bands are used to divide the frequency bands and obtain the frequency band boundary parameters. The filtered and normalized impedance spectrum data is then divided into sub-bands according to the frequency band boundary parameters. For any sub-band, the frequency corresponding to the maximum impedance amplitude within the corresponding frequency band is taken as the peak position of the impedance amplitude. Take the logarithm of the frequency to obtain the logarithmic frequency; after selecting a predetermined number of discrete sampling points in each sub-band, use least squares linear regression to fit the phase-frequency relationship, and use the slope as the phase change rate. For each sub-band, the difference between the impedance amplitude of the next frequency and the current frequency within the sub-band is used as the amplitude difference at the current frequency. When the amplitude difference at the previous frequency is greater than zero and the amplitude difference at the current frequency is not greater than zero, if the impedance amplitude at the current frequency is greater than the predefined peak significance threshold, the current frequency is used as a local peak candidate point; when the product of the reactance components at the current frequency and the next frequency is less than zero, the candidate frequency is obtained by linear interpolation; based on the impedance amplitude at the adjacent frequencies of the candidate frequency, the impedance amplitude of the corresponding candidate frequency is obtained by linear interpolation. Local peak candidate points and candidate frequencies are used as candidate resonance points to construct a candidate resonance point set; when there are multiple candidate resonance points in the candidate resonance point set, the one with the largest amplitude is taken as the main resonance point frequency of the frequency band. The peak positions, phase change rates, and main resonance frequencies extracted from each frequency band are sequentially concatenated to form a fixed-dimensional multidimensional feature vector.

4. A broadband impedance spectrum pattern recognition method based on Bayesian optimization as described in claim 1, characterized in that: The process of calculating the physical consistency index in S2 is as follows: Acquire the filtered impedance spectrum data and the peak position, phase change rate and main resonance frequency of different sub-bands under the data measurement parameter set; for each sub-band, calculate the corresponding reference reactance component based on the resistance component of the peak position and the main resonance frequency through Cramer-Kronig transform. Based on the resistance component of each frequency in the sub-band, the phase of the corresponding frequency is determined, and the phase-frequency relationship is fitted by least squares linear regression to determine the reference phase change rate. For any sub-band, the reactance component at the peak position and main resonance frequency is subtracted from the reference reactance component at the corresponding frequency, and then divided by the mean of the reference reactance components at the corresponding frequency in all sub-bands, and then squared to obtain the reactance error at the corresponding frequency. The difference between the phase change rate of the corresponding sub-band and the reference phase change rate is taken as the numerator, and the mean of the reference phase change rates in all sub-bands is taken as the denominator. The square of the ratio of the numerator to the denominator is taken as the phase change error. The average of the reactance error and phase change error at the peak position and main resonance frequency of all sub-bands is taken, the square root is taken, and then normalized. The normalized value is subtracted from 1 to obtain the physical consistency index.

5. A broadband impedance spectrum pattern recognition method based on Bayesian optimization as described in claim 1, characterized in that: The process of calculating the signal-to-noise ratio in S2 is as follows: For each sampling frequency, the difference between the impedance amplitude of the original impedance spectrum data and the filtered impedance spectrum data at the corresponding frequency is taken as the noise amplitude at the corresponding frequency. The average of the squared impedance amplitudes of the original impedance spectrum data at all sampling frequencies is taken as the numerator, and the average of the squared noise amplitudes at all sampling frequencies is taken as the denominator. The logarithm of the ratio of the numerator to the denominator is taken, multiplied by a predefined signal-to-noise ratio coefficient, and then normalized. The signal-to-noise ratio index is obtained by subtracting the normalized value from 1.

6. A broadband impedance spectrum pattern recognition method based on Bayesian optimization as described in claim 1, characterized in that: The process of calculating the characteristic stability index in S2 is as follows: Multiple different noises were added to the original impedance spectrum data, and the peak position, phase change rate, and main resonance frequency with different noise disturbances were obtained based on the data side parameter set. For each sub-band, the mean and standard deviation of the peak position, phase change rate, and main resonance frequency of the corresponding band under all noise disturbances were calculated. The ratio of the standard deviation to the mean was used as the coefficient of variation, and the mean of the coefficients of variation of the peak position, phase change rate, and main resonance frequency was subtracted from 1 to obtain the characteristic stability index.

7. A broadband impedance spectrum pattern recognition method based on Bayesian optimization as described in claim 1, characterized in that: The process of adjusting the data-side parameter set in S2: All initial data-side parameter sets and corresponding comprehensive evaluation indicators are used as observation data to form an observation dataset; Determine the maximum value of the comprehensive evaluation index in the observation data as the optimal comprehensive evaluation index; In each round of optimization iteration, based on the observation dataset at the current iteration number, the distribution of the comprehensive evaluation index is fitted by Gaussian process regression; the expected improvement acquisition function is adopted to determine the set of data-side parameters when the comprehensive evaluation index is maximized within the predefined parameter space; Take the comprehensive evaluation index corresponding to the data-side parameter set under the current iteration number, and add the data-side parameter set and the corresponding comprehensive evaluation index as new observation samples to the observation dataset; when the maximum iteration number is reached or the difference between the optimal comprehensive evaluation index under adjacent iteration numbers is less than a predetermined difference threshold, output the data-side parameter set with the maximum comprehensive evaluation index in the current observation dataset as the optimal data-side parameter set.

8. A broadband impedance spectrum pattern recognition method based on Bayesian optimization as described in claim 1, characterized in that: The process for calculating the confidence interval and drift health of the comprehensive evaluation index in S4 is as follows: Collect the most recent impedance spectrum samples using a rolling window of length W, and calculate the comprehensive evaluation index of the corresponding impedance spectrum samples under the optimal data side parameter set; The comprehensive evaluation index of the impedance spectrum samples is resampled to obtain the confidence interval of the comprehensive evaluation index; The maximum mean difference between the impedance spectrum samples and the training samples is calculated and normalized. The drift health is obtained by subtracting the normalized value from 1.

9. A broadband impedance spectrum pattern recognition method based on Bayesian optimization as described in claim 1, characterized in that: The process for calculating the confidence level of the combined diagnosis in S4 is as follows: Based on the trained fault classification model, the class probability of the output impedance spectrum sample under different fault categories is calculated. For all fault categories, the logarithm of the class probability under different fault categories is multiplied by the corresponding class probability and then averaged. The average value is then subtracted from 1 to obtain the prediction entropy and certainty. The comprehensive evaluation index is normalized, and the normalized comprehensive evaluation index, prediction entropy certainty, and drift health are multiplied by a predefined exponential coefficient to obtain the joint diagnostic credibility of the corresponding impedance spectrum sample.

10. A broadband impedance spectrum pattern recognition method based on Bayesian optimization as described in claim 1, characterized in that: The process of correcting model parameters in S4 is as follows: When the confidence level of the joint diagnosis is not greater than a predetermined confidence threshold, a feedback sample is constructed. When the upper limit of the confidence level of the comprehensive evaluation index of the impedance spectrum samples is less than the predetermined confidence threshold, the data-side parameter set is corrected based on the updated training dataset; the hyperparameters of the fault classification model are updated based on the updated training dataset until the joint diagnostic confidence of the impedance spectrum samples is greater than the predetermined confidence threshold, at which point the correction stops.

11. A broadband impedance spectrum pattern recognition method based on Bayesian optimization as described in claim 10, characterized in that: The process of constructing feedback samples is as follows: Obtain the fault category label and actual spatial location of the impedance spectrum sample; use the fault category label and the corresponding original impedance spectrum data of the cable as feedback samples; Calculate the classification error between the predicted fault category and the fault category label of the feedback sample; use the difference between the predicted spatial location and the actual spatial location as the numerator and the difference between the upper and lower limits of the spatial location as the denominator to calculate the positioning error. Subtract the confidence level of the joint diagnosis from 1 to obtain the sample unreliability; subtract the drift health from 1 to obtain the sample drift. The sample selection weights are obtained by weighting the sample unreliability, sample drift, classification error, and localization error of the feedback samples and limiting them to a predetermined range. When the sample selection weight is greater than the predetermined selection threshold, the corresponding feedback sample is added to the training dataset.

12. A broadband impedance spectrum pattern recognition system based on Bayesian optimization, using the method described in any one of claims 1-11, characterized in that, include: The data acquisition module collects and processes the broadband impedance spectrum of the cable to generate a training dataset. The feature extraction module performs smoothing filtering and normalization on each sample in the training dataset, extracts features using a frequency band partitioning strategy, and combines them into a multi-dimensional feature vector. The frequency band partitioning strategy and the filtering window size are used as the data-side parameter set to initialize multiple sets of data-side parameters. Physical consistency, signal-to-noise ratio, and feature stability indices are calculated. The logarithms of each index are taken and weighted to obtain the exponent of a predefined exponential function, yielding a comprehensive evaluation index. Using the comprehensive evaluation index as the objective, the data-side parameter set is optimized using Bayesian methods to determine the multi-dimensional feature vector under the optimal data-side parameter set. The fault category identification module inputs multi-dimensional feature vectors into a predefined fault classification model to train the model; based on the trained model, it identifies the fault type from the impedance spectrum data of the cable under test. The fault classification model correction module generates impedance spectrum samples based on the impedance spectrum data of the cable under test and calculates the corresponding comprehensive evaluation index, confidence interval of the comprehensive evaluation index and drift health. Based on the fault classification model, the class probability of the output sample is calculated, and the prediction entropy and certainty are determined. Based on the comprehensive evaluation indicators of the samples, drift health, and predictive entropy certainty, the credibility of joint diagnosis is calculated and the model parameters are corrected.

13. A computer device comprising a memory and a processor, wherein the memory stores a computer program, characterized in that, When the processor executes the computer program, it implements the steps of the method according to any one of claims 1 to 11.

14. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by a processor, it implements the steps of the method according to any one of claims 1 to 11.

Images