Switch cabinet life prediction method and device, and program product

Abstract

The invention provides a switch cabinet life prediction method and device, and a program product, and the method comprises the steps: collecting a prediction model based on LSTM improvement, and enabling the prediction model to comprise a feature engineering module, an intelligent door control module, and an output module, and carrying out the fusion of the collected state data of a plurality of switch cabinets through the feature engineering module; an intelligent gating module obtains the hidden state of each moment by adopting LSTM for the fusion feature, and finally an output module obtains the attention weight of the hidden state of each moment through an attention mechanism, and carries out the weighted fusion of the hidden state of each moment based on the attention weight to obtain a global attention feature. And the residual life is obtained through full connection layer processing. The prediction method is clear in process and high in operability, the residual life of the switch cabinet can be accurately evaluated through the steps of data preprocessing, model training, verification and the like, a reliable decision basis is provided for operation and maintenance personnel of power equipment, the operation and maintenance cost is reduced, and the operation reliability of a power system is improved.

Description

Technical Field

[0001] This invention belongs to the field of data processing and power equipment condition assessment technology, specifically relating to a method, device, and program product for predicting the lifespan of switchgear. Background Technology

[0002] As a key electrical device in a power system, switchgear plays a vital role in power distribution, line protection, and control. Its operating status directly affects the safety, stability, and reliability of the power system. With the increasing years of operation of the power system, switchgear will gradually exhibit deterioration phenomena such as insulation aging, mechanical component wear, and increased contact resistance under the combined effects of electrical, thermal, mechanical stress, and environmental factors (such as humidity, pollution, and temperature changes).

[0003] Currently, methods for assessing the remaining life of switchgear can be mainly divided into three categories: traditional experience-based assessment, physical model-based assessment, and data-driven assessment. Traditional experience-based assessment relies on the on-site experience of maintenance personnel and historical equipment failure data. The assessment results are highly subjective, have low accuracy, and are difficult to adapt to complex and changing operating environments. Physical model-based assessment describes the degradation process by establishing physicochemical equations for equipment degradation, but it requires a deep understanding of the internal degradation mechanisms of the equipment, and the model parameters are difficult to determine precisely, resulting in poor adaptability to complex operating conditions. Data-driven assessment, utilizing algorithms such as machine learning and deep learning, assesses the remaining life by uncovering hidden patterns in equipment operating status data. It has advantages such as not requiring explicit knowledge of degradation mechanisms and strong adaptability, and has become a research hotspot in recent years.

[0004] Long Short-Term Memory (LSTM) networks, as a classic time-series data processing model, can effectively capture long-short-term dependencies in time-series data and have been widely used in equipment remaining life assessment. However, when using traditional LSTM to process switchgear status data, the data contains a large amount of noise (such as sensor errors, electromagnetic interference, etc.); traditional LSTM is sensitive to noisy data, which can easily lead to a decrease in the model's generalization ability. Summary of the Invention

[0005] To address the shortcomings of existing technologies, this invention provides a method, device, and program product for predicting the lifespan of switchgear, which solves the noise resistance problem of traditional LSTM and improves the accuracy of switchgear lifespan prediction.

[0006] The present invention achieves the above-mentioned technical objectives through the following technical means.

[0007] A method for predicting the lifespan of switchgear, employing the following switchgear lifespan prediction model, includes:

[0008] Feature Engineering Module: Fusion of various switchgear status data collected, and output of fused features arranged in time sequence to the intelligent door control module;

[0009] Intelligent gating module: LSTM is used to obtain the hidden state at each time step based on the fused features;

[0010] Output module: The attention weights of the hidden states at each time step are obtained through the attention mechanism. The hidden states at each time step are weighted and fused based on the attention weights to obtain the global attention features. The remaining lifetime is then obtained through the fully connected layer.

[0011] Furthermore, in the output module:

[0012] By using causal convolution, multiple hidden states at different scales at various time points are extracted. subscript in the formula For scaling variables, subscript It is a time variable;

[0013] Hidden state based on the output of the intelligent gating module and the extracted hidden states at multiple different scales High-dimensional features are obtained by splicing. ;

[0014] Based on high-dimensional features Attention weights at each time step are generated using a parameter generation network. Attention bias Note the scoring vector ;

[0015] Based on attention weights and attention bias ,right Perform a linear transformation to obtain ;

[0016] right Perform feature cross processing to obtain ;

[0017] right and Perform hybrid activation to obtain gating coefficients at various scales. ;

[0018] Based on gating coefficient And attention rating vector Obtain single-scale scores at each scale A multi-scale comprehensive score is obtained by weighted summation. Multi-scale comprehensive scoring The attention weights are obtained by normalization. .

[0019] Furthermore, in the output module, global attention features The calculation is as follows:

[0020]

[0021]

[0022] In the formula, The maximum sampling time, This represents the globally average hidden state.

[0023] Furthermore, in the output module, the Softmax function is used for multi-scale comprehensive scoring. Perform normalization to obtain attention weights :

[0024]

[0025] In the formula, This is the maximum sampling time.

[0026] Furthermore, in the output module:

[0027] High-dimensional features The splicing also includes feature contribution factors. Scene state factors, including degradation phase. and noise intensity ;

[0028] Feature contribution factor The calculation is as follows:

[0029]

[0030] In the formula, and For adaptive weights, subscripts Same meaning , for Different feature sources at different times The characteristic variance between them This is the maximum sampling time;

[0031] Degraded phase The calculation is as follows:

[0032] Calculate the cumulative percentage increase of partial discharge, insulation resistance, and loop resistance respectively:

[0033]

[0034] In the formula, subscript The values ​​1, 2, and 3 represent the partial discharge quantity, insulation resistance, and circuit resistance, respectively.

[0035] The percentage of the overall cumulative increment is calculated as follows: :

[0036]

[0037] Calculate the degradation phase as follows: :

[0038]

[0039] Noise intensity The calculation is as follows:

[0040]

[0041]

[0042]

[0043] In the formula, The absolute value of the residual. This is the smoothed value of the moving average. This is the output of the forget gate in the intelligent gate control module.

[0044] Furthermore, in the output module:

[0045] Feature cross processing is as follows:

[0046]

[0047] In the formula, " indicates element-wise multiplication. For feature-crossing networks;

[0048] Hybrid activation is:

[0049]

[0050] In the formula, The SiLU activation function is used. It is the hyperbolic tangent activation function;

[0051] Calculate the comprehensive gating coefficient as follows: :

[0052]

[0053]

[0054]

[0055] In the formula, It is the sigmoid activation function. Use the GELU activation function;

[0056] Single-scale rating The calculation is as follows:

[0057]

[0058] Multi-scale comprehensive score The calculation is as follows:

[0059]

[0060] in The scale weights are calculated as follows:

[0061]

[0062] In the formula, Same meaning , As the baseline hidden state, This is the cosine similarity function.

[0063] Furthermore, the intelligent gate control module includes:

[0064] Characteristic change rate factor calculate:

[0065]

[0066] In the formula, For the fusion feature, Characteristic rate of change factor, subscript For time variables, To avoid constants with a denominator of 0;

[0067] Forgotten Gate:

[0068]

[0069] In the formula, for The output of the forget gate at all times It is the sigmoid activation function. Here is the forget gate weight matrix. for The hidden state at all times Forget gate bias term; This is the rate of change weighting coefficient in the forgetting gate, and its value ranges from [0.1, 0.5].

[0070] Input Gate:

[0071]

[0072] In the formula, for The output of the input gate at any time. The input gate weight matrix, For input gate bias terms; This is the rate-of-change weighting coefficient in the input gate, and its value ranges from [0.1, 0.5].

[0073] Cell state:

[0074]

[0075]

[0076] In the formula, for Cellular state at any given moment for The state of candidate cells at any given time. The hyperbolic tangent activation function is used. This is the cell state weight matrix. This is a cell state bias term;

[0077] Output gate:

[0078]

[0079]

[0080] In the formula, for The output of the output gate at all times. for The hidden state at all times This is the output gate weight matrix. This is the output gate bias term.

[0081] Furthermore, in the feature engineering module, the collected data on various switchgear statuses are fused as follows:

[0082]

[0083] In the formula, As a feature of fusion, and For each feature source feature value, corresponding to each switch cabinet status data, the subscript is... and Indicates the feature source number, subscript For time variables, The total number of feature sources, and For each feature source, there are adaptive fusion weights, and... , This represents the weight coefficient for the interaction item, with a value range of [0.05, 0.2].

[0084] The collected switchgear status data includes:

[0085] Electrical characteristic data: partial discharge quantity, loop resistance, insulation resistance, three-phase current / voltage;

[0086] Environmental characteristic data: ambient temperature, relative humidity, and degree of pollution;

[0087] Mechanical characteristic data: operating mechanism action time, vibration amplitude;

[0088] The preprocessing of the above data includes:

[0089] Data cleaning: Outliers in the original data are removed using the 3σ criterion, and missing values ​​are filled using linear interpolation.

[0090] Normalization: The min-max normalization method is used;

[0091] Time series alignment: Based on timestamps, time series data from different sources are aligned to the same timeline.

[0092] A computer device, including a memory and a processor;

[0093] The memory is used to store computer programs;

[0094] The processor is used to execute the computer program and, in executing the computer program, implement the above-described switchgear life prediction method.

[0095] A computer program product includes a computer program that, when executed by a processor, implements the above-described switchgear life prediction method.

[0096] The beneficial effects of this invention are as follows:

[0097] (1) This invention provides a method for predicting the lifespan of switchgear, wherein a temporal attention mechanism is introduced at the output of the model to assign weights to the hidden states of the LSTM output, highlighting the contribution of key temporal features, reducing the impact of noise data on the evaluation results, thereby significantly improving the model's noise resistance and evaluation accuracy.

[0098] (2) Compared to the traditional LSTM, which has a fixed weight update mechanism for the forget gate, input gate, and output gate and fails to dynamically adjust according to the characteristics of different degradation stages of the switchgear, making it difficult to accurately capture key information of different degradation stages. This invention introduces a feature change rate factor into the LSTM's forget gate and input gate to dynamically adjust the gating output, enabling the model to adaptively adjust the information forgetting and update strategy according to the feature change rate of different degradation stages of the switchgear, thereby enhancing the model's ability to capture features of different degradation stages and improving the model's adaptability.

[0099] (3) The present invention adds a multi-source feature fusion module (i.e. feature engineering module) to the model and adopts an adaptive weighted fusion strategy to fuse the electrical, environmental and mechanical multi-source features of the switch cabinet. It makes full use of the complementary information of different types of features for the remaining life assessment and solves the problem that the traditional LSTM relies on only a single feature, resulting in insufficient assessment accuracy.

[0100] (4) The prediction method of the present invention has a clear process and strong operability. Through data preprocessing, model training and verification, it can accurately assess the remaining life of switchgear, provide reliable decision-making basis for power equipment operation and maintenance personnel, and help reduce operation and maintenance costs and improve the reliability of power system operation. Attached Figure Description

[0101] Figure 1 This is a flowchart of the switchgear life prediction method of the present invention;

[0102] Figure 2 This is a structural diagram of the intelligent door control module in the switchgear life prediction model of this invention;

[0103] Figure 3 This is a structural diagram of the output module in the switchgear life prediction model of the present invention;

[0104] Figure 4 The figures show a comparison of the evaluation performance of the prediction model of this invention with that of the conventional LSTM model; (a) shows the comparison of the evaluation performance of the remaining life of the switchgear; and (b) shows the comparison of the prediction error distribution. Detailed Implementation

[0105] The embodiments of the present invention are described in detail below. Examples of the embodiments are shown in the accompanying drawings. The embodiments described below with reference to the accompanying drawings are exemplary and intended to explain the present invention, but should not be construed as limiting the present invention.

[0106] I. Technical Solution

[0107] like Figure 1 As shown, this embodiment provides the following method for predicting the remaining lifespan of switchgear:

[0108] 1. Multi-source status data acquisition and preprocessing for switchgear

[0109] 1.1 Multi-source status data during the operation of the switchgear is collected through a sensor array deployed on the switchgear, including but not limited to:

[0110] (1) Electrical characteristic data: partial discharge quantity, loop resistance, insulation resistance, three-phase current / voltage;

[0111] (2) Environmental characteristic data: ambient temperature, relative humidity, and degree of pollution;

[0112] (3) Mechanical characteristic data: operating mechanism action time and vibration amplitude.

[0113] 1.2 The collected raw data is preprocessed, including data cleaning, normalization, and time-series alignment, to obtain a standardized multi-source time-series feature dataset, wherein:

[0114] 1) Data cleaning

[0115] Outliers in the original data were removed using the 3σ criterion; for missing values, linear interpolation was used to fill them in.

[0116] 2) Normalization

[0117] The min-max normalization method is used to map feature data of different dimensions to the [0,1] interval.

[0118] 3) Timing alignment

[0119] Using timestamps as a reference, time series data from different sources are aligned to the same time axis to obtain multi-source time series feature sequences of uniform length.

[0120] 2. Switchgear life prediction model

[0121] For the preprocessed multi-source time-series feature data, an improved LSTM-based prediction model is used to predict the lifespan of the switchgear. The prediction model consists of three main modules, from front to back: ① Feature engineering module, ② Intelligent gate control module, and ③ Output module.

[0122] 2.1 Feature Engineering Module

[0123] To address the issue of inconsistent dimensions and varying contributions of multi-source heterogeneous data (electrical, mechanical, environmental, etc.) from switchgear, this invention designs an end-to-end adaptive weighted fusion method. The weight coefficients are not manually set but are automatically learned through backpropagation during model training. This enables data-driven feature importance discrimination, dynamically capturing subtle changes in the impact of various features on remaining lifespan under different operating conditions (such as high temperature and high load in summer, and low temperature in winter), thereby maximizing the complementary value of multi-source data.

[0124] The formula for adaptive weighted fusion of multi-source features is:

[0125]

[0126] In the formula, As a feature of fusion, and The subscripts represent the feature values ​​of each feature source (after preprocessing). and Indicates the feature source number (e.g., specifying the feature "partial discharge quantity" as number 1), subscript express time, The total number of feature sources, and For each feature source, there are adaptive fusion weights, and... Its specific value is obtained through back-learning during model training, and the initial value can be set to... , The interaction term weight coefficient has a value range of [0.05, 0.2] and is determined by the validation set optimization.

[0127] 2.2 Intelligent gate control module

[0128] like Figure 2 As shown, this invention creatively introduces a feature change rate factor into the forget and input gates, abandoning the fixed gating update strategy of traditional LSTM. This allows the model to not only focus on the "current state" of the switchgear but also perceive "how fast its state changes." This mechanism can dynamically identify the inflection point from slow aging to accelerated degradation, enabling it to adaptively adjust the memory and forgetting strategies according to the actual health decline rate of the equipment, significantly improving the model's accuracy in assessing the adaptability of different degradation stages.

[0129] Specifically as follows:

[0130] (1) Characteristic rate of change factor

[0131] Calculate the characteristic rate of change factor as follows: :

[0132]

[0133] In the formula, for The characteristic rate of change factor at time, Used to avoid a denominator of 0.

[0134] (2) Gate of Oblivion

[0135] Calculate the output of the forget gate as follows: :

[0136]

[0137] In the formula, for The output of the forget gate at all times It is the sigmoid activation function. Here is the forget gate weight matrix. for The hidden state at all times Forget gate bias term; is the rate of change weighting coefficient in the forgetting gate, with a value range of [0.1, 0.5], used to adjust the influence of the perceptual factor (i.e., the feature rate of change factor).

[0138] (3) Input gate

[0139] Calculate the output of the input gate as follows: :

[0140]

[0141] In the formula, for The output of the input gate at any time. The input gate weight matrix, For input gate bias terms; is the rate of change weighting coefficient in the input gate, and its value ranges from [0.1, 0.5].

[0142] (4) Cell state

[0143] Update cell status as follows:

[0144]

[0145]

[0146] In the formula, for Cellular state at any given moment for The state of candidate cells at any given time. The hyperbolic tangent activation function is used. This is the cell state weight matrix. This is a cell state bias term.

[0147] (5) Output gate

[0148] Calculate the output of the input gate as follows: and hidden state :

[0149]

[0150]

[0151] In the formula, for The output of the output gate at all times. for The hidden state at all times This is the output gate weight matrix. This is the output gate bias term.

[0152] 2.3 Output Module

[0153] Reference Figure 3 As shown, this embodiment introduces a temporal attention mechanism at the model output. This mechanism can automatically assign high weights to key time points in the historical sequence that are most relevant to the end of life (such as multiple abnormal operations or continuous insulation degradation stages), while ignoring invalid or noise information caused by transient interference (such as electromagnetic noise or false alarms). This makes the model evaluation results more stable and reliable, effectively reduces false alarm and false alarm rates, and improves robustness in complex industrial environments.

[0154] (1) Multi-scale hidden states

[0155] This invention sets three different time scales—short, medium, and long—and extracts the hidden states at each scale through causal convolution to avoid future information leakage.

[0156]

[0157] In the formula, For scale The hidden state below, For causal convolution, For the window length, in this embodiment The values ​​3, 7, and 30 correspond to short, medium, and long time scales, respectively. and These are the first two causal convolutions. Group weights and biases.

[0158] (2) Characteristic contribution factor

[0159] Based on feature-adaptive weights and temporal variance, the contribution ratio of features to the current state at different times is quantified:

[0160]

[0161] In the formula, for The characteristic contribution factor at time, and For adaptive weights, subscripts Same meaning It is only used in the summation formula with external [variables]. Make distinctions. It is the variance function. The calculation is Different feature sources at different times The characteristic variance between them This represents the length of the sampling period.

[0162] (3) Scene state factor

[0163] Introduce scenario variables that are strongly related to the task, including: degradation phase. Noise intensity Specifically, it is calculated from the actual data of the task scenario, where:

[0164] (3.1) Degraded phase ( The time interval (in seconds) is calculated using three characteristic quantities: partial discharge quantity, insulation resistance, and loop resistance.

[0165] Let the feature source numbers of the three be... If the values ​​are 1, 2, and 3 respectively, then the three... Eigenvalues ​​at time 1 Recorded in order as , , .

[0166] Calculate the cumulative increment percentage of each of the above three factors:

[0167]

[0168] In the formula, This formula is used to count only positive increments; the partial discharge quantity, insulation resistance, and loop resistance are calculated using the above formula. Cumulative increment percentage at any given moment , , .

[0169] The cumulative incremental percentages of the three factors are weighted and summed according to the following weighting to obtain the overall cumulative incremental percentage. :

[0170]

[0171] Finally, the degraded phase is obtained using the following mapping function. :

[0172]

[0173] (3.2) Noise intensity ( The time (time) is calculated as follows:

[0174] Calculate the smoothing value of the moving average :

[0175]

[0176] In the formula, the sliding window size is 10.

[0177] Calculate the absolute value of the residual :

[0178]

[0179] Calculate noise intensity :

[0180]

[0181] (4) Multi-source input splicing

[0182] The multi-dimensional information obtained above is concatenated into high-dimensional features. ( time):

[0183]

[0184] (5) Dynamic weights and biases

[0185] This invention abandons fixed weights and biases, and adopts a parameter generation network based on the current high-dimensional features. The following parameters are generated dynamically:

[0186]

[0187]

[0188]

[0189] In the formula, , , They are respectively Attention weights, attention biases, and attention score vectors at each moment. , , These are the parameter generation networks used to generate the three dynamic parameters mentioned above.

[0190] (6) Linear transformation

[0191] Based on the above and Hidden states at each scale Perform a linear transformation as :

[0192]

[0193] (7) Feature Cross

[0194] Follow the steps below Perform feature crossing processing to obtain :

[0195]

[0196] In the formula, " indicates element-wise multiplication. It is a feature cross-network.

[0197] (8) Hybrid activation

[0198] For different scales The gating coefficient is obtained by performing mixed activation as follows. :

[0199]

[0200] In the formula, The SiLU activation function is used to enhance gradients. Used to enhance nonlinearity.

[0201] (9) Gated modulation

[0202] Calculate the gate control coefficient of the noise suppression gate as follows: :

[0203]

[0204] Calculate the gate control factor of the deteriorated and reinforced door as follows: :

[0205]

[0206] In the formula, Use the GELU activation function;

[0207] Calculate the comprehensive gating coefficients at each scale as follows: :

[0208]

[0209] In the above process, The larger the value, the stronger the gating effect, and the better it can suppress noise during scoring. The larger the value, the stronger the gating effect, which can amplify the scoring at critical stages.

[0210] (10) Multiscale scoring

[0211] Calculate scores for short, medium, and long scales separately, and aggregate them based on scale weights:

[0212] Single-scale rating :

[0213]

[0214] In the formula, for transpose;

[0215] Scale weight :

[0216]

[0217] In the formula, Same meaning , The preset baseline hidden state, The cosine similarity function; The hidden state can be obtained by preparing a batch of switchgear test samples in advance, measuring the hidden state of each sample at the end of its lifespan, and taking the average value as... .

[0218] Multi-scale comprehensive score :

[0219]

[0220] (11) Attention weights

[0221] Using the Softmax function, Multi-scale comprehensive score at each time point within the time period After normalization, the attention weights at each time step are obtained. :

[0222]

[0223] In the formula, It is an exponential function.

[0224] (12) Residual connection

[0225] Based on attention weights The attention features at each time step are obtained by summing the hidden state at each time step with the global average hidden state. :

[0226]

[0227]

[0228] In the formula, For the globally average hidden state, the residual term It is used to supplement global basic information and avoid feature distortion caused by excessively high or low weights at a certain moment (for example, during the stable operation of the switch cabinet, the features at a single moment are not abnormal, but the global average can retain the basic state).

[0229] Will Attention characteristics at different times within the time period Summing yields the global attention features. :

[0230]

[0231] (13) Remaining life prediction

[0232] Global attention features Input a fully connected layer and calculate the remaining lifetime. :

[0233]

[0234] In the formula, and These are the weights and biases of the fully connected layer, respectively.

[0235] 3. Model Training

[0236] For the switchgear life prediction model constructed above, during training:

[0237] 1) The loss function uses the root mean square error;

[0238] 2) The Adam optimizer is used to update the model parameters. The initial learning rate is set to 0.001, and the learning rate is adjusted by a learning rate decay strategy, specifically, the learning rate decays to 0.9 every 100 iterations.

[0239] II. Testing and Verification

[0240] This test uses a 10kV switchgear as the evaluation object, and the specific implementation process is as follows:

[0241] 1. Multi-source status data acquisition and preprocessing for switchgear

[0242] 1.1 Data Acquisition: By deploying partial discharge sensors, temperature sensors, humidity sensors, vibration sensors, and current / voltage sensors inside the switchgear, the operating status data of the switchgear was collected for 6 consecutive months. The sampling frequency was 1 time / hour, and a total of 4320 sets of data were obtained. The collected features included: partial discharge quantity (pC), loop resistance (mΩ), insulation resistance (MΩ), three-phase current (A), three-phase voltage (kV), ambient temperature (°C), relative humidity (%), vibration amplitude (mm / s), and operating mechanism action time (ms).

[0243] 1.2 Data Cleaning: Outliers in each feature were removed using the 3σ criterion, resulting in the removal of 128 sets of outlier data. For the 36 missing sets of data, linear interpolation was used to fill in the missing data.

[0244] 1.3 Normalization: The min-max normalization method is used to map all feature data to the [0,1] interval;

[0245] 1.4 Time alignment: Based on the timestamp, the data collected by different sensors are aligned to a time interval of 1 hour to obtain a multi-source time-series feature sequence with a length of 4228.

[0246] 2. Prediction Model

[0247] 2.1 Feature Engineering Module: Number of feature sources n=9, initial weight coefficients, optimal weights are obtained through adaptive learning in the subsequent training process (e.g., the weight of the ambient temperature feature is increased to 0.18 under high temperature conditions in summer and adjusted to 0.12 under low temperature conditions in winter); interaction term weight coefficients;

[0248] 2.2 Intelligent Gating Module: The number of neurons in the hidden layer is set to 64, and the rate of change weight coefficient is... , The activation functions used are sigmoid and tanh, and the initial values ​​of the weight matrix and bias terms are initialized using Xavier. The feature change rate factor can accurately identify the inflection point of accelerated insulation aging that appears in the 4th month of this switchgear.

[0249] 2.3 Output Module: Attention Scoring Vector The attention weight matrix has a dimension of 64. The dimension is 64×64, with bias terms. The dimension is 64; through attention weight allocation, high weight (average weight 0.35) is assigned to key time-series data that continue to deteriorate insulation after the 4th month, and low weight (average weight 0.08) is assigned to transient abnormal data caused by electromagnetic interference; the output dimension of the fully connected layer is 1, corresponding to the remaining lifetime assessment value.

[0250] 3. Model Training and Validation

[0251] 3.1 Dataset partitioning: The 4228 preprocessed multi-source time-series feature sequences were divided into a training set (2959 sets) and a validation set (1269 sets) in a 7:3 ratio. The fused feature sequences were used as input, and the actual remaining lifespan of the switchgear (obtained through a full life cycle test) was used as the label.

[0252] 3.2 Model Training: RMSE was used as the loss function, and the Adam optimizer was used to update the parameters. The initial learning rate was 0.001, the number of iterations was 500, and the learning rate was decayed to 0.9 every 100 iterations.

[0253] 3.3 Model Validation: The trained model was validated using a validation set. By adjusting hyperparameters such as the number of hidden layer neurons (trying 32, 64, and 128) and the rate of change weight coefficient (trying 0.1-0.5), the optimal model parameters were finally determined: 64 hidden layer neurons, at which point the RMSE of the validation set was 0.052.

[0254] 4. Remaining life assessment of switchgear

[0255] The latest 100 sets of preprocessed multi-source status data of the 10kV switchgear to be evaluated were input into the trained improved LSTM model. The model output a remaining lifetime assessment value of 8.2 years, with an error of 3.5% compared with the actual remaining lifetime (8.5 years). The assessment accuracy meets the engineering requirements.

[0256] Comparative Example: Using a traditional LSTM model to assess the remaining lifetime of the same set of data, the validation set RMSE was 0.098. The estimated remaining lifetime of the switchgear under assessment was 7.5 years, with an error of 11.8% compared to the actual value. Specific comparison results are as follows: Figure 4 As shown in the figure. The results show that the improved LSTM model of this invention has significantly better evaluation accuracy than the traditional LSTM model, and truly has the potential to replace traditional empirical methods and simple model methods, providing a directly deployable technical tool for predictive maintenance of power systems.

[0257] III. Devices, storage media, and software products

[0258] 1. Based on the same inventive concept as the above-described switchgear life prediction method, this application also provides an electronic device, which includes a processor and a memory, wherein computer-readable code is stored in the memory, and when the computer-readable code is executed by the processor, the switchgear life prediction method of the present invention is implemented.

[0259] The memory includes a non-volatile storage medium and internal memory; the non-volatile storage medium can store the operating system and computer-readable code. The computer-readable code includes program instructions that, when executed, cause the processor to perform the switchgear lifetime prediction method. The processor provides computational and control capabilities to support the operation of the entire electronic device. The memory provides an environment for the execution of the computer-readable code in the non-volatile storage medium, which, when executed by the processor, causes the processor to perform the switchgear lifetime prediction method.

[0260] It should be understood that a processor can be a central processing unit, other general-purpose processors, digital signal processors, application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs) or other programmable logic devices, transistor logic devices, discrete hardware components, etc. Among them, a general-purpose processor can be a microprocessor or any conventional processor.

[0261] 2. This application also provides a readable storage medium, which may be an internal storage unit of the electronic device described in the foregoing embodiments, such as the hard disk or memory of the computer device. The readable storage medium may also be an external storage device of the electronic device, such as a plug-in hard disk, smart memory card, or secure digital card equipped on the electronic device.

[0262] 3. This application also provides a computer program product, including a computer program or instructions, which, when executed by a processor, implement the switchgear life prediction method of the present invention.

[0263] This invention is not limited to the above-described embodiments. Any obvious improvements, substitutions, or modifications that can be made by those skilled in the art without departing from the essence of this invention are within the scope of protection of this invention.

Claims

1. A switchgear lifetime prediction method, characterized by: The switch cabinet life prediction model comprises: A feature engineering module: fusing various switch cabinet state data collected, and outputting fused features arranged in time sequence; Intelligent gate module: for fusion features, use LSTM to obtain the hidden state at each time ; Output module: Obtains data at each time step through an attention mechanism. Attention weights ,based on For each moment Weighted fusion yields global attention features, which are then passed through a fully connected layer to obtain the remaining lifetime; where: The hidden states of multiple different scales at each time are extracted by causal convolution , subscript is a scale variable, is a time variable; Based on and a plurality of different scales , splicing to get high-dimensional features ; Based on , the parameter generation network is used to generate the attention weight at each time , attention bias , attention score vector ; based on and to performing a linear transformation, resulting in ; To perform feature intersection processing to obtain ; To and perform mixed activation to obtain the gating coefficients at each scale ; Based on and Obtain single-scale scores at each scale , and then obtain a multi-scale comprehensive score by weighted sum , and then obtain a multi-scale comprehensive score by weighted sum , and then obtain a multi-scale comprehensive score by weighted sum .

2. The switchgear life prediction method according to claim 1, characterized in that: The output module, global attention feature The calculation is: In the formula, is the maximum sampling time, is the global average hidden state.

3. The switchgear life prediction method according to claim 1, characterized in that: In the output module, a Softmax function is used to score the multi-scale comprehensive score Normalization is performed to obtain attention weights : In the formula, is the maximum sampling time.

4. The switchgear life prediction method according to claim 1, characterized in that: In the output module: high-dimensional features splicing also includes a feature contribution factor and a scene state factor, the scene state factor including a degradation phase and a noise intensity ; Feature contribution degree factor The calculation is: wherein and are adaptive weights, subscript have the same meaning as , is the feature variance between different feature sources at time , is the maximum sampling time; Deterioration phase The calculation is: The cumulative increment proportions of local discharge amount, insulation resistance, and loop resistance are respectively calculated: In the formula, , the subscript The values 1, 2, 3 represent partial discharge, insulation resistance, and loop resistance in turn. The integrated cumulative delta ratio is calculated as follows : The degradation phase is calculated as follows : Noise intensity The calculation is: In the formula, is a residual absolute value, is a sliding average smoothing value, is an output quantity of a forgetting gate in the intelligent gating module.

5. The switchgear life prediction method according to claim 4, characterized in that: In the output module: The feature cross processing is: wherein, represents element-wise multiplication, is a feature cross network; The mixed activation is: wherein is a SiLU activation function, is a hyperbolic tangent activation function; The integrated gating coefficient is calculated as follows : wherein sigmoid activation function, GELU activation function; Single scale score Computed as: Multiscale composite score Computed as: wherein is a scale weight, calculated as: wherein have the same meaning as , is the baseline hidden state, is the cosine similarity function.

6. The switchgear life prediction method according to claim 1, characterized in that: The intelligent gate module comprises: Characteristic change rate factor Calculation: wherein is the fusion feature, is the feature rate of change factor, subscript is the time variable, is a constant to avoid division by zero. A forget gate: wherein, is the output of the forget gate at the time t, is a sigmoid activation function, is a forget gate weight matrix, is the hidden state at the time t, is a forget gate bias term; is a forget gate rate of change weight coefficient, which takes a value in the range [0.1, 0.5]. An input gate: In the formula, is the output quantity of the time input gate, is the input gate weight matrix, is the input gate bias term; is the change rate weight coefficient in the input gate, and the value range is [0.1, 0.5]. A cell state: wherein is cell state at time step t, is candidate cell state at time step t, is a hyperbolic tangent activation function, is a cell state weight matrix, is a cell state bias term; An output gate: wherein is the output of the output gate at time t, is the hidden state at time t, is the output gate weight matrix, is the output gate bias term.

7. The switchgear life prediction method according to claim 1, characterized in that: In the feature engineering module, the various switch cabinet state data collected are fused as follows: In the formula, is the fusion feature, and is the feature value of each feature source, corresponding to each switch cabinet state data respectively, subscript and indicate the feature source number, subscript is a time variable, is the total number of feature sources, and is the adaptive fusion weight corresponding to each feature source, and has , is the interaction term weight coefficient, the value range is [0.05, 0.2]. The switch cabinet state data collected comprise: Electrical feature data: local discharge amount, loop resistance, insulation resistance, three-phase current / voltage; Environmental feature data: environmental temperature, relative humidity, contamination degree; Mechanical feature data: operation mechanism action time, vibration amplitude; The preprocessing of the above data comprises: Data cleaning: adopting a 3σ criterion to eliminate abnormal values in original data, and adopting a linear interpolation method to fill in missing values; Normalization: adopting a min-max normalization method; Time sequence alignment: taking a time stamp as a benchmark, and aligning time sequence data from different sources to the same time axis.

8. A computer apparatus, characterized by: comprise a memory and a processor; The memory is used to store a computer program; The processor is used to execute the computer program and realize the switch cabinet life prediction method according to any one of claims 1-7 when executing the computer program.

9. A computer program product, characterised in that: comprise a computer program, which realizes the switch cabinet life prediction method according to any one of claims 1-7 when executed by a processor.

Images