A method and apparatus for monitoring cable insulation

Abstract

The application discloses a cable insulation monitoring method and device, the method comprises: acquiring the pulse current signal, dielectric constant signal and environmental parameter signal of the target cable; signal processing is performed on the pulse current signal; environmental response modeling and conditional correction are performed on the pulse current signal and the dielectric constant signal according to the environmental parameter signal; the corrected dielectric constant signal is subjected to timing modeling and feature extraction; multi-source fusion is performed based on the corrected pulse current feature, dielectric constant timing feature and environmental parameter feature to generate a fusion feature; the fusion feature is input into a classification model, and an insulation state result of the target cable is output. Through correction, fusion and intelligent classification analysis of multi-source information, the application overcomes the defects of insufficient stability and accuracy of results caused by relying on single signal static diagnosis, and can realize real-time, accurate and robust monitoring of the cable insulation state, and is suitable for operation sites and intelligent operation and maintenance scenes of power cables.

Description

Technical Field

[0001] This application relates to the field of power fault detection technology, and in particular to a cable insulation monitoring method and device. Background Technology

[0002] As the main carrier of power transmission and distribution in power systems, the insulation performance of power cables directly affects the safe and stable operation of the power system. With the increase in the service life of cables, the insulation material may degrade due to factors such as electrothermal aging, moisture absorption, and partial discharge, ultimately leading to breakdown accidents. How to achieve online monitoring and diagnosis of cable insulation status is a core technical issue in the field of power operation and maintenance.

[0003] Traditional cable insulation monitoring technologies primarily rely on the acquisition and analysis of single signals, such as detection based on partial discharge pulse signals or measurement based on dielectric parameters. While these methods can reflect the cable insulation status to some extent, they typically suffer from the following shortcomings: single signals are easily affected by operating environment interference, resulting in insufficient stability of diagnostic results; it is difficult to dynamically characterize insulation degradation trends; the lack of comprehensive utilization of multi-source information leads to limited diagnostic accuracy; and the use of fixed thresholds or static features makes it difficult to adapt to the dynamic changes in signal statistical characteristics under complex operating environments.

[0004] Therefore, it is necessary to propose a cable insulation monitoring method and device that can integrate multiple feature information and improve diagnostic accuracy and real-time performance. Summary of the Invention

[0005] The embodiments of this application provide a cable insulation monitoring method and apparatus to solve the problems that cable insulation monitoring is mostly based on a single signal and the diagnostic results are not stable and accurate enough.

[0006] To address the aforementioned technical problems, embodiments of this application disclose the following technical solutions: Firstly, a cable insulation monitoring method is provided, comprising: acquiring a pulse current signal, a dielectric constant signal, and an environmental parameter signal of a target cable; filtering and amplifying the pulse current signal, and converting the filtered and amplified pulse current signal into a digital signal; performing environmental response modeling and condition correction on the pulse current signal and the dielectric constant signal based on the environmental parameter signal to compensate for the influence of environmental changes on the pulse current signal and the dielectric constant signal, thereby acquiring corrected pulse current signals and dielectric constant signals; performing time-series feature modeling on the corrected dielectric constant signal to acquire dielectric constant time-series features; performing normalization processing on the environmental parameter signal to acquire environmental parameter features; performing multi-source fusion based on the pulse current features corresponding to the corrected pulse current signal, the dielectric constant time-series features, and the environmental parameter features to generate fused features; inputting the fused features into a classification model and outputting the insulation state result of the target cable.

[0007] Furthermore, acquiring the pulse current signal, dielectric constant signal, and environmental parameter signal of the target cable includes: setting a high-frequency current sensor on the grounding wire of the shield layer of the target cable to sense the partial discharge pulse current during the operation of the target cable and outputting the pulse current signal; collecting the operating voltage and leakage current of the target cable to calculate the equivalent capacitance, and converting the relative dielectric constant of the insulation layer of the target cable according to the equivalent capacitance, and outputting the dielectric constant signal according to the time series of the relative dielectric constant; at least collecting the temperature parameter, humidity parameter, and load current parameter of the target cable during operation as the environmental parameter signal; wherein, the pulse current signal characterizes the transient characteristics of the partial discharge current signal of the target cable, the dielectric constant signal characterizes the scalar sequence of the relative dielectric constant changing with time, and the environmental parameter signal at least characterizes the temperature, humidity, and load current of the operating environment of the target cable.

[0008] Furthermore, the environmental response modeling includes: establishing an environmental response model based on the temperature, humidity, and load current in the environmental parameter signals to characterize the nonlinear coupling relationship between environmental conditions and the pulse current signal and the dielectric constant signal; compensating for amplitude drift, phase shift, or baseline drift of the pulse current signal and the dielectric constant signal according to the environmental response model, and outputting a corrected signal; wherein the environmental response model is established based on multiple linear regression, support vector regression, or neural network regression algorithms.

[0009] Furthermore, the conditional correction includes: generating an environmental parameter feature vector based on the environmental parameter signal; inputting the environmental parameter feature vector into a parameter generation network to obtain scaling and bias coefficients for correcting the pulse current signal and the dielectric constant signal; and performing conditional affine transformations on the pulse current signal and the dielectric constant signal respectively based on the scaling and bias coefficients.

[0010] Furthermore, the filtering and amplification of the pulse current signal includes: performing bandpass filtering on the pulse current signal to filter out power frequency and noise, and inputting the filtered pulse current signal into an adjustable gain amplifier for amplification.

[0011] Furthermore, the step of performing time-series feature modeling on the corrected dielectric constant signal includes: performing denoising and detrending processing on the corrected dielectric constant signal in time series form, and inputting it into a time-series modeler; the time-series modeler is any one of a recurrent neural network, a time-series convolutional network, or a converter network based on an attention mechanism.

[0012] Furthermore, the normalization processing of the environmental parameter signals includes: performing standardization operations on the temperature parameter, the humidity parameter, and the load current parameter respectively to generate environmental parameter features; the standardization operation includes scaling transformation based on the median and median absolute deviation, or scaling transformation based on the mean and standard deviation.

[0013] Furthermore, the fusion feature is a numerical feature vector, and the multi-source fusion is performed by concatenating at least two types of feature vectors. The at least two types of feature vectors include: a pulse current feature vector extracted from the pulse current signal after digitization, which is used to characterize at least one of the amplitude, repetition rate, time-domain statistics, or frequency-domain statistics of the partial discharge of the target cable; a dielectric constant time-series feature vector obtained by modeling the dielectric constant signal through time-series features; and an environmental parameter feature vector obtained by normalizing the environmental parameter signal.

[0014] Furthermore, before the vector concatenation is performed, conditional normalization or linear scaling bias transformation is performed on the pulse current feature vector and the dielectric constant timing feature vector based on the environmental parameter feature vector to generate a corrected feature vector.

[0015] Furthermore, the step of inputting the fused features into the classification model and outputting the insulation status result of the target cable includes: inputting the fused features into the classification model to obtain a probability distribution belonging to a preset state label set; and outputting the insulation status result of the target cable based on the probability distribution and a preset threshold. The classification model is a neural network model containing a convolutional neural network and a fully connected layer, or an ensemble learning model based on a gradient boosting tree, and the classification model is deployed in an edge computing terminal to achieve real-time output of the insulation status result of the target cable.

[0016] Secondly, a cable insulation monitoring device is provided, comprising: an acquisition unit for acquiring pulse current signals, dielectric constant signals, and environmental parameter signals of a target cable; a signal processing unit for filtering and amplifying the pulse current signals and converting them into digital signals; a correction unit for performing environmental response modeling and condition correction on the pulse current signals and dielectric constant signals based on the environmental parameter signals, to output corrected pulse current signals and dielectric constant signals; a modeling unit for performing time-series feature modeling on the corrected dielectric constant signals, to output dielectric constant time-series features; a normalization unit for normalizing the environmental parameter signals, to output environmental parameter features; a fusion unit for generating fused features based on the pulse current features corresponding to the pulse current signals, the dielectric constant time-series features, and the environmental parameter features; and a classification unit for inputting the fused features into a classification model and outputting the insulation status result of the target cable.

[0017] Furthermore, the acquisition unit includes a high-frequency current sensor, a dielectric constant detection module, a temperature and humidity sensor, and a load current sensor.

[0018] Furthermore, the classification unit includes a classification model deployed in an edge computing terminal. The signal processing unit, correction unit, and classification unit are all deployed in the edge computing terminal. The classification unit includes a classification model, which is a neural network model containing a convolutional neural network or a fully connected layer, or an ensemble learning model based on a gradient boosting tree. It is used to output a probability distribution belonging to a preset state label set, and to determine the insulation status of the target cable based on the probability distribution and a preset threshold.

[0019] The above-mentioned technical solutions have at least the following advantages or beneficial effects: By introducing environmental response modeling and condition correction mechanisms, the pulse current signal and dielectric constant signal can be adaptively adjusted according to the real-time acquired environmental parameters, compensating for amplitude drift, phase shift, and baseline drift caused by environmental changes, so that the signals acquired under different operating conditions are statistically aligned, thereby significantly improving the stability and accuracy of insulation condition diagnosis; on this basis, by simultaneously acquiring pulse current signals, dielectric constant signals, and environmental parameter signals, and combining the multi-source fusion analysis of the corrected signals, the defects of single signal and static signal diagnosis being susceptible to interference and having insufficient stability and accuracy of results can be effectively overcome, thereby improving the reliability and accuracy of cable insulation condition monitoring. Attached Figure Description

[0020] The specific embodiments of this application will be described in detail below with reference to the accompanying drawings, so that the technical solution and its beneficial effects of this application will be readily apparent.

[0021] Figure 1 A flowchart illustrating an exemplary cable insulation monitoring method provided in this application.

[0022] Figure 2 A schematic diagram of the frame of the cable insulation monitoring device 100 provided in this application.

[0023] Explanation of reference numerals in the attached figures: 100, cable insulation monitoring device; 101, target cable; 110, acquisition unit; 120, signal processing unit; 130, correction unit; 140, modeling unit; 150, normalization unit; 160, fusion unit; 170, classification unit. Detailed Implementation

[0024] To make the objectives, technical solutions, and beneficial effects of this application clearer, the following detailed description, in conjunction with the accompanying drawings and specific embodiments, further illustrates this application. It should be understood that the specific embodiments described in this specification are merely for explaining this application and are not intended to limit it.

[0025] In this application, unless otherwise expressly specified and limited, the term "pulse current signal" refers to the transient current signal generated by partial discharge activity during the operation of the target cable. This signal is typically a high-frequency pulse in the nanosecond to microsecond range, which can reflect the transient discharge characteristics of insulation defects.

[0026] In the description of this application, it is also necessary to explain some terms related to circuit signal processing. The term "high frequency" generally refers to the electromagnetic signal frequency band above the power frequency and its harmonics. In this application, it mainly refers to pulse signals in the range of hundreds of kilohertz to tens of megahertz caused by partial discharge. "Low frequency" generally refers to the frequency band close to or below the power frequency, often including 50Hz or 60Hz power frequency components and their low-order harmonics in power cable operation. The term "high-pass filtering" refers to using a circuit to allow signals above a certain cutoff frequency to pass while suppressing components below that frequency. The term "low-pass filtering" refers to allowing signals below a certain cutoff frequency to pass while suppressing high-frequency components. The term "band-pass filtering" refers to allowing signals within a certain frequency range to pass while simultaneously suppressing low-frequency and high-frequency components outside that range. In this application, band-pass filtering is often used to retain the effective frequency band where partial discharge pulses are located. The term "amplification" refers to the process of using an amplifier circuit to increase the amplitude of an input signal, thereby enhancing the measurability and adaptability of the signal for subsequent processing. The term "analog-to-digital conversion" refers to the process of converting analog signals into digital signals using an analog-to-digital converter. The resulting digital signals can be used by computers or digital circuits for feature extraction and modeling. The term "baseline" refers to the stable reference level corresponding to the signal output under conditions of no significant events or disturbances during signal measurement. For pulse current signals, the baseline is the zero level or static background level between discharge pulses. When temperature, humidity, load current, or measurement circuit parameters change, this reference level may shift or drift slowly, a phenomenon known as baseline drift. For dielectric constant signals, the baseline represents the average trend value of its time series, reflecting the stable reference level of the dielectric constant under conditions of no insulation degradation. The "baseline correction" or "baseline compensation" described in this application refers to adjusting the DC offset or slow trend of the signal through environmental response modeling or signal processing methods to return it to the standard reference level, thereby ensuring the accuracy of signal amplitude and phase measurements.

[0027] In this application, further explanation is needed regarding certain terms used in the design of power cables. The term "grounding lead location" refers to the point where the cable shield is connected to the ground grid via a grounding lead. This location is typically situated at the cable termination or joint and is used to safely introduce shield current and leakage current into the ground grid. The term "matching termination" refers to a resistor or circuit structure installed in the transmission path of a high-frequency pulse signal to ensure that the transmission line matches the input impedance of the acquisition circuit. This structure absorbs signal energy and reduces reflections to guarantee the authenticity of the signal waveform. The term "protective limiting network" refers to a clamping or limiting circuit, such as a combination of resistors, capacitors, diodes, or transient voltage suppression devices, configured at the input of the signal acquisition circuit. This circuit bypasses or limits transient overvoltages or overshoot signals exceeding normal amplitudes within a safe range to prevent damage to subsequent amplifiers or analog-to-digital converters.

[0028] Partial discharge is a common and representative form of insulation defect in power cables. Due to defects, moisture, or aging within or on the surface of the cable insulation material, localized electrical breakdown occurs. This phenomenon does not immediately lead to overall insulation breakdown, but it creates transient high-frequency current pulses in the defective area. As the cable operates, partial discharge activity gradually intensifies, leading to further degradation of the insulation material and potentially causing overall insulation breakdown. Traditional cable insulation monitoring often relies primarily on partial discharge pulse signals, acquiring signals by installing high-frequency current sensors on the cable shield grounding wire. However, this method is highly susceptible to environmental interference, and the detected signals often have weak amplitudes and low signal-to-noise ratios, making it difficult to consistently reflect the insulation status and effectively monitor the dynamic process of insulation degradation.

[0029] Therefore, this application proposes a cable insulation monitoring method that, based on partial discharge signal detection, further introduces dielectric constant signal and environmental parameter signal, and uses an algorithm to process the influence of dynamic changes in environmental conditions on partial discharge signal and dielectric constant signal.

[0030] Figure 1 This is a flowchart illustrating an exemplary cable insulation monitoring method provided in this application. Figure 1As shown, the method includes the following steps: S1 acquiring the pulse current signal, dielectric constant signal, and environmental parameter signal of the target cable 101; S2 filtering and amplifying the pulse current signal, and converting the filtered and amplified pulse current signal into a digital signal; S3 modeling and correcting the environmental response of the pulse current signal and dielectric constant signal based on the environmental parameter signal; S4 modeling the time-series characteristics of the corrected dielectric constant signal to obtain the dielectric constant time-series characteristics; S5 normalizing the environmental parameter signal to obtain environmental parameter characteristics; S6 performing multi-source fusion based on the pulse current characteristics, dielectric constant time-series characteristics, and environmental parameter characteristics corresponding to the corrected pulse current signal to generate fused features; S7 inputting the fused features into a classification model and outputting the insulation status result of the target cable. There is no strict requirement for the order of steps S4 and S5. This exemplary method improves the accuracy and robustness of cable insulation status diagnosis through multi-source information influence correction, fusion, and classification analysis, thereby better meeting the needs of power systems for online monitoring of cable insulation.

[0031] Based on the above exemplary cable insulation monitoring method, this application correspondingly proposes an exemplary cable insulation monitoring device 100.

[0032] Figure 2 This is a schematic diagram of the frame of the cable insulation monitoring device 100 provided in this application. Figure 2As shown, the device includes: a data acquisition unit 110 for acquiring pulse current signals, dielectric constant signals, and environmental parameter signals of the target cable 101; a signal processing unit 120 for filtering and amplifying the pulse current signals acquired by the data acquisition unit 110 and converting them into digital signals; and a correction unit 130 for performing environmental response modeling and condition correction on the pulse current signals output by the signal processing unit 120 and the dielectric constant signals output by the data acquisition unit 110 based on the environmental parameter signals acquired by the data acquisition unit 110, and outputting the corrected dielectric constant signals and pulse current signals. The signal processing unit 120, the correction unit 130, and the classification unit 170 are all deployed in an edge computing terminal. The modeling unit 140 is used to model the timing features of the corrected dielectric constant signal output by the correction unit 130 and output the timing features of the dielectric constant. The normalization unit 150 is used to normalize the environmental parameter signals acquired by the acquisition unit 110 and output the environmental parameter features. The fusion unit 160 is used to generate fused features based on the pulse current features, dielectric constant timing features, and environmental parameter features corresponding to the corrected pulse current signal. The classification unit 170 is used to input the fused features into a classification model and output the insulation state result of the target cable 101. In one feasible embodiment, the signal processing unit 120, the correction unit 130, and the classification unit 170 are all deployed in an edge computing terminal. The classification unit 170 includes a classification model, which is a neural network model containing a convolutional neural network or a fully connected layer, or an ensemble learning model based on a gradient boosting tree. It outputs a probability distribution belonging to a preset state label set and determines the insulation state result of the target cable 101 based on the probability distribution and a preset threshold.

[0033] The following is combined with Figure 1 and Figure 2 Based on the above exemplary process and apparatus, the specific implementation of each step and the specific device selection of each unit in the apparatus will be further explained.

[0034] In step S1, the acquisition unit 110 may include a high-frequency current sensor installed on the grounding wire of the shield layer of the target cable 101 to sense partial discharge pulse current and output a pulse current signal. Optionally, the acquisition unit 110 may also include a voltage acquisition module and a current acquisition module to acquire the operating voltage and leakage current of the target cable 101, and then calculate the equivalent capacitance and convert it to obtain the relative permittivity, thereby forming a permittivity signal. In addition, the acquisition unit 110 may also be configured with a temperature sensor, a humidity sensor, and a load current acquisition module to form environmental parameter signals to comprehensively reflect the operating environment of the target cable 101.

[0035] In step S1 above, to obtain the pulse current signal caused by partial discharge activity, a feasible implementation is to loop a high-frequency current sensor around the grounding wire of the shield layer of the target cable 101, so that it can sense the discharge transient current without changing the primary loop topology. The high-frequency current sensor can be a broadband current transformer or a Rogowski coil, preferably a device with a wide amplitude-frequency response and low phase distortion, and is connected to the front-end acquisition circuit in the acquisition unit 110 through a shielded coaxial cable. The input terminal of the front-end acquisition circuit is equipped with a matching termination and a protection limiting network to suppress reflection and overshoot. To reduce the superposition of power frequency leakage inductance and external electromagnetic field coupling, the high-frequency current sensor is installed close to the grounding lead and reliably equipotentially connected to the primary ground. If necessary, a metal shield is installed outside the sensor and grounded at one end. For a three-phase system, high-frequency current sensors can be arranged separately on the shielding grounding wires of each phase cable or at the three-phase combined grounding branch, so as to improve the detection capability of partial discharge pulses while maintaining electrical safety. Before the analog signal output by the high-frequency current sensor enters the front-end bandpass link and adjustable gain stage of the acquisition circuit, amplitude and polarity verification are performed, and a unified sampling clock identifier is written for subsequent timestamp alignment, thereby ensuring the time domain registration of data from different channels.

[0036] To obtain the dielectric constant signal, the acquisition unit 110 acquires the operating voltage and leakage current without changing the operating conditions of the target cable 101. The operating voltage is preferentially taken from the secondary side of the voltage transformer, or an insulated and calibrated capacitive voltage divider is used to bring the high-voltage side voltage to a safe measurement range. The leakage current can be obtained by a zero-sequence current transformer through a set of through-holes on the outside of the three-phase conductors of the target cable 101, or by sensing the return current in the shielding grounding branch using an isolated current sensing unit. To obtain the capacitive component Ic for equivalent capacitance calculation, in step S1, phase-sensitive detection or synchronous demodulation methods can be used to project the leakage current onto an orthogonal axis leading the power frequency voltage by 90°. After filtering out the active component and random noise, the effective value of Ic is obtained. Then, the equivalent capacitance C is calculated according to C=Ic / (ωU), where ω is the power frequency angular frequency and U is the effective voltage value. Combining this with the geometric parameters of the target cable 101, for example, according to a cylindrical capacitor model, the equivalent capacitance of the cable is: ;in, The vacuum permittivity (approximately 8.854 × 10⁻⁶) −12 F / m), It is the relative permittivity, reflecting the ability of an insulating material to respond to an electric field. For cable length, Where is the conductor radius, It is the outer radius of the insulation (the inner diameter from the cable shaft to the shielding layer). Characterize geometric factors. The relative permittivity can be calculated using the above formula. And construct according to sampling time sequence If the time series is a time series, then the time series can be analyzed through time- The curve reflects the dynamic changes of the insulating material. A slow, prolonged increase indicates insulation aging. A sudden anomaly indicates that the insulation layer may be damp or broken down. To improve robustness in weak signal scenarios, an auxiliary test signal can be optionally applied to the neutral point or shielded branch using a small amplitude, single-frequency, or narrowband injection method without interrupting power supply. The corresponding response current component is extracted by phase-locked amplification, and the obtained equivalent capacitance and relative permittivity are then compared. The dielectric constant signal sequence is also included; the above narrowband injection method is an optional measure in this embodiment and does not affect the acquisition of the dielectric constant signal under the condition of relying only on the operating voltage and leakage current.

[0037] To acquire environmental parameter signals, the acquisition unit 110 arranges environmental sensors near the terminal, joint, or inside the switchgear of the target cable 101 to reflect the influence of external conditions on the insulation characteristics of the target cable 101. To accurately reflect environmental impact, at least the temperature, humidity, or load current parameters of the target cable 101 are acquired. Temperature parameters can be simultaneously measured inside and outside the terminal accessory housing of the target cable 101 using a surface-mounted platinum resistance thermometer or a digital temperature sensor to distinguish between the internal temperature rise of the terminal accessory housing and the ambient temperature. Humidity parameters are preferably measured using a capacitive humidity sensor placed in the airflow area of ​​the terminal cavity or cabinet to detect the risk of moisture absorption and condensation. Load current is measured outside the conductor of the target cable 101 using a primary-side current transformer, an open-type Hall sensor, or a Rogowski coil, with the measurement bandwidth covering the main power frequency band and retaining necessary low-order harmonic components to facilitate subsequent characterization of thermal stress levels along with temperature parameters. At least one of the above-mentioned temperature, humidity, and load current parameters is acquired to form an environmental parameter signal; preferably, all three are acquired simultaneously to improve the accuracy of environmental impact assessment.

[0038] To ensure the consistency of multi-source data in the time domain, the acquisition unit 110 timestamps the data from each channel using the same reference clock, and shapes jitter and short-term sample loss through local buffering, outputting aligned pulse current signals, dielectric constant signals and environmental parameter signals to provide synchronous input for the processing in steps S2 to S7.

[0039] In step S2, the signal processing unit 120 filters and amplifies the pulse current signal acquired by the high-frequency current sensor and converts the processed signal into a digital signal. Specifically, filtering can be achieved using a bandpass filter, preserving the frequency band of the partial discharge pulse while filtering out power frequency and its low-order harmonic components, as well as irrelevant high-frequency interference. The bandpass filter can be constructed by cascading a high-pass filter and a low-pass filter. The former is used to suppress power frequency background and low-frequency noise, while the latter is used to eliminate frequency components higher than the sampling bandwidth, thereby ensuring that the signal entering the subsequent circuit maintains high fidelity in both the time and frequency domains. The filtered pulse current signal is then processed by an amplification circuit, preferably an adjustable gain amplifier, which can automatically or manually adjust the amplification factor according to the amplitude of the input signal, so as to ensure that small signals are detectable while avoiding amplifier saturation caused by large signals. To further improve the signal-to-noise ratio, a low-noise preamplifier stage can be optionally set in the amplification circuit, and DC isolation and bias adjustment can be introduced at the output to ensure the stability of the input of the subsequent analog-to-digital converter. The filtered and amplified signal enters the analog-to-digital converter for sampling and digitization. The sampling rate of the analog-to-digital converter (ADC) should be selected based on the upper cutoff frequency of the bandpass filter to satisfy the Nyquist sampling theorem and avoid spectral aliasing. Optionally, the ADC can be a successive approximation type or a high-speed pipelined structure to achieve a balance between different cost and resolution requirements. The digitized data will carry a unified clock identifier and be aligned on the time axis with the dielectric constant signal and environmental parameter signals output from other acquisition channels, providing synchronous input for subsequent feature modeling and fusion analysis.

[0040] In step S3, the correction unit 130 receives the environmental parameter signal, pulse current signal, and dielectric constant signal output by the acquisition unit 110, and establishes an environmental response model based on the temperature, humidity, and load current in the environmental parameter signal. This model characterizes the nonlinear coupling relationship between environmental conditions and the pulse current signal and the dielectric constant signal. The environmental response model aligns the statistical distributions of the pulse current signal and the dielectric constant signal under different temperature, humidity, and load conditions to a unified feature domain, thereby eliminating the interference of amplitude drift, phase shift, and baseline drift caused by environmental changes on subsequent feature extraction and discrimination. Specifically, environmental response modeling first establishes a mapping relationship between the environmental parameter signal and the deviations of the other two types of signals. For example, an environmental vector composed of temperature, humidity, and load current is used as the independent variable, and quantitative indicators such as the amplitude deviation, repetition rate deviation, baseline shift, and phase shift of the pulse current signal, as well as the mean drift, trend term slope, and short-term fluctuation amplitude of the dielectric constant time series are used as dependent variables. Model identification is completed offline based on historical data. In this embodiment, the environmental response model can be implemented using multiple linear regression, support vector regression, or shallow neural networks with two to three layers. During the identification process, overfitting is suppressed through cross-validation and regularization constraints, and the generalization ability is evaluated using an independent validation set across different seasons and load current ranges. Once the above modeling process is completed, the environmental response model is obtained. Its inputs are the current environmental vector and the original pulse current signal and dielectric constant signal, or short-window statistics for each input value. The outputs are the compensation amounts to be applied to each signal, including amplitude scaling factors, DC bias correction, and phase correction. The corrected pulse current signal and dielectric constant signal are unified to the standard state corresponding to the reference conditions in terms of mean, variance, and phase reference.

[0041] In actual operation, environmental response modeling can be performed online at the edge computing terminal. The correction unit 130 receives time-aligned data from the acquisition unit 110 and the signal processing unit 120. For the pulse current signal, it first calculates the baseline level and energy index within a short-time sliding window. Then, it inputs the window statistics along with the synchronously arriving temperature, humidity, and load current into the environmental response model to obtain the amplitude compensation and baseline compensation for the current window. Phase alignment is then performed in the phase domain using the reference voltage phase as a benchmark. For the dielectric constant signal, after constructing the time series, slowly varying statistics and trend terms are extracted using a longer time window. These quantities, along with the current environmental vector, are input into the environmental response model, outputting compensation for the mean drift and slow trend of the dielectric constant. This eliminates the systematic offset caused by the environment without masking short-time real fluctuations. To avoid overcompensation, the system sets engineering limits for each compensation channel. For example, the amplitude scaling, DC bias and phase correction are all limited by the calibrated safety boundaries. When the environmental response model output exceeds the boundaries or the confidence level is insufficient, the correction unit will revert to a conservative mode that only performs baseline regression and temperature drift coefficient correction, and record the status for subsequent maintenance and retraining.

[0042] In the conditional correction stage, the correction unit 130 uses a parameter generation network to map the environmental parameter signals into scaling and bias coefficients related to the current operating condition, and performs conditional affine transformations on the pulse current signal and the dielectric constant signal. In this embodiment, the parameter generation network uses a small neural network containing one or two fully connected layers. The input is a normalized environmental vector, and the output is two sets of parameters, corresponding to the scaling and bias coefficients of the pulse current signal and the dielectric constant signal, respectively. The correction unit 130 performs the transformation on the input signal samples in a windowed manner, that is, it performs x′=γ⋅x+β on the corresponding time window samples of the pulse current signal and the dielectric constant sequence, where the scaling coefficient γ is a dimensionless scaling coefficient, the bias coefficient β is dimensionless with x′, and the scaling coefficient γ and the bias coefficient β are updated in real time with the environmental vector. Through the above conditional affine transformation, the input signals under different environmental conditions are statistically aligned to a unified feature domain, specifically manifested as the mean of the corrected short window being close to zero, the variance being within a predetermined range, and the signal distribution distance under the reference condition being reduced. To ensure robustness of the criteria, a temperature scaling or confidence gating module can be added after the parameter generation network. When the environmental input is outside the training distribution or the generated parameters are unstable, the scaling intensity is automatically reduced and only the bias fine-tuning is retained, thereby avoiding distortion under extreme conditions. To improve online adaptability, step S3 above also introduces a lightweight incremental update mechanism. The correction unit continuously monitors the residual drift and distribution offset of the corrected signal without interrupting inference. When the residual offset exceeds a threshold within the sliding observation window, it triggers fine-tuning of the environmental response model or parameter generation network. Fine-tuning uses the most recent running data as samples, aims to maintain distribution alignment and avoid excessive changes to existing parameters, and achieves gradual correction through small-step optimization. This process is limited by the maximum update frequency and parameter change amplitude to ensure online stability. After dual processing of environmental response modeling and condition correction, the correction unit outputs the corrected pulse current signal and the corrected dielectric constant signal, which are used for subsequent pulse feature extraction and dielectric constant time series modeling, respectively. At the same time, the environmental vector, after being scaled by the normalization unit, participates in multi-source fusion as an independent environmental feature, so that the fusion and classification model receives stable features under a unified reference domain.

[0043] In this embodiment, step S3 fully considers synchronization and timing consistency. The correction unit aligns data frames from different channels with a unified clock and adopts a windowing strategy matching the sampling frequency, so that the short-window statistics of pulse current and the long-window statistics of dielectric constant form a nested relationship on the time axis and match it with the environmental vector at the same moment. To reduce computational overhead, the environmental response model and parameter generation network are both implemented using fixed-point quantization and vectorization and placed in the same processor or coprocessor of the edge computing terminal. The computational latency is controlled within a single frame window, ensuring real-time serial connection with the downstream S4 and S5 stages. Through the above process, step S3 completes the environmental adaptive alignment of the two types of signals without changing the primary system operating state, providing a stable and consistent input basis for subsequent feature modeling, fusion, and classification, and maintaining diagnostic accuracy and robustness under different seasons, loads, and humidity and heat conditions.

[0044] In the exemplary step S3 described above, by introducing an environmental response modeling and conditional correction mechanism, adaptive compensation for environmental impacts at the digital signal level is achieved, fundamentally overcoming the "environmental sensitivity" problem commonly found in existing cable insulation monitoring technologies that rely on hardware circuits. Traditional methods typically assume that the statistical characteristics of pulse current signals and dielectric constant signals remain unchanged under various operating conditions, ignoring the dynamic influence of environmental factors such as temperature, humidity, and load current on the polarization behavior of the insulation medium and the characteristics of the sensing link. This leads to detection threshold drift and increased false positive rates under high temperature, high humidity, or heavy load conditions. In contrast, this embodiment explicitly establishes a nonlinear mapping relationship between environmental parameters and signal deviations through environmental response modeling, and compensates for signal amplitude, phase, and baseline based on real-time environmental data during the online phase, unifying the signal distribution under different operating conditions to the same reference domain. Furthermore, conditional correction utilizes a parameter generation network to generate scaling and bias coefficients in real time based on environmental characteristics, performing conditional affine transformations on the signal to ensure that the stability and comparability of the feature distribution are maintained even when the environment changes rapidly or exceeds the training distribution. Therefore, this embodiment can not only dynamically offset the coupling effect of external environmental changes on the partial discharge signal and the dielectric constant signal, but also maintain the reliability of subsequent feature extraction and classification, achieving adaptive and environment-independent characteristics that traditional static models cannot possess. The signal-level environmental correction strategy differs from schemes that only use sample equalization or threshold correction during model training, further improving the system's diagnostic accuracy, robustness, and long-term stability in complex operating scenarios, and has outstanding engineering application value.

[0045] In step S4, the modeling unit 140 performs time-series feature modeling on the corrected dielectric constant signal to extract dynamic features that characterize the degradation trend of cable insulation. Specifically, the corrected dielectric constant signal is first preprocessed into a time series, which reflects the continuous change of the relative dielectric constant of the target cable 101 over time. To ensure the stability and accuracy of the time-series feature modeling results, the time series needs to undergo denoising and detrending processing before being input into the modeler. Denoising can be performed using methods such as moving average, wavelet thresholding, or empirical mode decomposition to reduce high-frequency fluctuations caused by voltage sensing errors or random interference; detrending is used to remove low-frequency drift caused by slowly changing environmental factors, such as diurnal temperature cycles, long-term humidity increases, or long-term high load currents, so that the remaining sequence better highlights the variation patterns of the insulation material.

[0046] After obtaining the preprocessed dielectric constant time series, modeling unit 140 inputs it into a time series modeler for feature extraction. The time series modeler can employ a recurrent neural network (RNN), such as a Long Short Term Memory (LSTM) network or a gated recurrent unit (GRU), to capture the dependence of the dielectric constant on long time scales; it can also employ a temporal convolutional network (TCN) to extract multi-scale dynamic features through one-dimensional convolution and dilated convolution structures; or it can employ an attention-based transformer network to analyze the correlation within the sequence from a global perspective using a self-attention structure. In practical implementation, one or more modeling methods can be flexibly selected according to computational resources and diagnostic accuracy requirements. Through the above process, the time series of the dielectric constant signal is mapped to a dielectric constant time series feature vector. This vector contains both the temporal evolution trend of the dielectric constant signal and retains the overall distribution information of the signal in a statistical sense, such as average level, fluctuation range, skewness, kurtosis, and frequency domain energy distribution. The dielectric constant timing characteristics will then be used together with the pulse current characteristics and environmental parameter characteristics in the multi-source fusion to generate more comprehensive fusion characteristics.

[0047] In step S5, the normalization unit 150 normalizes the environmental parameter signals to eliminate differences between different units and value ranges, ensuring that different environmental parameters are comparable and equally important during feature fusion. Preferably, the environmental parameter signals include at least temperature, humidity, and load current. These signals typically have different physical units and numerical scales in their original measurements; for example, temperature is expressed in degrees Celsius, humidity in percentage, and load current in amperes. If these raw values ​​are directly input into the classification model, parameters with larger numerical ranges may have excessive weight in model training and inference, weakening the role of other parameters. To address this issue, the normalization unit 150 performs standardization operations on the aforementioned environmental parameters. Specifically, a standardization method based on mean and standard deviation can be used, that is, subtracting the mean of each parameter within a certain time window from its current value, and then dividing by the standard deviation, so that the mean of the processed parameter distribution is close to 0 and the standard deviation is close to 1. In this way, the parameters are unified in numerical scale, facilitating subsequent processing. Alternatively, a robust standardization method based on the median and median absolute deviation can be used. This method is insensitive to outliers and can maintain the stability of the parameter distribution in the presence of sudden noise or extreme operating conditions, thereby enhancing the robustness of the features.

[0048] The following example illustrates the standardization process. During a given operating cycle, the temperature value collected by the temperature sensor fluctuates between 20°C and 80°C, the humidity value measured by the humidity sensor varies between 40% and 95%, and the load current fluctuates between 50A and 500A. If directly input into the model without normalization, the load current's numerical scale would be much larger than that of the temperature and humidity, creating a bias in the model calculation. After standardization, the numerical ranges of all three types of signals are compressed to a distribution close to zero mean and uniform variance. For example, the standardized temperature value may vary between -1 and 1, and the standardized values ​​of humidity and load current remain on the same order of magnitude, thus avoiding the unbalanced impact on diagnostic results caused by different parameters having different numerical values.

[0049] In practical applications, the standardization method can be flexibly selected based on the on-site operating conditions and data characteristics. For example, in scenarios where environmental parameters are stable over a long period and outliers are few, standardization based on the mean and standard deviation can be used first; while in scenarios with large environmental fluctuations or where there may be occasional outliers in the measurement, robust standardization is preferred. The normalized temperature, humidity, and load current characteristics will be output as environmental parameter characteristics, and will participate in subsequent multi-source fusion along with the pulse current and dielectric constant time-series characteristics.

[0050] In step S6, the fusion unit 160 integrates feature information from different sources to generate fused features that can be used as input for subsequent classification models. Specifically, after filtering, amplification, and digitization, the pulse current signal can be used to extract a feature vector characterizing partial discharge activity. This feature vector may include parameters such as the statistical distribution of pulse amplitude, discharge repetition rate, pulse interval time, and spectral energy distribution. After completing time-series feature modeling, the corrected dielectric constant signal forms a dielectric constant time-series feature vector, which reflects the changing trend of the dielectric constant over time. After normalization, the environmental parameter signal yields an environmental parameter feature vector. This vector unifies data such as temperature, humidity, and load current to the same numerical scale, enabling it to maintain an appropriate weight distribution with the influencing factors represented by other feature vectors during model training and inference.

[0051] The fusion unit 160 concatenates various feature vectors in a predetermined order to generate a numerical feature vector as the fused feature. The predetermined order can be fixed based on the feature source; for example, pulse current features can be concatenated first, followed by dielectric constant timing features, and finally environmental parameter features, to maintain data structure consistency and resolvability. The fused feature simultaneously includes partial discharge characteristics during cable operation, insulation medium timing features, and the impact of environmental conditions on insulation performance in the vector space, thus providing a more comprehensive and robust input data foundation for subsequent classification models.

[0052] In step S6, the classification unit 170 receives the fused features output by the fusion unit 160 and inputs them into the classification model to determine the insulation state of the cable. In this application, the classification model primarily performs nonlinear mapping and pattern recognition on multi-source features, and its output is the classification result of the insulation state of the target cable 101. Specifically, after receiving the fused features, the classification model obtains the probability distribution of the fused features belonging to a preset state label set through forward inference. The state label set can include categories such as "healthy," "mildly aged," and "severely aged," and can also be expanded to include more refined levels of state classification according to application requirements.

[0053] After obtaining the probability distribution, the classification unit 170 determines the insulation status of the target cable 101 based on preset threshold rules. For example, when the probability value of a certain category exceeds the threshold of 0.7, the target cable 101 is classified as belonging to that category. If the probability distribution is relatively balanced, the final status label can be further determined through methods such as confidence level calibration. Optionally, in addition to outputting the category label, the classification model can also simultaneously output the confidence score, risk index, or remaining life prediction value of the cable insulation status, so as to facilitate more detailed decisions by maintenance personnel.

[0054] Specifically, the classification model can employ a deep neural network structure containing convolutional neural networks (CNNs) and fully connected layers to automatically extract spatial patterns and nonlinear relationships from the fused features; alternatively, it can utilize an ensemble learning model based on gradient boosting decision trees (GBDTs) to leverage its advantages in structured data classification tasks. To ensure the real-time and independent nature of the diagnosis, the classification model can be deployed on an edge computing terminal, enabling feature processing and model inference to be completed on-site at the cable operation site, thereby reducing reliance on remote servers and communication links. In this way, the classification unit 170 can output the insulation status results of the target cable in real time, achieving rapid detection and early warning of cable insulation degradation.

[0055] The following section further explains the discrimination process of the classification model.

[0056] First, let's explain the fusion features. The statistical and spectral quantization vector of the pulse current signal, the temporal feature vector of the dielectric constant signal, and the normalized feature vector of the environmental parameters are sequentially or according to fixed rules concatenated into a numerical vector f∈ℝ^M of length M. Since all components have been scaled and time-aligned, this vector can be directly used as input to the classification model. The classification model maps this vector to a probability distribution p(y|f) on the set of state labels, and then outputs the state result or risk indicator based on this distribution.

[0057] When a CNN is used in the classification model, f is treated as a one-dimensional sequence input to a one-dimensional convolutional layer. The convolutional kernel slides along the vector dimension, which is equivalent to a linear combination of local receptive fields on the feature axis, enabling it to automatically learn the co-occurrence patterns and local interactions of "neighboring feature blocks". For example, a convolution with a kernel length of 3 or 5 can simultaneously "see" the boundary regions from pulse current feature blocks and dielectric constant time-series feature blocks, thereby capturing the discrimination signal brought about by the coupling changes between the two. Multi-layer stacking, dilated convolutions, or residual connections can expand the receptive field, enabling the network to perceive both local dependencies and long-distance dependencies across sub-blocks. After convolution, batch normalization and non-linear activation are usually applied to stabilize training, followed by one-dimensional global average pooling to compress the variable-length or multi-channel convolutional features into a fixed-length representation. Finally, one or more fully connected layers output to a softmax layer. ;in, It is a set of real-valued vectors output by the last layer of the model (usually a fully connected layer). The numerator represents the score of the k-th category. Map the fraction to a positive number, the denominator The sum of probabilities for all categories is guaranteed to be 1. The probability of each state label is obtained from the above formula. During the training phase, cross-entropy is used as the loss function, combined with L2 regularization or early stopping to suppress overfitting. During deployment, the computational structure of CNN is regular and has high parallelism, making it suitable for real-time discrimination with fixed-point or half-precision inference on edge computing terminals.

[0058] When the classification model employs a deep neural network structure containing fully connected layers, namely a Multilayer Perceptron (MLP), f is directly fed into several fully connected layers, each followed by a nonlinear activation function, such as ReLU, GELU, or SiLU. Batch normalization or layer normalization can be used to stabilize the gradients, and dropout or L2 constraints can be applied to improve generalization ability. This structure does not rely on local weight sharing in convolutions but learns high-order combination relationships between arbitrary features through dense connections, making it suitable for handling global interactions of "structured vectors." To enhance the representation of sub-blocks from different sources, linear embedding layers can be set at the input for three types of sub-blocks: pulse current features, dielectric constant time-series features, and environmental parameter features. Then, feature-level residual fusion is performed in the intermediate layers, preserving the discriminative power within sub-blocks while allowing the network to automatically learn cross-modal nonlinear interactions at higher levels. The output also uses a softmax layer to obtain multi-class probability distributions, and threshold strategies and confidence calibration, such as temperature scaling, are used to convert the probabilities into final state labels or graded warnings. The advantages of the above structure are its simplicity, natural friendliness to vector inputs, and low latency even on edge devices with limited computing power.

[0059] When GBDT is used in the classification model, a set of shallow decision trees are used as base learners. Gradient boosting is used to iteratively fit the residuals or negative gradients of the previous round, gradually approaching the optimal decision boundary in the function space. For multi-class tasks, log loss based on a softmax layer or a one-to-many strategy can be used. Each tree learns split points on the feature axis to divide the sample space into several regions. The leaf nodes of each region store the scores or log probabilities for each category. The overall prediction after integration is the weighted sum of the outputs of all trees. Since GBDT automatically learns nonlinear interactions on one-dimensional numerical features through gradient-driven learning, it can achieve good results with only a minimal number of feature vectors. During training, the bias-variance tradeoff is controlled by hyperparameters such as learning rate, subsampling, maximum depth, and number of leaf nodes. During inference, the fused feature vector f is sequentially fed into each tree, accumulated to obtain the scores for each category, mapped to probabilities via softmax or a logistic function, and then combined with a threshold to output the state result.

[0060] Regardless of the classification model used, for the scenario described in this application, the key lies in fusing feature vectors to simultaneously encode the intensity and repetition pattern of partial discharge pulses, the slow drift and abrupt changes in dielectric constant over time, and the instantaneous levels of environmental stresses such as temperature, humidity, and load. CNNs learn "pattern fragments" on the feature axis using one-dimensional convolutions, making them particularly sensitive to local coupling across sub-blocks; MLPs learn global combinations through layer-by-layer nonlinear mappings, exhibiting strong representation capabilities for complex high-order correlations; GBDT approximates nonlinear boundaries with piecewise constant functions, often demonstrating good generalization and interpretability when the sample size is limited or the feature dimension is moderate. When continuous quantization indicators are required, the softmax probability can be used to obtain a risk score through risk mapping or coupled with a degradation curve model to obtain a remaining lifetime estimate. All three types of classifiers can be deployed lightweight on edge computing terminals, meeting the real-time and reliability requirements of the field through point quantization, tree model pruning, or small convolutional / fully connected structures.

[0061] In summary, the cable insulation monitoring method and device provided in this application construct fused features and utilize a classification model to intelligently determine the insulation state by jointly acquiring, correcting, and fusing pulse current signals, dielectric constant signals, and environmental parameter signals. Compared with existing static detection methods that rely on a single signal, the method in this application can effectively improve the accuracy and robustness of diagnosis while ensuring real-time performance. It is suitable for online monitoring and early warning in cable operation sites, and can maintain monitoring accuracy and long-term stability, especially in complex operating environments such as high temperature and humidity and frequent load fluctuations.

[0062] It should be noted that the specific embodiments described above are only used to illustrate the technical solutions of this application, and are not intended to limit this application. For those skilled in the art, various modifications and changes can be made without departing from the spirit and scope of the technical solutions of this application, and all of these should fall within the protection scope of this application.

Claims

1. A method for monitoring cable insulation, characterized in that, include: Acquire the pulse current signal, dielectric constant signal, and environmental parameter signal of the target cable; The pulse current signal is filtered and amplified, and the filtered and amplified pulse current signal is converted into a digital signal; Based on the environmental parameter signals, an environmental response model is established using multiple linear regression, support vector regression, or neural network regression algorithms. The amplitude drift, phase drift, or baseline drift of the pulse current signal and the dielectric constant signal are compensated according to the environmental response model to characterize the impact of environmental changes on the pulse current signal and the dielectric constant signal. An environmental parameter feature vector is generated based on the environmental parameter signal, and input into a parameter generation network to obtain scaling and bias coefficients for correcting the pulse current signal and the dielectric constant signal. Based on the scaling and bias coefficients, conditional affine transformations are performed on the pulse current signal and the dielectric constant signal respectively to obtain the corrected pulse current signal and dielectric constant signal. The time-series characteristics of the corrected dielectric constant signal are modeled to obtain the time-series characteristics of the dielectric constant. The environmental parameter signals are normalized to obtain environmental parameter characteristics; Multi-source fusion is performed based on the pulse current characteristics corresponding to the corrected pulse current signal, the dielectric constant timing characteristics, and the environmental parameter characteristics to generate fused features; The fused features are input into the classification model and the target cable insulation status result is output.

2. The cable insulation monitoring method as described in claim 1, characterized in that, The acquisition of the pulse current signal, dielectric constant signal, and environmental parameter signals of the target cable includes: A high-frequency current sensor is installed on the grounding wire of the shield layer of the target cable to sense the partial discharge pulse current during the operation of the target cable and output the pulse current signal. The operating voltage and leakage current of the target cable are collected to calculate the equivalent capacitance, and the relative permittivity of the insulation layer of the target cable is calculated based on the equivalent capacitance. The relative permittivity time series is output as the permittivity signal. At least the temperature, humidity, and load current parameters of the target cable during operation are collected as environmental parameter signals. The pulse current signal characterizes the transient characteristics of the partial discharge current signal of the target cable, the dielectric constant signal characterizes the scalar sequence of the relative dielectric constant changing over time, and the environmental parameter signal characterizes at least the temperature, humidity, and load current of the operating environment of the target cable.

3. The cable insulation monitoring method as described in claim 1 or 2, characterized in that, The filtering and amplification of the pulse current signal includes: performing bandpass filtering on the pulse current signal to remove power frequency and noise, and inputting the filtered pulse current signal into an adjustable gain amplifier for amplification.

4. The cable insulation monitoring method as described in claim 1 or 2, characterized in that, The time-series feature modeling of the corrected dielectric constant signal includes: The corrected dielectric constant signal in time series form is denoised and detrended, and then input into the time series modeler; The temporal modeler is any one of a recurrent neural network, a temporal convolutional network, or an attention-based transformer network.

5. The cable insulation monitoring method as described in claim 2, characterized in that, The normalization process for the environmental parameter signal includes: The temperature parameter, humidity parameter, and load current parameter are standardized respectively to generate environmental parameter characteristics; the standardization operation includes scaling transformation based on median and median absolute deviation, or scaling transformation based on mean and standard deviation.

6. The cable insulation monitoring method as described in claim 1 or 2, characterized in that, The fusion feature is a numerical feature vector, and the multi-source fusion is performed by concatenating at least two types of feature vectors. The at least two types of feature vectors include: The pulse current feature vector corresponding to the corrected pulse current signal is used to characterize at least one of the amplitude, repetition rate, time-domain statistics, or frequency-domain statistics of the partial discharge of the target cable. The dielectric constant time-series feature vector is obtained by modeling the corrected dielectric constant signal using time-series features; The environmental parameter feature vector is obtained by normalizing the environmental parameter signal.

7. The cable insulation monitoring method as described in claim 6, characterized in that, Before the vector concatenation is performed, conditional normalization or linear scaling bias transformation is performed on the pulse current feature vector and the dielectric constant timing feature vector based on the environmental parameter feature vector to generate a corrected feature vector.

8. The cable insulation monitoring method as described in claim 1, characterized in that, The step of inputting the fused features into the classification model and outputting the target cable insulation state result includes: The fused features are input into the classification model to obtain the probability distribution of belonging to a preset state label set; Based on the probability distribution and the preset threshold, the insulation status result of the target cable is output; The classification model is a neural network model containing a convolutional neural network or a fully connected layer, or an ensemble learning model based on a gradient boosting tree, and the classification model is deployed in an edge computing terminal to achieve real-time output of the insulation status results of the target cable.

9. A cable insulation monitoring device, characterized in that, include: The acquisition unit is used to acquire the pulse current signal, dielectric constant signal, and environmental parameter signal of the target cable; The signal processing unit is used to filter and amplify the pulse current signal and convert it into a digital signal; The correction unit is configured to establish an environmental response model based on multiple linear regression, support vector regression, or neural network regression algorithms according to the environmental parameter signal, and to compensate for amplitude drift, phase drift, or baseline drift of the pulse current signal and the dielectric constant signal according to the environmental response model. It is also configured to generate an environmental parameter feature vector based on the environmental parameter signal, input it into a parameter generation network to obtain scaling and bias coefficients for correcting the pulse current signal and the dielectric constant signal, and to perform conditional affine transformations on the pulse current signal and the dielectric constant signal respectively according to the scaling and bias coefficients to output the corrected pulse current signal and dielectric constant signal. The modeling unit is used to model the timing characteristics of the corrected dielectric constant signal to output the timing characteristics of the dielectric constant. The normalization unit is used to normalize the environmental parameter signal to output environmental parameter features; A fusion unit is used to generate fused features based on the pulse current characteristics corresponding to the corrected pulse current signal, the dielectric constant timing characteristics, and the environmental parameter characteristics; The classification unit is used to input the fused features into the classification model and output the insulation status result of the target cable.

10. The cable insulation monitoring device as described in claim 9, characterized in that, The acquisition unit includes a high-frequency current sensor, a dielectric constant detection module, a temperature and humidity sensor, and a load current sensor.

11. The cable insulation monitoring device as described in claim 9, characterized in that, The signal processing unit includes a bandpass filter, an adjustable gain amplifier, and an analog-to-digital converter. The fusion unit is used to concatenate the pulse current feature vector, the dielectric constant timing feature vector, and the environmental parameter feature vector to form a numerical feature vector as the fused feature.

12. The cable insulation monitoring device as described in claim 9, characterized in that, The signal processing unit, correction unit, and classification unit are all deployed in the edge computing terminal. The classification unit includes a classification model, which is a neural network model containing a convolutional neural network or a fully connected layer, or an ensemble learning model based on a gradient boosting tree. The classification model is used to output the probability distribution of belonging to a preset state label set, and to determine the insulation status of the target cable based on the probability distribution and a preset threshold.

Images