Transform-based air compressor fault prediction method and system

Abstract

The invention discloses an air compressor fault prediction method based on transform, relates to the crossing field of large models and production and manufacturing, and solves the technical problems that an existing air compressor fault prediction mode is difficult to effectively integrate data, poor in dynamic adaptability, poor in model interpretation and low in calculation efficiency, and a prediction window is difficult to meet prediction requirements. The method comprises a data acquisition step: acquiring operation data of the air compressor; a data preprocessing step: preprocessing the operation data to obtain operation characteristics; and a multi-modal data fusion step: carrying out multi-modal data fusion on the operation features to obtain fusion features. The invention further discloses an air compressor fault prediction system based on the transformer. The dynamic adaptive capacity of the prediction model is improved, the prediction time window is optimized to cope with sudden and gradual change faults at the same time, and the interpretability of the model is enhanced to support decision making.

Description

Technical Field

[0001] The present invention relates to the intersection of large models and production manufacturing, and more specifically, to a transformer-based air compressor fault prediction method and system. Background Art

[0002] Air compressor fault prediction is a crucial topic in the industrial field. It can effectively avoid downtime losses, reduce maintenance costs, extend equipment life, and improve safety and energy efficiency. Optimizing equipment management through data-driven decision-making is a key means to achieve intelligent operation and maintenance. At present, a variety of statistical methods, machine learning methods, and deep learning methods are applied to air compressor fault detection and prediction scenarios. Statistical methods: such as control charts, Z-scores, Tukey's Range Test and other methods are used to identify data points that deviate from the normal range. Its advantages are simple calculations, easy implementation, and suitable for simple anomaly detection tasks. However, its limitations are that it is difficult to handle complex nonlinear relationships and high-dimensional data, and it is sensitive to noise. Traditional machine learning methods such as support vector machines (SVM), random forests, gradient boosting trees and other algorithms are widely used in fault classification and prediction. They can handle nonlinear relationships and have good generalization capabilities. However, feature engineering relies on expert experience and it is difficult to automatically extract deep features. Deep learning methods such as autoencoders, variational autoencoders (VAE) and generative adversarial networks (GANs) are used for complex time series and high-dimensional data anomaly detection. It can automatically learn the deep features of data and is suitable for processing large-scale complex data. However, it requires a large amount of training data, high computing resource requirements, and poor model interpretability. Time series analysis methods such as ARIMA models, LSTM (Long Short-Term Memory Networks) and models based on self-attention mechanisms (such as Transformer) are used to process anomaly detection of time series data. It can capture the temporal dependencies of data and is suitable for processing long-term and short-term time patterns. However, it has limited predictive capabilities for sudden failures and is difficult to integrate multi-source heterogeneous data.

[0003] Although these methods have achieved certain results in the field of air compressor fault prediction, the following challenges still exist: 1) Data fusion problem: Existing methods often find it difficult to effectively integrate multi-source heterogeneous data (such as sensor data, maintenance records, environmental parameters, etc.).

[0004] 2) Dynamic adaptability: The industrial environment is complex and changeable. The operating status and failure mode of equipment may change over time. Existing models lack sufficient dynamic adaptability.

[0005] 3) Prediction window optimization: A fixed prediction window is difficult to meet the prediction requirements for both sudden failures and gradual failures.

[0006] 4) Model interpretability: Many deep learning models behave as “black boxes” and it is difficult to provide an explainable basis for decision making.

[0007] 5) Computational efficiency: The real-time performance of some complex models is difficult to meet the needs of industrial sites. Summary of the invention

[0008] The technical problem to be solved by the present invention is to provide a transformer-based air compressor fault prediction method and system in response to the shortcomings of the existing technology, so as to solve the technical problems that the existing air compressor fault prediction method is difficult to effectively integrate data, has poor dynamic adaptability, the prediction window is difficult to meet the prediction requirements, has poor model interpretability and low computational efficiency.

[0009] The present invention discloses a transformer-based air compressor fault prediction method, which comprises the following steps: data collection: obtaining operation data of the air compressor;

[0010] Data preprocessing step: preprocessing the operation data to obtain operation characteristics;

[0011] Multimodal data fusion step: fusing the operating features into multimodal data to obtain fused features;

[0012] Dynamic adaptive window prediction step: adjusting the fault prediction model according to the fusion feature, and outputting fault warning information through the fault prediction model.

[0013] As a further improvement, in the multimodal data fusion step, the method of performing multimodal data fusion on the operation features to obtain fusion features is:

[0014] S11: Encode the time series feature of the operation feature to obtain a time series feature vector;

[0015] S12: Encode the text feature of the running feature to obtain a text feature vector;

[0016] S13: performing weighted combination of the time series feature vector and the text feature vector through a multi-head attention mechanism to obtain a weighted time series feature vector and a weighted text feature vector;

[0017] S14: Performing weighted fusion on the weighted temporal features and the weighted text features through attention weights to obtain fused features.

[0018] Further, in S13, the time series feature vector and the text feature vector are weightedly combined by the multi-head attention mechanism to obtain the expressions of the weighted time series feature vector and the weighted text feature vector:

[0019]

[0020] Q=W q X 时序 ; K = W k X 文本 ; V = W v X 文本 ;

[0021] Among them, W q , W k and W v is the learnable weight matrix, is the scaling factor, Softmax is the function to calculate the attention distribution, X 文本 is the text feature vector, X 时序 is the time series feature vector.

[0022] Furthermore, the preliminary fusion features are optimized through residual connection and normalization layer to obtain the final fusion feature expression:

[0023] X 融合 =Concat(Attention(Q1,K1,V1),…,Attention(Q i ,K i ,V i ))W o ;

[0024] Among them, Concat means concatenating the outputs of multiple attention heads, W0 is the weight matrix of linear transformation, and X 融合 is the final fusion feature, and i is a natural number greater than 1.

[0025] Furthermore, in S14, the weighted time series feature vector and the weighted text feature vector are concatenated into a unified feature vector, the unified feature vector is fused and dimensionally reduced through a fully connected layer to obtain a preliminary fusion feature, and the preliminary fusion feature is optimized through a residual connection and a normalization layer to obtain a final fusion feature.

[0026] Furthermore, in the dynamic adaptive window prediction step, the method for adjusting the fault prediction model according to the fusion feature is:

[0027] S21: analyzing the fusion feature and evaluating the current operating state change of the air compressor to obtain a current operating state score;

[0028] S22: selecting an optimal window length and an optimal prediction frequency of a fault prediction model through the current operating status score and the reinforcement learning strategy;

[0029] S23: adjusting the window length and prediction frequency of the fault prediction model through the reinforcement learning strategy;

[0030] S24: adjusting the fault prediction model using the adjusted window length and the adjusted prediction frequency.

[0031] Furthermore, in S23, the expression for selecting the optimal window length and prediction frequency of the fault prediction model through the current operating state score and the reinforcement learning strategy is: Q(s,a)←Q(s,a)+α(r+γmax a′ Q(s′,a′)-Q(s,a));

[0032] Among them, Q (s,a) is the value function for selecting action a in state s, α is the learning rate, γ is the discount factor, r is the immediate reward after executing the action, s' is the new state after executing the action, a' is all possible actions in state s' after executing the action, and max α’ Q(s',a') means finding an action from all possible actions so that Q(s',a') takes the maximum value.

[0033] Furthermore, in S24, the method for adjusting the window length of the fault prediction model and the prediction frequency of the fault prediction model by the reinforcement learning strategy is to obtain the window length W of the new state after the action is executed. t-1 and the predicted frequency of the new state F t-1 , the optimal window length W t and the window length W of the new state t-1 The window variation ΔW is obtained by subtraction, and the best prediction frequency F t and the predicted frequency of the new state F t-1 The frequency variation ΔF is obtained by difference, the window length of the fault prediction model is adjusted by the window variation ΔW, and the prediction frequency of the fault prediction model is adjusted by the frequency variation ΔF.

[0034] Furthermore, the method for preprocessing the operating data to obtain operating characteristics is that the operating data includes time series data, and basic statistical characteristics, time characteristics and frequency domain characteristics are extracted from the preprocessed time series data, and the basic statistical characteristics, time characteristics and frequency domain characteristics are used as time series characteristics.

[0035] A transformer-based air compressor fault prediction system, the system comprising: a data acquisition module: used for acquiring operating data of the air compressor;

[0036] Data preprocessing module: used for preprocessing the operation data to obtain operation characteristics;

[0037] Multimodal data fusion module: used for performing multimodal data fusion on the operation characteristics to obtain fusion characteristics;

[0038] Dynamic adaptive window prediction module: used to adjust the fault prediction model according to the fusion feature, and output fault warning information through the fault prediction model.

[0039] Beneficial Effects

[0040] The advantages of the present invention are:

[0041] The present invention obtains the operating data of the air compressor, pre-processes the operating data to obtain operating characteristics, fuses the operating characteristics to obtain fused characteristics, adjusts the fault prediction model according to the fused characteristics, and outputs fault warning information through the fault prediction model; thereby improving the dynamic adaptability of the prediction model, optimizing the prediction time window to simultaneously cope with sudden and gradual faults, and enhancing the interpretability of the model to support decision-making; at the same time, the present invention also focuses on enhancing the modeling ability of complex nonlinear relationships, improving computing efficiency and real-time performance, enhancing the generalization ability of the model to adapt to different equipment and environments, and effectively handling data imbalance and sparsity problems; by solving the above-mentioned technical problems, the prediction accuracy can be significantly improved, false alarms and missed alarms can be reduced, and strong technical support can be provided for the intelligent operation and maintenance and predictive maintenance of industrial equipment, thereby helping enterprises reduce maintenance costs, extend equipment life, improve production efficiency and safety, and create significant economic benefits. BRIEF DESCRIPTION OF THE DRAWINGS

[0042] Figure 1 It is a flow chart of the air compressor fault prediction method of the present invention;

[0043] Figure 2 It is a flow chart of preprocessing time series data and text data of the present invention;

[0044] Figure 3 It is a multimodal data fusion flow chart of the present invention;

[0045] Figure 4 A flow chart of adjusting a fault prediction model for the fusion features of the present invention;

[0046] Figure 5 This is a collection category diagram of the time series data of the present invention. DETAILED DESCRIPTION

[0047] The present invention will be further described below in conjunction with the embodiments, but it does not constitute any limitation to the present invention. Any limited number of modifications made by anyone within the scope of the claims of the present invention are still within the scope of the claims of the present invention.

[0048] See also Figure 1-Figure 5 , a transformer-based air compressor fault prediction method of the present invention is,

[0049] Data collection steps: Obtain time series data and text data from the air compressor operation data.

[0050] In the present invention, various sensors can be used to obtain Figure 5 Each data in .

[0051] Data preprocessing: The collected data is divided into time series digital data and text data. Time series data sources: ambient temperature sensor, operating load sensor, pressure sensor, flow meter, etc.

[0052] Data preprocessing step: preprocess the time series data and extract the time series features of the motion features from the preprocessed time series data, and preprocess the text data to obtain the text features of the motion features.

[0053] like Figure 2 As shown, the time series data is preprocessed:

[0054] Data cleaning: Remove obvious outliers, such as erroneous data caused by sensor failure, and use statistical methods (such as IQR or interquartile range or 3σ rule) to detect outliers in time series data.

[0055] Missing value filling: Use time series interpolation (linear interpolation or filling in previous and next values) to maintain data continuity and ensure the integrity of the predicted time series.

[0056] Standardization: The data of each sensor is processed by a normalization layer, using a Min-Max normalization layer or Z-score normalization to unify the scale of the output of different sensors.

[0057] The method of extracting time series features from preprocessed time series data is to extract basic statistical features, time features and frequency domain features from the preprocessed time series data and use the basic statistical features, time features and frequency domain features as time series features. The purpose is to enhance the model's ability to capture data patterns.

[0058] Among them, the basic statistical features include: extracting the mean, variance, maximum value, minimum value, etc. within each dynamic window.

[0059] Temporal characteristics: Extract trend, periodicity and seasonality characteristics based on load curves and pressure data.

[0060] Frequency domain features: Frequency domain features are extracted through fast Fourier transform (FFT) to help identify periodic anomalies.

[0061] The data source of text data: non-time-series text data such as inspection sheets and maintenance records.

[0062] The method for preprocessing text data to obtain text features is:

[0063] Text cleaning: Remove unnecessary symbols, spaces and other noise content from text data, and perform spelling check and normalization.

[0064] Word segmentation and stop word filtering: Perform word segmentation on text data, remove common stop words, and retain keywords related to device status.

[0065] Text embedding encoding: Use pre-trained word vector models (such as Word2Vec, BERT) to convert text data into numerical vector representations to prepare for subsequent fusion.

[0066] Dimension alignment: Align and adjust the dimensions of features from different sources to ensure that the input fusion features are consistent with the model requirements in terms of dimension and format.

[0067] After completing data preprocessing, the processed time series and text features will be used as input and enter the multimodal data fusion module. The features of each data source will be dynamically weighted and fused through the adaptive multi-head attention mechanism to improve the prediction accuracy.

[0068] Multimodal data fusion step: Perform multimodal data fusion on time series features and text features to obtain fused features.

[0069] like Figure 3 As shown in the figure, the method of multimodal data fusion of time series features and text features to obtain fusion features is:

[0070] S11: Encode the time series features through a convolutional neural network or a recurrent neural network to obtain a time series feature vector. Use a convolutional neural network (CNN) or a recurrent neural network (RNN) to encode the time series features and extract local patterns and time dependencies. Convert the extracted features into a high-dimensional vector.

[0071] The expression for encoding the time series features to obtain the time series feature vector is:

[0072] X 时序 =CNN / RNN(X t );

[0073] Among them, X t is the time series feature, CNN / RNN represents the extraction of time series features by convolutional neural network or recurrent neural network, and Xtimeseries is the time series feature vector.

[0074] S12: Encode the text features through a pre-trained language model to obtain a text feature vector. Use a pre-trained language model (such as BERT, Word2Vec) to encode the text data and extract implicit information related to the device operating status in the text. Convert the text features into a vector representation that matches the time series features.

[0075] The expression for encoding text features to obtain text feature vectors is:

[0076] X 文本 =Embedding(T);

[0077] Where T is the text input, Embedding represents the text embedding operation through the pre-trained model (such as BERT), and X 文本 is the text feature vector.

[0078] S13: The time series feature vector and the text feature vector are weightedly combined through the multi-head attention mechanism to obtain the weighted time series feature vector and the weighted text feature vector.

[0079] The adaptive multi-head attention mechanism is used to capture the complex relationship between multimodal features. By calculating the attention weights of each modal feature, the contribution of different data sources to the fusion features is dynamically adjusted.

[0080] Multi-head processing: Multiple attention heads are processed in parallel to learn different relationships between temporal features and text features.

[0081] Weighted calculation: Calculate the importance of each modal feature to the final output, and use the attention weight to perform a weighted combination of the features.

[0082] Attention Matrix: Generate an attention matrix to focus on the most critical feature combinations for fault prediction through a soft attention mechanism, suppressing the impact of noisy data.

[0083] The expressions of the weighted time series feature vector and the weighted text feature vector obtained by weighted combination of the time series feature vector and the text feature vector through the multi-head attention mechanism are:

[0084]

[0085] Q=W q X 时序 ; K = W k X 文本 ; V = W v X 文本 ;

[0086] Among them, W q , W k and W v is the learnable weight matrix, is the scaling factor, Softmax is the function to calculate the attention distribution, X 文本 is the text feature vector, X 时序 is the time series feature vector.

[0087] S14: Concatenate the weighted time series feature vector and the weighted text feature vector into a unified feature vector, fuse and reduce the dimension of the unified feature vector through a fully connected layer to obtain preliminary fusion features, and optimize the preliminary fusion features through residual connections and normalization layers to obtain the final fusion features, thereby enhancing the stability and generalization ability of the model.

[0088] The expression of the final fusion feature obtained by optimizing the preliminary fusion features through residual connection and normalization layer is:

[0089] X 融合 =Concat(Attention(Q1,K1,V1),…,Attention(Q 13 ,K 13 ,V 13 ))W o ;

[0090] Among them, Concat means concatenating the outputs of multiple attention heads, W0 is the weight matrix of linear transformation, and X 融合 is the final fusion feature.

[0091] The fused high-dimensional feature vector is used as the input of the prediction module to provide comprehensive information support for fault prediction.

[0092] Dynamic adaptive window prediction step: adjust the fault prediction model according to the fusion characteristics, and output fault warning information through the fault prediction model.

[0093] like Figure 4 As shown in Figure 2, the method for adjusting the fault prediction model based on the fusion features is:

[0094] S21: Obtain the current operating state of the air compressor, and evaluate the current operating state of the air compressor according to the fusion features to obtain a current operating state score of the air compressor.

[0095] Analyze the fused features and evaluate the changes in the current operating status to decide whether the prediction window needs to be adjusted. The expression for the current operating status score is as follows:

[0096] S t =f(X 融合 (t));

[0097] Among them, S t is the state score at the current moment, which is calculated by function f based on the fusion feature X 融合 Calculated.

[0098] Perform time domain and frequency domain analysis on the input features to identify significant changes in the operating status. Based on the magnitude and frequency of the changes, preliminarily evaluate the rationality of the current window settings.

[0099] S22: Formulate reinforcement learning strategies based on the current running status score.

[0100] The optimal time window length and prediction frequency are dynamically selected through reinforcement learning strategies.

[0101] State representation: The current running state is represented by the feature score S t express.

[0102] Action space: includes adjusting the window length (increase, decrease, remain unchanged) and adjusting the prediction frequency.

[0103] Reward function: rewards are given based on prediction accuracy and response speed. The optimization goal is to maximize prediction accuracy and minimize prediction delay.

[0104] The expression for formulating the reinforcement learning strategy based on the current running state score is: Q(s,a)←Q(s,a)+α(r+γmax a′ Q(s′,a′)-Q(s,a));

[0105] Among them, Q (s,a) is the value function for selecting action a in state s, α is the learning rate, γ is the discount factor, r is the immediate reward after executing the action, s' is the new state after executing the action, a' is all possible actions in state s' after executing the action, and max α’ Q(s',a') means finding an action from all possible actions so that Q(s',a') takes the maximum value.

[0106] S23: The window length of the fault prediction model and the prediction frequency of the fault prediction model are adjusted through the reinforcement learning strategy to adapt to the current operating status.

[0107] The method for adjusting the window length and prediction frequency of the fault prediction model by using the reinforcement learning strategy is to obtain the window length W of the new state after the action is executed. t-1 and the predicted frequency of the new state F t-1 , the optimal window length W t and the window length W of the new state t-1 The window change ΔW is obtained by subtraction, and the expression is: W t =W t-1 +ΔW, the best predicted frequency F t and the predicted frequency of the new state F t-1 The frequency change ΔF is obtained by subtraction, and the expression is F t =F t-1 +ΔF, adjust the window length of the fault prediction model by the window change ΔW, and adjust the prediction frequency of the fault prediction model by the frequency change ΔF.

[0108] S24: The fault prediction model is adjusted using the adjusted window length and the adjusted prediction frequency, and the input data range of the fault prediction model is updated to ensure that the adjusted window can accurately reflect the latest state changes.

[0109] Output fault warning information based on the dynamically adjusted window prediction. Evaluate the adjusted prediction results to ensure that the prediction performance under the new configuration is improved. Provide prediction results and corresponding adjustment suggestions to support operation and maintenance decisions.

[0110] A transformer-based air compressor fault prediction system, which includes:

[0111] Data acquisition module: used to obtain the operating data of the air compressor.

[0112] Data preprocessing module: used for preprocessing the operation data to obtain operation characteristics.

[0113] Multimodal data fusion module: deeply fuses time series data (such as pressure, temperature, load, etc.) and text data (such as inspection records, maintenance logs) to improve the accuracy of the model's prediction of air compressor failures. This module uses an adaptive multi-head attention mechanism to dynamically weight features from different sources to achieve effective information integration and feature complementarity. The preprocessed data is input into the multimodal data fusion module in the form of feature vectors to ensure that the fusion model can effectively learn the relationship between the features.

[0114] Dynamic adaptive window prediction module: optimizes the accuracy and response speed of air compressor fault prediction by adjusting the length and frequency of the prediction window in real time. The core of this module is to adjust the prediction window according to real-time data characteristics and operating status by combining reinforcement learning or Bayesian optimization methods to better capture different types of fault signals, especially sudden faults and progressive faults.

[0115] The above is only a preferred embodiment of the present invention. It should be pointed out that for those skilled in the art, several modifications and improvements can be made without departing from the structure of the present invention, which will not affect the effect of the implementation of the present invention and the practicality of the patent.

Claims

1. A transformer-based air compressor fault prediction method, characterized in that: The method comprises the following steps: obtaining the operating data of the air compressor; Data preprocessing step: preprocessing the operation data to obtain operation characteristics; Multimodal data fusion step: fusing the operating features into multimodal data to obtain fused features; Dynamic adaptive window prediction step: adjusting the fault prediction model according to the fusion feature, and outputting fault warning information through the fault prediction model.

2. The air compressor fault prediction method based on transformer according to claim 1 is characterized in that: In the multimodal data fusion step, the method of performing multimodal data fusion on the operation features to obtain fusion features is: S11: Encode the time series feature of the operation feature to obtain a time series feature vector; S12: Encode the text feature of the running feature to obtain a text feature vector; S13: performing weighted combination of the time series feature vector and the text feature vector through a multi-head attention mechanism to obtain a weighted time series feature vector and a weighted text feature vector; S14: Performing weighted fusion on the weighted temporal features and the weighted text features through attention weights to obtain fused features.

3. The air compressor fault prediction method based on transformer according to claim 2 is characterized in that: In S13, the time series feature vector and the text feature vector are weightedly combined by the multi-head attention mechanism to obtain the weighted time series feature vector and the weighted text feature vector. The expressions are: Q=W q X 时序 ;K=W k X 文本 ;V=W v X 文本 ; Among them, W q , W k and W v is the learnable weight matrix, is the scaling factor, Softmax is the function to calculate the attention distribution, X 文本 is the text feature vector, X 时序 is the time series feature vector.

4. The air compressor fault prediction method based on transformer according to claim 2 is characterized in that: In S14, the weighted time series feature vector and the weighted text feature vector are concatenated into a unified feature vector, the unified feature vector is fused and dimensionally reduced through a fully connected layer to obtain a preliminary fusion feature, and the preliminary fusion feature is optimized through a residual connection and a normalization layer to obtain a final fusion feature.

5. The air compressor fault prediction method based on transformer according to claim 4 is characterized in that: The expression of the final fusion feature obtained by optimizing the preliminary fusion feature through residual connection and normalization layer is: X 融合 =Concat(Attention(Q1,K1,V1),…,Attention(Q i ,K i V i ))W o ; Among them, Concat means concatenating the outputs of multiple attention heads, W0 is the weight matrix of linear transformation, and X 融合 is the final fusion feature, and i is a natural number greater than 1.

6. The transformer-based air compressor fault prediction method according to claim 1, characterized in that: In the dynamic adaptive window prediction step, the method of adjusting the fault prediction model according to the fusion feature is: S21: analyzing the fusion feature and evaluating the current operating state change of the air compressor to obtain a current operating state score; S22: selecting an optimal window length and an optimal prediction frequency of a fault prediction model through the current operating status score and the reinforcement learning strategy; S23: adjusting the window length and prediction frequency of the fault prediction model through the reinforcement learning strategy; S24: adjusting the fault prediction model using the adjusted window length and the adjusted prediction frequency.

7. The air compressor fault prediction method based on transformer according to claim 6 is characterized in that: In S23, the expression for selecting the optimal window length and prediction frequency of the fault prediction model through the current operating state score and the reinforcement learning strategy is: Q(s,a)←Q(s,a)+α(r+γmax α′ Q(s′, α′)-Q(s,a)); Among them, Q (s,a) is the value function for selecting action a in state s, α is the learning rate, γ is the discount factor, r is the immediate reward after executing the action, s' is the new state after executing the action, a' is all possible actions in state s' after executing the action, and max α’ Q(s',a') means finding an action from all possible actions so that Q(s',a') takes the maximum value.

8. The transformer-based air compressor fault prediction method according to claim 6, characterized in that: In S23, the method for adjusting the window length and prediction frequency of the fault prediction model by the reinforcement learning strategy is to obtain the window length W of the new state after the action is executed. t-1 and the predicted frequency of the new state F t-1 , the optimal window length W t and the window length W of the new state t-1 The window variation ΔW is obtained by subtracting the optimal prediction frequency F t and the predicted frequency of the new state F t-1 The frequency variation ΔF is obtained by difference, the window length of the fault prediction model is adjusted by the window variation ΔW, and the prediction frequency of the fault prediction model is adjusted by the frequency variation ΔF.

9. The transformer-based air compressor fault prediction method according to claim 1, characterized in that: The method for preprocessing the operating data to obtain operating characteristics is that the operating data includes time series data, basic statistical characteristics, time characteristics and frequency domain characteristics are extracted from the preprocessed time series data, and the basic statistical characteristics, time characteristics and frequency domain characteristics are used as time series characteristics.

10. A transformer-based air compressor fault prediction system, characterized in that: The system includes: a data acquisition module: used to obtain the operating data of the air compressor; Data preprocessing module: used for preprocessing the operation data to obtain operation characteristics; Multimodal data fusion module: used for performing multimodal data fusion on the operation characteristics to obtain fusion characteristics; Dynamic adaptive window prediction module: used to adjust the fault prediction model according to the fusion feature, and output fault warning information through the fault prediction model.

Images