A method for intelligent operation and maintenance risk early warning of power transmission equipment in harsh environments

Abstract

This invention belongs to the technical field of power system monitoring and relates to an intelligent operation and maintenance risk early warning method for transmission equipment in harsh environments. The method includes the following steps: synchronously collecting sensor and meteorological data; calculating a dynamic benchmark based on an association model and differentially generating state residual signals; monitoring vibration residuals and triggering a hardware interrupt when conditions are met to synchronously capture high-frequency waveforms of leakage current and temperature during fault transients; extracting feature vectors through dimensionality reduction encoding of the data and generating residual coupling fingerprints by combining them with timestamps; mapping the fingerprints to state trajectories and calculating their geometric similarity to standard templates in a fault topology library to generate targeted early warnings; and dynamically adjusting the acquisition parameters and updating the association model based on the early warning results. This invention solves the problem that existing transmission equipment monitoring methods struggle to isolate environmental interference, leading to low early warning reliability and difficulty in accurately characterizing faults.

Description

Technical Field

[0001] This invention belongs to the technical field of power system monitoring and relates to a method for intelligent operation and maintenance risk early warning of power transmission equipment in harsh environments. Background Technology

[0002] In the field of power transmission network operation and maintenance, ensuring the stable operation of lines and their auxiliary equipment under various environmental conditions is a core task. Especially in coastal, windy, humid, or industrially polluted areas, power transmission equipment such as insulators, conductors, and connecting hardware are exposed to harsh environments for a long time. Their operating status is easily affected by a combination of factors such as wind load, humidity changes, salt spray corrosion, and dirt accumulation, which can induce mechanical loosening, decreased insulation performance, and even flashover tripping.

[0003] To prevent such incidents, the industry typically employs regular manual inspections and online monitoring systems. Conventional online monitoring systems usually deploy one or more sensors on the equipment, such as those for vibration, temperature, or leakage current, and determine whether the equipment is malfunctioning by setting static alarm thresholds; that is, an alarm is generated when the sensor's measured value exceeds a preset fixed threshold.

[0004] However, the aforementioned conventional monitoring methods have technical limitations in practical applications. They treat real-time sensor readings as an absolute reflection of the device's condition, often ignoring the modulation effect of external environmental factors on sensor readings. For example, strong winds or high humidity can cause vibration amplitude or leakage current reference values ​​to increase. Static thresholds are difficult to dynamically adapt to these environmental changes, leading to invalid alarms in severe weather. Alternatively, raising thresholds to reduce false alarm rates can make early latent faults difficult to detect. Furthermore, existing technologies often lack in-depth fusion analysis and causal correlation judgment between data sources, making it difficult to qualitatively diagnose the root cause of faults. Continuous high-frequency acquisition also generates a large amount of redundant data, putting pressure on the storage, power consumption, and data transmission of edge devices. Summary of the Invention

[0005] To address the aforementioned problems, this invention provides a method for intelligent operation and maintenance risk early warning of power transmission equipment in harsh environments.

[0006] A method for intelligent operation and maintenance risk early warning of power transmission equipment in harsh environments includes the following steps:

[0007] S1. Synchronously collect multi-source heterogeneous sensor data of power transmission equipment and environmental meteorological data around the line. Calculate dynamic environmental reference signal based on pre-trained equipment status and environmental factor correlation model. Perform differential operation on multi-source heterogeneous sensor data and dynamic environmental reference signal to generate equipment status residual signal containing vibration residual component.

[0008] S2. Continuously monitor the vibration residual component in the equipment status residual signal, generate a hardware interrupt trigger command when the preset trigger condition is met, and in response to the hardware interrupt trigger command, synchronously capture the high-frequency waveform data of leakage current and temperature during the fault transient process.

[0009] S3. Perform dimensionality reduction feature extraction on the captured high-frequency waveform data and vibration residual components to extract feature vectors characterizing electrical, mechanical and thermodynamic properties, and encapsulate them with shared absolute timestamps to generate residual coupling fingerprints.

[0010] S4. Map the residual coupled fingerprint to a multi-dimensional feature space to form a device state trajectory, calculate the geometric similarity between the device state trajectory and the standard state trajectory template in the preset fault causal topology library, and generate targeted early warning instructions including fault types accordingly.

[0011] S5. Based on the fault type and confidence level in the targeted early warning command, dynamically adjust the trigger judgment parameter or high-frequency acquisition window duration in step S2, and use the data that caused the early warning to update the parameters of the equipment status and environmental factor correlation model in step S1.

[0012] A further aspect of the present invention, step S1, includes the following steps:

[0013] The edge computing unit synchronously acquires mechanical vibration signals, insulator leakage current signals, and distributed fiber optic temperature measurement signals as multi-source heterogeneous sensing data, and simultaneously reads wind speed, humidity, and salt density as environmental meteorological data.

[0014] The acquired wind speed, humidity and salt density are input into the equipment status and environmental factor correlation model to calculate and output a dynamic environmental reference signal that includes reference vibration amplitude, reference leakage current value and reference temperature value.

[0015] The real-time measured values ​​of mechanical vibration signal, insulator leakage current signal and distributed optical fiber temperature measurement signal are respectively subjected to real-time differential operation with the corresponding reference vibration amplitude, reference leakage current value and reference temperature value to obtain the equipment status residual signal containing vibration residual component, leakage current residual component and temperature residual component.

[0016] A further aspect of the present invention, step S2, includes the following steps:

[0017] The vibration residual component is defined as the main trigger source of the early warning logic, and the energy index or amplitude index of the vibration residual component is calculated within the sliding time window.

[0018] The calculated energy index or amplitude index is compared with the dynamic noise threshold. When the dynamic noise threshold is exceeded, it is determined to be an abnormal mechanical impact event, and a level transition signal is output as a hardware interrupt trigger instruction.

[0019] Upon receiving a hardware interrupt trigger command, the leakage current and temperature acquisition controller configured in slave mode immediately wakes up from the low-power polling state and starts a high-frequency burst acquisition window of preset duration to record the raw waveform data of leakage current and temperature data.

[0020] A further aspect of the present invention, step S3, includes the following steps:

[0021] Wavelet packet decomposition is performed on the high-frequency waveform data of leakage current to calculate the energy distribution ratio of each characteristic frequency band, and the sample entropy value of the waveform data is calculated and combined to form the current energy fingerprint vector.

[0022] Fast Fourier transform is performed on the vibration residual components within the high-frequency burst acquisition window to extract the dominant frequency, harmonic component frequencies, and energy spectral density, thus forming a vibration feature vector.

[0023] The temperature data sequence within the high-frequency burst acquisition window is linearly fitted to obtain the temperature gradient change rate. The current energy fingerprint vector, vibration feature vector, and temperature gradient change rate are bound to the timestamp when the hardware interrupt trigger instruction is generated to form a residual coupled fingerprint.

[0024] A further aspect of the present invention, step S4, includes the following steps:

[0025] The feature vectors in the residual coupled fingerprint are analyzed, and the feature points are connected in time order in the multidimensional feature space to construct the device state trajectory.

[0026] Traverse the fault cause-effect topology library and calculate the Hausdorff distance between the point set of the device state trajectory and the point set of each standard state trajectory template.

[0027] Select the minimum Hausdorff distance. When it is less than the preset confidence threshold, the match is considered successful. The fault type represented by the corresponding standard state trajectory template and the confidence level calculated based on the distance are encapsulated into a targeted early warning instruction.

[0028] A further aspect of the present invention, step S5, includes the following steps:

[0029] Analyze the targeted early warning command; if the fault type is an early latent fault and the confidence level is within the preset range, generate strategy adjustment parameters.

[0030] The edge computing unit adjusts its parameters according to the strategy, reducing the value of the corresponding dynamic noise threshold and increasing the duration of the high-frequency burst acquisition window;

[0031] The cloud server counts the number of consecutive targeted early warning commands of the same type in the same area. If the number exceeds the threshold, the historical environmental meteorological data that caused the early warning and the residual signal of equipment status are extracted as incremental training samples to update the weight coefficients of environmental factors in the correlation model between equipment status and environmental factors.

[0032] A further aspect of the present invention is that the dynamic noise threshold is determined by: obtaining an energy index sequence of vibration residual components over a historical period, calculating the moving average and moving standard deviation of the sequence, and using the sum of the moving average and the moving standard deviation by a preset multiple as the current dynamic noise threshold.

[0033] A further aspect of this invention is that the high-frequency burst acquisition window employs a circular buffer backtracking mechanism, specifically including: maintaining a circular buffer inside the acquisition controller; automatically freezing and backtracking historical data of a preset length before the interruption time in the buffer when a hardware interruption trigger instruction is received; and continuously acquiring evolution data after the interruption time to jointly constitute complete high-frequency waveform data.

[0034] A further aspect of this invention involves calculating the sample entropy value as follows: setting the embedding dimension m and the similarity tolerance r, counting the number of matches between vector templates of length m whose distance is less than the similarity tolerance r, and counting the number of matches between vector templates of length m+1 whose distance is less than the similarity tolerance r, and calculating the negative natural logarithm of the ratio of the two to obtain the sample entropy value.

[0035] A further aspect of this invention involves introducing a weight anchor point constraint mechanism during the online update of weight coefficients. Specifically, when adjusting weight coefficients using the online gradient descent algorithm, the absolute difference between the updated weight coefficients and the initial weight coefficients of the model is constrained to not exceed a preset maximum allowable offset.

[0036] In summary, the present invention has the following beneficial technical effects:

[0037] 1. By constructing a dynamic reference signal that is correlated with the environmental state in real time and generating a device state residual signal containing vibration residual components, differential processing of real-time sensor measurements is achieved. This technique utilizes a pre-trained correlation model to quantify and remove background fluctuations caused by environmental changes such as wind speed, humidity, and salt density, enabling the generated residual signal to more accurately reflect the state changes of the equipment itself. This processing method reduces the probability of false alarms caused by environmental fluctuations and is also beneficial for detecting early, weak fault characteristics under strong background noise, thus improving the accuracy of early warning.

[0038] 2. A fault transient process synchronous acquisition mechanism based on master-slave hardware interlocking is adopted. This mechanism keeps the system in a low-power polling mode under normal conditions. Only when the vibration residual signal, which serves as the primary trigger source, becomes abnormal will a hardware interrupt instruction wake up the slave acquisition controller and initiate high-frequency burst acquisition. This event-driven acquisition method ensures that complete waveform data of multiple physical fields, including electrical, mechanical, and thermal fields, can be captured synchronously within the critical time window of the fault occurrence. While preserving key transient information, it reduces the system's average power consumption and data transmission load, alleviating the resource consumption problem caused by continuous high-frequency acquisition.

[0039] 3. By performing dimensionality reduction encoding on synchronously captured data to generate multidimensional residual coupled fingerprints, and performing geometric matching between state trajectories and fault topology templates, lightweight feature extraction of complex raw waveform data is achieved. By analyzing the real-time state trajectories formed by fingerprint sequences in a multidimensional feature space and calculating their geometric similarity to standard templates representing specific fault evolution patterns, this invention can perform pattern recognition based on the temporal logic of fault development and the coupling relationship of multiple physics fields. This helps to qualitatively identify specific fault types, realizing the transformation from single numerical alarms to targeted early warnings, and providing data support for subsequent operation and maintenance decisions.

[0040] 4. A closed-loop feedback and dynamic adjustment strategy based on early warning results was established. The system can automatically generate strategy adjustment parameters and distribute them to edge computing units based on the fault type and confidence level of the early warning. This allows for dynamic lowering of the trigger threshold and adjustment of data acquisition duration for early-stage latent faults. Simultaneously, this mechanism uses continuously occurring early warning data of the same type as samples to iteratively update the environmental factor correlation model. This adaptive adjustment mechanism enhances the system's adaptability to changes in operating conditions in specific areas, helping to maintain the long-term effectiveness and sensitivity of the early warning system. Attached Figure Description

[0041] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the accompanying drawings used in the description of the embodiments or the prior art will be briefly introduced below. The drawings are used to provide a further understanding of the present invention.

[0042] Figure 1 A flowchart illustrating an embodiment of this application is disclosed.

[0043] Figure 2 Structural schematic diagrams of embodiments of this application are disclosed. Detailed Implementation

[0044] The following is in conjunction with the appendix Figure 1 - Figure 2 A preferred description of the present invention is provided below.

[0045] See attached document Figure 1 This invention proposes a method for intelligent operation and maintenance risk early warning of power transmission equipment in harsh environments, comprising the following steps:

[0046] S1. Synchronously collect multi-source heterogeneous sensor data of power transmission equipment and environmental meteorological data around the line. Calculate dynamic environmental reference signal based on pre-trained equipment status and environmental factor correlation model. Perform differential operation on multi-source heterogeneous sensor data and dynamic environmental reference signal to generate equipment status residual signal containing vibration residual component.

[0047] S2. Continuously monitor the vibration residual component in the equipment status residual signal, generate a hardware interrupt trigger command when the preset trigger condition is met, and in response to the hardware interrupt trigger command, synchronously capture the high-frequency waveform data of leakage current and temperature during the fault transient process.

[0048] S3. Perform dimensionality reduction feature extraction on the captured high-frequency waveform data and vibration residual components to extract feature vectors characterizing electrical, mechanical and thermodynamic properties, and encapsulate them with shared absolute timestamps to generate residual coupling fingerprints.

[0049] S4. Map the residual coupled fingerprint to a multi-dimensional feature space to form a device state trajectory, calculate the geometric similarity between the device state trajectory and the standard state trajectory template in the preset fault causal topology library, and generate targeted early warning instructions including fault types accordingly.

[0050] S5. Based on the fault type and confidence level in the targeted early warning command, dynamically adjust the trigger judgment parameter or high-frequency acquisition window duration in step S2, and use the data that caused the early warning to update the parameters of the equipment status and environmental factor correlation model in step S1.

[0051] In one embodiment of the present invention, step S1 includes the following steps:

[0052] The edge computing unit synchronously acquires mechanical vibration signals, insulator leakage current signals, and distributed fiber optic temperature measurement signals as multi-source heterogeneous sensing data, and simultaneously reads wind speed, humidity, and salt density as environmental meteorological data.

[0053] The acquired wind speed, humidity and salt density are input into the equipment status and environmental factor correlation model to calculate and output a dynamic environmental reference signal that includes reference vibration amplitude, reference leakage current value and reference temperature value.

[0054] The real-time measured values ​​of mechanical vibration signal, insulator leakage current signal and distributed optical fiber temperature measurement signal are respectively subjected to real-time differential operation with the corresponding reference vibration amplitude, reference leakage current value and reference temperature value to obtain the equipment status residual signal containing vibration residual component, leakage current residual component and temperature residual component.

[0055] Specifically, the edge computing unit utilizes its integrated multi-channel analog-to-digital converter interface to simultaneously acquire raw electrical signals output from vibration acceleration sensors, insulator leakage current sensors, and distributed fiber optic temperature measurement systems deployed on power transmission equipment. The multi-source heterogeneous sensor data mentioned here refers to measurement data streams acquired from sensors with different physical principles, and whose data formats and sampling rates may vary. In this embodiment, this specifically manifests as mechanical vibration signals, insulator leakage current signals, and distributed fiber optic temperature measurement signals.

[0056] The real-time operating system embedded within the edge computing unit uses a preset period to meet the sampling frequency requirements of key signals. It samples and quantizes the aforementioned raw electrical signals, generating timestamp-aligned vibration waveform sequences, leakage current value sequences, and temperature value sequences. The preset period, for example, is 10 ms. This 10 ms (i.e., 100 Hz) is based on the coverage requirements of typical mechanical vibration frequencies of transmission lines. Since the aerobatic vibration frequency of transmission lines is typically between 5 Hz and 100 Hz, and the galloping frequency is between 0.1 Hz and 3 Hz, according to the Nyquist sampling theorem, a 10 ms sampling period is sufficient to recover signals in these frequency bands without distortion. The edge computing unit maintains a high-precision global clock with an error controlled within ±10 μs to ensure strict alignment of the vibration, current, and temperature data on the time axis.

[0057] Simultaneously, the edge computing unit can also synchronously read real-time environmental meteorological data from weather stations deployed around the transmission line via its serial communication interface. This data includes wind speed scalars in meters per second, relative humidity scalars in percentage units, and equivalent salt density scalars in milligrams per square centimeter. The real-time environmental meteorological data referred to here are physical quantity measurements directly related to the meteorological conditions at the location of the power transmission equipment operation and that change over time; in this step, these specifically include wind speed, relative humidity, and equivalent salt density.

[0058] The edge computing unit then calls its locally stored device state-environment factor correlation model, trained based on historical data, and feeds real-time environmental meteorological data as input feature vectors into the model. The device state-environment factor correlation model is a mathematical model pre-trained using machine learning methods. Its function is to establish a mapping relationship between environmental variables and sensor readings under normal device conditions. Before deployment, this model has been trained offline using sensor data marked as "normal" in a historical database and its corresponding meteorological data as a training set. In this embodiment, the device state-environment factor correlation model is implemented using a random forest regression algorithm. Based on the input wind speed, humidity, and salt density, it calculates and outputs a set of dynamic environmental reference signals that match the current environmental conditions and characterize the expected sensor readings of the device under fault-free conditions. The dynamic environmental reference signal here refers to a time-varying sequence of reference values ​​calculated by the correlation model based on real-time environmental input, used to characterize the expected "healthy" readings of the device sensors under the current specific environmental conditions. This dynamic environmental reference signal includes a reference vibration amplitude corresponding to the vibration waveform sequence, a reference leakage current value corresponding to the leakage current value sequence, and a reference temperature value corresponding to the temperature value sequence.

[0059] Subsequently, the execution thread of the edge computing unit performs real-time differential operations on the instantaneous amplitude of the vibration waveform sequence, the instantaneous value of the leakage current sequence, and the instantaneous value of the temperature sequence, respectively, with the reference values ​​of the corresponding physical quantities in the dynamic environmental reference signal. This differential operation is performed at each data sampling timestamp, that is, subtracting the corresponding dynamic reference value from the measured instantaneous value, thereby generating a new set of time series data, which is defined as the equipment state residual signal. The so-called equipment state residual signal refers to the signal sequence obtained after filtering out the environmental reference component through differential operation, whose numerical changes are theoretically caused only by abnormalities or faults in the equipment's own state; the equipment state residual signal specifically includes vibration residual components, leakage current residual components, and temperature residual components.

[0060] It is worth noting that the data synchronization cycle of the edge computing unit can be set to 10 ms, which is based on the Nyquist sampling theorem for the common mechanical fault frequency of power transmission equipment (typically below 100 Hz). The typical range of input environmental parameters for the correlation model used to calculate the dynamic environmental baseline is: wind speed 0 to 30 m / s, humidity 10% to 100%, and salt density 0.01 to 0.3 mg / cm³. 2 The sliding time window used for the difference operation is set to a preset duration sufficient to smooth short-term fluctuations and calculate meaningful residual trends, such as 60 seconds.

[0061] It should be noted that the random forest model was configured with the following key hyperparameters before deployment to balance prediction accuracy and edge computing overhead: the number of decision trees was set to 50 to 100, and the maximum depth was limited to 10 to 15 layers. The input feature vectors were standardized, including wind speed (normalized to 0-1), relative humidity (normalized to 0-1), and isosalt density (normalized after logarithmic transformation). The model's training set consisted of historical monitoring data from the past 12 months for this route, covering seasonal variations. The generation logic for the "equipment status residual signal" is as follows:

[0062]

[0063] in, for The measured value of the time sensor. for The environment vector at any given time. The nonlinear mapping function constructed for the trained correlation model is called the model operator. The reason for setting the sliding time window length for the difference operation to 60 seconds is that changes in environmental parameters (such as humidity and wind speed) typically have minute-level inertia, while sudden equipment failures (such as hardware breakage) are sub-second-level. A 60-second window can smooth out random white noise from the sensor while also accurately capturing residual shifts caused by changes in the equipment's state, which last longer than the inertia of environmental changes.

[0064] For example, suppose the edge computing unit synchronously acquires a set of sensor data at a certain moment: the instantaneous value of the vibration acceleration sensor is 0.15 g, the instantaneous value of the leakage current sensor is 2.1 mA, and the instantaneous value of the fiber optic temperature measuring point is 35.6℃. Simultaneously, the real-time environmental meteorological data read are: wind speed 8 m / s, humidity 85%, and equivalent salt density 0.1 mg / cm³. 2At this point, wind speed (8), humidity (85%), and salt density (0.1%) are used as input vectors, and the equipment state-environmental factor correlation model is invoked. Assuming the model learns from historical data, under the current environment, the normal vibration amplitude baseline is 0.12 g, the normal leakage current baseline is 1.8 mA, and the normal temperature baseline is 34.5℃. Therefore, the model outputs the dynamic environmental baseline signal as follows: baseline vibration amplitude 0.12 g, baseline leakage current value 1.8 mA, and baseline temperature value 34.5℃. Next, differential calculations are performed: the vibration residual component is 0.15 g minus 0.12 g, equaling 0.03 g; the leakage current residual component is 2.1 mA minus 1.8 mA, equaling 0.3 mA; and the temperature residual component is 35.6℃ minus 34.5℃, equaling 1.1℃. This generates the equipment state residual signal corresponding to this moment, which includes the vibration residual component 0.03 g, the leakage current residual component 0.3 mA, and the temperature residual component 1.1℃. Store this set of residual data along with the timestamp into a cache queue for subsequent processing steps.

[0065] In one embodiment of the present invention, step S2 includes the following steps:

[0066] The vibration residual component is defined as the main trigger source of the early warning logic, and the energy index or amplitude index of the vibration residual component is calculated within the sliding time window.

[0067] The calculated energy index or amplitude index is compared with the dynamic noise threshold. When the dynamic noise threshold is exceeded, it is determined to be an abnormal mechanical impact event, and a level transition signal is output as a hardware interrupt trigger instruction.

[0068] Upon receiving a hardware interrupt trigger command, the leakage current and temperature acquisition controller configured in slave mode immediately wakes up from the low-power polling state and starts a high-frequency burst acquisition window of preset duration to record the raw waveform data of leakage current and temperature data.

[0069] Specifically, during the system initialization phase, the data stream path for processing the vibration residual component of the mechanical vibration signal can be logically set as the main trigger source via firmware configuration of the edge computing unit. The main trigger source is the signal source selected to first detect anomalies and trigger subsequent collaborative acquisition logic; in this embodiment, it is the mechanical vibration signal processing channel. Correspondingly, the enable pins of the acquisition controller processing the leakage current signal and the acquisition controller processing the distributed fiber optic temperature measurement signal are configured in slave mode. Slave mode here is a hardware or firmware operating state where the device's data acquisition behavior is not autonomous and periodic, but rather activated by an external interrupt signal, i.e., controlled by the interrupt output of the main trigger source.

[0070] The main monitoring thread within the edge computing unit continuously reads the vibration residual component sequence from the device status residual signal, using a preset time window, such as 100 ms. For the data within each time window, the main monitoring thread calculates its root mean square value as an energy indicator. The maximum absolute value of the vibration residual component within the window is recorded as the amplitude index. Maintain dynamic noise threshold. This threshold is a time-adaptive numerical threshold designed to distinguish between background environmental fluctuations and genuine abnormal events. It is set based on the statistical characteristics of recent historical residual data, using the mean plus three standard deviations to cover 99.7% of normally distributed background noise; signals exceeding this range are considered abnormal. This threshold value is based on a preset historical time period, such as the vibration residual energy over the past 60 seconds. The historical sequence is dynamically updated by calculating the sum of its mean and the standard deviation of a preset multiple, such as three times.

[0071] It should be noted that, The specific calculation logic is not simple statistics, but rather a dynamic weighting based on a sliding window. The specific formula is:

[0072]

[0073] in, It's the past The moving average of the vibrational residual energy over a period of 60 seconds. This is the corresponding moving standard deviation. (Coefficient) Set to a floating-point number between 3 and 4.5. Parameter Selection criteria: Under the assumption of a conventional normal distribution, A value of 3 can cover 99.7% of the background noise. However, considering the non-faulty gusts of wind in the field, this embodiment preferably uses a lower value to reduce the false alarm rate. The value is set to 3.5. This means that only when the instantaneous energy of the vibration residual exceeds 3.5 times the standard deviation of the background noise baseline will it be considered a significant anomaly.

[0074] At the end of each time window, the main monitoring thread will display the currently calculated energy index. or amplitude index With dynamic noise threshold Compare. Once or More than one of them Upon receiving the signal, the edge computing unit immediately determines that an abnormal mechanical impact event has occurred. Subsequently, the edge computing unit generates and sends a hardware interrupt trigger command, transitioning from a low to a high level, through a specific pin of its general-purpose input / output interface. This hardware interrupt trigger command is essentially a digital electrical signal used to achieve microsecond-level response synchronization control at the hardware level, thereby ensuring that the slave device can be woken up instantly. The leakage current and temperature acquisition controller configured in slave mode has its enable pin connected to this interrupt pin via circuitry. When the level transition of the hardware interrupt trigger command is captured by the slave-mode acquisition controller, its internal state machine is immediately woken up from a low-power polling state.

[0075] Upon wake-up, the controller forcibly interrupts its regular low-frequency sampling task and starts an internal timer. This timer defines a period that is a high-frequency burst acquisition window of preset duration. During this window, the sensor operates at a sampling rate much higher than in normal mode, aiming to capture detailed waveforms of fault transients. For example, the window duration... The preset time is 200ms. Within this window, the leakage current acquisition controller switches the sampling rate of its analog-to-digital converter from a normal low sampling rate to a preset high sampling rate, such as switching from 1 Hz to 1 kHz, continuously acquiring and buffering the raw waveform data of the leakage current. Simultaneously, the acquisition controller of the distributed fiber optic temperature measurement system increases its demodulation scanning frequency to the highest rate, densely acquiring temperature data from the measurement points. Through this mechanism, the system synchronously captures data up to and including the time from the moment the hardware interrupt trigger instruction is generated. Complete high-frequency waveform data of leakage current and high-frequency temperature sampling sequence associated with abnormal mechanical impact events within a time window. Window duration. The setting of 200 ms is a typical value based on statistics of the duration of typical mechanical impact events on power transmission equipment and the resulting electrical and thermal responses, which is usually sufficient to cover most transient processes.

[0076] It should be noted that, to prevent the loss of the first few milliseconds of data due to wake-up delay—the "dead zone" problem—the slave controller maintains a circular buffer. In low-power mode, data is written only at a low frequency (e.g., 1 Hz), but the hardware retains the most recent 500 ms of raw sampling points. When an interrupt signal is received (rising edge triggered, response delay <50 μs), the controller not only immediately switches to 1 kHz high-frequency sampling but also automatically freezes and rewinds the circular buffer to the time before the interrupt. (For example, 50 ms) of data. Therefore, the final generated waveform data includes 50 ms of precursor data before the fault and 150 ms of evolution data after the fault, for a total length of 200 ms. This is important for analyzing whether the fault is sudden (such as a lightning strike, with an extremely steep waveform leading edge) or gradual (such as a flashover, with a preceding leakage current rise).

[0077] For example, based on the aforementioned example, the edge computing unit continuously monitors the vibration residual component sequence. Assume that within a 100 ms time window, the vibration residual component values ​​read are 0.03 g, 0.028 g, 0.032 g, etc. Calculate the root mean square value of the data within this time window. The value is 0.03 g. Simultaneously, query the currently dynamically updated dynamic noise threshold value. It is 0.02 g. Due to current energy indicators... (0.03g) greater than (0.02 g) Immediately, an abnormal mechanical impact event is detected. Subsequently, the control pin output level of the edge computing unit changes, generating a hardware interrupt trigger instruction. Upon receiving this interrupt signal, the leakage current acquisition controller in slave mode is immediately awakened. It interrupts its normal task of sampling once per second, starts its internal timer, and begins a continuous sampling period. A high-frequency burst acquisition window of 200 ms is defined. Within this window, the controller acquires the leakage current at a high-speed sampling rate of 1 kHz. Assuming that at the interrupt trigger time t0, the original residual component of the leakage current is 0.3 mA, under high-speed acquisition, a pulse waveform with an amplitude of 15 mA may be captured 2 ms after t0, and the complete decaying oscillation waveform will be recorded in the subsequent window. Simultaneously, the temperature acquisition controller is also synchronously awakened and records temperature changes in high-frequency mode, possibly capturing temperature gradient data that increases by 0.5 °C within tens of milliseconds, starting from 35.6 °C. These timing data captured within the high-frequency burst acquisition window are temporarily buffered, awaiting further processing.

[0078] In one embodiment of the present invention, step S3 includes the following steps:

[0079] Wavelet packet decomposition is performed on the high-frequency waveform data of leakage current to calculate the energy distribution ratio of each characteristic frequency band, and the sample entropy value of the waveform data is calculated and combined to form the current energy fingerprint vector.

[0080] Fast Fourier transform is performed on the vibration residual components within the high-frequency burst acquisition window to extract the dominant frequency, harmonic component frequencies, and energy spectral density, thus forming a vibration feature vector.

[0081] The temperature data sequence within the high-frequency burst acquisition window is linearly fitted to obtain the temperature gradient change rate. The current energy fingerprint vector, vibration feature vector, and temperature gradient change rate are bound to the timestamp when the hardware interrupt trigger instruction is generated to form a residual coupled fingerprint.

[0082] Specifically, the data processing thread of the edge computing unit performs dimensionality reduction feature extraction on the complete high-frequency waveform data of the leakage current acquired within the high-frequency burst acquisition window in step S2. First, a wavelet packet decomposition algorithm is applied to the waveform data. Wavelet packet decomposition is a time-frequency analysis method that simultaneously decomposes a signal layer by layer in both high and low frequency bands, suitable for capturing the detailed features of transient impact signals. A suitable wavelet basis function for capturing transient signal features is selected, such as the 'db4' wavelet basis function, for complete decomposition at a preset number of layers, for example, 4 layers. The 'db4' (Daubechies 4) wavelet basis is chosen because its asymmetry is highly similar to the waveform features of partial discharge and arc pulses in transmission lines. The specific technical basis for setting the decomposition layer to 4 layers is: at a sampling rate of 1 kHz, the Nyquist frequency is 500 Hz. After 4 layers of decomposition, the signal is divided into 16 sub-bands, each with a bandwidth of approximately 31.25 Hz.

[0083] Low-frequency nodes (0-62.5 Hz): contain power frequency (50 Hz) components. Changes in their energy proportion reflect fluctuations in the fundamental amplitude of the leakage current, corresponding to a decrease in insulation resistance.

[0084] Intermediate frequency nodes (62.5-187.5 Hz): typically correspond to the higher harmonics of the power supply.

[0085] High-frequency nodes (above 187.5 Hz): Abnormal energy increases in this frequency band are usually a characteristic indicator of localized arcing or corona discharge on the surface of insulators.

[0086] Therefore, eigenvectors It's not just about numbers; it directly corresponds to different physical failure mechanisms.

[0087] Next, the sum of squares of the coefficients for each sub-band is calculated to obtain the energy value of each band. Then, the ratio of energy in each sub-band to the total energy is calculated. The formula is: This generates a vector consisting of 16 energy distribution ratio values. Energy distribution ratio It characterizes the concentration of signal energy in different frequency bands, and the energy distribution patterns under normal and fault conditions are usually different.

[0088] Simultaneously, the sample entropy of the complete high-frequency waveform data of the leakage current is calculated; this metric is used to quantify the time series complexity. A preset embedding dimension *m* and similarity tolerance *r* are set, for example, *m* is 2 and *r* is 0.2 times the signal standard deviation. These are commonly used parameter settings for processing such engineering signals, used to distinguish between random noise and deterministic chaotic modes. The length is calculated... In the sequence, the distance between template vectors is less than The sample entropy value can be obtained by taking the logarithmic difference of the probabilities. Finally, the multidimensional energy distribution ratio vector is compared with the sample entropy value. They are spliced ​​together to form a predetermined dimension, such as a 17-dimensional current energy fingerprint vector.

[0089] For the vibration residual component in the equipment status residual signal, within the time period corresponding to the high-frequency burst acquisition window, a data segment of a preset duration, centered on the hardware interrupt trigger instruction generation time, is selected, for example, 200 ms. After applying a Hanning window to this data segment, a Fast Fourier Transform with a predetermined number of points is performed, for example, 2048 points, to meet the frequency resolution requirements. The Hanning window is used to reduce spectral leakage. From this spectrum, the frequency component with the largest amplitude is extracted as the dominant frequency. The dominant frequency component in a vibration signal is usually associated with the fault characteristic frequency; identifying the presence of significant spectral peaks at integer multiples of these frequencies is crucial for determining harmonic components. It is a frequency component that is an integer multiple of the main frequency, and its presence often indicates nonlinear or impulsive characteristics; and the signal energy spectral density integral value is calculated in a preset key frequency band, such as the 0 to 100 Hz band. It is used to quantify the total vibration energy within a specific frequency band. The 0-100 Hz band is defined based on the fact that the aerobatic vibrations of transmission lines are mainly concentrated in the 5-100 Hz range, while severe icing galloping frequencies are typically <3 Hz. This is achieved by calculating the integrated energy within this frequency band. and clock speed It can be distinguished as follows:

[0090] High frequency low amplitude ( ): Normal light breeze vibration;

[0091] Low frequency high amplitude ( Dangerous wire dancing;

[0092] Full-band noise (no obvious) Impact response caused by external impact or hardware breakage.

[0093] This generates a vibrational characteristic vector that includes the dominant frequency, harmonic components, and energy spectral density values.

[0094] In addition, the slope of the linear fit of the high-frequency temperature sampling sequence synchronously captured within the high-frequency burst acquisition window is calculated as the rate of change of the temperature gradient. It is obtained by linearly fitting the temperature sequence within the time window using the least squares method, and is used to characterize the rate of temperature rise.

[0095] Finally, the edge computing unit combines the current energy fingerprint vector, vibration feature vector, and temperature gradient change rate. They are bound to a shared absolute timestamp that they belong to, which is determined by the interruption time in step S2, and encapsulated into a structured data object. This object is defined as a lightweight multidimensional residual coupled fingerprint and uploaded through the communication module.

[0096] The shared absolute timestamp originates from the system clock that generates the hardware interrupt trigger instruction in step S2, ensuring strict synchronization of electrical, mechanical, and thermal characteristics on the timeline. The lightweight multidimensional residual coupled fingerprint is a data structure that encapsulates multidimensional features and timestamps. Compared to the original high-frequency waveform data, its data volume is significantly reduced, making it suitable for low-bandwidth transmission.

[0097] The formula for calculating sample entropy involved in this step is as follows:

[0098]

[0099] in, In terms of tolerance Below, the length is The number of matches between vector templates, In terms of tolerance Below, the length is The number of matches between vector templates. This represents the length of the time series.

[0100] For example, complete high-frequency waveform data of leakage current (including a 15 mA pulse and oscillation decay process) captured within a 200 ms window is processed. First, a 4-layer 'db4' wavelet packet decomposition is performed, assuming that the calculated sub-band energy ratio representing 125-250Hz is obtained. The value is abnormally high at 0.35, while the ratios of other sub-bands are relatively low. Simultaneously, the sample entropy of this waveform data was calculated. The value is 0.8. This forms the current energy fingerprint vector, whose key components can be represented as [...]. =0.35, ..., =0.8]. For the vibration residual signal segment within the same time window (based on the original 0.03 g residual and its subsequent changes), after performing FFT, it is assumed that the dominant frequency is extracted. The frequency is 85 Hz, and its second harmonic peak exists at 170 Hz. The integral value of the energy spectral density in the 0-100 Hz frequency band is... 0.05 g 2 / Hz. This forms the vibration characteristic vector as [...]. =85 Hz, =170 Hz, =0.05 g 2 / Hz]. For the synchronously captured temperature sequence (increasing from 35.6℃), calculate the rate of change of the temperature gradient within the 200 ms window. The rate of change is 2.5℃ / s. Finally, the above current energy fingerprint vector, vibration feature vector, and temperature gradient change rate of 2.5℃ / s are bound to the absolute timestamp of the interruption moment and encapsulated into a multidimensional residual coupled fingerprint to complete the data upload preparation.

[0101] In one embodiment of the present invention, step S4 includes the following steps:

[0102] The feature vectors in the residual coupled fingerprint are analyzed, and the feature points are connected in time order in the multidimensional feature space to construct the device state trajectory.

[0103] Traverse the fault cause-effect topology library and calculate the Hausdorff distance between the point set of the device state trajectory and the point set of each standard state trajectory template.

[0104] Select the minimum Hausdorff distance. When it is less than the preset confidence threshold, the match is considered successful. The fault type represented by the corresponding standard state trajectory template and the confidence level calculated based on the distance are encapsulated into a targeted early warning instruction.

[0105] Specifically, during initialization, the cloud-based early warning server has a pre-installed and loaded fault causal topology library in its internal storage. This library is a pre-built knowledge base where each fault mode is defined not only with static characteristics but also with typical paths or sequences of feature evolution over time. The library stores multiple predefined typical fault modes in a data structure, each represented as a standard state trajectory template in a multi-dimensional feature space composed of electrical, mechanical, and thermal features. For example, the "insulator flashover" template is defined as a trajectory consisting of multiple feature points arranged in chronological order: the starting point is an increase in humidity residual; subsequent points represent a continuous increase in the energy of the leakage current residual in the low-frequency band; and the trajectory endpoint represents a sudden change in the high-frequency pulse current residual and the local temperature residual.

[0106] The multidimensional feature space is an abstract mathematical space whose coordinate axes correspond to the various feature dimensions extracted from the multidimensional residual coupled fingerprint, such as the low-frequency energy ratio of leakage current, the dominant vibration frequency, and the temperature gradient. The standard state trajectory template is a pre-stored path in the library, compiled by engineers based on historical fault case analysis, physical simulation, or domain knowledge, and entered into the system as a sequence of coordinate points before deployment.

[0107] The cloud-based early warning server continuously monitors and receives multidimensional residual coupled fingerprint data packets uploaded from various edge computing units via a network interface. For each received data packet, the server first parses out the encapsulated current energy fingerprint vector, vibration feature vector, temperature gradient rate of change, and shared absolute timestamp. Then, the server connects the parsed feature points in time order, plotting them on the horizontal axis and the component values ​​of the aforementioned feature vectors on the vertical axis, thereby mapping a real-time device status trajectory composed of discrete points in the multidimensional feature space. The real-time device status trajectory is the path depicted in this multidimensional space by mapping the feature values ​​from continuously uploaded fingerprint data packets in their chronological order.

[0108] Subsequently, the server traverses the fault cause-effect topology library, sequentially reading each standard state trajectory template. For each template, the server calculates the Hausdorff distance between the real-time device state trajectory and the template to quantify the similarity. Hausdorff distance is a mathematical method used to measure the similarity between two sets of points; it captures the differences between two trajectories in overall shape and spatial location. The Euclidean distance used in the calculation... This refers to the straight-line distance between two feature points in a multidimensional feature space. The calculation process is as follows: First, calculate the Euclidean distance from each point on the real-time trajectory to all points on the template trajectory. Take the minimum distance for each point, and then take the maximum value among these minimum distances, denoted as . Similarly, calculate the maximum value among the minimum distances from each point on the template trajectory to all points on the real-time trajectory, denoted as . The final distance of Hausdorf Take the larger of the two values.

[0109] After calculating the distances to all templates in the library, the server finds the minimum Hausdorff distance. and its corresponding standard state trajectory template. The server will Compared with the preset confidence threshold Comparison, confidence threshold This is a preset similarity threshold, typically ranging from 0.7 to 0.9 (after normalization). The specific value needs to be determined statistically based on the distribution of correct and incorrect matches in historical test data. It is used to control the strictness of the warning. It should be noted that... The value is not arbitrarily set, but determined based on ROC (Receiving Controller Operating Characteristic) curve analysis. During system development, a test set containing 1000 historical samples (500 normal and 500 of various faults) can be replayed.

[0110] When the threshold is set to 0.9, the fault detection rate (TPR) can reach 98%, but the false alarm rate (FPR) also rises to 15%; when the threshold is set to 0.7, the false alarm rate drops to 1%, but some early, minor faults are missed. Therefore, this embodiment adopts a tiered threshold strategy:

[0111] High confidence warning ( When the Hausdorff distance normalized similarity is >0.75, a red alert is triggered and pushed to the operations and maintenance personnel.

[0112] Low confidence level concern ( The system does not directly trigger an alarm, but instead enters a closed-loop observation mode, i.e., step S5, to actively verify by adjusting parameters.

[0113] In one embodiment of the present invention, if Less than or equal to If the real-time device status trajectory is sufficiently similar to the standard status trajectory template, the server will determine that the current device status is sufficiently similar to the fault type represented by the standard status trajectory template with the highest matching degree, and generate a structured targeted early warning instruction. This instruction includes a structured message containing fault qualitative, location, and quantitative confidence scores to drive subsequent closed-loop operation and maintenance actions. This instruction includes at least the specific fault type name matched, the device location identifier reported by the edge computing unit, and the calculated Hausdorff distance. This confidence level value is then sent to the operations and maintenance terminal via a message queue or API call.

[0114] It should be noted that, to make the solution more feasible, the fault causal topology library is actually stored as a multi-dimensional time series matrix in the cloud database. Taking insulator flashover as an example, it contains 5 key time steps ( to eigenvectors of )

[0115] (Precursor stage): [Vibration=0, leakage current low frequency ratio=0.1, temperature gradient=0], characterizing a high humidity environment.

[0116] (Development stage): [Vibration=0, leakage current low frequency ratio=0.4, temperature gradient=0.1], indicating that the dirt layer is damp and the resistance decreases.

[0117] (Active Period): [Vibration=0, leakage current high frequency ratio=0.6, sample entropy=1.5], indicating the presence of local dry arcing and chaotic waveform.

[0118] (Flashover period): [Vibration = Impact, Leakage current full frequency band = Max, Temperature gradient = 5.0], indicating local arc series connection and a sudden temperature rise.

[0119] The real-time state trajectory is a line formed by connecting a series of fingerprints uploaded from the edge in a multi-dimensional space. The Hausdorff distance is the maximum mismatch between this "measured line" and the aforementioned "standard line".

[0120] The Hausdorff distance calculation formula involved in this step is as follows:

[0121]

[0122] in, A set of points representing the real-time trajectory of device status. A set of points representing a standard state trajectory template. Representative point With point The Euclidean distance between them. The supremum represents the maximum value. The infimum represents the minimum value.

[0123] For example, the cloud-based early warning server receives a multidimensional residual coupled fingerprint and parses its characteristics: the energy ratio of the 125-250 Hz sub-band in the current energy fingerprint vector. It is 0.35. It is 0.8; in the vibration characteristic vector 85Hz 170 Hz 0.05g 2 / Hz; Rate of change of temperature gradient The temperature is 2.5℃ / s; the timestamp is t1. The server maps the current feature value to a point P1 in multidimensional space. Assume that the previous timestamp t0 (from an earlier fingerprint) has already been mapped to point P0. Connecting P0 and P1 forms a real-time device state trajectory. The server retrieves an "insulator flashover" template from the fault causal topology library. This template consists of points Q0 (features: high humidity, low current energy), Q1 (features: continuously increasing low-frequency current energy), and Q2 (features: high-frequency current energy, temperature abrupt change). The server calculates the Hausdorff distance between the real-time trajectory point set {P0, P1} and the template trajectory point set {Q0, Q1, Q2}. First, it calculates the Euclidean distance from P0 to {Q0, Q1, Q2}, assuming the minimum distance is to... The value is 0.2; calculate the Euclidean distance from P1 to {Q0, Q1, Q2}, assuming the minimum distance is to Q2, with a value of 0.15; then Next, calculate the distance from Q0 to {P0, P1}, assuming the minimum value is 0.5; the minimum distance from Q1 to {P0, P1} is 0.2; and the minimum distance from Q2 to {P0, P1} is 0.15. Ultimately, Hausdorf was far from... Assuming the system has a preset confidence threshold. Since the calculated value of 0.5 is less than 0.6, the server determines that the match is successful. The server associates the current state with the "insulator flashover" fault type and generates a targeted early warning instruction, with the following content: Fault type = "insulator flashover", Location = "XX line #32 tower A phase", Confidence level = 0.5.

[0124] In one embodiment of the present invention, step S5 includes the following steps:

[0125] Analyze the targeted early warning command; if the fault type is an early latent fault and the confidence level is within the preset range, generate strategy adjustment parameters.

[0126] The edge computing unit adjusts its parameters according to the strategy, reducing the value of the corresponding dynamic noise threshold and increasing the duration of the high-frequency burst acquisition window;

[0127] The cloud server counts the number of consecutive targeted early warning commands of the same type in the same area. If the number exceeds the threshold, the historical environmental meteorological data that caused the early warning and the residual signal of equipment status are extracted as incremental training samples to update the weight coefficients of environmental factors in the correlation model between equipment status and environmental factors.

[0128] Specifically, after the cloud-based early warning server generates and outputs a targeted early warning command, its internal policy management module starts simultaneously. This module first parses the targeted early warning command, extracting the fault type string and confidence level value. Next, the module queries a pre-set policy mapping table, which defines the adjustment policy codes and specific parameters corresponding to different fault types and confidence level ranges. Based on the query results, the cloud-based early warning server generates a set of structured policy adjustment parameters and sends them to the source edge computing unit that generated the early warning via a secure network link (such as the MQTT protocol). The policy adjustment parameters are a set of instruction data generated and sent from the cloud to guide edge devices in changing their operating parameters. These parameters include the specific adjustment object (e.g., which sensor's threshold), the adjustment operation (e.g., multiplication, increment), and the adjustment amount (e.g., coefficient α, increment). ).

[0129] After receiving the policy instruction, the edge computing unit parses and executes adjustments. If the policy adjustment parameters indicate that the current warning corresponds to an early latent fault, which is a qualitative classification referring to warning events with low feature visibility, moderate confidence levels, but exhibiting a specific fault pattern, the determination is usually based on keywords in the fault type name combined with the confidence level range, such as "early" or "minor." For example, if the fault type is "early contamination accumulation in insulators" and the confidence level is in the moderate range, such as 0.4-0.7, the edge computing unit performs the following operations: for sensors directly related to the fault, such as leakage current sensors, the corresponding dynamic noise threshold value... The original value will be multiplied by an attenuation factor less than 1. (e.g., 0.8), thereby reducing the trigger sensitivity; where the attenuation coefficient is... This is used to linearly reduce the dynamic noise threshold, with a typical value range of 0.5 to 0.9. The specific value is set according to the latency of the fault; a smaller value indicates a more aggressive increase in sensitivity. Simultaneously, the duration of the high-frequency burst acquisition window is also adjusted. Increase by one increment from the default 200 ms (For example, 100 ms) to extend the acquisition time for weak transient signals. The window duration increment... This is used to extend the high-frequency acquisition time, with a typical value range of 50 ms to 300 ms, to ensure more complete coverage of the fault transient process. The specific logic for parameter adjustment is as follows:

[0130] Attenuation coefficient This means reducing the dynamic noise threshold by 20%. 0.8 was chosen instead of 0.1 to prevent oversensitivity from causing the system to be continuously triggered by light breezes and deplete the battery. The lower limit of this coefficient is hardcoded to 0.5 by the system, effectively doubling the maximum sensitivity.

[0131] Window Increment Increasing the sampling duration is to cover a more complete "electric-thermal-mechanical" coupling process. For example, after an insulator experiences arc discharge (electrical signal), there is a physical delay of tens of milliseconds in heat conduction to the fiber optic sensor (thermal signal). Extending the sampling time by 100 ms effectively ensures the capture of the delayed temperature change characteristics, thereby improving the confidence of multidimensional feature matching.

[0132] If the cloud-based early warning server receives a preset number of targeted early warning instructions of the same type from the same power transmission equipment location identifier within a preset statistical period, such as 24 hours, such as two or more times, it will automatically raise the risk level identifier R of the equipment location from the initial value (such as R=1) by one level (such as increasing it to R=2). Specifically, the risk level identifier R is an integer variable used to quantify the degree of risk accumulation at a specific equipment location. The initial value is 1, and it is incremented by 1 for each consecutive occurrence of the same type of early warning. This identifier can be used for priority ranking and allocation of operation and maintenance resources.

[0133] Meanwhile, the model management component of the cloud-based early warning server adaptively optimizes the device status-environmental factor correlation model applied to the region based on these continuous early warning samples. This process involves incrementally training the existing machine learning model using new early warning sample data. The optimization process includes: using newly collected environmental and residual data that led to the early warning as the incremental training set, and employing online gradient descent to fine-tune the weight coefficients of environmental factors in the correlation model, such as humidity and salt density, making the model more sensitive to the environmental conditions that cause this type of failure, thereby generating a more accurate dynamic environmental baseline signal.

[0134] It should be noted that online gradient descent is one of the commonly used algorithms for implementing this incremental update, and will not be described in detail in this embodiment.

[0135] Through the above mechanism, the acquisition sensitivity and capture capability of the edge computing unit can be enhanced in a targeted manner, while the baseline model of the cloud early warning server can be continuously optimized. Together, they form a closed-loop system that dynamically adjusts based on actual operation and maintenance feedback. The closed-loop system describes the complete data flow and control flow loop from early warning generation to policy issuance, parameter adjustment, and model optimization. Its goal is to enable the system performance to continuously improve with the accumulation of operation and maintenance experience.

[0136] The simplified formula for online updating of weight coefficients involved in this step is as follows:

[0137]

[0138] in, This represents the old weighting coefficient of a certain environmental factor in the equipment status-environmental factor correlation model. This represents the updated weighting coefficients. The learning rate is a preset small positive number that controls the step size of each update. It is usually set to a small fixed value to avoid oscillations, such as 0.001-0.1, with a typical value of 0.01. Represents the loss function Regarding weight The gradient of the loss function measures the difference between the dynamic baseline predicted by the model based on the current weights and the actual residual signal that led to the warning.

[0139] It should be noted that although the formula provides an update mechanism, in practical applications, to prevent catastrophic drift of model weights due to sensor malfunctions, this embodiment may introduce weight anchor point constraints. That is: .in, These are the initial weights when the model is manufactured. This is the maximum allowable offset (e.g., 0.2). This means that no matter how the online learning adjusts, the weight of the humidity factor on the leakage current will not become negative or exceed the range allowed by physical laws, thus ensuring the physical interpretability and safety of the system.

[0140] For example, the cloud-based early warning server generates a targeted early warning command with a fault type of "insulator flashover" and a confidence level of 0.5. After parsing, the policy management module determines, according to the mapping table, that this is a medium-confidence "insulator flashover" early warning, corresponding to an early latent fault policy. The server generates policy adjustment parameters: {Target sensor: "leakage current", Operation: "threshold multiplication", Coefficient} 0.8; Operation: "Add Window", Increment :100 ms}. This command is sent to the edge computing unit located at "XX line #32 tower A phase". After receiving it, the unit sets the dynamic noise threshold value of its original leakage current vibration residual channel. Multiplying by 0.8 yields a new, more sensitive threshold value. Simultaneously, the duration of the high-frequency burst acquisition window is adjusted. The timeout period is set from 200 ms to 300 ms. Assume that within the next 24 hours, the same edge computing unit reports another warning matching "insulator flashover." The cloud-based warning server detects consecutive warnings of the same type and automatically raises the risk level indicator R for that location from 1 to 2. Simultaneously, the server uses the environmental data (high humidity, medium salt density) and observed abnormal residual features corresponding to these two warning events as incremental samples to initiate optimization of the equipment status-environmental factor correlation model applied to this area. Assume the weight of the humidity factor in the original model is... =0.5, using the learning rate =0.01, the gradient value is obtained based on the gradient calculation. If the value is 0.3, then the new weight is calculated according to the formula. Through multiple iterations, the model's representation of the impact of humidity was fine-tuned, and the dynamic environmental benchmark signal generated in similar high-humidity environments in the future will be more accurate, thus potentially highlighting abnormal residuals earlier.

[0141] See appendix Figure 2This invention also proposes an intelligent operation and maintenance risk early warning system for power transmission equipment in harsh environments, comprising the following modules:

[0142] The data synchronization and residual generation module is used to synchronously collect multi-source heterogeneous sensor data of power transmission equipment and environmental meteorological data around the line. Based on the pre-trained equipment status and environmental factor correlation model, it calculates dynamic environmental reference signal and performs differential operation on multi-source heterogeneous sensor data and dynamic environmental reference signal to generate equipment status residual signal containing vibration residual component.

[0143] The event synchronization capture module is used to continuously monitor the vibration residual component in the equipment status residual signal. When the preset trigger conditions are met, it generates a hardware interrupt trigger command and, in response to the hardware interrupt trigger command, synchronously captures the high-frequency waveform data of leakage current and temperature during the fault transient process.

[0144] The coupled fingerprint construction module is used to perform dimensionality reduction feature extraction operations on the captured high-frequency waveform data and vibration residual components, extract feature vectors that characterize electrical, mechanical and thermodynamic properties, and encapsulate them with shared absolute timestamps to generate residual coupled fingerprints.

[0145] The trajectory matching and early warning module is used to map the residual coupling fingerprint to a multi-dimensional feature space to form the device state trajectory, calculate the geometric similarity between the device state trajectory and the standard state trajectory template in the preset fault causal topology library, and generate targeted early warning instructions including fault types accordingly.

[0146] The closed-loop strategy adaptive module dynamically adjusts the trigger judgment parameters or high-frequency acquisition window duration in the event synchronization capture module based on the fault type and confidence level in the targeted early warning command, and uses the data that caused the early warning to update the parameters of the equipment status and environmental factor correlation model in the data synchronization and residual generation module.

[0147] Each of the modules can be implemented in whole or in part through software, hardware, or a combination thereof. It supports hardware embedded in or independent of the processor in the computer device, and also supports software stored in the memory of the computer device, so that the processor can call and execute the operations corresponding to each of the above modules.

[0148] The above embodiments are only used to illustrate the technical solutions of the present invention, and are not intended to limit it. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention, and should all be included within the protection scope of the present invention.

Claims

1. A method for intelligent operation and maintenance risk early warning of power transmission equipment in harsh environments, characterized in that, Includes the following steps: S1. Synchronously collect multi-source heterogeneous sensor data of power transmission equipment and environmental meteorological data around the line. Calculate dynamic environmental reference signal based on pre-trained equipment status and environmental factor correlation model. Perform differential operation on multi-source heterogeneous sensor data and dynamic environmental reference signal to generate equipment status residual signal containing vibration residual component. S2. Continuously monitor the vibration residual component in the equipment status residual signal, generate a hardware interrupt trigger command when the preset trigger conditions are met, and respond to the hardware interrupt trigger command to synchronously capture the high-frequency waveform data of leakage current and temperature during the fault transient process; including: defining the vibration residual component as the main trigger source of the early warning logic, and calculating the energy index or amplitude index of the vibration residual component within the sliding time window; comparing the calculated energy index or amplitude index with the dynamic noise threshold, and determining an abnormal mechanical impact event when the dynamic noise threshold is exceeded, and outputting a level transition signal as a hardware interrupt trigger command; the leakage current and temperature acquisition controller configured as a slave mode wakes up from the low-power polling state immediately after receiving the hardware interrupt trigger command, and starts a high-frequency burst acquisition window of preset duration to record the original waveform data of leakage current and temperature data; S3. Perform dimensionality reduction feature extraction on the captured high-frequency waveform data and vibration residual components to extract feature vectors characterizing electrical, mechanical and thermodynamic properties, and encapsulate them with shared absolute timestamps to generate residual coupling fingerprints. S4. Map the residual coupling fingerprint to a multi-dimensional feature space to form a device state trajectory, calculate the geometric similarity between the device state trajectory and the standard state trajectory templates in the preset fault causal topology library, and generate targeted early warning instructions including fault types accordingly; including: parsing the feature vectors in the residual coupling fingerprint, connecting feature points in the multi-dimensional feature space in chronological order to construct the device state trajectory; traversing the fault causal topology library, calculating the Hausdorff distance between the point set of the device state trajectory and the point set of each standard state trajectory template; selecting the smallest Hausdorff distance, and when it is less than a preset confidence threshold, determining that the match is successful, and encapsulating the fault type represented by the corresponding standard state trajectory template and the confidence level calculated based on the distance into a targeted early warning instruction; S5. Based on the fault type and confidence level in the targeted early warning command, dynamically adjust the trigger judgment parameters or high-frequency acquisition window duration in step S2, and use the data that caused the early warning to update the parameters of the equipment status and environmental factor association model in step S1, including: parsing the targeted early warning command; if the fault type belongs to an early latent fault and the confidence level is within a preset range, then generate strategy adjustment parameters; the edge computing unit adjusts the parameters according to the strategy, reduces the value of the corresponding dynamic noise threshold, and increases the duration of the high-frequency burst acquisition window; the cloud server counts the number of consecutive targeted early warning commands of the same type in the same area; if the number exceeds the threshold, then extract the historical environmental meteorological data and equipment status residual signal that caused the early warning as incremental training samples, and update the weight coefficients of environmental factors in the equipment status and environmental factor association model.

2. The intelligent operation and maintenance risk early warning method for power transmission equipment in harsh environments according to claim 1, characterized in that, Step S1 includes the following steps: The edge computing unit synchronously acquires mechanical vibration signals, insulator leakage current signals, and distributed fiber optic temperature measurement signals as multi-source heterogeneous sensing data, and simultaneously reads wind speed, humidity, and salt density as environmental meteorological data. The acquired wind speed, humidity and salt density are input into the equipment status and environmental factor correlation model to calculate and output a dynamic environmental reference signal that includes reference vibration amplitude, reference leakage current value and reference temperature value. The real-time measured values ​​of mechanical vibration signal, insulator leakage current signal and distributed optical fiber temperature measurement signal are respectively subjected to real-time differential operation with the corresponding reference vibration amplitude, reference leakage current value and reference temperature value to obtain the equipment status residual signal containing vibration residual component, leakage current residual component and temperature residual component.

3. The method for intelligent operation and maintenance risk early warning of power transmission equipment in harsh environments according to claim 1, characterized in that, Step S3 includes the following steps: Wavelet packet decomposition is performed on the high-frequency waveform data of leakage current to calculate the energy distribution ratio of each characteristic frequency band, and the sample entropy value of the waveform data is calculated and combined to form the current energy fingerprint vector. Fast Fourier transform is performed on the vibration residual components within the high-frequency burst acquisition window to extract the dominant frequency, harmonic component frequencies, and energy spectral density, thus forming a vibration feature vector. The temperature data sequence within the high-frequency burst acquisition window is linearly fitted to obtain the temperature gradient change rate. The current energy fingerprint vector, vibration feature vector, and temperature gradient change rate are bound to the timestamp when the hardware interrupt trigger instruction is generated to form a residual coupled fingerprint.

4. The intelligent operation and maintenance risk early warning method for power transmission equipment in harsh environments according to claim 1, characterized in that, The dynamic noise threshold is determined as follows: obtain the energy index sequence of the vibration residual components over a historical period, calculate the moving average and moving standard deviation of the sequence, and use the sum of the moving average and the moving standard deviation by a preset multiple as the current dynamic noise threshold.

5. The intelligent operation and maintenance risk early warning method for power transmission equipment in harsh environments according to claim 1, characterized in that, The high-frequency burst acquisition window adopts a circular buffer backtracking mechanism, which includes: the acquisition controller maintains a circular buffer internally, and when a hardware interrupt trigger command is received, it automatically freezes and backtracks the historical data of a preset length before the interrupt time in the buffer, and continuously acquires the evolution data after the interrupt time, which together constitute complete high-frequency waveform data.

6. The method for intelligent operation and maintenance risk early warning of power transmission equipment in harsh environments according to claim 3, characterized in that, The sample entropy is calculated as follows: set the embedding dimension m and the similarity tolerance r, count the number of matches between vector templates of length m whose distance is less than the similarity tolerance r, and the number of matches between vector templates of length m+1 whose distance is less than the similarity tolerance r, and calculate the negative natural logarithm of the ratio of the two to obtain the sample entropy.

7. The intelligent operation and maintenance risk early warning method for power transmission equipment in harsh environments according to claim 1, characterized in that, During the online update of weight coefficients, a weight anchor constraint mechanism is introduced, which specifically includes: when adjusting weight coefficients using the online gradient descent algorithm, the absolute difference between the updated weight coefficients and the initial weight coefficients of the model is constrained to not exceed the preset maximum allowable offset.

Images