A method and system for real-time monitoring of dynamic response of mechanical characteristics of high-voltage power equipment

Abstract

The present application relates to the high-voltage power technical field, especially relates to a kind of high-voltage power equipment mechanical characteristic dynamic response real-time monitoring method and system, comprising: obtaining standardization row procedure sequence and coil current sequence, according to standardization row procedure sequence and coil current sequence construct multi-scale time sequence characteristic tensor, extract depth response mode vector and point-by-point reconstruction residual sequence by lightweight long short-term memory network, according to depth response mode vector and point-by-point reconstruction residual sequence obtain mechanical characteristic health degree comprehensive score, compare mechanical characteristic health degree comprehensive score with adaptive early warning threshold line, judge whether to trigger early trend early warning or current abnormal alarm, the present application can realize the high sensitivity detection of early weak fault, and track equipment normal aging by adaptive adjustment early warning threshold, by analyzing health degree score trend, early warning potential progressive degradation, improve adaptability and monitoring reliability.

Description

Technical Field

[0001] This invention relates to the field of high-voltage power technology, and in particular to a method and system for real-time monitoring of the dynamic response of the mechanical characteristics of high-voltage power equipment. Background Technology

[0002] High-voltage power equipment (such as circuit breakers) is a critical safety component in the power system. The reliability of its mechanical characteristics directly affects the success of opening and closing operations, thereby affecting the stability and safety of the power grid. For online monitoring of the mechanical characteristics of high-voltage equipment, the industry generally adopts a method based on the comparison of preset feature thresholds. This method compares the collected signals and features with preset global features to determine whether an alarm should be triggered.

[0003] The mechanical characteristics of high-voltage equipment constitute a continuous and dynamic physical process. Its health status is not only reflected in a few discrete global feature points, but also contained in the dynamic response details of the entire operation process. Early mechanical faults (such as slight jamming of the mechanism, insufficient lubrication, and micro-wear of components) often first manifest as subtle distortions in local time segments during the dynamic response process. These subtle changes may not yet be clearly reflected in the global feature values, leading to missed detections. At the same time, the slow performance drift caused by normal wear and environmental temperature changes during long-term operation can easily lead to false alarms due to the rigidity of fixed thresholds. Therefore, monitoring systems that rely on preset feature thresholds are insufficient in terms of sensitivity for early fault detection and adaptability to normal aging of equipment. Summary of the Invention

[0004] The main objective of this invention is to provide a method and system for real-time monitoring of the dynamic response of the mechanical characteristics of high-voltage power equipment, aiming to solve the technical problems mentioned in the background art.

[0005] This invention proposes a method for real-time monitoring of the dynamic response of the mechanical characteristics of high-voltage power equipment, comprising the following steps:

[0006] Acquire the stroke-time curve signal and the opening / closing coil current signal of a single opening or closing response of high-voltage equipment, and obtain a standardized stroke sequence and coil current sequence based on the stroke-time curve signal and the opening / closing coil current signal;

[0007] A multi-scale temporal feature tensor is constructed based on the standardized travel sequence and the coil current sequence;

[0008] The multi-scale temporal feature tensor is input into a lightweight long short-term memory network. The lightweight long short-term memory network performs temporal encoding and temporal reconstruction processing on the multi-scale temporal feature tensor to obtain a deep response pattern vector and a point-by-point reconstruction residual sequence.

[0009] A comprehensive score for mechanical property health is obtained based on the deep response pattern vector and the point-by-point reconstructed residual sequence.

[0010] Obtain the adaptive warning threshold, compare the comprehensive score of mechanical characteristic health with the adaptive warning threshold, and determine whether an early trend warning or a current abnormal alarm is triggered.

[0011] Preferably, the step of constructing a multi-scale time-series feature tensor based on the standardized travel sequence and the coil current sequence includes:

[0012] Based on the standardized travel sequence, the instantaneous motion velocity sequence is calculated using the center difference method; based on the coil current sequence, the coil current change rate sequence is calculated using the center difference method.

[0013] The transient coupling strength sequence is obtained based on the instantaneous motion velocity sequence and the coil current change rate sequence;

[0014] Based on the standardized travel sequence, the instantaneous motion velocity sequence, and the transient coupling strength sequence, a first channel feature vector is constructed in chronological order;

[0015] The second channel feature vector is obtained by calculating the first-order backward difference of the first channel feature vector along the time axis;

[0016] The first channel feature vector and the second channel feature vector are stacked along the time dimension to construct a multi-scale temporal feature tensor.

[0017] Preferably, the step of inputting the multi-scale temporal feature tensor into a lightweight long short-term memory network, and performing temporal encoding and temporal reconstruction processing on the multi-scale temporal feature tensor through the lightweight long short-term memory network to obtain a deep response pattern vector and a pointwise reconstructed residual sequence includes:

[0018] The lightweight long short-term memory network calculates the state flow contribution value of the first channel feature vector at each time step in the multi-scale temporal feature tensor to the input gate, forget gate and output gate.

[0019] The lightweight long short-term memory network calculates the trend flow contribution value of the second channel feature vector at each time step in the multi-scale temporal feature tensor to the input gate, forget gate and output gate.

[0020] By fusing the state flow contribution value and the trend flow contribution value, the input gate activation value, the forget gate activation value, and the output gate activation value are generated;

[0021] Obtain the dynamic response long-term state vector of the previous time step, and update the dynamic response long-term state vector of the current time step according to the forget gate activation value, the input gate activation value and the dynamic response long-term state vector of the previous time step to obtain the updated dynamic response long-term state vector.

[0022] The dynamic response memory vector of the current time step is obtained based on the output gate activation value and the updated dynamic response long-term state vector of the current time step.

[0023] The time-step iterative calculation is performed, and the dynamic response memory vector of the final time step in the multi-scale temporal feature tensor is used as the deep response pattern vector to complete the temporal coding process.

[0024] Synchronously, based on the dynamic response memory vector of each time step in the multi-scale temporal feature tensor, temporal reconstruction processing is performed through the temporal autoencoder of the lightweight long short-term memory network to obtain the reconstructed feature vector. The point-by-point reconstructed residual sequence is obtained based on the mean square error between the reconstructed feature vector and the first channel feature vector.

[0025] Preferably, the step of obtaining the comprehensive score of mechanical characteristic health based on the deep response pattern vector and the point-by-point reconstructed residual sequence includes:

[0026] Establish a health response pattern library for high-voltage equipment, which stores historical health status deep response pattern vectors.

[0027] The deviation of the response pattern is obtained based on the deep response pattern vector and the health response pattern library;

[0028] Calculate the reconstruction anomaly index based on the point-by-point reconstruction residual sequence;

[0029] Based on the pattern deviation and the reconstruction anomaly index, the comprehensive score of the mechanical characteristic health is obtained through adaptive weighted fusion.

[0030] Preferably, the step of obtaining the adaptive warning threshold and comparing the comprehensive score of mechanical characteristic health with the adaptive warning threshold to determine whether an early trend warning or a current abnormal alarm is triggered includes:

[0031] A comprehensive health score of the historical mechanical characteristics of a healthy high-voltage device is obtained from several responses to form a health score set, and the mean and standard deviation of the health score set are calculated.

[0032] Using the set of health scores as a dynamic health score sequence, a comprehensive score of the mechanical characteristic health of the high-voltage equipment for each normal response during monitoring is obtained and added to the dynamic health score sequence. Starting from the mean and standard deviation, the estimated mean and estimated standard deviation of the dynamic health score sequence are updated using an exponential weighted moving average algorithm.

[0033] The cumulative number of responses from the high-voltage equipment is obtained, a threshold relaxation factor is obtained based on the cumulative number of responses, and an adaptive early warning threshold is obtained based on the threshold relaxation factor, the estimated mean, and the estimated standard deviation.

[0034] The overall score of the mechanical health characteristics of this response is compared with the adaptive early warning threshold.

[0035] If the comprehensive score of the mechanical characteristic health is not greater than the adaptive early warning threshold, then it is further determined whether an early trend warning is triggered.

[0036] If the overall health score of the mechanical characteristics is greater than the adaptive early warning threshold, then the current abnormal alarm is triggered.

[0037] Preferably, if the comprehensive score of the mechanical characteristic health is not greater than the adaptive warning threshold, the step of further determining whether to trigger an early trend warning includes:

[0038] Obtain a comprehensive health score of the mechanical characteristics based on a number of responses within a preset period to construct a health score sequence;

[0039] Based on the health score sequence, the least squares method is used to fit its linear trend information within a preset sliding time window, and the slope of the health deterioration trend is calculated.

[0040] Determine whether the slope of the health deterioration trend exceeds a preset trend sensitivity threshold for a consecutive preset number of times. If it does, trigger an early trend warning.

[0041] This invention also provides a real-time monitoring system for the dynamic response of mechanical characteristics of high-voltage power equipment, comprising:

[0042] The information acquisition module is used to acquire the stroke-time curve signal and the opening / closing coil current signal of a single opening or closing response of high-voltage equipment, and to acquire a standardized stroke sequence and coil current sequence based on the stroke-time curve signal and the opening / closing coil current signal;

[0043] The feature construction module is used to construct a multi-scale time-series feature tensor based on the standardized travel sequence and the coil current sequence;

[0044] The inference module is used to input the multi-scale temporal feature tensor into a lightweight long short-term memory network, and perform temporal encoding and temporal reconstruction processing on the multi-scale temporal feature tensor through the lightweight long short-term memory network to obtain a deep response pattern vector and a pointwise reconstructed residual sequence.

[0045] Anomaly scoring module is used to obtain a comprehensive score of mechanical characteristic health based on the deep response pattern vector and the point-by-point reconstruction residual sequence;

[0046] The judgment and decision module is used to obtain the adaptive warning threshold, compare the comprehensive score of mechanical characteristic health with the adaptive warning threshold, and determine whether to trigger an early trend warning or a current abnormal alarm.

[0047] Preferably, the anomaly scoring module includes:

[0048] A health mode library construction unit is used to establish a health response mode library for high-voltage equipment, wherein the health response mode library stores historical health status deep response mode vectors.

[0049] The pattern deviation calculation unit is used to obtain the response pattern deviation based on the deep response pattern vector and the healthy response pattern library;

[0050] A reconstruction anomaly index calculation unit is used to calculate the reconstruction anomaly index based on the point-by-point reconstruction residual sequence.

[0051] An adaptive scoring fusion unit is used to obtain a comprehensive score of mechanical characteristic health by adaptive weighted fusion based on the mode deviation degree and the reconstruction anomaly index.

[0052] The present invention also provides a computer device, including a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to implement the steps of a method for real-time monitoring of the dynamic response of the mechanical characteristics of high-voltage power equipment.

[0053] The present invention also provides a computer-readable storage medium having a computer program stored thereon, wherein the computer program, when executed by a processor, implements the steps of a method for real-time monitoring of the dynamic response of the mechanical characteristics of high-voltage power equipment.

[0054] The beneficial effects of this invention are as follows: This invention extracts deep response pattern vectors and reconstructs residual sequences point by point from the dynamic response process of the mechanical characteristics of high-voltage equipment through a lightweight long short-term memory network, thereby achieving high-sensitivity detection of early minor faults. Furthermore, it tracks the normal aging of equipment by adaptively adjusting the early warning threshold and predicts potential progressive degradation by analyzing the health score trend, thus improving adaptability and monitoring reliability. Attached Figure Description

[0055] Figure 1 This is a schematic diagram of a method flow according to an embodiment of this application.

[0056] Figure 2 This is a schematic diagram of the system structure according to an embodiment of this application.

[0057] The objectives, features, and advantages of this invention will be further explained in conjunction with the embodiments and with reference to the accompanying drawings. Detailed Implementation

[0058] It should be understood that the specific embodiments described herein are merely illustrative of the invention and are not intended to limit the invention.

[0059] like Figure 1 As shown, this application provides a method for real-time monitoring of the dynamic response of the mechanical characteristics of high-voltage power equipment, including the following steps:

[0060] S1. Obtain the original timing signal of a single opening or closing response of the high-voltage equipment. The original timing signal includes the stroke-time curve signal collected by the linear displacement sensor and the opening / closing coil current signal collected by the coil current transformer. Based on the stroke-time curve signal and the opening / closing coil current signal, perform signal alignment and fixed-frequency resampling based on the response start point to obtain a standardized stroke sequence and coil current sequence with strictly synchronized time axis and fixed length.

[0061] S2, construct a multi-scale time-series feature tensor based on the standardized travel sequence and the coil current sequence;

[0062] S3, input the multi-scale temporal feature tensor into a lightweight long short-term memory network (lightweight LSTM), and perform temporal encoding and temporal reconstruction processing on the multi-scale temporal feature tensor through the lightweight long short-term memory network to obtain a deep response pattern vector and a point-by-point reconstruction residual sequence;

[0063] S4. Obtain a comprehensive score for mechanical property health based on the deep response pattern vector and the point-by-point reconstruction residual sequence;

[0064] S5, obtain the adaptive warning threshold, compare the comprehensive score of mechanical characteristic health with the adaptive warning threshold, and determine whether an early trend warning or a current abnormal alarm is triggered.

[0065] As described in steps S1-S5 above, constructing a multi-scale temporal feature tensor enables the creation of a high-information-density dynamic process representation. Traditional methods directly extract global statistics such as travel time and average velocity, losing local temporal information of the process. This invention, while preserving the complete time series structure, integrates tensors with more dimensions of physical information, providing ample data support for achieving high-sensitivity detection. Lightweight long short-term memory (LSTM) network processing encodes the multi-scale feature tensor of the entire dynamic process into a fixed-dimensional deep response pattern vector. This deep response pattern vector captures global response pattern features. Simultaneously, the lightweight LSM network reconstructs the input features of each time step from its internal state. By comparing the reconstructed values ​​with the true values, a point-by-point reconstruction residual sequence is obtained. The deep response pattern vector is used to evaluate the health of the overall response pattern, while the point-by-point reconstruction residual sequence is sensitive to anomalies in local time segments. The combination of these two methods constitutes a dual anomaly detection capability covering both global and local anomalies. By calculating a comprehensive health score for mechanical characteristics, a unified quantitative health index can be generated. This index integrates information from two aspects: overall pattern deviation (based on deep response pattern vector) and local anomaly clustering (based on point-by-point reconstruction of residual sequences). It provides a quantifiable and comparable indicator, enabling a more objective measurement of the health level of a single operation's dynamic response. Adaptive threshold warnings enable dynamically evolving intelligent decision-making. Unlike fixed thresholds, adaptive warning thresholds are dynamically calculated based on the equipment's recent historical health scores. They automatically adjust to the slow performance drift caused by normal aging and environmental changes. When the equipment performance is in a new steady state of slow degradation but still within the normal range, the threshold is widened accordingly, effectively suppressing false alarms caused by normal drift. An alarm is only triggered when the score deviates significantly from the equipment's current normal fluctuation range. This solves the problems of insufficient sensitivity in detecting early minor faults and poor adaptability to normal performance drift in high-voltage equipment mechanical characteristic monitoring.

[0066] In one embodiment, the step of constructing a multi-scale time-series feature tensor based on the standardized travel sequence and the coil current sequence includes:

[0067] S21, Based on the standardized travel sequence, calculate the instantaneous motion velocity sequence using the center difference method; based on the coil current sequence, calculate the coil current change rate sequence using the center difference method.

[0068] S22, Obtain the transient coupling strength sequence based on the instantaneous motion velocity sequence and the coil current change rate sequence;

[0069] Formula for calculating transient coupling strength:

[0070]

[0071] In the formula, Indicates time step The transient coupling strength (used to quantify the matching relationship between the change in electromagnetic driving force and the change in mechanical motion acceleration at each time step). Indicates time step instantaneous velocity, It represents the absolute maximum value of instantaneous velocity (derived from an instantaneous motion velocity sequence). Indicates at time step The rate of change of coil current, This represents the absolute maximum value of the rate of change of the coil current (derived from the coil current rate of change sequence). Symbol function (used to preserve information about the direction of current change);

[0072] The transient coupling strengths of all time steps are normalized to the interval [-1, 1] after passing through the maximum and minimum values, resulting in a transient coupling strength sequence (the purpose is to eliminate the amplitude influence caused by inherent parameter differences between different operations and devices, so that the features focus on the relative shape and change law of the waveform, and improve the generalization ability of the features). Formula:

[0073]

[0074] In the formula, Indicates time step The normalized value of transient coupling strength, This represents the minimum transient coupling strength. This represents the maximum value of the transient coupling strength;

[0075] S23, based on the standardized travel sequence, the instantaneous motion velocity sequence, and the transient coupling strength sequence, construct a first channel feature vector in chronological order, wherein the first channel feature vector is represented at each time step as follows: ,in For time step The travel value (normalized value, derived from the standardized travel sequence). For time step The instantaneous velocity (normalized value). For time step The normalized value of transient coupling strength;

[0076] S24, by calculating the first-order backward difference of the first channel feature vector along the time axis, the second channel feature vector is obtained. The second channel feature vector is represented at each time step as follows: ,in, These are the first-order difference eigenvalues ​​of the travel value. The first-order difference eigenvalue of instantaneous motion velocity. The first-order difference eigenvalue of the transient coupling strength;

[0077] S25, stack the first channel feature vector and the second channel feature vector along the time dimension to construct a multi-scale temporal feature tensor.

[0078] As described in steps S21-S25 above, dynamic features can be extracted by calculating the instantaneous motion velocity sequence and the coil current change rate sequence. The instantaneous motion velocity and the coil current change rate describe the rate of change of state and are directly related to the dynamic characteristics of the system. The velocity reflects the inertia of mechanical motion, and the current change rate reflects the rate of establishment and disappearance of electromagnetic force. Early mechanical faults (such as increased friction and jamming) often first affect the acceleration of the mechanism or the smoothness of motion. This information is more directly contained in the velocity sequence and its changes, rather than in the displacement sequence. Therefore, extracting the instantaneous motion velocity sequence and the coil current change rate sequence elevates the analysis dimension from state to state change, enhancing the feature's ability to capture dynamic anomalies. By acquiring transient coupling strength sequences, the dynamic coordination between electromechanical systems can be quantified. This aims to create features that directly characterize the instantaneous matching relationship between changes in electromagnetic driving force and mechanical motion response. According to electromagnetic principles, the electromagnetic force of the operating mechanism is approximately proportional to the square of the current, and its trend is related to the rate of change of the current. The acceleration of mechanical motion is related to the derivative of the velocity. Therefore, the transient coupling strength at each time step constitutes a scalar reflecting the combined effect of motion and driving force changes. Using the transient coupling strength sequence, the instantaneous coordination between electromagnetic energy and mechanical energy during the conversion process can be directly monitored. Many early faults (such as changes in mechanism clearance or local lubrication failure) may have a slight impact on individual stroke or current curves, but they will disrupt the original precise time delay coordination or force-motion relationship. This disruption will manifest as local distortion, shifts in characteristic points (such as zero crossings or extreme points), or abnormal fluctuations on the waveform of the transient coupling strength sequence. It is more sensitive to early coordination faults, which cannot be provided by traditional static features or single signal analysis.

[0079] By constructing the first-channel feature vector, multi-physical-dimensional state information can be integrated, encompassing information on mechanical displacement, mechanical velocity, and electromechanical coupling strength. This allows for a comprehensive physical state description at each time step. Constructing the second-channel feature vector introduces state change trend information, reflecting the changes in each physical quantity between adjacent time steps—the change trend or instantaneous rate of change. This is more helpful in detecting dynamic anomalies. For example, a segment with normal velocity but near-zero acceleration may indicate motion obstruction; a sudden increase in coupling strength may indicate an instantaneous breakdown of coordination. The combination of the first and second-channel feature vectors provides dual information. Constructing a multi-scale temporal feature tensor forms a structured model input, using a [time step count, feature count, channel count] structure. This fully preserves the time-series information throughout the entire operation cycle and provides a 2x3 feature matrix at each time point. This provides a foundation for lightweight long short-term memory networks to learn and distinguish healthy dynamic patterns from various early abnormal dynamic patterns, thereby improving early fault detection capabilities from a data perspective.

[0080] In one embodiment, the step of inputting the multi-scale temporal feature tensor into a lightweight long short-term memory network, and performing temporal encoding and temporal reconstruction processing on the multi-scale temporal feature tensor through the lightweight long short-term memory network to obtain a deep response pattern vector and a pointwise reconstructed residual sequence includes:

[0081] S31, the lightweight long short-term memory network calculates its state flow contribution value to the input gate, forget gate and output gate based on the first channel feature vector of each time step in the multi-scale temporal feature tensor;

[0082] Formula for the state flow contribution of the input gate:

[0083]

[0084] In the formula, This represents the state flow contribution value of the input gate. Indicates the weight of the trip group. Indicates time step The travel value, Indicates the grouping weights for motion speed. Indicates time step instantaneous velocity, The grouping weights represent the transient coupling strength. Indicates time step The normalized value of transient coupling strength, The dynamic response memory weight matrix represents the state flow. Indicates time step The dynamic response memory vector, This represents the dot product operation. The bias term represents the state flow. This represents the Sigmoid activation function;

[0085] The state flow contribution values ​​of the forget gate and the output gate adopt the same grouping and weighting form as the state flow contribution values ​​of the input gate, with only the weight parameters and bias parameters being different.

[0086] S32, the lightweight long short-term memory network calculates the trend flow contribution value of the second channel feature vector of each time step in the multi-scale temporal feature tensor to the input gate, forget gate and output gate.

[0087] Formula for the trend flow contribution value of the input gate:

[0088]

[0089] In the formula, This represents the trend flow contribution value of the input gate. Indicates the grouping weights of the rate of change of travel. express The first-order difference eigenvalue of the travel value at the time step (representing the rate of displacement change). The weights represent the grouping of instantaneous velocity change rates. Indicates time step The first-order difference eigenvalue of the instantaneous velocity (representing the instantaneous rate of change of velocity). The weights represent the grouping weights based on the rate of change of coupling strength. Indicates time step The first-order difference eigenvalue of the transient coupling strength (representing the rate of change of the transient coupling strength). The dynamic response memory weight matrix represents the trend flow. Indicates time step The dynamic response memory vector, This represents the dot product operation. The bias term representing the trend flow;

[0090] The trend flow contribution values ​​of the forget gate and the output gate adopt the same grouped weighting form as the trend flow contribution values ​​of the input gate, with only the weight parameters and bias parameters being different.

[0091] S33, merge the state flow contribution value and the trend flow contribution value to generate the input gate activation value, forget gate activation value, and output gate activation value, using the following formula (the forget gate activation value and output gate activation value are fused using the same method as the input gate):

[0092]

[0093] In the formula, Indicates time step The input gate activation value, This represents the fusion weight coefficient (initial value set to 0.5, optimized through backpropagation during training). This represents the state flow contribution value of the input gate. This represents the trend flow contribution value of the input gate. This indicates element-wise multiplication;

[0094] S34, obtain the dynamic response long-term state vector of the previous time step, update the dynamic response long-term state vector of the current time step according to the forget gate activation value, the input gate activation value and the dynamic response long-term state vector of the previous time step, and obtain the updated dynamic response long-term state vector (the update formula adopts the LSTM standard calculation process).

[0095] S35, obtain the dynamic response memory vector of the current time step based on the output gate activation value and the updated dynamic response long-term state vector of the current time step;

[0096] S36, iteratively calculate by time step, and use the dynamic response memory vector of the final time step in the multi-scale temporal feature tensor as the deep response pattern vector to complete the temporal coding process;

[0097] S37. Synchronously, based on the dynamic response memory vector of each time step in the multi-scale temporal feature tensor, temporal reconstruction processing is performed through the temporal autoencoder of the lightweight long short-term memory network to obtain the reconstructed feature vector. The point-by-point reconstructed residual sequence is obtained based on the mean square error between the reconstructed feature vector and the first channel feature vector.

[0098] As described in steps S31-S37 above, the lightweight long short-term memory network adopts a two-stream grouped weighted processing architecture to perform temporal encoding and reconstruction of the input dynamic response process information to obtain high-order features for health assessment. By calculating the grouped weighted contribution values ​​of the state stream and trend stream, the single-path input processing of the standard long short-term memory network can be decoupled into processing of state features. ) and trend characteristics ( The independent weighted evaluation of two parallel information streams is used. State features reflect the response state at a specific moment, while trend features reflect the instantaneous changes in state. Processing these two features separately enables the network to simultaneously perceive the position and momentum of the dynamic process, enhancing its ability to detect fault modes (such as acceleration anomalies and response abrupt changes) and ensuring that both state and trend information can directly and independently influence the gating logic. Simultaneously, by providing... The lightweight long short-term memory network is equipped with independent grouping weight parameters for each of its first-order difference settings, enabling it to learn the differentiated contributions of different physical quantities and their rates of change to the input gate, forget gate, and output gate decisions. This design forces the network to distinguish the role of different types of physical information in modeling dynamic response processes, rather than performing homogenized mixing.

[0099] By fusing the state flow contribution value and the trend flow contribution value through element-wise multiplication, the final input gate activation value, forget gate activation value, and output gate activation value are generated. This makes the final activation strength of the gating depend on the synergistic effect of the two information flow contributions. This mechanism requires that the state information and trend information reach consistency in the gating decision. For example, a significant input signal needs to be supported by both state evaluation and trend evaluation. This design improves the robustness of the gating decision and can suppress false activation caused by noise or interference in a single information flow, making the updating of information in the cell state more prudent.

[0100] Following the standard update process of Long Short-Term Memory (LSTM) networks, the generated gated activation values ​​are used to iteratively update the long-term dynamic response state vector (cell state, initial value of zero) and the dynamic response memory vector (hidden state, initial value of zero). The dynamic response memory vector at the final time step of the sequence is output as the deep response pattern vector. The update mechanism of the long-term dynamic response state vector allows the network to selectively retain contextual information spanning multiple time steps. This information is used to characterize a long-term characteristic of the mechanical dynamic response of high-voltage equipment (such as motion smoothness in the middle of the stroke, phase coordination of current and velocity, and characteristic vibration frequency in the buffer phase). For high-voltage equipment (such as high-voltage circuit breakers...), this method is particularly useful for... The mechanical response process of a circuit device is such that the characteristics of subsequent stages (such as buffering) are often influenced by the state of the earlier stages (such as startup and acceleration). This mechanism enables lightweight long short-term memory networks to model such physical correlations across time steps, providing the necessary historical context for generating globally representative pattern vectors. The deep response pattern vector is a compression and abstraction of the dynamic response process represented by the entire multi-scale temporal feature tensor. It is a short-term response feature output at the current time step for the next time step or final decision, such as the current motion efficiency, the current electromagnetic-mechanical coupling synchronization deviation and the degree of anomaly compared with the history, providing an efficient data representation for subsequent pattern similarity comparison.

[0101] By utilizing the dynamic response memory vector generated at each time step, an additional temporal autoencoder reconstructor (a lightweight fully connected layer is set as the temporal autoencoder reconstructor for the lightweight long short-term memory network) attempts to reconstruct the first channel feature vector corresponding to that time step. The point-by-point reconstruction residual sequence is obtained by calculating the mean squared error, thus constructing an anomaly detection channel based on the reconstruction error. The temporal autoencoder reconstructor is trained on a healthy response pattern library (described later) to learn a standard mapping from the dynamic response memory vector to the first channel feature vector. When the input multi-scale temporal feature tensor contains anomaly dynamic patterns not seen during training, the internal analysis logic of the lightweight long short-term memory network, based on healthy patterns in the healthy response pattern library, cannot accurately reconstruct the anomaly dynamic pattern, leading to an increase in the reconstruction error at the corresponding time point. The point-by-point reconstruction residual sequence... Using time as an index, it is possible to directly locate and quantify which (or which) time segments deviate from the healthy pattern during the dynamic response process. This has high sensitivity for detecting local and non-global defects such as transient sudden failures (e.g., momentary jamming) and phased anomalies (e.g., slow startup). Therefore, pointwise reconstruction of the residual sequence can sensitively indicate the degree of deviation of local time segments from the learned healthy pattern during the dynamic response process, providing the ability to characterize transient or local anomalies. Moreover, the "point" anomaly information provided by the pointwise reconstruction residual sequence and the "surface" overall health information provided by the deep response fingerprint vector can form a strong complement. The combination of the two ensures that the system has a strong characterization and detection capability for both global performance degradation and local temporal distortion, thereby improving the detection rate of early complex faults.

[0102] It should be noted that after the dynamic response memory vector is output at each time step of the lightweight long short-term memory network, a lightweight fully connected layer is connected in parallel as a temporal autoencoder reconstructor. The input dimension of this lightweight fully connected layer is equal to the dimension of the dynamic response memory vector, and the output dimension is 3 (corresponding to...). During training, data from the healthy response pattern library is used, and a multi-task learning approach is adopted to train the lightweight long short-term memory network synchronously. The training loss function includes a reconstruction loss term. During monitoring, reconstruction calculation is performed synchronously after calculating the dynamic response memory vector at each time step, and the reconstruction error of that time step is generated in real time. During output, after traversing the entire sequence, the reconstruction errors of all time steps are arranged in order to obtain the pointwise reconstruction residual sequence.

[0103] In one embodiment, the step of obtaining a comprehensive score for mechanical property health based on the deep response pattern vector and the point-by-point reconstructed residual sequence includes:

[0104] S41, Establish a health response pattern library for high-voltage equipment (the establishment of the health response pattern library depends on the known health status data of the equipment in the early stage of commissioning or after major overhaul, and is composed of response data that are manually confirmed or automatically marked as "healthy"). The health response pattern library stores historical health status deep response pattern vectors.

[0105] S42, obtain the response pattern deviation based on the deep response pattern vector and the healthy response pattern library, using the following formula:

[0106]

[0107] In the formula, Indicates the deviation of the response pattern. Represents the depth response pattern vector. This represents the mean vector of all historical health state deep response pattern vectors in the health response pattern library. This represents the dot product operation;

[0108] S43, Calculate the reconstruction anomaly index based on the point-by-point reconstructed residual sequence, using the following formula:

[0109]

[0110] In the formula, Indicates the reconstruction of the anomaly index. This represents the mean of the point-by-point reconstructed residual sequence. This represents the standard deviation of the point-by-point reconstructed residual sequence. This represents the kurtosis (normalized value) of the pointwise reconstructed residual sequence. When the value is 0, the residual has no fluctuation, and the reconstruction anomaly index is 0;

[0111] S44, Based on the mode deviation and the reconstruction anomaly index, the comprehensive score of the mechanical characteristic health is obtained through adaptive weighted fusion, and the calculation formula is as follows:

[0112]

[0113] In the formula, This represents the overall score for the health of mechanical properties. Indicates the deviation of the response pattern. Indicates the reconstruction of the anomaly index. The health assessment fusion weights are dynamically adjusted based on the intra-class average distance of the sample vectors in the health response pattern library. The intra-class average distance reflects the natural fluctuation range of the device's deep response pattern vectors in a healthy state; the larger the intra-class average distance, the higher the weight. The smaller the value, the better to balance the impact of individual differences in different high-voltage equipment.

[0114] As described in steps S41-S44 above, establishing a health response pattern library is to create a set of deep response pattern vectors for the monitored high-voltage equipment to store its historical health status, which can obtain an individualized health benchmark. Due to the inherent differences between equipment, a uniform absolute health standard may lead to misjudgment. The health response pattern library records the dynamic patterns of the equipment itself during the health period, providing a reference benchmark that matches the characteristics of the equipment for calculating the deviation of the response pattern, thereby eliminating or reducing the systematic evaluation bias introduced by individual differences.

[0115] The response mode deviation primarily measures the difference in direction between the deep response mode vector and the mean vector. For the deep response mode vector, its magnitude may be affected by global factors (such as small fluctuations in operating energy) that are unrelated to core mechanical failures. The response mode deviation allows this indicator to focus more on assessing whether the shape or structure of the dynamic response mode has changed, rather than simply responding to global amplitude changes, thereby improving its robustness in detecting mode shifts that characterize system performance degradation.

[0116] Calculating the reconstruction anomaly index can assess the anomaly characteristics of the reconstruction residuals, where the first term... Larger The numerical value indicates that the mean of the reconstruction error is relatively high, while its standard deviation is relatively low, meaning the error exhibits a consistently high and stable pattern rather than random fluctuations. This suggests the existence of a systematic and consistent local pattern mismatch. (Second term) kurtosis This reflects the tail thickness of the residual distribution; a larger kurtosis value indicates the existence of a few extremely large residual values ​​(abnormal peaks). The function performs a monotonically increasing nonlinear transformation on the kurtosis, which, while retaining its ability to characterize sudden and transient anomalies, prevents the index from being overly sensitive to extreme values ​​through the characteristics of the logarithmic function, thus enhancing robustness. When the residuals exhibit a systematically high skewness (large first term), or when there are significant anomalous spikes (large second term), or both, the reconstruction anomaly index will show a high value. Therefore, the reconstruction anomaly index can effectively respond to various types of local temporal distortions.

[0117] The comprehensive score is obtained through adaptive weighted fusion. When the intra-class average distance of the health mode is large, it indicates that the normal dynamic response mode of the device itself has large fluctuations. At this time, the deviation of the response mode is more easily affected by such normal fluctuations, and its ability to distinguish between abnormalities and normal fluctuations decreases. Therefore, the value of the health assessment fusion weight is reduced, that is, the weight of the response mode deviation is reduced in the final score, while the weight of the reconstruction anomaly index, which is more sensitive to local anomalies and may be less affected by the inherent fluctuations of the overall mode, is relatively increased. Conversely, when the intra-class distance is small, the value of the health assessment fusion weight is increased, and more trust is placed on the response mode deviation. This mechanism enables the system to automatically adjust its trust in the two types of evidence, "overall mode deviation" and "local temporal distortion," based on the mode stability of the monitored object. This achieves individualized optimization of the assessment strategy and improves the accuracy and consistency of the comprehensive score of mechanical characteristic health in different application scenarios.

[0118] In one embodiment, the step of obtaining the adaptive warning threshold, comparing the comprehensive score of mechanical characteristic health with the adaptive warning threshold, and determining whether an early trend warning or a current abnormal alarm is triggered includes:

[0119] S51, obtain a comprehensive health score of the historical mechanical characteristics of a healthy high-voltage equipment for several responses to form a health score set, and calculate the mean and standard deviation of the health score set;

[0120] S52, using the health score set as a dynamic health score sequence, obtain the comprehensive score of mechanical characteristic health of the high-voltage equipment for each normal response (referring to the response that does not trigger early trend warning and current abnormal alarm) during monitoring, and add it to the dynamic health score sequence. Starting from the mean and standard deviation, use the exponential weighted moving average algorithm to update the estimated mean and estimated standard deviation of the dynamic health score sequence.

[0121] S53, obtain the cumulative number of responses from the high-voltage equipment, obtain a threshold relaxation factor based on the cumulative number of responses, and obtain an adaptive early warning threshold based on the threshold relaxation factor, the estimated mean, and the estimated standard deviation, using the following formula:

[0122]

[0123] In the formula, Indicates the adaptive warning threshold. This represents the estimated mean. This represents the estimated standard deviation. The threshold relaxation factor (represented by the cumulative number of responses) a monotonically non-decreasing function, (Cumulative calculation starting from the time the high-voltage equipment was put into operation or the last major overhaul).

[0124] S54, compare the comprehensive score of the mechanical characteristic health of this response with the adaptive early warning threshold:

[0125] S541, if the comprehensive score of the mechanical characteristic health is not greater than the adaptive early warning threshold, then it is further determined whether an early trend warning is triggered.

[0126] S542, if the comprehensive score of mechanical characteristic health is greater than the adaptive early warning threshold, then the current abnormal alarm is triggered.

[0127] As described in steps S51-S54 above, an initial health benchmark can be established through a health score set. Specifically, when the high-voltage equipment is in a known health state (e.g., during initial operation or after major overhaul), a comprehensive score of the mechanical characteristic health of several normal responses is collected to form an initial health score set, and the mean and standard deviation of this set are calculated. Starting from the initial health score set, a dynamic health score sequence is constructed. In subsequent continuous monitoring, whenever the high-voltage equipment completes a response that is judged to be normal, the comprehensive score of the mechanical characteristic health of that response is added to the sequence. Then, using an exponentially weighted moving average algorithm, starting from the initial mean and standard deviation, the estimated mean and estimated standard deviation of the sequence are dynamically updated. The exponentially weighted moving average algorithm can give higher weight to recent data, so that the estimated mean can smoothly follow the slow drift of the performance center value caused by factors such as normal aging and environmental changes. The estimated standard deviation dynamically reflects the dispersion of recent normal scores. When the equipment is operating stably, this value is small; when the equipment operating conditions change and the normal fluctuation range of the score increases, this value will increase accordingly. This allows the system to adaptively adjust its tolerance to normal fluctuations, avoiding false alarms caused by increased fluctuations during periods of variable operating conditions.

[0128] In the formula for calculating the adaptive early warning threshold, the cumulative number of responses is a variable characterizing the cumulative wear and aging of the equipment's mechanical components. As the number of equipment operations increases (i.e., the mechanical lifespan is consumed), when judging whether its state is abnormal, a gradually widening tolerance multiple is allowed based on its current normal fluctuation range. The estimated mean reflects the current performance center, the estimated standard deviation reflects the current normal fluctuation, and the threshold relaxation factor reflects the lifespan consumption. This allows the adaptive early warning threshold to evolve with the service life of the high-voltage equipment, not only adapting to the short-term state fluctuations of the equipment, but also responding to its long-term performance degradation trend.

[0129] Finally, a decision is made based on a comparison between the adaptive warning threshold and the comprehensive score of mechanical characteristic health. When the comprehensive score of mechanical characteristic health is greater than the adaptive warning threshold, it indicates that the mechanical characteristics of the current operation have significantly deviated from the normal expected range defined based on its recent state and lifespan stage. The system triggers a high-confidence alarm for the current anomaly. Because the adaptive warning threshold has adaptive capabilities, this type of alarm is highly accurate and has a low probability of false alarms. When the comprehensive score of mechanical characteristic health is less than or equal to the adaptive warning threshold, it indicates that the current operation has not reached the severity level for an immediate alarm. At this time, further judgment is made on whether to trigger an early trend warning. The long-term trend of the comprehensive score of mechanical characteristic health is analyzed, thereby realizing the hierarchical and refined management of equipment risk status. Severe and immediate anomalies are directly intercepted and alarmed by the threshold, while potential and slowly developing risks are captured and warned by a more sensitive trend analysis mechanism.

[0130] In one embodiment, the step of further determining whether to trigger an early trend warning if the comprehensive score of the mechanical characteristic health is not greater than the adaptive warning threshold includes:

[0131] S5411, Obtain the comprehensive health score of the mechanical characteristics based on a number of response times within a preset period, so as to construct a health score sequence;

[0132] S5412, Based on the health score sequence, the least squares method is used to fit its linear trend information within a preset sliding time window, and the slope of the health deterioration trend is calculated.

[0133] S5413, determine whether the slope of the health degradation trend exceeds a preset trend sensitivity threshold for a consecutive preset number of times. If it does, trigger an early trend warning.

[0134] As described in steps S5411-S5413 above, by acquiring the comprehensive mechanical characteristic health score of several responses within a preset period and arranging them in chronological order to construct a health score sequence, the discrete single operation evaluation results can be transformed into a continuous dataset that can be used for time series analysis. This provides a basis for identifying slow changing trends, enabling the system to observe and quantify the evolution pattern of the score with the number of operations or the passage of time. Then, the least squares method is used to calculate the slope of the health deterioration trend. The least squares method provides the optimal estimate of the linear trend of the data within the window by minimizing the sum of squares of the distances from the data points to the fitted straight line. The slope of the health deterioration trend quantifies the average rate of change of the score within the selected sliding time window. A continuously positive slope indicates that there is a systematic upward trend in the comprehensive mechanical characteristic health score, that is, the mechanical performance shows a continuous tendency to deteriorate. The use of a sliding window for calculation ensures that the trend slope can dynamically reflect the latest data changes and capture new signs of deterioration in a timely manner, without being over-averaged by historical data.

[0135] Since the trend slope calculated in a single instance may be affected by occasional operational anomalies, measurement noise, or short-term operating condition fluctuations, this invention requires that the results of multiple consecutive calculations exceed the trend sensitivity threshold in order to effectively filter out such transient interferences and ensure that the warned trend has persistence and statistical significance, thereby reducing the possibility of false alarms. The trend sensitivity threshold defines the sensitivity of the system to the deterioration trend. Users can adjust the trend sensitivity threshold according to the importance of the equipment, reliability requirements, and maintenance strategies to achieve a balance between warning sensitivity and system stability. For example, the higher the importance of the equipment, the smaller the trend sensitivity threshold should be to minimize missed alarms.

[0136] When the condition of the continuous over-trend sensitivity threshold is met, even if the comprehensive score of the mechanical characteristic health of the current single operation has not reached the adaptive early warning threshold that triggers the current abnormal alarm, the system will issue an early trend warning. This provides maintenance personnel with an early, quantitative signal that the equipment performance is undergoing potential decline, making it possible to arrange preventive maintenance before the fault develops into a serious problem, which helps to avoid unplanned downtime and extend the service life of the equipment.

[0137] like Figure 2 As shown, the present invention also provides a real-time monitoring system for the dynamic response of the mechanical characteristics of high-voltage power equipment, comprising:

[0138] The information acquisition module is used to acquire the stroke-time curve signal and the opening / closing coil current signal of a single opening or closing response of high-voltage equipment, and to acquire a standardized stroke sequence and coil current sequence based on the stroke-time curve signal and the opening / closing coil current signal;

[0139] The feature construction module is used to construct a multi-scale time-series feature tensor based on the standardized travel sequence and the coil current sequence;

[0140] The inference module is used to input the multi-scale temporal feature tensor into a lightweight long short-term memory network, and perform temporal encoding and temporal reconstruction processing on the multi-scale temporal feature tensor through the lightweight long short-term memory network to obtain a deep response pattern vector and a pointwise reconstructed residual sequence.

[0141] Anomaly scoring module is used to obtain a comprehensive score of mechanical characteristic health based on the deep response pattern vector and the point-by-point reconstruction residual sequence;

[0142] The judgment and decision module is used to obtain the adaptive warning threshold, compare the comprehensive score of mechanical characteristic health with the adaptive warning threshold, and determine whether to trigger an early trend warning or a current abnormal alarm.

[0143] Preferably, the anomaly scoring module includes:

[0144] A health mode library construction unit is used to establish a health response mode library for high-voltage equipment, wherein the health response mode library stores historical health status deep response mode vectors.

[0145] The pattern deviation calculation unit is used to obtain the response pattern deviation based on the deep response pattern vector and the healthy response pattern library;

[0146] A reconstruction anomaly index calculation unit is used to calculate the reconstruction anomaly index based on the point-by-point reconstruction residual sequence.

[0147] An adaptive scoring fusion unit is used to obtain a comprehensive score of mechanical characteristic health by adaptive weighted fusion based on the mode deviation degree and the reconstruction anomaly index.

[0148] The present invention also provides a computer device, including a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to implement the steps of a method for real-time monitoring of the dynamic response of the mechanical characteristics of high-voltage power equipment.

[0149] The present invention also provides a computer-readable storage medium having a computer program stored thereon, wherein the computer program, when executed by a processor, implements the steps of a method for real-time monitoring of the dynamic response of the mechanical characteristics of high-voltage power equipment.

[0150] It should be noted that, in this document, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, apparatus, article, or method that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, apparatus, article, or method. Unless otherwise specified, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, apparatus, article, or method that includes that element.

[0151] The above description is merely a preferred embodiment of the present invention and does not limit the patent scope of the present invention. Any equivalent structural or procedural transformations made based on the content of the present invention's specification and drawings, or direct or indirect applications in other related technical fields, are similarly included within the patent protection scope of the present invention.

Claims

1. A method for real-time monitoring of the dynamic response of mechanical characteristics of high-voltage power equipment, characterized in that, Includes the following steps: Acquire the stroke-time curve signal and the opening / closing coil current signal of a single opening or closing response of high-voltage equipment, and obtain a standardized stroke sequence and coil current sequence based on the stroke-time curve signal and the opening / closing coil current signal; Based on the standardized travel sequence, the instantaneous motion velocity sequence is calculated using the center difference method; based on the coil current sequence, the coil current change rate sequence is calculated using the center difference method. The transient coupling strength sequence is obtained based on the instantaneous motion velocity sequence and the coil current change rate sequence; Based on the standardized travel sequence, the instantaneous motion velocity sequence, and the transient coupling strength sequence, a first channel feature vector is constructed in chronological order; The second channel feature vector is obtained by calculating the first-order backward difference of the first channel feature vector along the time axis; The first channel feature vector and the second channel feature vector are stacked along the time dimension to construct a multi-scale temporal feature tensor. The multi-scale temporal feature tensor is input into a lightweight long short-term memory network. The lightweight long short-term memory network performs temporal encoding and temporal reconstruction processing on the multi-scale temporal feature tensor to obtain a deep response pattern vector and a point-by-point reconstruction residual sequence. A comprehensive score for mechanical property health is obtained based on the deep response pattern vector and the point-by-point reconstructed residual sequence. Obtain the adaptive warning threshold, compare the comprehensive score of mechanical characteristic health with the adaptive warning threshold, and determine whether an early trend warning or a current abnormal alarm is triggered.

2. The method for real-time monitoring of dynamic response of mechanical characteristics of high-voltage power equipment according to claim 1, characterized in that, The step of inputting the multi-scale temporal feature tensor into a lightweight long short-term memory network, and performing temporal encoding and temporal reconstruction processing on the multi-scale temporal feature tensor through the lightweight long short-term memory network to obtain a deep response pattern vector and a pointwise reconstructed residual sequence includes: The lightweight long short-term memory network calculates the state flow contribution value of the first channel feature vector at each time step in the multi-scale temporal feature tensor to the input gate, forget gate and output gate. The lightweight long short-term memory network calculates the trend flow contribution value of the second channel feature vector at each time step in the multi-scale temporal feature tensor to the input gate, forget gate and output gate. By fusing the state flow contribution value and the trend flow contribution value, the input gate activation value, the forget gate activation value, and the output gate activation value are generated; Obtain the dynamic response long-term state vector of the previous time step, and update the dynamic response long-term state vector of the current time step according to the forget gate activation value, the input gate activation value and the dynamic response long-term state vector of the previous time step to obtain the updated dynamic response long-term state vector. The dynamic response memory vector of the current time step is obtained based on the output gate activation value and the updated dynamic response long-term state vector of the current time step. The time-step iterative calculation is performed, and the dynamic response memory vector of the final time step in the multi-scale temporal feature tensor is used as the deep response pattern vector to complete the temporal coding process. Synchronously, based on the dynamic response memory vector of each time step in the multi-scale temporal feature tensor, temporal reconstruction processing is performed through the temporal autoencoder of the lightweight long short-term memory network to obtain the reconstructed feature vector. The point-by-point reconstructed residual sequence is obtained based on the mean square error between the reconstructed feature vector and the first channel feature vector.

3. The method for real-time monitoring of dynamic response of mechanical characteristics of high-voltage power equipment according to claim 1, characterized in that, The step of obtaining the comprehensive score of mechanical property health based on the deep response pattern vector and the point-by-point reconstructed residual sequence includes: Establish a health response pattern library for high-voltage equipment, which stores historical health status deep response pattern vectors. The deviation of the response pattern is obtained based on the deep response pattern vector and the health response pattern library; Calculate the reconstruction anomaly index based on the point-by-point reconstruction residual sequence; Based on the pattern deviation and the reconstruction anomaly index, the comprehensive score of the mechanical characteristic health is obtained through adaptive weighted fusion.

4. The method for real-time monitoring of dynamic response of mechanical characteristics of high-voltage power equipment according to claim 1, characterized in that, The step of obtaining the adaptive early warning threshold and comparing the comprehensive score of mechanical characteristic health with the adaptive early warning threshold to determine whether an early trend warning or a current abnormal alarm is triggered includes: A comprehensive health score of the historical mechanical characteristics of a healthy high-voltage device is obtained from several responses to form a health score set, and the mean and standard deviation of the health score set are calculated. Using the set of health scores as a dynamic health score sequence, a comprehensive score of the mechanical characteristic health of the high-voltage equipment for each normal response during monitoring is obtained and added to the dynamic health score sequence. Starting from the mean and standard deviation, the estimated mean and estimated standard deviation of the dynamic health score sequence are updated using an exponential weighted moving average algorithm. The cumulative number of responses from the high-voltage equipment is obtained, a threshold relaxation factor is obtained based on the cumulative number of responses, and an adaptive early warning threshold is obtained based on the threshold relaxation factor, the estimated mean, and the estimated standard deviation. The overall score of the mechanical health characteristics of this response is compared with the adaptive early warning threshold. If the comprehensive score of the mechanical characteristic health is not greater than the adaptive early warning threshold, then it is further determined whether an early trend warning is triggered. If the overall health score of the mechanical characteristics is greater than the adaptive early warning threshold, then the current abnormal alarm is triggered.

5. The method for real-time monitoring of dynamic response of mechanical characteristics of high-voltage power equipment according to claim 4, characterized in that, If the comprehensive score of the mechanical characteristic health is not greater than the adaptive warning threshold, the step of further determining whether an early trend warning is triggered includes: Obtain a comprehensive health score of the mechanical characteristics based on a number of responses within a preset period to construct a health score sequence; Based on the health score sequence, the least squares method is used to fit its linear trend information within a preset sliding time window, and the slope of the health deterioration trend is calculated. Determine whether the slope of the health deterioration trend exceeds a preset trend sensitivity threshold for a consecutive preset number of times. If it does, trigger an early trend warning.

6. A real-time monitoring system for the dynamic response of mechanical characteristics of high-voltage power equipment, characterized in that, include: The information acquisition module is used to acquire the stroke-time curve signal and the opening / closing coil current signal of a single opening or closing response of high-voltage equipment, and to acquire a standardized stroke sequence and coil current sequence based on the stroke-time curve signal and the opening / closing coil current signal; The feature construction module is used to calculate the instantaneous motion velocity sequence using the central difference method based on the standardized travel sequence, and to calculate the coil current change rate sequence using the central difference method based on the coil current sequence; to obtain the transient coupling strength sequence based on the instantaneous motion velocity sequence and the coil current change rate sequence; to construct a first channel feature vector in chronological order based on the standardized travel sequence, the instantaneous motion velocity sequence, and the transient coupling strength sequence; to obtain a second channel feature vector by calculating the first-order backward difference of the first channel feature vector along the time axis; and to stack the first channel feature vector and the second channel feature vector along the time dimension to construct a multi-scale temporal feature tensor. The inference module is used to input the multi-scale temporal feature tensor into a lightweight long short-term memory network, and perform temporal encoding and temporal reconstruction processing on the multi-scale temporal feature tensor through the lightweight long short-term memory network to obtain a deep response pattern vector and a pointwise reconstructed residual sequence. Anomaly scoring module is used to obtain a comprehensive score of mechanical characteristic health based on the deep response pattern vector and the point-by-point reconstructed residual sequence; The judgment and decision module is used to obtain the adaptive warning threshold, compare the comprehensive score of mechanical characteristic health with the adaptive warning threshold, and determine whether to trigger an early trend warning or a current abnormal alarm.

7. The real-time monitoring system for dynamic response of mechanical characteristics of high-voltage power equipment according to claim 6, characterized in that, The anomaly scoring module includes: A health mode library construction unit is used to establish a health response mode library for high-voltage equipment, wherein the health response mode library stores historical health status deep response mode vectors. The pattern deviation calculation unit is used to obtain the response pattern deviation based on the deep response pattern vector and the healthy response pattern library; A reconstruction anomaly index calculation unit is used to calculate the reconstruction anomaly index based on the point-by-point reconstruction residual sequence. An adaptive scoring fusion unit is used to obtain a comprehensive score of mechanical characteristic health by adaptive weighted fusion based on the mode deviation degree and the reconstruction anomaly index.

8. A computer device comprising a memory and a processor, wherein the memory stores a computer program, characterized in that, When the processor executes the computer program, it implements the steps of the method according to any one of claims 1 to 5.

9. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by a processor, it implements the steps of the method according to any one of claims 1 to 5.

Images