Insulator operation state acoustic identification and prediction method based on SSA-LSTM

Abstract

The invention is suitable for the technical field of artificial intelligence, and provides an insulator operation state acoustic identification and prediction method based on SSA-LSTM, and the method comprises the steps: obtaining an acoustic emission signal time sequence of a to-be-evaluated insulator in an operation state, environment parameters of a micrometeorological environment, and a recent historical risk score sequence of the to-be-evaluated insulator; performing preprocessing and feature extraction on the obtained signals to obtain an acoustic feature vector, a context feature vector and a trend feature vector; fusing the acoustic feature vector, the context feature vector and the trend feature vector, and calculating a target dynamic risk score based on the fused features; and comparing the target dynamic risk score with a plurality of preset risk threshold intervals, and mapping the target dynamic risk score to a corresponding discrete risk level. The method effectively reduces the false alarm rate and the missing alarm rate, achieves the advanced early warning of the pollution flashover risk of the insulator, and meets the urgent demands of an intelligent power grid for the precise sensing of the state of a power transmission line and the intelligent operation and maintenance.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of artificial intelligence, in particular to an insulator running state acoustic identification and prediction method based on SSA-LSTM. BACKGROUND

[0002] In the power transmission line under complex climate and geographical environment, the surface of the insulator often degrades the insulating performance due to the accumulation of dust, dirt and other pollutants. When encountering extremely severe weather conditions such as continuous fog, the surface dirt will significantly increase the conductivity of the insulator after dissolving, and aggravate the risk of pollution flashover accident. The pollution flashover accident not only damages the insulator, but also causes secondary damage to other electrical equipment such as conductors and fittings, resulting in increased operation and maintenance costs, reduced reclosing success rate, and even large-scale power supply interruption in severe cases.

[0003] In the prior art, the evaluation of the transient elastic wave phenomenon of the contaminated insulator is mainly based on the threshold value judgment of the characteristic parameters or the pattern recognition method based on machine learning. However, the threshold value judgment is mostly set by humans, and an alarm is triggered when the monitoring value exceeds the threshold value. This method can only send a signal when the discharge has reached a serious level, and cannot predict the degradation trend of the insulator state. Machine learning and neural network algorithms can overcome the limitations of threshold algorithms, but existing algorithms are difficult to fully capture the most relevant, subtle and deep pattern information in the acoustic emission signal related to the insulator state, and have inherent limitations such as incomplete feature representation and weak model generalization ability. This easily leads to poor robustness, high false alarm and missed alarm rates in complex field environments, and cannot meet the urgent needs of smart grids for accurate perception and advanced warning of power transmission line status.

[0004] Therefore, an insulator running state acoustic identification and prediction method based on SSA-LSTM is proposed. SUMMARY

[0005] The present application provides an insulator running state acoustic identification and prediction method based on SSA-LSTM, which solves the problem of not being able to meet the urgent needs of smart grids for accurate perception and advanced warning of power transmission line status.

[0006] The present application provides an insulator running state acoustic identification and prediction method based on SSA-LSTM, which includes: Acoustic emission signal time series of the insulator to be evaluated in the running state, environmental parameters of the microclimate environment and recent historical risk score sequence of the insulator to be evaluated are obtained; the state parameters at least include equivalent salt density and insulator material type; The acoustic emission signal original sequence, the micro-meteorological environment parameters, the state parameters and the historical risk score sequence are respectively subjected to signal preprocessing and feature extraction; wherein: The acoustic emission signal original sequence is subjected to denoising processing, and an acoustic feature vector is extracted by using an LSTM neural network; The micro-meteorological environment parameters and the state parameters are jointly coded into a context feature vector; The historical risk score sequence is subjected to time series feature extraction to obtain a trend feature vector; if it is the first evaluation without historical data, the trend feature vector is set to a zero vector or a default initial value is used; The acoustic feature vector, the context feature vector and the trend feature vector are fused, and a target dynamic risk score is calculated based on the fused features; The target dynamic risk score is compared with a plurality of preset risk threshold intervals to map a corresponding discrete risk level.

[0007] Further, the denoising processing of the acoustic emission signal original sequence and the extraction of the acoustic feature vector by using the LSTM neural network comprise: The denoised acoustic signal sequence is subjected to standardization processing and is constructed into a sequence format for input of the LSTM network; The pre-trained LSTM acoustic feature extraction model is used to extract time series features from the standardized sequence to obtain an acoustic feature vector.

[0008] Further, the establishment of the pre-trained LSTM acoustic feature extraction model comprises the following steps: An acoustic emission signal sample set containing multiple insulator discharge modes is obtained, and an initial LSTM network is trained with the sample set, so that the initial LSTM network obtains a basic ability to identify insulator discharge modes from acoustic signal sequences, and the trained LSTM network is used as a pre-trained acoustic representation basic model; A feature adaptation module is connected to the output end of the pre-trained acoustic representation basic model to form a complete acoustic feature extraction model; The acoustic feature extraction model is fine-tuned by using acoustic emission signal data containing complex field noise, and in the fine-tuning process, the parameters of the pre-trained LSTM acoustic representation basic model are fixed, and only the parameters of the feature adaptation module are updated.

[0009] Further, in the model training stage, one or more data enhancement disturbances are applied to the acoustic emission signal original sequence, and the disturbances include adding random noise, time stretching or amplitude scaling; the acoustic signal sequence after applying the disturbance and the original acoustic signal sequence are jointly subjected to denoising and subsequent processing for acoustic feature extraction model training.

[0010] Further, the microclimate environment parameters and the state parameters are jointly encoded into a context feature vector, comprising: The continuous parameters are normalized, and the discrete parameters are one-hot encoded; All processed parameters are spliced into a comprehensive parameter vector; The comprehensive parameter vector is input into a fully connected neural network for nonlinear transformation and dimension reduction, and the context feature vector is output.

[0011] Further, the continuous parameters include humidity, temperature, and equivalent salt density, and the discrete parameters include weather type and insulator material type.

[0012] Further, the historical risk score sequence is subjected to time series feature extraction to obtain a trend feature vector, comprising: The historical risk score sequence is input into a pre-trained time series feature extraction model; the time series feature extraction model is a recurrent neural network; The historical risk score sequence is encoded using the time series feature extraction model, and a fixed-dimensional vector is output as the trend feature vector; Further, if the length of the historical risk score sequence is variable or there are missing values, length normalization or missing value filling processing is performed before inputting the model.

[0013] Further, the acoustic feature vector, the context feature vector, and the trend feature vector are fused, and a target dynamic risk score is calculated based on the fused features, comprising: The acoustic feature vector, the context feature vector, and the trend feature vector are spliced to generate a comprehensive feature vector; The comprehensive feature vector is input into a pre-trained risk assessment model, and the risk assessment model is a deep neural network containing multiple fully connected layers; The comprehensive feature vector is subjected to nonlinear transformation by the risk assessment model, and a continuous numerical value is output as the target dynamic risk score, the value range of the target dynamic risk score being [0, 1], and the larger the value, the higher the risk.

[0014] Further, the target dynamic risk score is compared with a plurality of preset risk threshold intervals to map to corresponding discrete risk levels, comprising: Three risk threshold values include a first threshold T1, a second threshold T2, and a third threshold T3, wherein 0 < T1 < T2 < T3 < 1; When the target dynamic risk score is less than T1, it is mapped to a normal risk level; When the target dynamic risk score is greater than or equal to T1 and less than T2, it is mapped to a pre-warning risk level; When the target dynamic risk score is greater than or equal to T2 and less than T3, it is mapped to an alarm risk level; When the target dynamic risk score is greater than or equal to T3, it is mapped to a maintenance risk level.

[0015] From the above technical solutions, the present application has the following advantages: The present application obtains the acoustic emission signal time sequence of the to-be-evaluated insulator in the running state, the environmental parameters of the micro-meteorological environment, and the recent historical risk score sequence of the to-be-evaluated insulator, and respectively performs signal preprocessing and feature extraction on the acoustic emission signal original sequence, the micro-meteorological environment parameters, the state parameters, and the historical risk score sequence. The acoustic feature vector, the context feature vector, and the trend feature vector are fused, and the target dynamic risk score is calculated based on the fused features. Finally, the target dynamic risk score is compared with the preset multiple risk threshold intervals to map to a discrete risk level. This method overcomes the defects of the existing threshold determination method that can only alarm after the event and cannot predict the trend, and the problem of weak generalization ability of the machine learning method due to incomplete feature representation, thereby significantly improving the accuracy and robustness of the state evaluation of the contaminated insulator, effectively reducing the false alarm and missed alarm rates, realizing the early warning of the insulator pollution flashover risk, and meeting the urgent needs of intelligent power grids for accurate perception and intelligent operation of transmission line states. BRIEF DESCRIPTION OF DRAWINGS

[0016] Figure 1 It is an embodiment flowchart of the insulator running state acoustic identification and prediction method based on SSA-LSTM in the present application; Figure 2 It is an architecture diagram of the insulator running state acoustic identification and prediction method based on SSA-LSTM in the present application; Figure 3 It is another embodiment flowchart of the insulator running state acoustic identification and prediction method based on SSA-LSTM in the present application; Figure 4 It is another embodiment flowchart of the insulator running state acoustic identification and prediction method based on SSA-LSTM in the present application; Figure 5 It is another embodiment flowchart of the insulator running state acoustic identification and prediction method based on SSA-LSTM in the present application; Figure 6 It is another embodiment flowchart of the insulator running state acoustic identification and prediction method based on SSA-LSTM in the present application. Detailed Implementation

[0017] The terms “first,” “second,” “third,” “fourth,” etc. (if present) in the specification and accompanying drawings of this application are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. It should be understood that such data can be interchanged where appropriate so that the embodiments of this application described herein can be implemented in orders other than those illustrated or described herein. Furthermore, the terms “comprising” and “corresponding to,” and any variations thereof, are intended to cover non-exclusive inclusion; for example, a process, method, system, product, or apparatus that comprises a series of steps or units is not necessarily limited to those steps or units explicitly listed, but may include other steps or units not explicitly listed or inherent to such processes, methods, products, or apparatus.

[0018] Example 1 The implementation method in this embodiment can be implemented in a system, on a server, or on a terminal; no specific limitation is made. The following section will describe the SSA-LSTM-based acoustic recognition and prediction method for insulator operating status from a system implementation perspective. Please refer to... Figures 1 to 6 The method provided in this application includes the following steps: S1. Obtain the time series of acoustic emission signals of the insulator to be evaluated under operating conditions, the environmental parameters of the local micro-meteorological environment, and the recent historical risk score sequence of the insulator to be evaluated; the state parameters include at least the equivalent salt density and the insulator material type; When insulators experience partial discharge or pollution flashover precursors during operation, transient elastic waves are generated. The acoustic emission signal time series carries crucial information about the discharge activity on the insulator surface and can be acquired in real time using acoustic sensors. High-sensitivity acoustic emission sensors are deployed near the insulator to be evaluated, and the sampling frequency is set to cover the typical frequency band of the discharge signal. The sensors continuously acquire acoustic signals during the insulator's operation, forming the original acoustic emission signal time series. The data acquisition equipment converts the analog signals into a digital sequence and transmits it to the processing unit. To ensure signal quality, environmental interference such as mechanical vibration must be avoided during the acquisition process.

[0019] Micrometeorological conditions such as humidity, temperature, wind speed, and precipitation directly affect the discharge behavior of polluted insulators. For example, high humidity accelerates the ionization of the pollution layer, increasing the risk of flashover. Environmental parameters are acquired in real time through micrometeorological monitoring stations or sensor networks installed near the insulators. These parameters include both continuous and discrete data. The data is typically stored in timestamp format and synchronized with acoustic emission signals to ensure time consistency.

[0020] The insulator's inherent condition parameters determine its pollution resistance, equivalent salt density quantifies surface contamination, and material type affects discharge characteristics. Equivalent salt density is obtained through periodic on-site sampling (cleaning the insulator surface and measuring solution conductivity) or indirectly based on online monitoring data (leakage current). The insulator material type is obtained directly from equipment management files or databases as a known attribute. These parameters are typically used as static or slowly varying data inputs and remain stable throughout the evaluation period.

[0021] Historical risk score sequences record recent risk trends for insulators, helping to capture patterns of condition degradation. The risk score sequence for the insulator over a recent period (one score point per day) is extracted from the historical database of the risk assessment system. If it is the first assessment (no historical data), the sequence is set to empty or uses the default initial value (zero vector), and will be dynamically updated in subsequent assessments.

[0022] All acquired data is aligned and integrated based on timestamps to ensure that acoustic emission signals, environmental parameters, state parameters, and historical scores are synchronized in time. After acquisition, the data is stored in a temporary buffer for subsequent preprocessing modules to access.

[0023] S2. Signal preprocessing and feature extraction are performed on the original acoustic emission signal sequence, micro-meteorological environment parameters, state parameters, and historical risk score sequence, respectively; where: S21. Denoise the original acoustic emission signal sequence and extract acoustic feature vectors using an LSTM neural network; Singular Spectrum Analysis (SSA) is employed for signal decomposition and reconstruction denoising. The SSA method does not require pre-defined basis functions and adaptively decomposes the signal into trend, periodic, and noise components. By discarding noise components and reconstructing effective components, efficient denoising is achieved. The specific process includes: constructing the trajectory matrix, singular value decomposition, grouping, and diagonal averaging reconstruction. After SSA denoising, a significantly improved signal-to-noise ratio acoustic signal sequence is obtained.

[0024] The denoised signal is directly input into a Long Short-Term Memory (LSTM) network to extract high-level acoustic feature vectors in an end-to-end manner. This method avoids the subjectivity of traditional manual feature design, enabling the model to autonomously discover the deep temporal patterns in the acoustic signal most relevant to the insulator's state. The specific steps are as follows: S211. Standardize the denoised acoustic signal sequence and construct it into the sequence format for the LSTM network input; Specifically, the denoised one-dimensional time-domain acoustic signal sequence is normalized and divided into consecutively overlapping or non-overlapping time windows in chronological order. Each window contains data points of a fixed length, forming a sample sequence that can be processed by the LSTM network.

[0025] S212. Use a pre-trained LSTM acoustic feature extraction model to extract temporal features from the standardized sequence to obtain acoustic feature vectors; The pre-trained LSTM acoustic feature extraction model here is a deep LSTM network. Its pre-training process enables the model to learn to identify general temporal patterns of insulator discharge from acoustic signal sequences. The sequence is input into the LSTM network, which, through its internal gating mechanism, processes information and updates its internal state step by step, ultimately encoding the information of the entire sequence into the hidden state of the last time step. The output of this hidden state (or the aggregated hidden states from all time steps) is used as the acoustic feature vector characterizing the comprehensive properties of that acoustic signal segment.

[0026] The establishment of a pre-trained LSTM acoustic feature extraction model includes the following steps: 1. Obtain a sample set of acoustic emission signals containing multiple insulator discharge modes, and train an initial LSTM network with the sample set to enable the initial LSTM network to acquire the basic ability to identify insulator discharge modes from acoustic signal sequences. Use the trained LSTM network as the pre-trained acoustic representation basic model. 2. Connect a feature adaptation module (several layers of fully connected neural networks) to the output of the pre-trained acoustic representation basic model (i.e., LSTM network) to form a complete acoustic feature extraction model; 3. The acoustic feature extraction model is fine-tuned using acoustic emission signal data containing complex on-site noise. During the fine-tuning process, the parameters of the pre-trained LSTM acoustic representation base model are fixed, and only the parameters of the feature adaptation module are updated.

[0027] In a clean laboratory environment, acoustic emission signals of various typical modes, including no discharge, corona discharge, partial discharge, and arc discharge, were collected to construct a high-quality sample set. An LSTM network (i.e., the initial acoustic network) was trained using this dataset, aiming to minimize its discharge mode classification error. The trained LSTM network possesses the basic ability to identify insulator discharge modes from acoustic signal sequences, serving as a fixed pre-trained acoustic representation model. The pre-trained LSTM model was combined with a randomly initialized feature adaptation module to form a complete acoustic feature extraction model. Then, this combined model was fine-tuned using measured data containing complex field noise. During this process, the parameters of the LSTM model were frozen, and only the parameters of the feature adaptation module were updated. This strategy leverages the powerful temporal feature extraction capabilities learned by the base model on clean data while allowing the adaptation module to learn how to adjust features for real, noisy field environments, resulting in acoustic feature vectors that are both discriminative and robust.

[0028] In this embodiment, during the model training phase, one or more data augmentation perturbations are applied to the original acoustic emission signal sequence. The perturbations include adding random noise, performing temporal stretching, or amplitude scaling. The perturbated acoustic signal sequence and the original acoustic signal sequence are then subjected to denoising and subsequent processing together for training the acoustic feature extraction model.

[0029] Specifically, various controllable perturbations are applied to the original acoustic emission signal sequence used for training. For example, Gaussian white noise with different signal-to-noise ratios is added to simulate environmental interference, small-amplitude temporal stretching is applied to simulate changes in discharge rate, or amplitude scaling is applied to simulate sensor distance differences. These perturbated samples, together with the original samples, constitute the enhanced training set, which is then fed into the model for denoising and subsequent feature extraction training.

[0030] S22. Encode the micrometeorological environment parameters and state parameters together into a context feature vector; This step addresses the problem of unified representation of multi-source heterogeneous data, transforming static and dynamic background information that significantly impacts risk into a compact, numerical contextual feature vector, providing crucial environmental contextual information for subsequent fusion and judgment. Specifically, this is achieved through the following steps: S221. Normalize the continuous parameters and perform unique thermal encoding on the discrete parameters; whereby the continuous parameters include humidity, temperature, and equivalent salinity, and the discrete parameters include weather type and insulator material type. S222. Concatenate all processed parameters into a single composite parameter vector; S223. Input the comprehensive parameter vector into a fully connected neural network for nonlinear transformation and dimensionality reduction, and output the context feature vector.

[0031] Specifically, different types of parameters undergo standardization preprocessing to eliminate the influence of dimensions and adapt to neural network processing. For continuous parameters (humidity, temperature, equivalent salinity), their numerical ranges vary greatly; for example, humidity may range from 0-100%, and equivalent salinity may only be 0.01-0.5 mg / cm². Direct use of these parameters would lead to unstable model training. Therefore, a minimum and maximum normalization method is used to linearly scale the value of each parameter to the [0, 1] interval. For discrete parameters (weather type: sunny, rainy, foggy, etc.; insulator material type: porcelain, glass, composite silicone rubber, etc.), there is no inherent order relationship between them. One-hot encoding is used to create a binary vector for each possible value of each discrete parameter (the corresponding position is 1, and the rest are 0), thereby converting them into numerical form and avoiding the introduction of incorrect order relationships. After all parameters have been converted into numerical vectors, the normalized continuous parameter vectors corresponding to this timestamp and the one-hot encoded discrete parameter vectors are concatenated in a predetermined order along the feature dimension to form a high-dimensional comprehensive parameter vector. While this vector contains all the information, it is high-dimensional and may contain redundancy and complex interactions between features. Therefore, it is fed into a small, fully connected neural network. This network learns the complex interactions between different parameters through its non-linear activation function and automatically extracts the most effective feature combinations. Simultaneously, the network reduces the high-dimensional input to a pre-defined, lower-dimensional, dense vector, which is the final context feature vector.

[0032] S23. Extract time-series features from the historical risk score sequence to obtain a trend feature vector; if it is the first assessment and there is no historical data, set the trend feature vector to zero or use the default initial value; S231. The historical risk score sequence is used as input and fed into a pre-trained temporal feature extraction model; the temporal feature extraction model is a recurrent neural network; S232. Use a time-series feature extraction model to encode the historical risk score sequence and output a fixed-dimensional vector as a trend feature vector; If the length of the historical risk score sequence is variable or there are missing values, then length normalization or missing value imputation is performed before inputting it into the model.

[0033] The purpose of this step is to quantitatively extract the trend characteristics of risk from this time series, such as whether the risk is in a stable, rising, or falling channel, and the degree of change. This trend feature vector serves as a key temporal context for dynamic risk assessment, forming a more comprehensive judgment basis together with the acoustic and environmental characteristics of the current moment. For the first assessment or in cases where the historical sequence is empty due to data loss, the system sets the trend feature vector to a zero vector or a default initial value representing no historical trend to ensure the integrity and robustness of the assessment process. Specifically, this is achieved through a pre-trained time series model. The historical risk score sequence arranged in chronological order is input into a specially designed LSTM recurrent neural network. Through its internal gating mechanism, the network can effectively learn the long-term dependencies in the sequence, ultimately encoding the entire variable-length historical sequence into a fixed-dimensional trend feature vector rich in temporal information. In actual processing, to address the issues of inconsistent historical data lengths or missing values, preprocessing is performed before inputting the data into the model. This includes filling in missing values ​​using interpolation and normalizing the sequence to a uniform length (e.g., through truncation or zero padding) to ensure the standardization and stability of the model processing.

[0034] S3. The acoustic feature vector, context feature vector, and trend feature vector are fused, and the target dynamic risk score is calculated based on the fused features; Since acoustic features, contextual features, and trend features reflect the insulator's operating status from different dimensions (instantaneous discharge state, environmental background conditions, and historical evolution trends), no single feature can comprehensively assess the risk. Therefore, this step deeply fuses these three heterogeneous but complementary feature vectors and transforms them into an intuitive and quantitative risk indicator through a nonlinear mapping function, providing a direct basis for operation and maintenance decisions.

[0035] S31. Concatenate the acoustic feature vector, context feature vector, and trend feature vector to generate a comprehensive feature vector; Dimension of this acoustic feature vector Context feature vector dimension Trend feature vector dimension Three vectors are concatenated along the feature dimension in a specific order to generate a composite feature vector. Its mathematical representation is: in: This is a vector concatenation operation. The dimension is .

[0036] S32. Input the comprehensive feature vector into the pre-trained risk assessment model, which is a deep neural network containing multiple fully connected layers; S33. The comprehensive feature vector is transformed nonlinearly through the risk assessment model, and a continuous value is output as the target dynamic risk score. The target dynamic risk score ranges from [0,1], and the larger the value, the higher the risk.

[0037] The pre-trained risk assessment model is a specially designed deep feedforward neural network whose core function is to learn the complex mapping relationship from high-dimensional integrated features to risk scores. The model consists of an input layer, multiple hidden layers, and an output layer, and the computation process is as follows: 1. Input layer: Receives the synthesized feature vector. ; 2. Hidden Layers: Each hidden layer undergoes the following transformation: in: For the first The activation value of the layer, = , and The first The weight matrix and bias vector of the layer, As a non-linear activation function, through multiple such hidden layers, the model can progressively learn higher-order interactions and non-linear relationships between features.

[0038] 3. Output Layer: The output of the last hidden layer is passed to the output layer, which is a single neuron. The sigmoid activation function is used to compress the output value to the (0, 1) interval, which serves as the final risk score. : in: This is the output of the last hidden layer. and It refers to the weights and biases of the output layer; the sigmoid function ensures the output... It is a continuous value between 0 and 1, which intuitively represents the probability or degree of risk. The closer the value is to 1, the higher the risk.

[0039] S4. Compare the target dynamic risk score with multiple preset risk threshold intervals and map them to the corresponding discrete risk levels.

[0040] The implementation of risk level mapping is based on a preset threshold comparison principle. The system maintains a configuration that includes three key thresholds (T1, T2, T3), where 0 < T1 < T2 < T3 < 1. The determination of these thresholds is based on the statistical analysis of historical fault data, expert experience, and the distribution of model outputs under different operating conditions, aiming to best distinguish different operating state stages of insulators. The mapping process is an interval judgment logic: The system will calculate the target dynamic risk score and compare it with thresholds T1, T2, and T3 in sequence; according to the interval it belongs to, map it to the corresponding discrete risk level. The specific logic is as follows: When the target dynamic risk score is less than T1, it is mapped to the normal risk level; at this stage, the insulator is in good operating condition, there is no significant discharge risk, and only routine monitoring is required.

[0041] When the target dynamic risk score is greater than or equal to T1 and less than T2, it is mapped to the warning risk level; this stage indicates that early deterioration signs or risks of the insulator begin to accumulate, and it is necessary to strengthen the monitoring frequency, pay attention to environmental changes, and prepare contingency plans.

[0042] When the target dynamic risk score is greater than or equal to T2 and less than T3, it is mapped to the alarm risk level; this stage indicates that the discharge activity intensifies, the risk significantly increases, and there is a possibility of developing into a serious fault. It is necessary to generate a patrol work order and prepare to arrange for intervention inspection or cleaning.

[0043] When the target dynamic risk score is greater than or equal to T3, it is mapped to the maintenance risk level; this stage indicates that the insulator is in a high-risk state, and the possibility of flashover is extremely high. It is necessary to immediately take control measures (such as remotely adjusting the operating mode) and dispatch a maintenance team for emergency disposal.

[0044] The final output of this step is a clear and easy-to-understand discrete risk level label ("normal", "warning", "alarm", "maintenance"). This result will be directly presented to the grid operation and maintenance personnel and may trigger corresponding visual alarms in the centralized control center monitoring system, generate patrol work orders, and start an automated warning information push process.

[0045] It can be understood that those skilled in the art can, under the guidance of the above embodiments, combine various implementation manners in the above embodiments to obtain technical solutions of various implementation manners.

[0046] The above are only preferred embodiments of the present invention and are not intended to limit the present invention. Any modifications, equivalent replacements, and improvements made within the spirit and principle of the present invention shall be included in the protection scope of the present invention.

Claims

1. A method for acoustic identification and prediction of insulator operating status based on SSA-LSTM, characterized in that, include: The time series of acoustic emission signals of the insulator to be evaluated under operating conditions, the environmental parameters of the micrometeorological environment, and the recent historical risk score sequence of the insulator to be evaluated are obtained; the state parameters include at least the equivalent salt density and the insulator material type. Signal preprocessing and feature extraction are performed on the original acoustic emission signal sequence, the micro-meteorological environment parameters, the state parameters, and the historical risk score sequence, respectively; wherein: The original sequence of the acoustic emission signal is denoised, and acoustic feature vectors are extracted using an LSTM neural network. The micrometeorological environment parameters and the state parameters are jointly encoded into a context feature vector; The historical risk score sequence is subjected to time series feature extraction to obtain a trend feature vector; if it is the first assessment and there is no historical data, the trend feature vector is set to zero vector or the default initial value is used. The acoustic feature vector, the context feature vector, and the trend feature vector are fused, and the target dynamic risk score is calculated based on the fused features. The target dynamic risk score is compared with multiple preset risk threshold intervals and mapped to the corresponding discrete risk level.

2. The method for acoustic identification and prediction of insulator operating status based on SSA-LSTM according to claim 1, characterized in that, The denoising process of the original acoustic emission signal sequence and the extraction of acoustic feature vectors using an LSTM neural network include: The denoised acoustic signal sequence is standardized and constructed into a sequence format for the input of the LSTM network; A pre-trained LSTM acoustic feature extraction model is used to extract temporal features from the standardized sequence to obtain acoustic feature vectors.

3. The method for acoustic identification and prediction of insulator operating status based on SSA-LSTM according to claim 2, characterized in that, The establishment of the pre-trained LSTM acoustic feature extraction model includes the following steps: Acquire a sample set of acoustic emission signals containing multiple insulator discharge modes, and train an initial LSTM network with the sample set to enable the initial LSTM network to acquire the basic ability to identify insulator discharge modes from acoustic signal sequences. Use the trained LSTM network as the pre-trained acoustic representation basic model. A feature adaptation module is connected to the output of the pre-trained acoustic representation basic model to form a complete acoustic feature extraction model. The acoustic feature extraction model is fine-tuned using acoustic emission signal data containing complex ambient noise. During the fine-tuning process, the parameters of the pre-trained LSTM acoustic representation base model are fixed, and only the parameters of the feature adaptation module are updated.

4. The method for acoustic identification and prediction of insulator operating status based on SSA-LSTM according to claim 2 or 3, characterized in that, During the model training phase, one or more data augmentation perturbations are applied to the original acoustic emission signal sequence. These perturbations include adding random noise, performing temporal stretching, or amplitude scaling. The perturbated acoustic signal sequence is then combined with the original acoustic signal sequence for denoising and subsequent processing, which is then used to train the acoustic feature extraction model.

5. The method for acoustic identification and prediction of insulator operating status based on SSA-LSTM according to claim 1, characterized in that, The step of encoding the micrometeorological environment parameters and the state parameters together into a context feature vector includes: Continuous parameters are normalized, and discrete parameters are encoded using one-hot encoding. All processed parameters are concatenated into a single comprehensive parameter vector; The integrated parameter vector is input into a fully connected neural network for nonlinear transformation and dimensionality reduction, and the context feature vector is output.

6. The method for acoustic identification and prediction of insulator operating status based on SSA-LSTM according to claim 5, characterized in that, The continuous parameters include humidity, temperature, and equivalent salinity, while the discrete parameters include weather type and insulator material type.

7. The method for acoustic identification and prediction of insulator operating status based on SSA-LSTM according to claim 1, characterized in that, The step of extracting time-series features from the historical risk score sequence to obtain a trend feature vector includes: The historical risk score sequence is used as input and fed into a pre-trained temporal feature extraction model; the temporal feature extraction model is a recurrent neural network. The historical risk score sequence is encoded using a time-series feature extraction model, and a fixed-dimensional vector is output as the trend feature vector.

8. The method for acoustic identification and prediction of insulator operating status based on SSA-LSTM according to claim 7, if the length of the historical risk score sequence is variable or there are missing values, then length normalization or missing value imputation processing is performed before inputting it into the model.

9. The method for acoustic identification and prediction of insulator operating status based on SSA-LSTM according to claim 1, characterized in that, The process of fusing the acoustic feature vector, the context feature vector, and the trend feature vector, and calculating the target dynamic risk score based on the fused features, includes: The acoustic feature vector, the context feature vector, and the trend feature vector are concatenated to generate a comprehensive feature vector; The comprehensive feature vector is input into a pre-trained risk assessment model, which is a deep neural network containing multiple fully connected layers; The comprehensive feature vector is nonlinearly transformed by the risk assessment model to output a continuous value as the target dynamic risk score. The target dynamic risk score ranges from [0,1], with a larger value indicating a higher risk.

10. The method for acoustic identification and prediction of insulator operating status based on SSA-LSTM according to claim 1, characterized in that, The step of comparing the target dynamic risk score with multiple preset risk threshold intervals and mapping it to a corresponding discrete risk level includes: Three risk thresholds are preset, including a first threshold T1, a second threshold T2, and a third threshold T3, where 0 <T1<T2<T3<1; When the target dynamic risk score is less than T1, it is mapped to a normal risk level; When the target dynamic risk score is greater than or equal to T1 and less than T2, it is mapped to the early warning risk level; When the target dynamic risk score is greater than or equal to T2 and less than T3, it is mapped to an alarm risk level; When the target dynamic risk score is greater than or equal to T3, it is mapped to the maintenance risk level.

Images