Automatic identification system for cable fault waveform
By integrating data acquisition, deep learning and knowledge-driven fault reasoning systems, the problem of difficult cable fault waveform identification is solved, efficient and reliable cable fault location and early warning are achieved, and the accuracy and speed of cable fault detection are improved.
Patent Information
- Application Number
- CN202510607077.3
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-05-13
- Publication Date
- 2025-09-12
AI Technical Summary
Existing technologies cannot efficiently and reliably identify cable fault waveforms, which makes cable fault location difficult and consumes a lot of manpower and material resources. In addition, traditional methods are easily affected by cable parameters and transition resistance.
It adopts data acquisition unit, deep time-frequency domain feature extraction network unit, knowledge-driven fault inference engine unit and adaptive incremental learning unit, integrates high-precision sensors and filters, and combines deep learning and knowledge graph to realize automatic identification of cable fault waveforms.
It achieves high accuracy (98.5%) and efficient fault identification, with a fault location error of less than ±3m and a real-time processing delay of less than 500ms. It has fault early warning capabilities and improves the efficiency and reliability of cable fault location.
Smart Images
Figure CN120629801A_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the technical field of cable fault detection, and in particular to an automatic recognition system for cable fault waveforms. Background Art
[0002] As a vital component of the distribution network, power cables are widely used in power engineering construction due to their numerous advantages, including high power supply reliability, minimal environmental impact, and the fact that they do not occupy ground corridors. Their use is increasing, and their distribution is becoming increasingly complex. However, if underground cables fail due to potential problems such as mechanical damage, insulation moisture, and insulation aging, they can not only cause huge losses to the power company but also cause serious disruption to people's lives. Unlike overhead lines, which are directly observable, cables are deeply buried and concealed underground, and urban roads are primarily hard surfaces. Excavating cables extensively to locate faults often consumes significant manpower, material, and financial resources. Therefore, there is an urgent need to develop efficient and reliable cable fault location methods to detect and repair faults in a timely manner.
[0003] Locating cable faults requires pre-positioning and distance measurement, followed by precise location and pinpointing. Distance measurement roughly locates the fault area, while pinpointing confirms the specific location of the fault. Cable fault distance measurement methods primarily include the impedance method and the traveling wave method. The impedance method calculates the fault location by calculating a function equation relating the fault current and voltage to the distance. The traveling wave method calculates the fault distance by measuring the time it takes for the traveling wave signal to travel from the measuring end to the fault point and back. The impedance method is susceptible to cable parameters and transition resistance, leading to its declining application. However, the traveling wave method, with its simple principle and ease of operation, is gaining increasing popularity in engineering applications.
[0004] As the traveling wave ranging methods widely used on site, the pulse current method and the secondary pulse method are suitable for high-resistance faults and flashover faults in which high voltage is applied to the cable and the fault point can be broken down and discharged. The principle of the pulse current method is to use a high-voltage signal generator to apply high voltage to break down the fault, collect and record the time difference between the fault point breakdown discharge pulse and its first reflected pulse to calculate the fault distance. However, the high-voltage signal generator may not necessarily break down the fault point, and the non-breakdown discharge waveform cannot be used for ranging. The principle of the secondary pulse method is to first emit a low-voltage pulse to obtain a reflected waveform, and then use a high-voltage signal generator to apply high voltage to break down the fault, maintain the arc and emit another low-voltage pulse to obtain a reflected waveform. The divergence point between the waveform without arc and the waveform with arc corresponds to the fault location. However, when there is no breakdown, the two waveforms basically overlap and cannot be used for ranging. Therefore, it is necessary to design a more advanced automatic recognition system for cable fault waveforms. Summary of the Invention
[0005] The purpose of the present invention is to provide an automatic recognition system for cable fault waveforms to solve the technical problems existing in the background technology.
[0006] In order to achieve the above object, the technical solution adopted by the present invention is as follows:
[0007] An automatic cable fault waveform recognition system includes a data acquisition unit, a deep time-frequency domain feature extraction network unit, a knowledge-driven fault inference engine unit, and an adaptive incremental learning unit. The data acquisition unit is connected to the deep time-frequency domain feature extraction network unit, the deep time-frequency domain feature extraction network unit is connected to the knowledge-driven fault inference engine unit, and the knowledge-driven fault inference engine unit is connected to the adaptive incremental learning unit.
[0008] The data acquisition unit integrates high-precision current and voltage sensors, distributed fiber-optic temperature sensors, and partial discharge detection modules to achieve synchronous acquisition of multiple physical quantities. It uses a hardware-level digital phase-locked loop for synchronous sampling, combined with an adaptive notch filter to eliminate power frequency interference, and supports 5G / fiber-optic redundant transmission.
[0009] The deep time-frequency domain feature extraction network unit includes a hybrid model architecture and dynamic feature enhancement. The hybrid model architecture extracts time-frequency features through wavelet packet decomposition and channel attention convolution. LSTM-GRU captures temporal dynamics and combines it with Transformer to model long-range dependencies. Dynamic feature enhancement introduces a generative adversarial network to expand small sample fault data. The training difficulty in noisy environments is gradually increased through course learning.
[0010] The knowledge-driven fault reasoning engine unit includes a knowledge graph construction module and a hybrid reasoning mechanism module. The knowledge graph construction module integrates IEC standards, historical fault cases, equipment parameters, and environmental data to construct a dynamic knowledge graph containing causal relationships. Monte Carlo tree search combined with Bayesian networks triggers the expert rule base to perform confidence correction when the model is uncertain.
[0011] The adaptive incremental learning unit includes a federated learning architecture module and a meta-learning optimization module. The federated learning architecture module supports collaborative training of edge nodes in multiple regions, protecting data privacy while improving the model's generalization ability. The meta-learning optimization module uses the MAML algorithm to enable the model to adapt to new failure modes.
[0012] Furthermore, in the data acquisition unit, current and voltage sensors are used to measure the current and voltage values in the circuit in real time to ensure data accuracy and provide basic data support for the safe operation of the equipment. The distributed fiber optic temperature sensor uses the sensitivity of optical fiber to temperature to continuously measure the temperature distribution along the line, timely detect abnormal conditions such as local overheating, and prevent faults from occurring. The local discharge detection module is used to detect tiny discharge phenomena caused by insulation aging or damage inside the equipment. Discharge is a precursor to equipment failure. Timely detection can avoid major accidents. The digital phase-locked loop implemented with a dedicated chip tracks signal frequency changes in real time and dynamically adjusts the sampling clock phase to accurately align the sampling points with the signal waveform. The hardware processing delay is less than 1 microsecond, which can eliminate the signal distortion caused by traditional asynchronous sampling. It is especially suitable for high-frequency power or sensor signal acquisition. For 50 / 60Hz power frequency interference, the system has a built-in dual-mode filter: in normal mode, it locks the fixed interference frequency; when a frequency offset is detected, it automatically switches to FFT analysis and gradient descent algorithm to complete the modeling of the new interference frequency within 50ms, with a stopband attenuation of more than -60dB, effectively retaining the original signal characteristics, and the data is simultaneously transmitted through 5G NR MIMO antennas and multi-mode optical fiber transmission adopt an intelligent arbitration mechanism. Under normal circumstances, 5G bears 95% of the traffic (low latency mode), and the optical fiber serves as a verification channel. When anomalies such as base station switching / optical disconnection are detected, the protocol stack migration is completed within 3ms, ensuring end-to-end transmission reliability of >99.999%. Dual-link bandwidth aggregation supports 10Gbps-level real-time data backhaul.
[0013] Furthermore, in the deep time-frequency domain feature extraction network unit, wavelet packet decomposition is first used to decompose the signal into different frequency sub-bands to separate the key time-frequency components, and then channel attention convolution is used to enhance the feature response of important channels, focusing on local time-frequency patterns. Two recurrent units are connected in series or in parallel. LSTM captures long-term trends through a gating mechanism, and GRU strengthens short-term fluctuations with a lightweight structure to jointly analyze dynamic changes in time series. The self-attention layer directly models global dependencies, breaking through the long-range information attenuation limitations of RNN and associating long-distance time series events.
[0014] Furthermore, in the knowledge-driven fault reasoning engine unit, the IEC standard is used as a framework to unify the formats of equipment parameters and environmental data, define the semantic mapping between fault codes and processing procedures, transform historical fault cases into a chain relationship between cause and effect, establish cross-dimensional connections, form a conditional relationship between equipment parameters and environmental data, and cross-validate with the safety thresholds in the IEC standard. A real-time data stream interface is deployed. When the sensor detects that the ambient temperature rise rate exceeds the IEC specified value, the graph reasoning engine is automatically triggered to generate a risk prediction path. An incremental learning mechanism is used to update the weight parameters for every 100 new fault records. Combined with Monte Carlo simulation, the probability of equipment failure under different environmental combinations is predicted to ensure that the knowledge graph adapts to changes in the operating environment.
[0015] The Bayesian network quantifies the dependencies between variables through a probabilistic graphical model and dynamically updates the confidence of an event. When the two are combined, MCTS generates potential decision branches, and the Bayesian network calculates the probability distribution of each branch in real time, forming a closed loop of simulation and probability correction. When the Bayesian network detects that the model confidence is lower than the threshold, the system triggers the expert rule base. In power grid fault diagnosis, if the Bayesian network has doubts about the probability of a short circuit on a certain line, the expert rules will forcibly inject industry standard logic: if the current mutation exceeds the threshold and is accompanied by a sudden drop in voltage, it is directly judged as a short circuit fault, and the confidence is corrected to a high confidence interval. The mechanism integrates data-driven and knowledge-driven methods, which not only retains the flexibility of MCTS, but also avoids model bias through expert experience and improves the reliability of decision-making in key scenarios.
[0016] Furthermore, in the adaptive incremental learning unit, several geographically distributed edge nodes collaborate to train the AI model. Each node retains the original data locally and only uploads model parameter updates. The global model is integrated and optimized through an aggregation algorithm, avoiding the cross-regional transmission of sensitive data while leveraging the characteristics of data from multiple locations to improve model adaptability.
[0017] The AI model iteratively learns on a training set containing multiple fault modes, including traditional power grid faults and simulated faults of renewable energy grid connection. In each iteration, an inner loop adjusts the model parameters using specific samples for the current task, while an outer loop reversely optimizes the model's initial parameters based on the average performance of all tasks.
[0018] The present invention has the following beneficial effects due to the adoption of the above technical solution:
[0019] The fault identification accuracy of the system of the present invention is as high as 98.5%, which is much higher than the traditional 82%. The fault location error reaches ±3m, the real-time processing delay is less than 500ms, and the fault adaptation time is less than 30 minutes. It has achieved a technological leap from passive maintenance to active early warning. In the test of the State Grid's ultra-high voltage lines, it has achieved a 97.3% fault warning capability of 6 hours in advance. BRIEF DESCRIPTION OF THE DRAWINGS
[0020] Figure 1 It is a system block diagram of the present invention. DETAILED DESCRIPTION
[0021] To make the objectives, technical solutions, and advantages of the present invention more clearly understood, the present invention is further described in detail below with reference to the accompanying drawings and by way of preferred embodiments. However, it should be noted that many of the details listed in this specification are merely provided to help the reader gain a thorough understanding of one or more aspects of the present invention, and these aspects of the present invention can be practiced even without these specific details.
[0022] like Figure 1As shown, a system for automatically identifying cable fault waveforms includes a data acquisition unit, a deep time-frequency domain feature extraction network unit, a knowledge-driven fault inference engine unit and an adaptive incremental learning unit. The data acquisition unit is connected to the deep time-frequency domain feature extraction network unit, the deep time-frequency domain feature extraction network unit is connected to the knowledge-driven fault inference engine unit, and the knowledge-driven fault inference engine unit is connected to the adaptive incremental learning unit.
[0023] The data acquisition unit integrates high-precision current and voltage sensors, distributed fiber-optic temperature sensors, and partial discharge detection modules to achieve synchronous acquisition of multiple physical quantities. It uses a hardware-level digital phase-locked loop for synchronous sampling, combined with an adaptive notch filter to eliminate power frequency interference, and supports 5G / fiber-optic redundant transmission. In the data acquisition unit, current and voltage sensors are used to measure the current and voltage values in the circuit in real time, ensuring data accuracy and providing basic data support for the safe operation of the equipment. Distributed fiber optic temperature sensors use the sensitivity of optical fibers to temperature to continuously measure temperature distribution along the line, promptly detecting abnormal conditions such as local overheating and preventing failures. The partial discharge detection module is used to detect tiny discharges caused by insulation aging or damage inside the equipment. Discharge is a precursor to equipment failure, and timely detection can avoid major accidents. The digital phase-locked loop implemented with a dedicated chip tracks signal frequency changes in real time and dynamically adjusts the sampling clock phase to accurately align the sampling points with the signal waveform. The hardware processing delay is less than 1 microsecond, which can eliminate the signal distortion caused by traditional asynchronous sampling. It is particularly suitable for high-frequency power or sensor signal acquisition. To address 50 / 60Hz power frequency interference, the system has a built-in dual-mode filter: in normal mode, it locks the fixed interference frequency; when a frequency offset is detected, it automatically switches to FFT analysis and gradient descent algorithm to complete the modeling of the new interference frequency within 50ms, with a stopband attenuation of more than -60dB, effectively preserving the original signal characteristics. The data is also transmitted through 5G NR MIMO antennas and multi-mode optical fiber transmission adopt an intelligent arbitration mechanism. Under normal circumstances, 5G bears 95% of the traffic (low latency mode), and the optical fiber serves as a verification channel. When anomalies such as base station switching / optical disconnection are detected, the protocol stack migration is completed within 3ms, ensuring end-to-end transmission reliability of >99.999%. Dual-link bandwidth aggregation supports 10Gbps-level real-time data backhaul.
[0024] The deep time-frequency domain feature extraction network unit includes a hybrid model architecture and dynamic feature enhancement. The hybrid model architecture extracts time-frequency features through wavelet packet decomposition and channel-attention convolution. LSTM-GRU captures time series dynamics and combines it with Transformer to model long-term dependencies. Dynamic feature enhancement introduces a generative adversarial network to expand small-sample fault data. Training difficulty in noisy environments is gradually increased through course learning. The deep time-frequency domain feature extraction network unit first uses wavelet packet decomposition to decompose the signal into different frequency subbands, isolating key time-frequency components. Channel-attention convolution is then used to enhance the characteristic responses of important channels, focusing on local time-frequency patterns. Two recurrent units are connected in series or parallel. LSTM captures long-term trends through a gating mechanism, while GRU enhances short-term fluctuations with a lightweight structure. Together, they analyze time series dynamics. The self-attention layer directly models global dependencies, overcoming the long-range information attenuation limitations of RNNs and linking distant time series events.
[0025] The knowledge-driven fault inference engine unit includes a knowledge graph construction module and a hybrid inference mechanism module. The knowledge graph construction module integrates IEC standards, historical fault cases, equipment parameters, and environmental data to construct a dynamic knowledge graph containing causal relationships. Monte Carlo tree search, combined with Bayesian networks, triggers the expert rule base to perform confidence correction when the model is uncertain. The knowledge-driven fault inference engine unit uses the IEC standard as a framework to unify the formats of equipment parameters and environmental data, define the semantic mapping between fault codes and processing procedures, transform historical fault cases into a chain relationship between cause and effect, establish cross-dimensional connections, form conditional relationships between equipment parameters and environmental data, and cross-validate with safety thresholds in the IEC standard. A real-time data stream interface is deployed. When the sensor detects that the ambient temperature rise rate exceeds the IEC specified value, the graph inference engine is automatically triggered to generate a risk prediction path. Using an incremental learning mechanism, the weight parameters are updated for every 100 new fault records. Combined with Monte Carlo simulation, the probability of equipment failure under different environmental combinations is predicted, ensuring that the knowledge graph adapts to changes in the operating environment.
[0026] The Bayesian network quantifies the dependencies between variables through a probabilistic graphical model and dynamically updates the confidence of an event. When the two are combined, MCTS generates potential decision branches, and the Bayesian network calculates the probability distribution of each branch in real time, forming a closed loop of simulation and probability correction. When the Bayesian network detects that the model confidence is lower than the threshold, the system triggers the expert rule base. In power grid fault diagnosis, if the Bayesian network has doubts about the probability of a short circuit on a certain line, the expert rules will forcibly inject industry standard logic: if the current mutation exceeds the threshold and is accompanied by a sudden drop in voltage, it is directly judged as a short circuit fault, and the confidence is corrected to a high confidence interval. The mechanism integrates data-driven and knowledge-driven methods, which not only retains the flexibility of MCTS, but also avoids model bias through expert experience and improves the reliability of decision-making in key scenarios.
[0027] The adaptive incremental learning unit includes a federated learning architecture module and a meta-learning optimization module. The federated learning architecture module supports collaborative training of edge nodes in multiple regions, protecting data privacy while improving model generalization capabilities. The meta-learning optimization module uses the MAML algorithm to enable the model to adapt to new failure modes. In the adaptive incremental learning unit, the AI model is trained collaboratively by several geographically distributed edge nodes. Each node retains the original data locally and only uploads model parameter updates. The global model is integrated and optimized through an aggregation algorithm, avoiding the cross-regional transmission of sensitive data while leveraging the characteristics of data from multiple locations to improve model adaptability.
[0028] The AI model iteratively learns on a training set containing multiple fault modes, including traditional power grid faults and simulated faults of renewable energy grid connection. In each iteration, an inner loop adjusts the model parameters using specific samples for the current task, while an outer loop reversely optimizes the model's initial parameters based on the average performance of all tasks.
[0029] The system implements a three-layer innovative architecture of "physical layer perception enhancement + knowledge layer reasoning optimization + decision layer dynamic adaptation".
[0030] Matters not covered by the present invention are known technologies.
[0031] The above is only a preferred embodiment of the present invention. It should be pointed out that for ordinary technicians in this technical field, several improvements and modifications can be made without departing from the principles of the present invention. These improvements and modifications should also be regarded as within the scope of protection of the present invention.
Claims
1. An automatic identification system for cable fault waveforms, characterized by: It includes a data acquisition unit, a deep time-frequency domain feature extraction network unit, a knowledge-driven fault inference engine unit and an adaptive incremental learning unit. The data acquisition unit is connected to the deep time-frequency domain feature extraction network unit, the deep time-frequency domain feature extraction network unit is connected to the knowledge-driven fault inference engine unit, and the knowledge-driven fault inference engine unit is connected to the adaptive incremental learning unit. The data acquisition unit integrates high-precision current and voltage sensors, distributed fiber-optic temperature sensors, and partial discharge detection modules to achieve synchronous acquisition of multiple physical quantities. It uses a hardware-level digital phase-locked loop for synchronous sampling, combined with an adaptive notch filter to eliminate power frequency interference, and supports 5G / fiber-optic redundant transmission. The deep time-frequency domain feature extraction network unit includes a hybrid model architecture and dynamic feature enhancement. The hybrid model architecture extracts time-frequency features through wavelet packet decomposition and channel attention convolution. LSTM-GRU captures temporal dynamics and combines it with Transformer to model long-range dependencies. Dynamic feature enhancement introduces a generative adversarial network to expand small sample fault data. The training difficulty in noisy environments is gradually increased through course learning. The knowledge-driven fault reasoning engine unit includes a knowledge graph construction module and a hybrid reasoning mechanism module. The knowledge graph construction module integrates IEC standards, historical fault cases, equipment parameters, and environmental data to construct a dynamic knowledge graph containing causal relationships. Monte Carlo tree search combined with Bayesian networks triggers the expert rule base to perform confidence correction when the model is uncertain. The adaptive incremental learning unit includes a federated learning architecture module and a meta-learning optimization module. The federated learning architecture module supports collaborative training of edge nodes in multiple regions, protecting data privacy while improving the model's generalization ability. The meta-learning optimization module uses the MAML algorithm to enable the model to adapt to new failure modes.
2. The automatic cable fault waveform identification system according to claim 1 is characterized in that: In the data acquisition unit, current and voltage sensors are used to measure the current and voltage values in the circuit in real time, ensuring data accuracy and providing basic data support for the safe operation of the equipment. Distributed fiber optic temperature sensors use the sensitivity of optical fibers to temperature to continuously measure temperature distribution along the line, promptly detecting abnormal conditions such as local overheating and preventing failures. The partial discharge detection module is used to detect tiny discharges caused by insulation aging or damage inside the equipment. Discharge is a precursor to equipment failure, and timely detection can avoid major accidents. The digital phase-locked loop implemented with a dedicated chip tracks signal frequency changes in real time and dynamically adjusts the sampling clock phase to accurately align the sampling points with the signal waveform. The hardware processing delay is less than 1 microsecond, which can eliminate the signal distortion caused by traditional asynchronous sampling. It is particularly suitable for high-frequency power or sensor signal acquisition. To address 50 / 60Hz power frequency interference, the system has a built-in dual-mode filter: in normal mode, it locks the fixed interference frequency; when a frequency offset is detected, it automatically switches to FFT analysis and gradient descent algorithm to complete the modeling of the new interference frequency within 50ms, with a stopband attenuation of more than -60dB, effectively preserving the original signal characteristics. The data is also transmitted through 5G NR MIMO antennas and multimode optical fibers are used for transmission, and an intelligent arbitration mechanism is adopted. Under normal circumstances, 5G carries 95% of the traffic (low latency mode), and the optical fiber serves as a verification channel. When anomalies such as base station switching / fiber disconnection are detected, protocol stack migration is completed within 3ms, ensuring end-to-end transmission reliability of >99.999%. Dual-link bandwidth aggregation supports 10Gbps-level real-time data backhaul.
3. The automatic cable fault waveform identification system according to claim 1 is characterized in that: In the deep time-frequency domain feature extraction network unit, wavelet packet decomposition is first used to decompose the signal into different frequency sub-bands to separate the key time-frequency components. Then, channel attention convolution is used to enhance the feature response of important channels, focusing on local time-frequency patterns. Two recurrent units are connected in series or in parallel. LSTM captures long-term trends through a gating mechanism, and GRU strengthens short-term fluctuations with a lightweight structure to jointly analyze dynamic changes in time series. The self-attention layer directly models global dependencies, breaking through the long-range information attenuation limitations of RNN and associating long-distance time series events.
4. The automatic cable fault waveform identification system according to claim 1, characterized in that: The knowledge-driven fault reasoning engine unit uses the IEC standard as a framework to unify the formats of equipment parameters and environmental data, define semantic mappings between fault codes and processing procedures, transform historical fault cases into a chain relationship between cause and effect, establish cross-dimensional connections, form conditional relationships between equipment parameters and environmental data, and cross-validate with safety thresholds in the IEC standard. A real-time data stream interface is deployed. When the sensor detects that the ambient temperature rise rate exceeds the IEC specified value, the graph reasoning engine is automatically triggered to generate a risk prediction path. An incremental learning mechanism is used to update weight parameters for every 100 new fault records. Monte Carlo simulation is combined to predict the probability of equipment failure under different environmental combinations, ensuring that the knowledge graph adapts to changes in the operating environment. The Bayesian network quantifies the dependencies between variables through a probabilistic graphical model and dynamically updates the confidence of an event. When the two are combined, MCTS generates potential decision branches, and the Bayesian network calculates the probability distribution of each branch in real time, forming a closed loop of simulation and probability correction. When the Bayesian network detects that the model confidence is lower than the threshold, the system triggers the expert rule base. In power grid fault diagnosis, if the Bayesian network has doubts about the probability of a short circuit on a certain line, the expert rules will forcibly inject industry standard logic: if the current mutation exceeds the threshold and is accompanied by a sudden drop in voltage, it is directly judged as a short circuit fault, and the confidence is corrected to a high credibility interval. The mechanism integrates data-driven and knowledge-driven methods, which not only retains the flexibility of MCTS, but also avoids model bias through expert experience and improves the reliability of decision-making in key scenarios.
5. The automatic cable fault waveform identification system according to claim 1 is characterized in that: In the adaptive incremental learning unit, several geographically distributed edge nodes collaborate to train the AI model. Each node retains the original data locally and only uploads model parameter updates. The global model is integrated and optimized through an aggregation algorithm, avoiding the cross-regional transmission of sensitive data while leveraging data characteristics from multiple locations to improve model adaptability. The AI model iteratively learns on a training set containing multiple fault modes, including traditional power grid faults and simulated faults of renewable energy grid connection. In each iteration, an inner loop adjusts the model parameters using specific samples for the current task, while an outer loop reversely optimizes the model's initial parameters based on the average performance of all tasks.