Electric power communication line insulation deterioration early abnormity detection method and medium

Abstract

The invention provides an electric power communication line insulation degradation early-stage anomaly detection method and a medium, relates to the field of anomaly detection, and provides an EWED anomaly detection model for long-sequence-dependent electric power communication line data, and the EWED anomaly detection model is composed of a long-time dependence module, a frequency spectrum coding module and an anomaly judgment module. The identification capability of the model for early abnormal trends is enhanced by the long-time dependence module; the frequency spectrum coding module enhances the resolution of insulation degradation early-stage features; the anomaly judgment module improves the accuracy and response sensitivity of anomaly judgment, the interpretability and controllability of results are enhanced through probabilistic expression, and the system has the capacity of being flexibly adjusted to adapt to different operation environments through a threshold mechanism.

Description

Technical Field

[0001] This invention belongs to the field of anomaly detection, specifically relating to a method and medium for detecting early anomalies in the insulation degradation of power communication lines. Background Technology

[0002] As the fundamental support for power system dispatch control, relay protection, condition monitoring and information transmission, the reliability of power communication lines is directly related to the safety and stability of the power grid. With the continuous expansion of the power system scale and the improvement of the intelligence level, the amount of data and real-time requirements of communication lines have increased significantly. However, the lines are exposed to complex environments such as high temperature and humidity for a long time, and the insulation materials are prone to deterioration such as micro-cracks and interface debonding, forming early anomalies with strong concealment and uneven development speed, which lays hidden dangers for subsequent faults.

[0003] Currently, insulation condition detection of power communication lines mainly relies on manual inspections, periodic insulation tests, or threshold judgment methods based on a single physical quantity. These methods generally have limitations such as long detection cycles, insufficient real-time performance, and low sensitivity to subtle degradation characteristics. At the same time, the operating environment of communication lines is complex, and the insulation degradation process is multi-scale, covert, and gradual. Early degradation signals often manifest as weak disturbances, irregular fluctuations, or ambiguous features, making it difficult for traditional statistical analysis methods and univariate monitoring methods to achieve reliable identification. In addition, data acquisition is characterized by noise interference, strong non-stationarity, and coupling of multiple physical factors, making the early anomaly detection of insulation degradation even more challenging.

[0004] With the development of sensing technology, data acquisition equipment, and intelligent algorithms, intelligent early anomaly detection methods based on multi-dimensional operational characteristics have gradually become a research focus in the field of power communication. By introducing a data-driven anomaly modeling mechanism, a multi-scale feature model can be established for the normal behavior of communication lines under different loads, climates, and operating modes. Early weak degradation signals can be enhanced and extracted to improve detection sensitivity and accuracy. Deep learning is a method that can integrate multiple physical quantity features, has strong robustness, high real-time performance, and strong adaptability. In the context of early anomaly detection of insulation degradation in power communication lines, it can not only promptly identify potential hazards and reduce equipment maintenance costs, but also play an important role in improving the overall safety and reliability of power communication systems. Summary of the Invention

[0005] This invention provides a method and medium for early anomaly detection of insulation degradation in power communication lines. For power communication line data with long-sequence dependencies, an EWED anomaly detection model is proposed, which consists of a long-time dependency module, a spectrum coding module, and an anomaly judgment module.

[0006] The technical solution adopted by the present invention to achieve the above objectives specifically includes the following steps:

[0007] Collect data related to power communication lines and construct a dataset for preprocessing and partitioning.

[0008] Constructing a long-term dependency module: Using a sliding window, construct power communication lines and historical power communication line sequences, construct square and linear magnitude parts for local changes, and calculate the inertial diffusion kernel;

[0009] The global inertial memory field is calculated using the inertial diffusion kernel and the comprehensive characterization value. The time field within the window is obtained by combining the time position offset, the action factor, and the time gradient.

[0010] Constructing a spectrum coding module: Inputting a power communication line sequence, constructing an integer sequence using a discrete integer set, and calculating an integer grid sequence;

[0011] A standardized power communication line sequence is constructed based on the window time location sequence. After processing the adjacency and degree matrix calculation graph Laplace matrix, the quantization factor is output.

[0012] The modulus space mapping sequence is calculated based on the time position offset, and the indicator quantity is calculated by combining it with the integer grid point sequence; the resonance cumulative quantity is calculated.

[0013] Based on the calculation of local particulate coarsening index using time position offset and power communication line sequence, the trajectory increment and geometric consistency deviation index are processed, and the flexible attenuation factor is obtained by combining the basic attenuation constant.

[0014] The flexible matching factor is calculated by mapping the modulus space, the integer lattice sequence and the flexible attenuation factor. The flexible resonance intensity is obtained by summing the cumulative values ​​and then quantizing the maximum value to obtain the pseudo-spectral vector.

[0015] An anomaly detection module is constructed: the time field and pseudo-spectral vector within the splicing window are used as joint features, and the results of anomalies in power communication lines are judged by linear scoring, probabilization, and thresholding.

[0016] Preferably, relevant data of power communication lines are collected, including partial discharge high-frequency pulse data, leakage current data, harmonic component data, traveling wave characteristic measurement data, and hot spot temperature measurement data. The original dataset is constructed and preprocessed, and the dataset is divided into two parts in a ratio of 7:2:1.

[0017] Preferably, power communication line data is input. Construct a sliding window of length w as the sequence of power communication lines. The historical power communication line sequence is as follows For each time position in the historical power communication line sequence The local variation is obtained based on the difference between the values ​​of two adjacent historical power communication lines. The data is then quantified and characterized by constructing quadratic and linear components, respectively, and introducing quadratic and linear weights to obtain a comprehensive representation value for each historical time position. The specific mathematical model is as follows:

[0018] ;

[0019] In the formula, As a comprehensive characterization value, These are squared weights and linear weights, respectively. They are respectively the quadratic and linear components, and then based on the time position in the historical power communication line sequence. The interval between time positions t in the power communication line sequence is used to construct an exponentially decaying term, which is controlled by the historical information decay rate. Then, a linearly extending term is formed based on the same time interval, and its growth rate is adjusted using an inertia enhancement factor. Combining the exponentially decaying term and the linearly extending term yields the relationship between the current time position t and the historical time position t. The kernel value, specifically the mathematical model, is as follows:

[0020] ;

[0021] In the formula, It is an exponential function. For the current time position t and the historical time position The core value below, Historical information decay rate As an inertia enhancement factor, It is an exponentially decaying term. Assuming the linear extension term, the set of kernel values ​​for all time intervals is arranged to obtain the inertial diffusion kernel. .

[0022] Preferably, the input inertial diffusion nucleus With comprehensive characterization value , for each time position The corresponding inertial diffusion kernel and the comprehensive characterization value are multiplied term by term and accumulated in chronological order to obtain the global inertial memory field. The specific mathematical model is as follows:

[0023] ;

[0024] In the formula, For the current time position t and the historical time position The core value below, For the global inertial memory field, a corresponding time field gradient is constructed. At time position t, the global inertial memory field is compared and calculated with the global inertial memory field at the previous time position t-1 to form the time field gradient corresponding to that time position. The specific mathematical model is as follows:

[0025] ;

[0026] In the formula, For the time field gradient, For the global inertial memory field at the previous time position t-1, a time position offset is defined based on a constructed sliding window of length w. This offset is used to sequentially number each time position within the window relative to the window's starting time position, distinguishing the relative positional relationships of each time position within the window. A corresponding action factor is constructed for each position based on its offset i. The action factor is obtained by normalizing the time position offset i using the window length w. Then, based on the time field gradient, the corresponding time field gradients are obtained within the window according to the time position offset i. The action factor of each time position offset i is weighted and combined with the time field gradient, and accumulated to obtain the time field within the window. The specific mathematical model is as follows:

[0027] ;

[0028] In the formula, For the time field within the window, The length of the sliding window. Let be the gradient of the time field at time position t+i.

[0029] Preferably, the input power communication line sequence Using a set of discrete integers as a frequency index Construct a sequence of integers of length w for the frequency index, offset by the time position within the window. As independent variables, the time position offset and frequency index are calculated using integer transformation, and based on the integer modulus space parameter. By applying rounding constraints, an integer grid sequence is formed. The specific mathematical model is as follows:

[0030] ;

[0031] In the formula, It is an integer grid sequence. This is an integer modulo operation.

[0032] Preferably, a quantization factor is designed to perform translation and scale normalization on the power communication line sequences at each time position within the window, resulting in a standardized power communication line sequence. A similarity measure is performed on the standardized power communication line sequences between any two time points within the window to construct the corresponding adjacency matrix. The specific mathematical model is as follows:

[0033] ;

[0034] In the formula, This is an adjacency relation matrix. It is an exponential mapping function. As a similarity decay factor, For a standardized power communication line sequence at time position j, a degree matrix is ​​formed by calculating the metric values ​​at each time position based on the adjacency matrix. The specific mathematical model is as follows:

[0035] ;

[0036] In the formula, The degree matrix is ​​used as the basis for calculating the graph Laplacian matrix of the window using the two matrices. The specific mathematical model is as follows:

[0037] ;

[0038] In the formula, For the graph Laplacian matrix, the second and third eigenvalues ​​in its spectral structure are selected, and the difference ratio between the eigenvalues ​​is extracted to obtain an intermediate index. This index is then input into a nonlinear mapping unit for processing and range-limited processing based on upper and lower bound parameters. The output is the quantization factor corresponding to the current window. The specific mathematical model is as follows:

[0039] ;

[0040] In the formula, As a quantification factor, The lower bound parameter, The upper bound parameter, It is a nonlinear mapping function. This is the magnification factor. These are the second and third eigenvalues, respectively. For stable parameters, This is an intermediate indicator.

[0041] Preferably, the power communication line sequence at the start time position of the window is obtained based on the time position offset. An intermediate mapping quantity is obtained by linear scaling using a quantization factor. Then, integer modulo space parameters are used for rounding constraints to output the modulo space mapping sequence. The specific mathematical model is as follows:

[0042] ;

[0043] In the formula, It is a sequence of modulo space mappings. As a quantification factor, For intermediate mapping quantity, For integer modulo operations, For integer modulo space parameters, for integer lattice sequence The modulus space mapping sequence is compared one by one at the same offset position, and the indicative quantity is obtained based on the comparison result. The indicator values ​​are generated sequentially within a time offset i within the window, and each indicator value is accumulated to obtain the resonant cumulative value with respect to the frequency index k. The specific mathematical model is as follows:

[0044] ;

[0045] In the formula, This is the cumulative amount of resonance. This represents the length of the sliding window.

[0046] Preferably, a flexible attenuation factor is proposed, which applies to each time position offset i within the window in the power communication line sequence. Multiple neighboring value pairs with different intervals are selected, and the differences between neighboring value pairs are calculated to obtain multiple scale difference quantities. The scale difference quantities are combined with the bias parameter to process the difference quantities at different scales, thereby obtaining the local fractal roughness index. The specific mathematical model is as follows:

[0047] ;

[0048] In the formula, For local fractal roughness index, For bias parameters, This represents the scale difference calculated from neighboring value pairs at different intervals. These represent the logarithmic scales corresponding to the coarse-scale time interval and the fine-scale time interval, respectively. Then, the sequence of adjacent integer grid points... and Perform differential calculation of trajectory increment The amplitude relationship between adjacent trajectory increments is normalized, the difference ratio is calculated, and the offset is processed. This is then adjusted using an adjustment coefficient to obtain the geometric consistency deviation index. The specific mathematical model is as follows:

[0049] ;

[0050] In the formula, This is a geometric consistency deviation index. To and Adjacent trajectory increments, As the adjustment coefficient, combined with the basic attenuation constant, the local fractal roughness index and the geometric consistency deviation index are multiplicatively coupled to obtain the flexible attenuation factor. The specific mathematical model is as follows:

[0051] ;

[0052] In the formula, As a flexible attenuation factor, It is the basic attenuation constant.

[0053] Preferably, the modular space mapping sequence With integer grid sequence Offset difference calculation, combined with flexible attenuation factor After amplitude scaling and exponentialization, the flexible matching factor is obtained. The specific mathematical model is as follows:

[0054] ;

[0055] In the formula, As a flexible matching factor, For exponential functions, the sliding window length is... As the traversal range, the flexible matching factors within the window are cumulatively summed, and the resonance cumulative amount is calculated. As a discrete benchmark, the discrete resonance relationship is defined, providing a baseline reference for the flexible resonance intensity and a consistent alignment structure to obtain the flexible resonance intensity. The specific mathematical model is as follows:

[0056] ;

[0057] In the formula, The flexible resonance intensity is determined, and then the maximum value of the flexible resonance intensity is determined. Using the grid index as the traversal range, the flexible resonance intensity is transformed relative to the maximum value of the flexible resonance intensity to obtain a pseudo-spectral vector of length N. The specific mathematical model is as follows:

[0058] ;

[0059] In the formula, It is a pseudo-spectral vector. For grid index, The grid length is This represents the maximum value of the flexible resonance intensity.

[0060] Preferably, the time field within the input window With pseudo-spectral vector splicing as joint features The anomaly score of power communication lines is obtained using a linear scoring function. The specific mathematical model is as follows:

[0061] ;

[0062] In the formula, For scoring abnormalities in power communication lines, For learnable weight vectors, This is the bias term, and then the probability value of power communication line anomalies is output through a probabilistic mechanism. The specific mathematical model is as follows:

[0063] ;

[0064] In the formula, Probability value of power communication line anomalies For the Sigmoid function, As an exponential constant, a threshold mechanism is introduced to judge the probability value of power communication line anomalies based on a preset threshold, thereby obtaining the anomaly result of the power communication line. .

[0065] The technical effects and advantages provided by the present invention in the above technical solution are as follows:

[0066] 1. In the long-term dependency module, this invention implements sliding window modeling for power communication line sequences and historical sequences, introducing local changes, combinations of square and linear magnitudes, and exponential decay and linear extension terms constructed from time intervals. This allows the correlation between different time positions to be dynamically characterized by quantitative kernel values. The sequential arrangement of kernel values ​​forms an inertial diffusion kernel that can characterize the continuity of historical influences. Thus, in power communication line scenarios with slow changing trends, strong disturbance accumulation, and weak abrupt changes in electrical quantities due to external environmental influences, the time dependency can be continuously expressed across windows and scales. This promotes the capture of subtle degradation signs, improves the sensitivity and stability of insulation performance changes, enhances the ability to identify early abnormal trends, and enables the detection model to maintain robust time series understanding and abnormal response capabilities even under conditions of strong noise background, significant trend delay, and unclear early signals.

[0067] 2. In the spectrum coding module, this invention constructs integer sequences and integer grid sequences using discrete integer sets, enabling the temporal evolution characteristics of power communication lines to be re-expressed in discrete frequency form. Combined with the standardized sequence of window time positions, an adjacency and degree matrix is ​​constructed using similarity relationships, forming a graph Laplacian matrix. Structural change patterns within the window are extracted through eigenvalue differences, and quantization factors are output, achieving a consistent description of multi-timescale fluctuations. A modulus space mapping sequence is generated based on time position offsets, compared with the integer grid sequence to obtain an indicator quantity, and accumulated as a resonance cumulative quantity, allowing frequency-related changes to be aggregated and expressed. Furthermore, a flexible attenuation factor is constructed by fusing local fractal roughness and geometric consistency deviations. Flexible matching factors are obtained by flexibly adjusting offset differences, and the flexible resonance intensity is obtained through cumulative summation and quantized as a pseudo-spectral vector. This module can highlight hidden weak change patterns at the frequency domain level, enhancing the resolution of early insulation degradation characteristics. It enables multi-scale disturbances in complex time series to be aligned and compared within a unified frequency domain framework, thereby improving the sensitivity, stability, and robustness of anomaly detection.

[0068] 3. In the anomaly detection module, this invention concatenates the time field output within the window with a pseudo-spectral vector to form a joint expression that simultaneously includes long-term time-domain dependency features and frequency-domain structural resonance features. Then, a linear scoring function is used to map the joint features and output an anomaly score. Subsequently, a probabilistic mechanism is used to transform the score into an interpretable anomaly probability, and a threshold mechanism is combined to complete the final judgment. This module can unify and fuse the multi-scale, multi-structure, and multi-domain features extracted by the preceding modules, so that anomaly identification not only depends on instantaneous changes, but also comprehensively considers long-term trends, frequency-domain resonance behavior, and local disturbance intensity, thereby significantly improving the accuracy and response sensitivity of anomaly detection. The probabilistic expression enhances the interpretability and controllability of the results, while the threshold mechanism enables the system to flexibly adjust and adapt to different operating environments. Attached Figure Description

[0069] Figure 1 This is a step diagram of a method for detecting early abnormalities in the insulation of power communication lines.

[0070] Figure 2 This is a long-term dependency module diagram.

[0071] Figure 3 This is a diagram of the spectrum coding module.

[0072] Figure 4 This is an error distribution diagram of the EWED anomaly detection model.

[0073] Figure 5 This is a screenshot of the EWED anomaly detection model.

[0074] Figure 6This is a histogram showing the distribution of comprehensive judgment indicators for the EWED anomaly detection model. Detailed Implementation

[0075] This invention proposes a method for detecting early abnormalities in the insulation degradation of power communication lines, the steps of which are as follows: Figure 1 As shown, an EWED anomaly detection model is proposed for power communication line data with long-sequence dependencies. This model consists of a long-time dependency module, a spectrum coding module, and an anomaly judgment module. The long-time dependency module enhances the model's ability to identify early anomaly trends, enabling robust time-series understanding and anomaly response capabilities even under conditions of strong noise background, significant trend delays, and unclear early signals. The spectrum coding module enhances the resolution of early insulation degradation features, allowing multi-scale disturbances in complex time series to be aligned and compared within a unified frequency domain framework, thereby improving the sensitivity, stability, and robustness of anomaly detection. The anomaly judgment module improves the accuracy and response sensitivity of anomaly judgment, and the probabilistic expression strengthens the interpretability and controllability of the results. The threshold mechanism enables the system to flexibly adjust and adapt to different operating environments.

[0076] Collect data related to power communication lines and construct a dataset for preprocessing and partitioning.

[0077] Furthermore, power communication line data acquisition was conducted, including the simultaneous acquisition of partial discharge high-frequency pulse data, leakage current data, harmonic component data, traveling wave characteristic measurement data, and hotspot temperature measurement data. A unified time base, unified sampling strategy, and fixed parameter configuration were adopted to ensure that multi-source features could be used for degradation trend modeling and anomaly identification. For the partial discharge high-frequency pulse acquisition, an HFCT sensor with a center frequency of 5MHz, bandwidth of 100kHz–30MHz, sensitivity of 20mV / mA, and noise below 2mV was installed 5–10cm outside the insulator string fittings, performing high-speed sampling at a sampling frequency of 100MS / s and 12-bit quantization accuracy. A 200 A μs sampling window and a 500μs event recording length were used. The trigger threshold was set to a discharge pulse current of not less than 8mA, and the minimum event recording interval was set to 1ms. Calibration was performed using standard pulses of 10mA, 50mA, and 100mA. Leakage current was acquired using a sensor with a range of 0–50mA and an accuracy of ±0.5%FS, recorded at a sampling frequency of 1kHz and a resolution of 16bits. A 200ms RMS calculation window was used, and data was generated every 1s and reported every 10s. A leakage current RMS exceeding 5mA was set as a warning threshold, and exceeding 10mA was considered a serious degradation indicator. Harmonic component acquisition relied on the CT secondary side signal, with a sampling frequency of 12.8kHz and a resolution of 16bits. Sampling was performed with bit quantization precision. A 256-point FFT window was used to achieve a harmonic resolution of 0.195Hz. The amplitudes of the 3rd, 5th, 7th, 9th, and 11th harmonics within the 0–2kHz frequency band were recorded, and THD changes were monitored. When the THD exceeded 15% or the amplitude of any odd harmonic exceeded 5% of the fundamental frequency, it was marked as having discharge-type nonlinear characteristics. Traveling wave characteristic measurement used a traveling wave monitoring unit with a frequency band of 0.1–2 MHz, and GPS / BeiDou PPS signals were used to achieve a time synchronization accuracy of no more than 50ns. Traveling wave recordings were performed at a sampling frequency of 5MHz for 2ms. The trigger threshold was set to a traveling wave amplitude change of no less than 0.5kA / km, and the pre-trigger time was 100μs. The arrival time, wavefront steepness, and amplitude change rate of the traveling wave were recorded. The line degradation point was accurately located by using the multi-point traveling wave arrival time difference Δt. The overall positioning error was controlled within ±15m. Hot spot temperature measurement used an infrared thermometer or contact temperature sensor in the 8–14μm band, with an accuracy of ±1 The temperature of insulator fittings, tension clamps, and joints is continuously monitored at a sampling frequency of 1Hz and a recording period of 10s. The normal temperature of insulator fittings is generally 45–65℃. A temperature of 75℃ is a warning, and 85℃ is an abnormality. The temperature rise of tension clamps exceeding 80℃ triggers a warning, and exceeding 90℃ is considered abnormal. A temperature at the joint exceeding 70℃ indicates a risk of deterioration.

[0078] Build long-term dependent modules, such as Figure 2As shown, the module is as follows: it uses a sliding window to construct power communication lines and historical power communication line sequences, constructs square and linear order of magnitude parts for local changes, and calculates the inertial diffusion kernel.

[0079] Furthermore, input power communication line data Construct a sliding window of length w as the sequence of power communication lines. The historical power communication line sequence is as follows For each time position in the historical power communication line sequence The local variation is obtained based on the difference between the values ​​of two adjacent historical power communication lines. The data is then quantified and characterized by constructing quadratic and linear components, respectively, and introducing quadratic and linear weights to obtain a comprehensive representation value for each historical time position. The specific mathematical model is as follows:

[0080] ;

[0081] In the formula, As a comprehensive characterization value, These are squared weights and linear weights, respectively. They are respectively the quadratic and linear components, and then based on the time position in the historical power communication line sequence. The interval between time positions t in the power communication line sequence is used to construct an exponentially decaying term, which is controlled by the historical information decay rate. Then, a linearly extending term is formed based on the same time interval, and its growth rate is adjusted using an inertia enhancement factor. Combining the exponentially decaying term and the linearly extending term yields the relationship between the current time position t and the historical time position t. The kernel value, specifically the mathematical model, is as follows:

[0082] ;

[0083] In the formula, It is an exponential function. For the current time position t and the historical time position The core value below, Historical information decay rate As an inertia enhancement factor, It is an exponentially decaying term. Assuming the linear extension term, the set of kernel values ​​for all time intervals is arranged to obtain the inertial diffusion kernel. .

[0084] In this embodiment, the sliding window length is set to 60. The insulation degradation of power communication lines is a slowly changing anomaly, and its disturbances typically exhibit continuous weak fluctuations within dozens of sampling points. A window of 60 sampling points is suitable. The square weight and linear weight are set to 0.7 and 0.3, respectively. Early insulation degradation of power communication lines typically manifests as a small increase in fluctuation energy. The square-order part is more sensitive to amplitude changes, hence a higher weight is assigned. The linear part reflects the directional change trend of the signal, and its importance is lower than that of energy-type changes, therefore a lower weight is given. The historical information attenuation rate and inertia enhancement factor are set to 0.015 and 0.2, respectively. Early insulation degradation signals have weak disturbances and long-term correlation characteristics. An attenuation rate of 0.015 ensures the capture of gradual changes in insulation degradation. The insulation degradation process has inertial accumulation characteristics; setting the inertia enhancement factor to 0.20 avoids excessive enhancement leading to noise amplification.

[0085] The global inertial memory field is calculated using the inertial diffusion kernel and the comprehensive characterization value. The time field within the window is obtained by combining the time position offset, the action factor, and the time gradient.

[0086] Furthermore, input inertial diffusion nucleus With comprehensive characterization value , for each time position The corresponding inertial diffusion kernel and the comprehensive characterization value are multiplied term by term and accumulated in chronological order to obtain the global inertial memory field. The specific mathematical model is as follows:

[0087] ;

[0088] In the formula, For the current time position t and the historical time position The core value below, For the global inertial memory field, a corresponding time field gradient is constructed. At time position t, the global inertial memory field is compared and calculated with the global inertial memory field at the previous time position t-1 to form the time field gradient corresponding to that time position. The specific mathematical model is as follows:

[0089] ;

[0090] In the formula, For the time field gradient, For the global inertial memory field at the previous time position t-1, a time position offset is defined based on a constructed sliding window of length w. This offset is used to sequentially number each time position within the window relative to the window's starting time position, distinguishing the relative positional relationships of each time position within the window. A corresponding action factor is constructed for each position based on its offset i. The action factor is obtained by normalizing the time position offset i using the window length w. Then, based on the time field gradient, the corresponding time field gradients are obtained within the window according to the time position offset i. The action factor of each time position offset i is weighted and combined with the time field gradient, and accumulated to obtain the time field within the window. The specific mathematical model is as follows:

[0091] ;

[0092] In the formula, For the time field within the window, The length of the sliding window. Let be the gradient of the time field at time position t+i.

[0093] Construct a spectrum coding module, such as Figure 3 As shown, the module is: inputting a power communication line sequence, constructing an integer sequence through a discrete integer set, and calculating an integer grid sequence.

[0094] Furthermore, input power communication line sequence Using a set of discrete integers as a frequency index Construct a sequence of integers of length w for the frequency index, offset by the time position within the window. As independent variables, the time position offset and frequency index are calculated using integer transformation, and based on the integer modulus space parameter. By applying rounding constraints, an integer grid sequence is formed. The specific mathematical model is as follows:

[0095] ;

[0096] In the formula, It is an integer grid sequence. This is an integer modulo operation.

[0097] In this embodiment, the integer modulus space parameter is set to 97, which can avoid the periodic frequency components from exhibiting predictable repeating structures in the integer domain, and make the weak frequency domain offset corresponding to early insulation degradation more distinguishable in the mapped integer space.

[0098] A standardized power communication line sequence is constructed based on the window time location sequence. After processing the adjacency and degree matrix calculation graph Laplace matrix, the quantization factor is output.

[0099] Furthermore, a quantization factor is designed to perform translation and scale normalization on the power communication line sequences at each time position within the window, resulting in a standardized power communication line sequence. A similarity measure is performed on the standardized power communication line sequences between any two time points within the window to construct the corresponding adjacency matrix. The specific mathematical model is as follows:

[0100] ;

[0101] In the formula, This is an adjacency relation matrix. It is an exponential mapping function. As a similarity decay factor, For a standardized power communication line sequence at time position j, a degree matrix is ​​formed by calculating the metric values ​​at each time position based on the adjacency matrix. The specific mathematical model is as follows:

[0102] ;

[0103] In the formula, The degree matrix is ​​used as the basis for calculating the graph Laplacian matrix of the window using the two matrices. The specific mathematical model is as follows:

[0104] ;

[0105] In the formula, For the graph Laplacian matrix, the second and third eigenvalues ​​in its spectral structure are selected, and the difference ratio between the eigenvalues ​​is extracted to obtain an intermediate index. This index is then input into a nonlinear mapping unit for processing and range-limited processing based on upper and lower bound parameters. The output is the quantization factor corresponding to the current window. The specific mathematical model is as follows:

[0106] ;

[0107] In the formula, As a quantification factor, The lower bound parameter, The upper bound parameter, It is a nonlinear mapping function. This is the magnification factor. These are the second and third eigenvalues, respectively. For stable parameters, This is an intermediate indicator.

[0108] In this embodiment, the similarity attenuation factor is set to 0.65 to ensure that the weak degradation features are sufficiently sensitive while avoiding amplification of normal slight fluctuations; the lower bound parameter is set to 0.10 to weaken the extremely low amplitude disturbances caused by background noise, power frequency interference, and equipment quantization errors during the line sampling process, preventing low-amplitude random disturbances from entering the subsequent nonlinear adjustment process; the upper bound parameter is set to 3.0 to limit the similarity change within a reasonable range of typical weak degradation features in the early stage of insulation, ensuring that occasional spikes do not affect the overall trend structural judgment; the amplification factor is set to 5.0 to moderately enhance the mildly gradual variation characteristics that are usually present during the insulation degradation process of power communication lines, making the weak amplitude structure of early anomalies distinguishable during the calculation process; the stability parameter is set to 1×10. -6 It is used to maintain numerical stability when the denominator of the similarity ratio approaches a minimum value, and to avoid abnormal amplification of the ratio due to long-term stable operation.

[0109] The modulus space mapping sequence is calculated based on the time position offset, and the indicator quantity is calculated by combining it with the integer grid point sequence. The resonance cumulative quantity is then calculated.

[0110] Furthermore, based on the time position offset, the power communication line sequence at the start time position of the window is obtained. An intermediate mapping quantity is obtained through linear scaling using a quantization factor. Then, integer modulo space parameters are used for rounding constraints to output the modulo space mapping sequence. The specific mathematical model is as follows:

[0111] ;

[0112] In the formula, It is a sequence of modulo space mappings. As a quantification factor, For intermediate mapping quantity, For integer modulo operations, For integer modulo space parameters, for integer lattice sequence The modulus space mapping sequence is compared one by one at the same offset position, and the indicative quantity is obtained based on the comparison result. The indicator values ​​are generated sequentially within a time offset i within the window, and each indicator value is accumulated to obtain the resonant cumulative value with respect to the frequency index k. The specific mathematical model is as follows:

[0113] ;

[0114] In the formula, This is the cumulative amount of resonance. This represents the length of the sliding window.

[0115] In this embodiment, the integer modulus space parameter is set to 97, which can avoid the periodic frequency components from exhibiting predictable repeating structures in the integer domain, and make the weak frequency domain offset corresponding to early insulation degradation more distinguishable in the mapped integer space.

[0116] Based on the time position offset and power communication line sequence, a local fractal coarsening index is obtained. The trajectory increment and geometric consistency deviation index are processed, and the flexible attenuation factor is obtained by combining the basic attenuation constant.

[0117] Furthermore, a flexible attenuation factor is proposed, which applies to each time position offset i within the window in the power communication line sequence. Multiple neighboring value pairs with different intervals are selected, and the differences between neighboring value pairs are calculated to obtain multiple scale difference quantities. The scale difference quantities are combined with the bias parameter to process the difference quantities at different scales, thereby obtaining the local fractal roughness index. The specific mathematical model is as follows:

[0118] ;

[0119] In the formula, For local fractal roughness index, For bias parameters, This represents the scale difference calculated from neighboring value pairs at different intervals. These represent the logarithmic scales corresponding to the coarse-scale time interval and the fine-scale time interval, respectively. Then, the sequence of adjacent integer grid points... and Perform differential calculation of trajectory increment The amplitude relationship between adjacent trajectory increments is normalized, the difference ratio is calculated, and the offset is processed. This is then adjusted using an adjustment coefficient to obtain the geometric consistency deviation index. The specific mathematical model is as follows:

[0120] ;

[0121] In the formula, This is a geometric consistency deviation index. To and Adjacent trajectory increments, As the adjustment coefficient, combined with the basic attenuation constant, the local fractal roughness index and the geometric consistency deviation index are multiplicatively coupled to obtain the flexible attenuation factor. The specific mathematical model is as follows:

[0122] ;

[0123] In the formula, As a flexible attenuation factor, It is the basic attenuation constant.

[0124] In this embodiment, the bias parameter is set to 0.20. This setting ensures that very small disturbances are not perceived as structural offsets by the system, while also not masking the slight, persistent deviations that may occur in the early stages of insulation degradation. The adjustment coefficient is set to 1.35, which can maintain the enhancement of the trend under weak noise conditions, while avoiding excessive amplification of transient noise. The basic attenuation constant is set to 0.92, which enables the deviation information to maintain a smooth, continuous, and interpretable trajectory along the time axis, more accurately reflecting the weak but persistent evolution pattern in the early stages of insulation degradation.

[0125] The flexible matching factor is calculated by mapping the modulus space, the integer lattice sequence and the flexible attenuation factor. The flexible resonance intensity is obtained by summing the results and then quantizing the results according to the maximum value to obtain the pseudo-spectral vector.

[0126] Furthermore, for the modular space mapping sequence With integer grid sequence Offset difference calculation, combined with flexible attenuation factor After amplitude scaling and exponentialization, the flexible matching factor is obtained. The specific mathematical model is as follows:

[0127] ;

[0128] In the formula, As a flexible matching factor, For exponential functions, the sliding window length is... As the traversal range, the flexible matching factors within the window are cumulatively summed, and the resonance cumulative amount is calculated. As a discrete benchmark, the discrete resonance relationship is defined, providing a baseline reference for the flexible resonance intensity and a consistent alignment structure to obtain the flexible resonance intensity. The specific mathematical model is as follows:

[0129] ;

[0130] In the formula, The flexible resonance intensity is determined, and then the maximum value of the flexible resonance intensity is determined. Using the grid index as the traversal range, the flexible resonance intensity is transformed relative to the maximum value of the flexible resonance intensity to obtain a pseudo-spectral vector of length N. The specific mathematical model is as follows:

[0131] ;

[0132] In the formula, It is a pseudo-spectral vector. For grid index, The grid length is This represents the maximum value of the flexible resonance intensity.

[0133] An anomaly detection module is constructed: the time field and pseudo-spectral vector within the splicing window are used as joint features, and the results of anomalies in power communication lines are judged by linear scoring, probabilization, and thresholding.

[0134] Furthermore, the time field within the input window With pseudo-spectral vector splicing as joint features The anomaly score of power communication lines is obtained using a linear scoring function. The specific mathematical model is as follows:

[0135] ;

[0136] In the formula, For scoring abnormalities in power communication lines, For learnable weight vectors, This is the bias term, and then the probability value of power communication line anomalies is output through a probabilistic mechanism. The specific mathematical model is as follows:

[0137] ;

[0138] In the formula, Probability value of power communication line anomalies For the Sigmoid function, As an exponential constant, a threshold mechanism is introduced to judge the probability value of power communication line anomalies based on a preset threshold, thereby obtaining the anomaly result of the power communication line. .

[0139] In this embodiment, the bias term is set to 0.05. This value is small enough to avoid interfering with subsequent calculations. On the other hand, this value can keep the initial offset non-zero, so that the weak signal can form a traceable change trajectory during the structural quantization process. The threshold is set to 0.5, which can make the probabilistic quantity present a symmetrical state division when it is higher or lower than this value, thereby maintaining the neutral point between normal and abnormal.

[0140] Furthermore, the EWED anomaly detection model was written in Python, and the experiments were run on a Windows operating system. PyTorch was used as the framework in the CUDA 11.27 environment, and training was performed on a GeForce RTX 3090. The optimizer used was Adam, with an initial learning rate of 0.001, a training batch size of 64, and a training period of 100. The dataset consisted of 60 days of power communication line-related data, which was preprocessed and then input into the EWED anomaly detection model.

[0141] Furthermore, the error distribution plot, performance chart, and comprehensive judgment index distribution histogram of the EWED anomaly detection model are shown below. Figure 4 , 5As shown in Figure 6, Figure 4 The mean square error is concentrated near zero, the KDE curve is smooth with a single peak and is approximately symmetrical in the positive and negative directions, indicating that the model can stably characterize the time-varying law of the insulation of power communication lines. The long-term dependent structure effectively suppresses short-term noise, and the spectral structure analysis enhances the multi-scale feature extraction capability, resulting in a small overall prediction bias and demonstrating high accuracy and robustness. Figure 5 The model demonstrates the joint modeling capability of the long-term dependency module, the spectrum coding module, and the anomaly detection module. The long-term dependency module represents the temporal structure of the sequence through an inertial diffusion kernel and a global inertial memory field, keeping the probability of the normal phase in a low range. The spectrum coding module characterizes the structural features of the signal through integer grid sequence, quantization factor, and resonance cumulant, making the abnormal phase significantly different in the pseudo-spectral vector. The anomaly detection module fuses the window time field and the pseudo-spectral vector to output the anomaly probability. Therefore, as shown in the figure, the model can produce a significant jump in the degradation range and accurately identify the anomaly point when it exceeds the threshold, indicating that the method has good early detection capability and high robustness. Figure 6 The scores exhibit a clear two-segment distribution. The low-segment is concentrated in the 0.2–0.3 range, corresponding to the normal operation features extracted by the long-term dependency module through the inertial diffusion kernel and global inertial memory field, which makes the normal state form a stable cluster in the comprehensive score space. The high-segment is concentrated in the 0.6–0.75 range, which comes from the abnormal structure differences obtained by the spectrum coding module based on integer grid sequence, quantization factor and resonance accumulation, causing the degradation features to show a significant shift in the pseudo-spectral vector. The anomaly judgment module further separates the two score ranges by coupling the window time field and the pseudo-spectral vector. Therefore, this figure reflects the separability of the method for normal and degraded states, and verifies the effectiveness and discriminative ability of the model in the early detection of insulation degradation.

[0142] This embodiment also provides a storage medium on which a computer program is stored. When the program is executed by a processor, it implements the method for early anomaly detection of insulation degradation in power communication lines as proposed in the above embodiments. The storage medium can be implemented by any type of volatile or non-volatile storage device or a combination thereof, such as static random access memory, electrically erasable programmable read-only memory, erasable programmable read-only memory, programmable read-only memory, read-only memory, magnetic memory, flash memory, magnetic disk, or optical disk.

Claims

1. A method for detecting early abnormalities in insulation degradation of power communication lines, characterized in that, Includes the following steps: Collect data related to power communication lines and construct a dataset for preprocessing and partitioning. Constructing a long-term dependency module: Using a sliding window, construct power communication lines and historical power communication line sequences, construct square and linear magnitude parts for local changes, and calculate the inertial diffusion kernel; The global inertial memory field is calculated using the inertial diffusion kernel and the comprehensive characterization value. The time field within the window is obtained by combining the time position offset, the action factor, and the time gradient. Constructing a spectrum coding module: Inputting a power communication line sequence, constructing an integer sequence using a discrete integer set, and calculating an integer grid sequence; A standardized power communication line sequence is constructed based on the window time location sequence. After processing the adjacency and degree matrix calculation graph Laplace matrix, the quantization factor is output. The modulus space mapping sequence is calculated based on the time position offset, and the indicator quantity is calculated by combining it with the integer grid point sequence; the resonance cumulative quantity is calculated. Based on the calculation of local particulate coarsening index using time position offset and power communication line sequence, the trajectory increment and geometric consistency deviation index are processed, and the flexible attenuation factor is obtained by combining the basic attenuation constant. The flexible matching factor is calculated by mapping the modulus space, the integer lattice sequence and the flexible attenuation factor. The flexible resonance intensity is obtained by summing the cumulative values ​​and then quantizing the maximum value to obtain the pseudo-spectral vector. An anomaly detection module is constructed: the time field and pseudo-spectral vector within the splicing window are used as joint features, and the results of anomalies in power communication lines are judged by linear scoring, probabilization, and thresholding.

2. The method for detecting early abnormalities in insulation degradation of power communication lines according to claim 1, characterized in that, Input power communication line data, construct power communication line sequence and historical power communication line sequence using a sliding window, and obtain local change amount for each time position in the historical power communication line sequence based on the difference between the value of that position and its adjacent historical power communication line; Based on the local change, a quadratic and linear part are constructed, and a comprehensive characterization value is obtained by combining the quadratic weight and the linear weight respectively. Then, an exponential decay term is constructed based on the time interval between each time position in the historical power communication line sequence and the time position in the power communication line sequence. Combined with the linear extension term, the kernel value at the current time position and the historical time position is obtained by combining the two terms. The inertial diffusion kernel is obtained by arranging the set of kernel values ​​of all time intervals.

3. The method for detecting early abnormalities in insulation degradation of power communication lines according to claim 2, characterized in that, The inertial diffusion kernel and the comprehensive characterization value are multiplied term by term and accumulated in time order to construct the global inertial memory field. The time gradient is calculated by comparing the global inertial memory fields at adjacent time positions. The time position offset is defined according to the sliding window definition. It is used to sequentially number each time position inside the window relative to the starting time position of the window. The numbering is used to distinguish the relative position relationship of each time position inside the window. The action factor is obtained by normalizing it through the window length. Then, the action factor corresponding to the time offset is combined with the time gradient to obtain the time field inside the window.

4. The method for detecting early abnormalities in insulation degradation of power communication lines according to claim 2, characterized in that, Input a power communication line sequence, construct an integer sequence using a discrete integer set as a frequency index, and construct an integer sequence based on each frequency index. Using the time position within the window as the independent variable, calculate the time position and frequency index according to the integer transformation, and apply a rounding constraint based on the integer modulus space parameter to obtain an integer grid sequence.

5. The method for detecting early abnormalities in insulation degradation of power communication lines according to claim 3, characterized in that, The design quantification factor is used to perform translation and scale normalization on the power communication line sequences at each time position within the window to obtain standardized power communication line sequences. The similarity between the standardized power communication line sequences at any two time positions within the window is used to construct the corresponding adjacency matrix. The metric value at each time position is calculated based on this matrix to obtain the degree matrix. Then, the graph Laplacian matrix of the window is calculated through the adjacency matrix and the degree matrix. The second and third eigenvalues ​​in the spectral structure of the graph Laplacian matrix are selected, and the difference ratio between the eigenvalues ​​is extracted to obtain the intermediate index. Then, it is input into the nonlinear mapping unit for processing and range restriction processing is performed according to the upper and lower bound parameters. The quantization factor corresponding to the current window is output.

6. The method for detecting early abnormalities in insulation degradation of power communication lines according to claim 4, characterized in that, The power communication line sequence at the start time position of the window is obtained based on the time position offset. The intermediate mapping quantity is obtained by linear scaling through a quantization factor. Then, the integer modulus space parameter is used for rounding constraint to output the modulus space mapping sequence. The integer grid sequence and the modulus space mapping sequence are compared one by one at the same offset position. Based on the comparison result, the indicator is obtained. The indicator is generated sequentially with the time position offset within the window as the range. The indicator is then accumulated item by item to obtain the resonance accumulation for the frequency index k.

7. The method for detecting early abnormalities in insulation degradation of power communication lines according to claim 6, characterized in that, A flexible attenuation factor is proposed. For each time position offset within the window, multiple neighboring value pairs with different intervals are selected in the power communication line sequence. The difference between neighboring value pairs is calculated to obtain multiple scale difference quantities. The scale difference quantities are combined with the bias parameter to process the difference quantities at different scales to obtain the local fractal roughness index. The trajectory increment is calculated by differential calculation of adjacent integer grid point sequences. The amplitude relationship between adjacent trajectory increments is normalized, differentially proportionalized and offset processed, and adjusted in combination with the adjustment coefficient to obtain the geometric consistency deviation index. Then, combined with the basic attenuation constant, the local particulate roughness index and the geometric consistency deviation index are multiplicatively coupled to obtain the flexible attenuation factor.

8. The method for detecting early abnormalities in insulation degradation of power communication lines according to claim 6, characterized in that, The offset difference between the modulus space mapping sequence and the integer grid point sequence is calculated, and the amplitude scaling and exponentialization are performed in combination with the flexible attenuation factor to obtain the flexible matching factor. Using the sliding window length as the traversal range, the flexible matching factor within the window is accumulated and summed, and the accumulated resonance amount is used as a discrete benchmark to define the discrete resonance relationship, providing a baseline reference for the flexible resonance intensity and a consistent alignment structure to obtain the flexible resonance intensity. The maximum value of the flexible resonance intensity is determined, and the flexible resonance intensity is transformed relative to the maximum value of the flexible resonance intensity using the grid index as the traversal range to obtain the pseudo-spectral vector.

9. A method for detecting early abnormalities in insulation degradation of power communication lines according to claim 1, characterized in that, The time field and pseudo-spectral vector within the input window are concatenated to obtain joint features. A linear scoring function is used to obtain the power communication line anomaly score. Then, the power communication line anomaly probability value is output through a probabilistic mechanism. A threshold mechanism is introduced to judge the power communication line anomaly probability value according to a preset threshold, and the power communication line anomaly result is obtained.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by the processor, it implements the steps of the method for early abnormal detection of insulation degradation in power communication lines as described in any one of claims 1 to 9.

Images