A method and system for early cable fault detection based on adaptive fusion of multimodal features

Abstract

This application discloses a method and system for early cable fault detection based on adaptive fusion of multimodal features, relating to the field of cable fault detection technology. The method includes: monitoring and acquiring partial discharge signals of the cable using a current sensor to extract time-domain and frequency-domain features; performing early fault degree analysis to obtain time-domain and frequency-domain early degrees, processing to obtain a fused early degree, and performing cross-verification to obtain the early degree deviation; configuring a fault identification strategy to identify faults based on the time-domain and frequency-domain features to obtain cable fault detection results; and processing based on the fused early degree and the cable fault detection results to obtain the final early cable fault detection result. This solves the technical problem of existing technologies failing to fully consider feature uncertainty and decision reliability in early fault detection, leading to low accuracy in cable detection.

Description

Technical Field

[0001] This application relates to the field of cable fault detection technology, specifically to a method and system for early cable fault detection based on adaptive fusion of multimodal features. Background Technology

[0002] With the continuous expansion of urban power grids and the increasing proportion of cable lines in power transmission systems, the safety and reliability of cable operation has become a key factor in ensuring the stable operation of power systems. Early cable faults typically manifest as weak abnormalities such as partial discharge. If these are not identified and addressed in a timely manner, they can easily develop into serious faults such as insulation breakdown, causing widespread power outages and economic losses.

[0003] However, existing early fault detection technologies for cables do not fully consider the balance between characteristic uncertainty and decision reliability in early fault detection, and lack an effective mutual verification and calibration mechanism, resulting in low accuracy in cable detection. Summary of the Invention

[0004] This application provides a method and system for early cable fault detection using multimodal feature adaptive fusion, which solves the problem that existing technologies do not fully consider feature uncertainty and decision reliability in early fault detection, resulting in low accuracy in cable detection.

[0005] The technical solution to the above-mentioned technical problems in this application is as follows:

[0006] In a first aspect, this application provides a method for early cable fault detection based on adaptive fusion of multimodal features, the method comprising:

[0007] Partial discharge signals of the cable are monitored and collected using a current sensor, and time-domain and frequency-domain features are extracted.

[0008] Based on the time-domain and frequency-domain characteristics, cable fault early-age analysis is performed to obtain time-domain early-age and frequency-domain early-age characteristics. The results are then processed to obtain a fused early-age characteristic, and mutual verification is performed to obtain the early-age deviation.

[0009] Based on the fusion early degree and early degree deviation, a fault identification strategy is configured to perform fault identification on the time domain features and frequency domain features to obtain cable fault detection results.

[0010] Based on the early fusion rate and cable fault detection results, the early cable fault detection results are obtained.

[0011] Secondly, this application provides a cable early fault detection system based on multimodal feature adaptive fusion, including:

[0012] The feature extraction module is used to monitor and collect partial discharge signals of cables through a current sensor, and extract time-domain and frequency-domain features.

[0013] The deviation calculation module is used to perform cable fault early degree analysis based on the time domain characteristics and frequency domain characteristics, obtain the time domain early degree and frequency domain early degree, process them to obtain the fused early degree, and perform mutual verification to obtain the early degree deviation.

[0014] The fault identification module is used to configure a fault identification strategy based on the fusion early degree and early degree deviation, and to identify faults in the time domain features and frequency domain features to obtain cable fault detection results.

[0015] The result acquisition module is used to process the early cable fault detection results based on the fusion early degree and the cable fault detection results to obtain the cable early fault detection results.

[0016] This application provides one or more technical solutions, which have at least the following technical effects or advantages:

[0017] This application provides a method and system for early cable fault detection based on adaptive fusion of multimodal features. First, it monitors and collects partial discharge signals from the cable using a current sensor, extracting feature information from both the time and frequency domains to achieve comprehensive multimodal data acquisition. Second, it analyzes the early fault extent of the cable based on the time and frequency domain features, obtaining and processing the time and frequency domain early fault extents to obtain a fused early fault extent. Simultaneously, it calculates the early fault extent deviation through a mutual verification mechanism, effectively quantifying the correlation between feature uncertainty and decision reliability. Third, it adaptively configures the fault identification strategy based on the fused early fault extent and the early fault extent deviation, dynamically adjusting the number of participants in the fault identification network to achieve intelligent matching of detection resources and detection requirements. Finally, it comprehensively processes the fused early fault extent and the cable fault detection results, using a credibility coefficient to compensate and correct the detection results, thereby improving the accuracy and reliability of early cable fault detection.

[0018] Through the above technical solution, this application solves the problems of single feature detection being susceptible to noise interference and insufficient decision reliability by using an adaptive fusion and mutual verification calibration mechanism of multimodal features, thereby achieving the technical effect of improving the accuracy and robustness of early cable fault detection. Attached Figure Description

[0019] To more clearly illustrate the technical solutions in the embodiments of this application, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0020] Figure 1This is a flowchart illustrating the cable early fault detection method based on multimodal feature adaptive fusion provided in the embodiments of this application.

[0021] Figure 2 This is a schematic diagram of the structure of the cable early fault detection system with multimodal feature adaptive fusion provided in the embodiments of this application.

[0022] The components represented by each number in the attached diagram are explained below:

[0023] Feature extraction module 11, deviation calculation module 12, fault identification module 13, and result acquisition module 14. Detailed Implementation

[0024] This application provides a method and system for early cable fault detection based on multimodal feature adaptive fusion, which addresses the technical problem of low cable detection accuracy caused by the prior art's failure to fully consider feature uncertainty and decision reliability in early fault detection.

[0025] Example 1, as Figure 1 As shown in the embodiments of this application, a method for early cable fault detection based on adaptive fusion of multimodal features is provided, including:

[0026] S10: The partial discharge signal of the cable is monitored and collected by a current sensor, and the time-domain and frequency-domain features are extracted.

[0027] In this embodiment, a partial discharge signal of the cable is first acquired using a current sensor installed at the cable joint or terminal. A high-frequency, wideband sensor is used to capture the pulse current signal generated by the partial discharge. During the acquisition process, the sampling frequency is set to be no less than 100MHz to ensure that the waveform details of the discharge pulse can be completely recorded. The acquired raw signal is amplified, filtered, and preprocessed to form a monitoring signal sequence for subsequent feature extraction.

[0028] Then, the time-domain and frequency-domain features of the partial discharge signal of the cable are extracted. The time-domain features include the amplitude and the rate of rise of the waveform, such as the voltage change per microsecond; the frequency features include the signal strength of multiple frequency bands, such as 50Hz, 100Hz, and 1MHz.

[0029] Specifically, step S10 in the method includes:

[0030] The partial discharge pulse signal of the cable is monitored and collected by a current sensor to obtain the monitoring signal sequence;

[0031] The average pulse rise rate and pulse amplitude of the pulse signal within the monitoring signal sequence are extracted to obtain the time-domain characteristics.

[0032] Frequency domain features are extracted from the monitoring signal sequence to obtain frequency domain features.

[0033] In this embodiment, firstly, a monitoring signal sequence is obtained by monitoring and acquiring partial discharge pulse signals of the cable using a current sensor. The current sensor is a high-frequency current transformer, installed at the grounding wire of the cable's metallic sheath or at the cross-connection box, to achieve non-invasive detection of the partial discharge pulse current.

[0034] The sensor bandwidth is set to 1MHz to 100MHz, effectively covering the main frequency components of the partial discharge signal while suppressing power frequency and harmonic interference. During data acquisition, a high-speed data acquisition card is used for continuous sampling at a sampling rate of no less than 200MS / s, with a single acquisition duration set to 20ms to 50ms to ensure the capture of a complete discharge pulse sequence. The acquired raw signals are then processed sequentially through bandpass filtering, amplification, and analog-to-digital conversion to remove low-frequency noise and high-frequency interference, forming a standardized monitoring signal sequence.

[0035] Secondly, the average pulse rise rate and pulse amplitude of the pulse signals within the monitoring signal sequence are extracted. Specifically, pulse detection is performed on the monitoring signal sequence to identify all pulse events exceeding a preset threshold, and the peak time and amplitude of each pulse are recorded. For the detected pulse, the time interval from the 10% peak value to the 90% peak value is calculated, and then the pulse rise rate, i.e., the ratio of amplitude change to time interval, is obtained.

[0036] Furthermore, the pulse rise rates of all pulses are arithmetically averaged to obtain the average pulse rise rate. The pulse amplitudes of all pulses are then arithmetically averaged to obtain the average pulse amplitude. The average pulse rise rate and average pulse amplitude together constitute the time-domain characteristics, used to characterize the intensity and steepness of the partial discharge pulse.

[0037] Next, frequency domain features are extracted from the acquired monitoring signal sequence to obtain frequency intensity, which is then used as the frequency domain feature.

[0038] Specifically, frequency domain feature extraction is performed on the monitoring signal sequence, including:

[0039] Perform continuous wavelet transform on the monitoring signal sequence to obtain the time spectrum and filter out noise to obtain the basic frequency domain features;

[0040] Based on the aforementioned fundamental frequency domain features, multiple frequency intensities across multiple frequency bands are extracted as frequency domain features.

[0041] In this embodiment, firstly, a continuous wavelet transform is performed on the monitoring signal sequence, using the Morlet wavelet as the mother wavelet. The scale range is set to cover a frequency interval from 1MHz to 50MHz. Through convolution operations, the monitoring signal sequence is convolved with wavelet basis functions at different scales to obtain the time-frequency spectrum. Wavelet coefficients at each time point and scale are then calculated to form a two-dimensional time-frequency distribution matrix. Next, adaptive thresholding is applied to the time-frequency spectrum. A robust threshold estimation method based on median absolute deviation is used to set time-frequency coefficients below the threshold to zero, preserving effective time-frequency distribution information and obtaining the basic frequency domain features.

[0042] Secondly, based on the basic frequency domain characteristics, multiple characteristic frequency bands are divided, including the low frequency band from 1MHz to 5MHz, the mid frequency band from 5MHz to 20MHz, and the high frequency band from 20MHz to 50MHz. The integral value of the wavelet coefficient energy in each frequency band is calculated to obtain multiple frequency intensities.

[0043] Furthermore, the frequency intensity of each frequency band is arranged in order to form a frequency domain feature vector, which is used to characterize the energy distribution characteristics of the partial discharge signal in different frequency ranges.

[0044] S20: Perform cable fault early warning analysis based on the time domain characteristics and frequency domain characteristics to obtain the time domain early warning and frequency domain early warning, process them to obtain the fused early warning, and perform mutual verification to obtain the early warning deviation.

[0045] In this embodiment of the application, since the early fault point of the cable has just begun to form, the discharge channel is short, the gap impedance is high, the discharge signal is weak and the characteristics are not obvious, and the single-dimensional feature analysis is difficult to accurately assess the degree of fault development. Therefore, it is necessary to construct a multi-modal feature fusion mechanism to comprehensively quantify the early degree of cable fault from both time and frequency domains, and to evaluate the consistency and reliability of feature fusion through a mutual verification and identification mechanism.

[0046] Specifically, the time-domain earlyness and frequency-domain earlyness are adaptively fused to obtain the fused earlyness. Then, the time-domain earlyness and frequency-domain earlyness are mutually verified to determine whether they match, and the earlyness deviation is calculated.

[0047] Specifically, step S20 in the method includes:

[0048] Based on historical cable early fault monitoring data, the maximum pulse rise rate and the maximum high-frequency intensity of the early fault were extracted.

[0049] Extract the average pulse rise rate and high-frequency intensity of the high-frequency band within the time-domain and frequency-domain features;

[0050] The ratio of the average pulse rise rate to the maximum pulse rise rate is calculated as the time-domain earlyness.

[0051] The ratio of the high-frequency intensity to the maximum high-frequency intensity is calculated as the frequency domain earlyness.

[0052] The fusion early degree is calculated based on the time-domain early degree and the frequency-domain early degree.

[0053] In this embodiment of the application, firstly, a benchmark parameter library is constructed based on the cable early fault monitoring data over a historical period. Early fault cases that have been manually verified are selected as training samples. The maximum value of the pulse rise rate in all early fault cases is counted and recorded as the maximum pulse rise rate. At the same time, the maximum value of the frequency intensity in the high-frequency band is counted and recorded as the maximum high-frequency intensity.

[0054] Secondly, the average pulse rise rate is extracted from the time-domain characteristics acquired by current monitoring, and the high-frequency intensity corresponding to the high-frequency band is extracted from the frequency-domain characteristics. The average pulse rise rate reflects the average steepness characteristics of the current discharge pulse, and the high-frequency intensity reflects the energy concentration of the discharge signal in the high-frequency band. Both represent the activity level of partial discharge from two dimensions: time-domain waveform characteristics and frequency-domain energy distribution.

[0055] Next, the early stage in the time domain is calculated by dividing the average pulse rise rate by the maximum pulse rise rate to obtain a dimensionless ratio. The closer this ratio is to 1, the closer the steepness of the current discharge pulse is to the critical state of an early fault, and the higher the early stage in the time domain.

[0056] Similarly, to calculate the earlyness in the frequency domain, divide the high-frequency intensity by the maximum high-frequency intensity to obtain a quantitative value of the earlyness in the frequency domain dimension. The closer this ratio is to 1, the closer the proportion of high-frequency energy is to the typical characteristics of early faults, and the higher the earlyness in the frequency domain dimension.

[0057] Furthermore, the fusion early degree is calculated based on the time-domain early degree and the frequency-domain early degree, and an adaptive weighted fusion strategy is adopted to dynamically adjust the weight coefficients according to the relative magnitude of the two early degrees.

[0058] Specifically, when the difference between the time-domain early stage and the frequency-domain early stage is small, it indicates that the two dimensions have consistent characteristics, and equal-weighted fusion is used. When the difference is large, the weight of the larger early stage is increased, and the weight of the smaller early stage is decreased to suppress the interference of anomalous features. The formula for calculating the fused early stage is: Fusion Early Stage = Time-domain Early Stage × Time-domain Weight + Frequency-domain Early Stage × Frequency-domain Weight, where the weight coefficients are adaptively determined based on the difference in early stage to ensure that the fusion result can comprehensively reflect the early development state of cable faults.

[0059] For example, when the early time domain is 0.75 and the early time domain is 0.80, the difference between the two is 0.05, which is less than the preset threshold of 0.10. Therefore, the features are considered to be consistent, and equal weight fusion is adopted. The weights of the time domain and the frequency domain are both set to 0.5, and the early time of fusion is 0.75×0.5+0.80×0.5=0.775.

[0060] When the early time domain is 0.60 and the early time domain is 0.90, the difference between the two is 0.30, which is greater than the preset threshold. This is determined to be a feature inconsistency. At this time, the weight of the larger early time domain is increased. The weight of the frequency domain is adjusted to 0.7 and the weight of the time domain is adjusted to 0.3. The fused early time domain is 0.60×0.3+0.90×0.7=0.81, which prioritizes reflecting the high early time features displayed in the frequency domain dimension.

[0061] Finally, the early-age characteristics in the time domain and frequency domain are cross-validated to calculate the early-age deviation. Specifically, the absolute difference between the early-age characteristics in the time domain and the early-age characteristics in the frequency domain is calculated as the early-age deviation. The early-age deviation reflects the consistency between the evaluation results of the two dimensions of features. The smaller the deviation, the higher the reliability of feature fusion; the larger the deviation, the more likely there is feature uncertainty or abnormal interference, requiring corresponding adjustments to the subsequent fault identification strategy.

[0062] For example, when the time domain earlyness is 0.75 and the frequency domain earlyness is 0.72, the earlyness deviation is 0.03, indicating that the evaluation results of the two dimensions of features are highly consistent and the credibility of feature fusion is high. When the time domain earlyness is 0.85 and the frequency domain earlyness is 0.45, the earlyness deviation is 0.40, indicating that the frequency domain features may be affected by noise interference or signal attenuation, and there is a large uncertainty in feature fusion. It is necessary to reduce the confidence level of the fusion earlyness and adopt a conservative strategy in the fault identification stage.

[0063] S30: Based on the fusion early degree and early degree deviation, configure a fault identification strategy, perform fault identification on the time domain features and frequency domain features, and obtain cable fault detection results;

[0064] In this embodiment of the application, based on the fusion early degree and early degree deviation calculated above, a fault identification strategy is configured, and cable fault identification is performed through ensemble learning to obtain the cable fault rate.

[0065] The greater the early-stage fusion rate, the more networks are used to improve recognition accuracy; the greater the early-stage deviation, the less likely it is to be a cable fault, so fewer networks are used to save computing power.

[0066] Specifically, mutual verification is performed to identify early-rate biases, including:

[0067] Based on fault detection data that have verified the existence of early cable faults over a historical period, obtain the sample time-domain early degree set and the sample frequency-domain early degree set, and construct an early degree mapping table.

[0068] The time-domain early degree is input into the early degree mapping table to obtain the mapped frequency-domain early degree.

[0069] The deviation ratio between the mapped frequency domain early degree and the frequency domain early degree is calculated and used as the early degree deviation.

[0070] In this embodiment of the application, firstly, based on the fault detection data of cable early faults that have been manually verified and confirmed to exist within a historical period, the time-domain earlyness and frequency-domain earlyness of each sample are extracted, and sample time-domain earlyness sets and sample frequency-domain earlyness sets are constructed respectively.

[0071] Secondly, based on the sample time-domain early degree set and the sample frequency-domain early degree set, a mapping relationship between the time-domain early degree and the frequency-domain early degree is established, and an early degree mapping table is constructed. Specifically, a multinomial fitting method is used, with the time-domain early degree as the independent variable and the frequency-domain early degree as the dependent variable, to fit a mapping function, so that for a given time-domain early degree, the corresponding mapped frequency-domain early degree can be predicted through the mapping function.

[0072] Secondly, the calculated time-domain earlyness is input into the earlyness mapping table, and the mapped frequency-domain earlyness is calculated through the mapping function. This mapped frequency-domain earlyness represents the expected value of the frequency-domain earlyness under the current time-domain earlyness condition, based on the statistical regularity of historical data.

[0073] Next, the deviation ratio between the mapped frequency domain early degree and the actual frequency domain early degree is calculated. That is, the early degree deviation is calculated by "early degree deviation = |mapped frequency domain early degree - actual frequency domain early degree| ÷ mapped frequency domain early degree". This deviation ratio quantifies the degree of deviation between the actual frequency domain characteristics and historical statistical expectations. The larger the deviation ratio, the less the current frequency domain characteristics conform to the typical pattern of early faults, and the higher the uncertainty of feature fusion.

[0074] For example, when the time-domain early degree is 0.70, the mapped frequency-domain early degree obtained by querying the early degree mapping table is 0.72, while the actual frequency-domain early degree is 0.75. The early degree deviation is |0.72-0.75|÷0.72=0.042, which is a small deviation, indicating that the frequency-domain features conform to historical statistical patterns and the feature fusion has high reliability. If the actual frequency-domain early degree is 0.50, the early degree deviation is |0.72-0.50|÷0.72=0.306, which is a large deviation, indicating that the frequency-domain features deviate significantly from expectations, possibly due to noise interference or signal anomalies.

[0075] Specifically, step S30 in the method includes:

[0076] Acquire the number of integrated fault rate identifications;

[0077] Based on the fusion early degree and early degree deviation, a fault identification strategy is configured, wherein the fault identification strategy includes a configured fault rate identification number calculated in combination with the fault rate integrated identification number;

[0078] In the cable fault identification integrated intelligent body, the configured fault rate identification number of the cable fault identification network is randomly selected. The cable fault identification integrated intelligent body includes the cable fault identification network with integrated fault rate identification number. The cable fault identification network with integrated fault rate identification number is obtained by collecting sample time domain feature set, sample frequency domain feature set and sample binary fault detection result set, and performing multiple partitioning with replacement and machine learning supervised training.

[0079] The time-domain and frequency-domain features are input into the cable fault identification network with the selected number of fault rate identifications. The binary fault detection result with the selected number of fault rate identifications is output. The proportion of faults is calculated to obtain the fault rate, which is used as the cable fault detection result. The binary fault detection result includes whether there is a fault or not.

[0080] In this embodiment of the application, firstly, the number of integrated fault rate identifications is obtained. This number is preset according to the scale of historical monitoring data and computing resource constraints, and is usually set to an integer between 50 and 200, representing the total number of cable fault identification networks contained in the integrated intelligent body of cable fault identification.

[0081] Specifically, the construction process of the integrated intelligent agent for cable fault identification includes: collecting a set of temporal and frequency domain features of samples from a historical period, as well as a set of manually verified binary fault detection results. The binary fault detection results are represented by 0 to indicate the absence of a fault and 1 to indicate the presence of a fault. Using the Bootstrap sampling with replacement method, training subsets are randomly selected multiple times from the sample set. Each training subset undergoes supervised machine learning training to construct an independent cable fault identification network. During training, temporal and frequency domain features are used as input vectors, and the binary fault detection results are used as output labels. Gradient boosting decision trees are selected as the basic learner, and hyperparameter optimization ensures the generalization ability and diversity of each network. After training, the cable fault identification networks integrating fault rate and the number of identified faults are integrated into a comprehensive intelligent agent for cable fault identification. Each network operates in parallel and independently.

[0082] Secondly, the fault identification strategy is configured based on the fusion early stage and early stage deviation, and the number of fault rate identifications is dynamically determined. Specifically, the fusion early stage reflects the early development of cable faults, and the early stage deviation reflects the reliability of feature fusion. When the fusion early stage is high, it indicates a greater fault risk, requiring improved identification accuracy, thus increasing the number of fault rate identifications. When the early stage deviation is large, it indicates uncertainty in feature fusion and a lower actual fault probability, thus reducing the number of fault rate identifications to save computational resources.

[0083] Next, within the integrated intelligent system for cable fault identification, a cable fault identification network with a specified number of fault rates is randomly selected. The selection process employs uniform random sampling to ensure that each network has an equal probability of being selected, thus avoiding selection bias. The selected networks form a temporary identification subset for fault identification based on the current monitoring data.

[0084] Furthermore, the extracted time-domain and frequency-domain features are standardized to align with the distribution of the training data, and then input into the selected cable fault identification network with the configured fault rate identification count. Each network independently performs forward inference, outputting a binary fault detection result, either 0 or 1. The outputs of all selected networks are collected, and the number of networks with an output of 1 (indicating a fault) is counted. The proportion of faults is calculated as: Fault Rate = Number of networks with an output of 1 ÷ Number of configured fault rate identifications. This fault rate serves as the cable fault detection result, ranging from 0 to 1; a value closer to 1 indicates a higher confidence level of fault detection.

[0085] For example, when the number of fault rate identifications is configured to be 50, with 38 network outputs being 1 and 12 network outputs being 0, the fault rate is 38 ÷ 50 = 0.76, indicating a high probability of early cable fault risk. When the number of fault rate identifications is configured to be 30, with 5 network outputs being 1 and 25 network outputs being 0, the fault rate is 5 ÷ 30 = 0.167, indicating a lower fault risk, possibly due to normal discharge or interference signals.

[0086] Finally, based on the failure rate, the failure level is classified and warning decisions are made. When the failure rate is below 0.3, it is judged as a normal state and routine monitoring continues. When the failure rate is between 0.3 and 0.7, it is judged as a state of attention, the monitoring cycle is shortened and tracking is strengthened. When the failure rate is above 0.7, it is judged as a warning state, triggering on-site inspection and detailed diagnostic procedures.

[0087] Furthermore, based on the fusion early degree and early degree deviation, a fault identification strategy is configured, including:

[0088] Configure a first failure rate identification coefficient based on the aforementioned early fusion rate;

[0089] Calculate the configuration second failure rate identification coefficient based on the aforementioned early degree deviation;

[0090] Based on the first and second failure rate identification coefficients, and combined with the number of integrated failure rate identifications, the number of failure rate identifications is calculated as a failure identification strategy.

[0091] In this embodiment of the application, firstly, a first failure rate identification coefficient is configured according to the early fusion degree, that is, the early fusion degree is directly used as the first failure rate identification coefficient. This coefficient is used to quantify the degree of requirement for identification accuracy of the failure risk level. This coefficient reflects that the higher the failure risk, the stronger the requirement for identification accuracy.

[0092] Secondly, a second failure rate identification coefficient is calculated based on the early-degree deviation. This coefficient quantifies the impact of feature reliability on identification resources. A larger early-degree deviation indicates higher uncertainty in feature fusion and a lower probability of actual failure, thus requiring a reduction in identification resource investment.

[0093] Specifically, the second failure rate identification coefficient = 1 - early degree deviation. For example, when the early degree deviation is 0.05, the second failure rate identification coefficient is 0.95; when the early degree deviation is 0.30, the second failure rate identification coefficient is 0.70.

[0094] Next, combining the first and second failure rate identification coefficients with the number of integrated failure rate identifications, the configured failure rate identification quantity is calculated. The configured failure rate identification quantity is the average of the first and second failure rate identification coefficients, multiplied by the number of integrated failure rate identifications and rounded down.

[0095] For example, when the early fusion rate is 0.75 and the early fusion deviation is 0.05, the first fault rate identification coefficient is 0.75 and the second fault rate identification coefficient is 0.95, with an average of 0.85. Assuming the number of fault rate integrated identifications is 100, the number of fault rate identifications configured is 0.85 × 100 = 85, which is rounded down to 85, meaning that 85 cable fault identification networks are enabled for fault identification.

[0096] S40: Based on the early fusion rate and cable fault detection results, process to obtain the early cable fault detection results.

[0097] In this embodiment, the cable early fault detection result is obtained based on the above-mentioned fusion early degree and cable fault detection result. That is, the cable fault detection result is compensated according to the early degree deviation and the early degree deviation threshold to generate the compensated fault probability, and then the cable early fault detection result is output after processing.

[0098] Specifically, step S40 in the method includes:

[0099] Based on the fault detection data that have been verified to have early cable faults in the historical time period, the maximum early degree deviation is obtained and used as the early degree deviation threshold.

[0100] Based on the early-age deviation, early-age deviation threshold, fused early-age, and cable fault detection results, the cable early-age fault detection results are calculated and processed to obtain the cable early-age fault detection results.

[0101] In this embodiment of the application, firstly, based on the fault detection data of cable early faults that have been manually verified and confirmed in the historical time period, the early degree deviation distribution of each sample is statistically analyzed, and the maximum value of the early degree deviation is extracted as the early degree deviation threshold. This threshold represents the maximum possible deviation between the time domain features and the frequency domain features in the historically confirmed cable early fault samples.

[0102] Specifically, the historical fault sample set is traversed, the early degree deviation of each sample is calculated, and the early degree deviation threshold is determined by percentile statistics. Usually, the 95th percentile or the maximum value is selected to ensure that the threshold can cover the characteristic fluctuation range under most normal conditions.

[0103] For example, when the early degree deviation distribution of historical samples is 0.02, 0.05, 0.08, 0.12, 0.15, 0.18, and 0.25, the maximum early degree deviation of 0.25 is taken as the early degree deviation threshold.

[0104] Secondly, based on the currently calculated early-age deviation, early-age deviation threshold, fusion early-age, and cable fault detection results, the compensation fault probability is calculated, and then the cable early fault detection results are obtained.

[0105] The cable early fault detection result is calculated and processed based on the early degree deviation, early degree deviation threshold, fusion early degree, and cable fault detection result, including:

[0106] The ratio of the early-age deviation to the early-age deviation threshold is calculated to obtain the confidence coefficient. The cable fault detection results are then compensated to obtain the compensated fault rate.

[0107] The fusion early stage and the compensation failure rate are used as the results of early cable failure detection.

[0108] In this embodiment of the application, firstly, the confidence coefficient is calculated by "confidence coefficient = 1 - early degree deviation ÷ early degree deviation threshold". This coefficient is used to quantify the confidence level of the current feature fusion. The closer the confidence coefficient is to 1, the more reliable the feature fusion is and the higher the confidence of the cable fault detection result. The closer the confidence coefficient is to 0, the greater the uncertainty of feature fusion is and the greater the need to compensate and correct the cable fault detection result.

[0109] Specifically, when the early deviation is less than or equal to the early deviation threshold, the confidence coefficient is positive, indicating that the current feature deviation is within the normal historical fluctuation range, and the cable fault detection result is basically reliable, but appropriate compensation is needed according to the degree of deviation; when the early deviation is greater than the early deviation threshold, the confidence coefficient is negative, indicating that the current feature deviates significantly from the historical pattern, there is serious anomaly or noise interference, the confidence of the cable fault detection result is extremely low, and significant correction or even re-collection of data is required.

[0110] Secondly, the cable fault detection results are compensated based on the confidence coefficient to obtain the compensated fault rate. The formula for calculating the compensated fault rate is "Compensated fault rate = Fault rate within the cable fault detection results × Confidence coefficient".

[0111] Furthermore, the early detection result of cable faults is output by combining the early detection result of the early detection result of the cable faults and the compensated fault rate. The early detection result of the early detection result reflects the early development degree of cable faults, and the compensated fault rate reflects the fault probability after the confidence correction. The combination of the two provides maintenance personnel with a multi-dimensional basis for decision-making.

[0112] In summary, compared with the prior art, this application quantifies the deviation between actual frequency domain features and historical statistical expectations by constructing a mapping relationship between time domain earlyness and frequency domain earlyness, and dynamically configures the fault identification strategy based on the fused earlyness and earlyness deviation, thereby realizing the quantitative evaluation and adaptive adjustment of feature fusion reliability in the early fault detection process of cables.

[0113] This application provides a method for early cable fault detection based on adaptive fusion of multimodal features. First, partial discharge signals of the cable are monitored and acquired using a current sensor, extracting feature information from both the time and frequency domains. Second, early fault severity analysis is performed based on the time and frequency domain features to obtain the time-domain and frequency-domain early severity values, which are then processed to obtain a fused early severity value. Simultaneously, an early severity deviation is calculated using a mutual verification mechanism. Third, a fault identification strategy is adaptively configured based on the fused early severity value and the early severity deviation, dynamically adjusting the number of participants in the fault identification network. Finally, the fused early severity value and the cable fault detection results are comprehensively processed, and a reliability coefficient is used to compensate and correct the detection results.

[0114] Through the above technical solution, this application solves the problems of single feature detection being susceptible to noise interference and insufficient decision reliability by using an adaptive fusion and mutual verification calibration mechanism of multimodal features, thereby achieving the technical effect of improving the accuracy and robustness of early cable fault detection.

[0115] Example 2, as Figure 2As shown, based on the same inventive concept as the cable early fault detection method with multimodal feature adaptive fusion provided in Embodiment 1, this application also provides a cable early fault detection system with multimodal feature adaptive fusion, including:

[0116] Feature extraction module 11 is used to monitor and collect partial discharge signals of the cable through a current sensor, and extract time-domain features and frequency-domain features.

[0117] The deviation calculation module 12 is used to perform cable fault early degree analysis based on the time domain characteristics and frequency domain characteristics, obtain the time domain early degree and frequency domain early degree, process them to obtain the fused early degree, and perform mutual verification to obtain the early degree deviation.

[0118] The fault identification module 13 is used to configure a fault identification strategy based on the fusion early degree and early degree deviation, perform fault identification on the time domain features and frequency domain features, and obtain cable fault detection results.

[0119] The result acquisition module 14 is used to process the cable early fault detection result based on the fusion early degree and the cable fault detection result.

[0120] In one embodiment, the feature extraction module 11 is specifically used for:

[0121] The partial discharge pulse signal of the cable is monitored and collected by a current sensor to obtain the monitoring signal sequence;

[0122] The average pulse rise rate and pulse amplitude of the pulse signal within the monitoring signal sequence are extracted to obtain the time-domain characteristics.

[0123] Frequency domain features are extracted from the monitoring signal sequence to obtain frequency domain features.

[0124] The frequency domain feature extraction of the monitoring signal sequence includes:

[0125] Perform continuous wavelet transform on the monitoring signal sequence to obtain the time spectrum and filter out noise to obtain the basic frequency domain features;

[0126] Based on the aforementioned fundamental frequency domain features, multiple frequency intensities across multiple frequency bands are extracted as frequency domain features.

[0127] In one embodiment, the deviation calculation module 12 is specifically used for:

[0128] Based on historical cable early fault monitoring data, the maximum pulse rise rate and the maximum high-frequency intensity of the early fault were extracted.

[0129] Extract the average pulse rise rate and high-frequency intensity of the high-frequency band within the time-domain and frequency-domain features;

[0130] The ratio of the average pulse rise rate to the maximum pulse rise rate is calculated as the time-domain earlyness.

[0131] The ratio of the high-frequency intensity to the maximum high-frequency intensity is calculated as the frequency domain earlyness.

[0132] The fusion early degree is calculated based on the time-domain early degree and the frequency-domain early degree.

[0133] Furthermore, in one embodiment of the application, mutual verification is performed to obtain early degree deviation, including:

[0134] Based on fault detection data that have verified the existence of early cable faults over a historical period, obtain the sample time-domain early degree set and the sample frequency-domain early degree set, and construct an early degree mapping table.

[0135] The time-domain early degree is input into the early degree mapping table to obtain the mapped frequency-domain early degree.

[0136] The deviation ratio between the mapped frequency domain early degree and the frequency domain early degree is calculated and used as the early degree deviation.

[0137] Furthermore, the fault identification module 13 is specifically used for:

[0138] Acquire the number of integrated fault rate identifications;

[0139] Based on the fusion early degree and early degree deviation, a fault identification strategy is configured, wherein the fault identification strategy includes a configured fault rate identification number calculated in combination with the fault rate integrated identification number;

[0140] In the cable fault identification integrated intelligent body, the configured fault rate identification number of the cable fault identification network is randomly selected. The cable fault identification integrated intelligent body includes the cable fault identification network with integrated fault rate identification number. The cable fault identification network with integrated fault rate identification number is obtained by collecting sample time domain feature set, sample frequency domain feature set and sample binary fault detection result set, and performing multiple partitioning with replacement and machine learning supervised training.

[0141] The time-domain and frequency-domain features are input into the cable fault identification network with the selected number of fault rate identifications. The binary fault detection result with the selected number of fault rate identifications is output. The proportion of faults is calculated to obtain the fault rate, which is used as the cable fault detection result. The binary fault detection result includes whether there is a fault or not.

[0142] Furthermore, in one embodiment, a fault identification strategy is configured based on the fusion early degree and the early degree deviation, including:

[0143] Configure a first failure rate identification coefficient based on the aforementioned early fusion rate;

[0144] Calculate the configuration second failure rate identification coefficient based on the aforementioned early degree deviation;

[0145] Based on the first and second failure rate identification coefficients, and combined with the number of integrated failure rate identifications, the number of failure rate identifications is calculated as a failure identification strategy.

[0146] Furthermore, the result acquisition module 14 is specifically used for:

[0147] Based on the fault detection data that have been verified to have early cable faults in the historical time period, the maximum early degree deviation is obtained and used as the early degree deviation threshold.

[0148] Based on the early-age deviation, early-age deviation threshold, fused early-age, and cable fault detection results, the cable early-age fault detection results are calculated and processed to obtain the cable early-age fault detection results.

[0149] Further, based on the early-age deviation, early-age deviation threshold, fused early-age, and cable fault detection results, the cable early-age fault detection results are calculated and processed to obtain the following:

[0150] The ratio of the early-age deviation to the early-age deviation threshold is calculated to obtain the confidence coefficient. The cable fault detection results are then compensated to obtain the compensated fault rate.

[0151] The fusion early stage and the compensation failure rate are used as the results of early cable failure detection.

Claims

1. A method for early cable fault detection based on adaptive fusion of multimodal features, characterized in that, The method includes: Partial discharge signals of the cable are monitored and collected using a current sensor, and time-domain and frequency-domain features are extracted. Based on the aforementioned time-domain and frequency-domain characteristics, cable fault early-age analysis is performed to obtain time-domain early-age and frequency-domain early-age characteristics. These are then processed to obtain a fused early-age, and cross-verification is performed to determine the early-age deviation, including: Based on historical cable early fault monitoring data, the maximum pulse rise rate and the maximum high-frequency intensity of the early fault were extracted. Extract the average pulse rise rate and high-frequency intensity of the high-frequency band within the time-domain and frequency-domain features; The ratio of the average pulse rise rate to the maximum pulse rise rate is calculated as the time-domain earlyness. The ratio of the high-frequency intensity to the maximum high-frequency intensity is calculated as the frequency domain earlyness. The fusion early degree is calculated based on the time-domain early degree and the frequency-domain early degree. Based on the fusion early degree and early degree deviation, a fault identification strategy is configured to identify faults in the time domain features and frequency domain features, thereby obtaining cable fault detection results. Based on the early fusion rate and cable fault detection results, the early cable fault detection results are processed to obtain the following: Based on the fault detection data that have been verified to have early cable faults in the historical time period, the maximum early degree deviation is obtained and used as the early degree deviation threshold. Based on the early-age deviation, early-age deviation threshold, fused early-age, and cable fault detection results, the cable early-age fault detection results are calculated and processed to obtain the following: The ratio of the early-age deviation to the early-age deviation threshold is calculated to obtain the confidence coefficient. The cable fault detection results are then compensated to obtain the compensated fault rate. The fusion early stage and the compensation failure rate are used as the results of early cable failure detection.

2. The cable early fault detection method based on multimodal feature adaptive fusion according to claim 1, characterized in that, Partial discharge signals from cables are monitored and acquired using a current sensor, and time-domain and frequency-domain features are extracted, including: The partial discharge pulse signal of the cable is monitored and collected by a current sensor to obtain the monitoring signal sequence; The average pulse rise rate and pulse amplitude of the pulse signal within the monitoring signal sequence are extracted to obtain the time-domain characteristics. Frequency domain features are extracted from the monitoring signal sequence to obtain frequency domain features.

3. The cable early fault detection method based on multimodal feature adaptive fusion according to claim 2, characterized in that, Frequency domain feature extraction of the monitoring signal sequence includes: Perform continuous wavelet transform on the monitoring signal sequence to obtain the time spectrum and filter out noise to obtain the basic frequency domain features; Based on the aforementioned fundamental frequency domain features, multiple frequency intensities across multiple frequency bands are extracted as frequency domain features.

4. The cable early fault detection method based on multimodal feature adaptive fusion according to claim 1, characterized in that, Cross-validation identification yields early-age biases, including: Based on fault detection data that have verified the existence of early cable faults over a historical period, obtain the sample time-domain early degree set and the sample frequency-domain early degree set, and construct an early degree mapping table. The time-domain early degree is input into the early degree mapping table to obtain the mapped frequency-domain early degree. The deviation ratio between the mapped frequency domain early degree and the frequency domain early degree is calculated and used as the early degree deviation.

5. The cable early fault detection method based on multimodal feature adaptive fusion according to claim 1, characterized in that, Based on the fusion early degree and early degree deviation, a fault identification strategy is configured to perform fault identification on the time-domain features and frequency-domain features, resulting in cable fault detection results, including: Acquire the number of integrated fault rate identifications; Based on the fusion early degree and early degree deviation, a fault identification strategy is configured, wherein the fault identification strategy includes a configured fault rate identification number calculated in combination with the fault rate integrated identification number; In the cable fault identification integrated intelligent body, the configured fault rate identification number of the cable fault identification network is randomly selected. The cable fault identification integrated intelligent body includes the cable fault identification network with integrated fault rate identification number. The cable fault identification network with integrated fault rate identification number is obtained by collecting sample time domain feature set, sample frequency domain feature set and sample binary fault detection result set, and performing multiple partitioning with replacement and machine learning supervised training. The time-domain and frequency-domain features are input into the cable fault identification network with the selected number of fault rate identifications. The binary fault detection result with the selected number of fault rate identifications is output. The proportion of faults is calculated to obtain the fault rate, which is used as the cable fault detection result. The binary fault detection result includes whether there is a fault or not.

6. The cable early fault detection method based on multimodal feature adaptive fusion according to claim 5, characterized in that, Based on the fusion early stage and early stage deviation, a fault identification strategy is configured, including: Configure a first failure rate identification coefficient based on the aforementioned early fusion rate; Calculate the configuration second failure rate identification coefficient based on the aforementioned early degree deviation; Based on the first and second failure rate identification coefficients, and combined with the number of integrated failure rate identifications, the number of failure rate identifications is calculated as a failure identification strategy.

7. A cable early fault detection system based on multimodal feature adaptive fusion, characterized in that, The cable early fault detection method for performing the multimodal feature adaptive fusion according to any one of claims 1-6 includes: The feature extraction module is used to monitor and collect partial discharge signals of cables through a current sensor, and extract time-domain and frequency-domain features. The deviation calculation module is used to perform cable fault early degree analysis based on the time domain characteristics and frequency domain characteristics, obtain the time domain early degree and frequency domain early degree, process them to obtain the fused early degree, and perform mutual verification to obtain the early degree deviation. The fault identification module is used to configure a fault identification strategy based on the fusion early degree and early degree deviation, perform fault identification on the time domain features and frequency domain features, and obtain cable fault detection results. The result acquisition module is used to process the early cable fault detection results based on the fusion early degree and the cable fault detection results to obtain the cable early fault detection results.

Images