Water conservancy equipment life prediction and fault monitoring method, equipment and storage medium

Abstract

The invention relates to a water conservancy equipment service life prediction and fault monitoring method. The water conservancy equipment service life prediction and fault monitoring method comprises the steps of performing comprehensive digital modeling on water conservancy equipment, determining a fault sensitive area and a key monitoring point, generating a sensor deployment scheme, preprocessing analog and simulated sensor data, and ensuring that the data is comprehensive and targeted; performing multi-domain fusion feature extraction, deep feature learning and correlation analysis on the data, screening out important features to form a feature subset, and capturing equipment fault features in all directions; a mixed life prediction model is constructed, parameter initialization, pre-training and formal training are completed, a multivariate Gaussian mixture model and a deep belief network are constructed, and then a fault monitoring model is obtained; sensor data arranged according to a deployment scheme is obtained in real time, features are extracted after preprocessing, real-time feature vectors are input into a life prediction model and a fault monitoring model respectively, life prediction and fault monitoring of the water conservancy equipment are achieved, powerful decision support is provided for equipment maintenance, reliable operation of the water conservancy equipment is guaranteed, and fault loss is reduced.

Description

Technical Field

[0001] The present invention relates to the technical field of water conservancy equipment maintenance, and in particular to a method, equipment and storage medium for predicting the life of water conservancy equipment and monitoring faults. Background Art

[0002] With the vigorous development of water conservancy, water conservancy equipment plays an increasingly critical role in water resource allocation, flood prevention and disaster reduction, hydropower generation, etc. The scale of water conservancy projects continues to expand and move towards refined management, which puts extremely stringent requirements on the stable and efficient operation of water conservancy equipment. The level of equipment maintenance is directly related to the overall benefits and safety of water conservancy projects.

[0003] At present, in the process of collecting data on the operation of water conservancy equipment, the means are relatively simple and lack systematic planning. In most cases, only a small number of specific types of sensors are relied on, such as only arranging pressure sensors and temperature sensors, while the collection of information in other dimensions such as vibration, flow, acoustics, strain torque, etc. is neglected. This collection method makes it impossible to fully obtain the operating status information of the equipment, and the relevant data of many key areas and potential fault points are omitted. Since it is difficult to perceive the operating status of the equipment in all directions, many potential fault hazards are difficult to be detected in time in the early stages, which poses a huge risk to the safe and stable operation of water conservancy equipment.

[0004] In the field of life prediction and fault monitoring, most of the existing methods are based on simple statistical or empirical models. Such models are seriously insufficient in adaptability when facing complex operating conditions and equipment aging processes. The operating conditions of water conservancy equipment are complex and changeable, and are affected by a variety of factors such as water flow velocity, water level fluctuations, water quality differences, and external environmental changes. At the same time, the equipment will gradually age during long-term operation, and its performance and operating status will also change accordingly. However, the existing models cannot effectively cope with these complex situations. When processing multi-source heterogeneous data, the existing models lack effective integration and utilization methods. Multi-source heterogeneous data covers data of different types, formats, and frequencies, such as data from different sensors, as well as unstructured data such as equipment operation logs and maintenance records. Since the existing models cannot effectively fuse and analyze these data, a large amount of valuable information is wasted, and it is impossible to provide a comprehensive and accurate basis for life prediction and fault monitoring. In dealing with the nonlinear and time-varying characteristics of equipment, the existing models also perform poorly. The operating status of water conservancy equipment does not change linearly, but has highly nonlinear and time-varying characteristics. Existing models are difficult to accurately capture and characterize these dynamic changes, resulting in low prediction accuracy and frequent false alarms and missed reports. This not only makes it difficult for maintenance personnel to accurately judge the actual operating status of the equipment, but may also lead to unnecessary maintenance operations, increase maintenance costs, and fail to promptly discover real hidden faults, seriously threatening the safe operation of water conservancy projects.

[0005] To sum up, the existing methods of water conservancy equipment operation data collection, life prediction and fault monitoring have many defects, which are difficult to meet the urgent needs of refined management of modern water conservancy projects. An innovative technical means is urgently needed to improve the maintenance level of water conservancy equipment and ensure the safe and stable operation of water conservancy projects. Summary of the invention

[0006] In view of this, the purpose of the present invention is to provide a method, device and storage medium for life prediction and fault monitoring of water conservancy equipment, so as to solve the above-mentioned problems in the prior art to a certain extent.

[0007] According to a first aspect of an embodiment of the present invention, a method for predicting life span and fault monitoring of water conservancy equipment is provided, comprising:

[0008] S11, conduct comprehensive digital modeling of the water conservancy equipment to be tested, determine the fault-sensitive areas and key monitoring points according to the equipment model, and generate a sensor deployment plan; perform simulation according to the constructed equipment model and sensor deployment plan to obtain simulated sensor data, and pre-process the sensor data;

[0009] S12, extracting multi-domain fusion features from the sensor data to obtain multi-domain fusion features, performing deep feature learning and feature correlation analysis on the multi-domain fusion features in sequence, and placing features whose feature importance is greater than a threshold into a feature subset;

[0010] S13, constructing a hybrid life prediction model architecture, initializing and pre-training parameters of the hybrid life prediction model; using a feature subset to formally train the pre-trained hybrid life prediction model to obtain a trained hybrid life prediction model;

[0011] S14, construct a multivariate Gaussian mixture model, and perform parameter estimation on the multivariate Gaussian mixture model according to a preset data set of normal operation status of the equipment; calculate the Mahalanobis distance threshold and probability density threshold of each feature dimension in the feature subset according to the parameter estimation result; construct a deep belief network for fault feature extraction and classification; and obtain a fault monitoring model based on the multivariate Gaussian mixture model and the deep belief network;

[0012] S15, acquiring sensor data in real time, wherein the sensor is deployed according to the sensor deployment plan; preprocessing the sensor data, extracting features from the preprocessed sensor data using step S12 to obtain a real-time feature vector; inputting the real-time feature vector into a hybrid life prediction model to perform life prediction, and inputting the real-time feature vector into a fault monitoring model to perform fault monitoring.

[0013] Preferably, in step S11, it also includes: constructing a collection frequency adaptive adjustment model using a machine learning algorithm according to a dynamic collection strategy of the working condition and environment of the water conservancy equipment to be tested;

[0014] In step S15, when the sensor data is acquired in real time, the method further includes: inputting the sensor data acquired in real time into the acquisition frequency adaptive adjustment model to obtain the sensor acquisition frequency, and controlling the acquisition frequency of each sensor according to the sensor acquisition frequency.

[0015] Preferably, in step S12, it also includes:

[0016] Construct a multi-domain fusion feature extraction framework, and use the multi-domain fusion feature extraction framework to extract multi-domain fusion features from the sensor data to obtain multi-domain fusion features; the multi-domain fusion feature extraction framework uses a combination of VMD and HHT in the time domain, a WPT and FFT joint analysis method in the frequency domain, and SST and GST tools in the time-frequency domain;

[0017] A deep learning AE model is constructed to perform deep feature learning on multi-domain fusion features; the deep learning AE model includes an encoder and a decoder. When training the deep learning AE model, a GAN adversarial training mechanism is introduced to construct a discriminator for distinguishing original and reconstructed features;

[0018] Feature correlation analysis was performed based on a method combining the maximum information coefficient and the Pearson correlation coefficient;

[0019] The feature importance evaluation method is used to obtain the importance score of each feature, and the features with feature importance scores greater than the threshold are placed in the feature subset.

[0020] Preferably, in step S12, it also includes:

[0021] The validity of the selected feature subset is verified using a leave-one-out method or a K-fold cross-validation method. If the verification fails, step S12 is optimized and adjusted.

[0022] Preferably, in step S13, it also includes:

[0023] The hybrid life prediction model architecture is composed of a convolutional neural network, a recurrent neural network and its variant, a long short-term memory network, a gated recurrent unit, and a support vector regression;

[0024] During formal training: the hybrid life prediction model is iteratively trained using feature subsets, and the parameters of the hybrid life prediction model are optimized using a parameter optimization method based on an improved particle swarm optimization algorithm; the improved particle swarm optimization algorithm, based on the particle swarm optimization algorithm, introduces an adaptive inertia weight adjustment strategy and a dynamic neighborhood search mechanism, and dynamically adjusts the inertia weight according to the distance between the current position of the particle and the historical optimal position and the distribution of the entire particle swarm;

[0025] During the iterative training process, the objective function is the weighted combination loss function that minimizes the mean square error and the mean absolute error between the predicted value and the true value.

[0026] Preferably, in step S14,

[0027] Parameter estimation is performed on the multivariate Gaussian mixture model, including: estimating GMM model parameters using an improved EM algorithm based on hierarchical clustering and variational inference.

[0028] Preferably, preprocessing the sensor data includes:

[0029] Build a deep neural network anomaly detection model, introduce sliding windows and dynamic threshold mechanisms, segment new sensor data into fixed-length windows, calculate statistical features and compare them with dynamic thresholds, and mark potential abnormal data;

[0030] Cluster analysis is performed on the abnormal data to remove abnormal values ​​that meet the abnormal conditions to obtain a cleaned data set.

[0031] Preferably, after obtaining the cleaned data set, the method further includes:

[0032] Perform real-time probability distribution estimation on the cleaned data set to determine the distribution type and parameters of each dimension of data;

[0033] According to the data distribution type, the corresponding normalization transformation method is selected to normalize the data set to obtain a normalized data set; during the normalization process, the time series trend of the data set is modeled and predicted, and when it is detected that the data has a preset trend change, the normalization parameters are dynamically adjusted.

[0034] According to a second aspect of an embodiment of the present invention, there is provided a hydraulic equipment life prediction and fault monitoring device, comprising:

[0035] A main controller, and a memory connected to the main controller;

[0036] a memory in which program instructions are stored;

[0037] The main controller is used to execute program instructions stored in the memory to perform any of the above methods.

[0038] According to a third aspect of an embodiment of the present invention, a computer-readable storage medium is provided, storing a computer program, and when the computer program is executed by a processor, the method described in any one of the above is implemented.

[0039] The technical solution provided by the embodiments of the present invention may have the following beneficial effects:

[0040] It can be understood that the present invention can comprehensively digitally model water conservancy equipment, determine fault-sensitive areas and key monitoring points, generate sensor deployment plans, and pre-process simulated sensor data to ensure that the data is comprehensive and targeted; perform multi-domain fusion feature extraction, deep feature learning and correlation analysis on the data, screen out important features to form feature subsets, and capture equipment failure characteristics in all directions; construct a hybrid life prediction model and complete parameter initialization, pre-training and formal training, construct a multivariate Gaussian mixture model and a deep belief network, and then obtain a fault monitoring model; obtain sensor data arranged according to the deployment plan in real time, extract features after pre-processing, and input real-time feature vectors into the life prediction model and the fault monitoring model respectively, to achieve life prediction and fault monitoring of water conservancy equipment, provide strong decision-making support for equipment maintenance, ensure the reliable operation of water conservancy equipment, and reduce failure losses.

[0041] It is to be understood that both the foregoing general description and the following detailed description are exemplary and explanatory only and are not restrictive of the invention. BRIEF DESCRIPTION OF THE DRAWINGS

[0042] The accompanying drawings, which are incorporated in and constitute a part of this specification, illustrate embodiments consistent with the invention and, together with the description, serve to explain the principles of the invention.

[0043] Figure 1 It is a schematic diagram of steps of a method for predicting the life of water conservancy equipment and monitoring faults according to an exemplary embodiment;

[0044] Figure 2 The present invention is a flowchart showing a method for predicting the life of water conservancy equipment and monitoring faults according to an exemplary embodiment. DETAILED DESCRIPTION

[0045] Exemplary embodiments will be described in detail herein, examples of which are shown in the accompanying drawings. When the following description refers to the drawings, the same numbers in different drawings represent the same or similar elements unless otherwise indicated. The embodiments described in the following exemplary embodiments do not represent all embodiments consistent with the present invention. Instead, they are merely examples of devices and methods consistent with some aspects of the present invention as detailed in the appended claims.

[0046] In one embodiment, see Figure 1 , providing a method for predicting the life of water conservancy equipment and monitoring faults, comprising:

[0047] S11. Conduct comprehensive digital modeling of the water conservancy equipment to be tested, determine fault-sensitive areas and key monitoring points based on the equipment model, and generate a sensor deployment plan; perform simulation based on the constructed equipment model and sensor deployment plan to obtain simulated sensor data, and pre-process the sensor data.

[0048] S12, performing multi-domain fusion feature extraction on the sensor data to obtain multi-domain fusion features, performing deep feature learning and feature correlation analysis on the multi-domain fusion features in sequence, and placing features whose feature importance is greater than a threshold into a feature subset.

[0049] S13, constructing a hybrid life prediction model architecture, initializing and pre-training the parameters of the hybrid life prediction model; using a feature subset to formally train the pre-trained hybrid life prediction model to obtain a trained hybrid life prediction model.

[0050] S14. Construct a multivariate Gaussian mixture model, and perform parameter estimation on the multivariate Gaussian mixture model according to a preset data set of normal equipment operation status; calculate the Mahalanobis distance threshold and probability density threshold of each feature dimension in the feature subset according to the parameter estimation results; construct a deep belief network for fault feature extraction and classification; and obtain a fault monitoring model based on the multivariate Gaussian mixture model and the deep belief network.

[0051] S15, acquiring sensor data in real time, wherein the sensor is deployed according to the sensor deployment plan; preprocessing the sensor data, extracting features from the preprocessed sensor data using step S12 to obtain a real-time feature vector; inputting the real-time feature vector into a hybrid life prediction model to perform life prediction, and inputting the real-time feature vector into a fault monitoring model to perform fault monitoring.

[0052] It can be understood that this embodiment can fully digitally model the water conservancy equipment, determine fault-sensitive areas and key monitoring points, generate sensor deployment plans, pre-process the simulated sensor data, and ensure that the data is comprehensive and targeted; perform multi-domain fusion feature extraction, deep feature learning and correlation analysis on the data, screen out important features to form feature subsets, and capture equipment failure characteristics in all directions; construct a hybrid life prediction model and complete parameter initialization, pre-training and formal training, construct a multivariate Gaussian mixture model and a deep belief network, and then obtain a fault monitoring model; obtain sensor data arranged according to the deployment plan in real time, extract features after pre-processing, and input the real-time feature vectors into the life prediction model and the fault monitoring model respectively, to achieve life prediction and fault monitoring of water conservancy equipment, provide strong decision-making support for equipment maintenance, ensure the reliable operation of water conservancy equipment, and reduce failure losses.

[0053] In practice, see Figure 2 ,In step S11, it mainly includes multi-source heterogeneous data collection planning, real-time adaptive data cleaning, dynamic intelligent data normalization, data quality assessment and feedback optimization, and data compression and storage optimization.

[0054] In the multi-source heterogeneous data collection planning, water conservancy equipment system modeling and simulation technology is used to conduct comprehensive digital modeling of the overall structure, working principle and operation process of water conservancy equipment, and accurately identify all potential fault-sensitive areas and key performance monitoring points.

[0055] According to the modeling results, a detailed sensor deployment plan is formulated, using various types of sensors including but not limited to pressure, temperature, vibration, flow, acoustics, strain torque, etc., to ensure the collection of equipment operation data from different physical dimensions and achieve a full range of perception of the equipment status. The selection of sensors follows the principles of high precision, high reliability, wide range and low power consumption, while considering their compatibility and scalability to adapt to the complex and changing operating environment of water conservancy equipment and possible future upgrade needs.

[0056] It should be noted that in step S11, it also includes: according to the dynamic collection strategy of the working condition and environment of the water conservancy equipment to be tested, using the machine learning algorithm to build an adaptive adjustment model of the collection frequency; in step S15, when acquiring sensor data in real time, it also includes: inputting the sensor data acquired in real time into the adaptive adjustment model of the collection frequency to obtain the sensor collection frequency, and controlling the collection frequency of each sensor according to the sensor collection frequency.

[0057] Establish a dynamic data collection strategy based on equipment operating conditions and environmental conditions. By real-time monitoring of equipment load changes, water flow velocity, water level fluctuations, water quality parameters, and external environmental temperature, humidity, air pressure and other factors, use machine learning algorithms to build an adaptive adjustment model for the collection frequency. For example, when the equipment is under high load, harsh environment or critical operation stage, the collection frequency is automatically increased to 100Hz or even higher to capture key details; and during stable low-load periods, the collection frequency is appropriately reduced to about 10Hz, while ensuring data integrity while reducing data redundancy and storage pressure, thereby obtaining a targeted and timely original data set D = {d1, d2, d3, …, d n}, where n = f × T, each data point d i Contains sensor measurements in multiple dimensions, such as d i =[p i ,t i ,v i ,q i ,a i ,s i ,…], respectively representing pressure, temperature, vibration amplitude, flow rate, acoustic intensity, strain value, etc.

[0058] In real-time adaptive data cleaning, it should be noted that a deep neural network anomaly detection model is constructed, and a sliding window and dynamic threshold mechanism is introduced to segment new sensor data into fixed-length windows when it flows in, calculate statistical features and compare them with dynamic thresholds, and mark potential abnormal data; cluster analysis is performed on the abnormal data, and outliers that meet the abnormal conditions are eliminated to obtain a cleaned data set.

[0059] In specific practice, we build an anomaly detection model based on a deep neural network, and use a large amount of historical normal data and a small amount of known abnormal data for supervised learning training. The model structure combines a multi-layer convolutional neural network (CNN) with a long short-term memory network (LSTM). CNN is used to automatically extract local feature patterns of data, while LSTM can capture the time series correlation and long-term dependency of data, so as to accurately identify abnormal patterns in the data.

[0060] A real-time anomaly detection mechanism based on sliding windows and dynamic thresholds is introduced. As new data continues to flow in, the data is segmented using a sliding window of fixed length. The statistical features of the data, such as mean, standard deviation, skewness, kurtosis, etc., are calculated in each window and compared with the dynamic thresholds learned based on historical data. When the statistical features of a data point or data segment exceed the threshold range, it is marked as potential anomaly data.

[0061] For the marked potential abnormal data, cluster analysis is further performed using the density-based spatial clustering algorithm (DBSCAN), and data points that are far away from the normal data cluster center but have a low density are identified as true outliers and removed to obtain a cleaned high-quality data set D clean This real-time adaptive data cleaning method can process massive amounts of real-time data quickly and accurately, effectively avoiding misjudgment and missed judgment problems caused by fixed thresholds or static models, and significantly improving the quality and reliability of data.

[0062] In dynamic intelligent data normalization, it should be noted that the probability distribution of the cleaned data set is estimated in real time to determine the distribution type and parameters of each dimensional data; according to the data distribution type, the corresponding normalization transformation method is selected to normalize the data set to obtain a normalized data set; during the normalization process, the time series trend of the data set is modeled and predicted, and when it is detected that the data has a preset trend change, the normalization parameters are dynamically adjusted.

[0063] In practice, a normalization method based on dynamic estimation of data distribution is adopted. First, the cleaned data set D clean Perform real-time probability distribution estimation and use non-parametric methods such as kernel density estimation (KDE) to determine the distribution type (such as normal distribution, gamma distribution, uniform distribution, etc.) and parameters (such as mean, variance, shape parameter, scale parameter, etc.) of each dimensional data.

[0064] According to the data distribution type, select the appropriate normalization transformation method. For normally distributed data, use the improved adaptive Z-score normalization method to dynamically adjust the normalization parameters according to the real-time mean and standard deviation of the data to avoid poor normalization results due to slight changes in data distribution; for non-normally distributed data, such as gamma distribution, first convert it to an approximate normal distribution through Box-Cox transformation or Johnson transformation, and then perform normalization.

[0065] At the same time, a normalization adjustment mechanism based on time series trend analysis is introduced. Linear regression, exponential smoothing and other methods are used to model and predict the time series trend of the data. When obvious trend changes (such as upward trend, downward trend or periodic fluctuations) are detected in the data, the normalization parameters are dynamically adjusted accordingly to ensure that the normalized data can truly reflect the trend of the equipment's operating status changes, and the normalized data set D is obtained. normalized , providing a stable and reliable data foundation for subsequent data analysis and model training.

[0066] In data quality assessment and feedback optimization, a comprehensive data quality assessment indicator system is established, including multiple dimensions such as data accuracy, completeness, consistency, timeliness, and relevance. By comparing and analyzing the known equipment physical model, historical data statistical laws, and industry standard data, the data quality of each dimension is quantitatively evaluated to obtain a data quality score Q = {q1, q2, …, q m}, where m is the number of evaluation metrics.

[0067] Based on the data quality score, the reinforcement learning algorithm is used to formulate an optimization strategy for data collection and preprocessing. If the data accuracy score is low, the model will automatically adjust the sensor calibration parameters or replace the sensor that may be faulty; if the integrity score is insufficient, it will check whether there are any missing data collection links and take corresponding supplementary measures; if there is a problem with consistency, the data from different sources will be resynchronized and calibrated. Through continuous evaluation and optimization, the data quality is continuously improved, forming a closed-loop feedback control system to ensure the efficiency and reliability of the entire data collection and preprocessing process.

[0068] In data compression and storage optimization, a data compression method based on wavelet transform and principal component analysis (PCA) is used. First, the normalized data set D normalized Perform wavelet transform to decompose the data into wavelet coefficients of different frequencies, then use PCA to reduce the dimension of the wavelet coefficient matrix, extract the main characteristic components, and retain the key coefficients that can represent most of the data information, so as to achieve efficient data compression. Design an intelligent data storage architecture to store data in different storage media such as cache, solid-state drive (SSD), and hard disk drive (HDD) according to the time series characteristics, importance level, and access frequency of the data. For recent high-frequency access data and key data, store them in cache and SSD to ensure fast data reading and writing; for historical data and low-frequency access data, store them in HDD and use distributed file system and data archiving technology to achieve long-term reliable data storage and efficient management, and greatly reduce data storage costs and management complexity while ensuring data availability.

[0069] In step S12, it includes: building a multi-domain fusion feature extraction framework, deep feature learning and automatic encoding, feature correlation analysis and screening, sorting and selection based on feature importance, feature validity verification and optimization, and establishing a feature adaptive update mechanism.

[0070] In the construction of the multi-domain fusion feature extraction framework, it should be noted that the multi-domain fusion feature extraction framework is constructed, and the multi-domain fusion feature extraction framework is used to perform multi-domain fusion feature extraction on the sensor data to obtain multi-domain fusion features; the multi-domain fusion feature extraction framework adopts the combination of VMD and HHT in the time domain, the WPT and FFT joint analysis method in the frequency domain, and the SST and GST tools in the time and frequency domains.

[0071] In specific practice, in terms of time domain feature extraction, a method based on variational mode decomposition (VMD) and Hilbert-Huang transform (HHT) is introduced. First, VMD is used to adaptively decompose the original signal into multiple modal components with different center frequencies and bandwidths. Each modal component represents the local characteristics of the signal at different time scales. Then, HHT is applied to each modal component for further analysis to extract features such as instantaneous frequency, instantaneous amplitude, and marginal spectrum. These features can effectively reflect the time-varying and nonlinear characteristics of the signal, which is especially important for the detection of early weak faults in equipment.

[0072] When extracting frequency domain features, a joint analysis method based on wavelet packet transform (WPT) and fast Fourier transform (FFT) is used. WPT can decompose the signal into different frequency sub-bands, perform FFT transform on the signal of each sub-band, and obtain detailed spectrum information, thereby accurately extracting the energy distribution, frequency center of gravity, spectral kurtosis and other features of the signal in different frequency bands. These features are of great value for identifying the fault type and fault severity of the equipment.

[0073] The time-frequency domain fusion feature extraction uses advanced time-frequency analysis tools such as synchronous compression transform (SST) and generalized S transform (GST). SST can represent the signal with high resolution on the time-frequency plane, highlighting the time-frequency energy distribution characteristics of the signal; GST further improves the accuracy and flexibility of time-frequency analysis by modifying the S transform. By extracting these joint features in the time-frequency domain, such as time-frequency, time-frequency peak, and time-frequency ridge, the time-varying and frequency-varying characteristics of the signal can be fully characterized, providing richer and more accurate information for equipment fault diagnosis.

[0074] In deep feature learning and automatic encoding, it should be noted that it includes: constructing a deep learning AE model to perform deep feature learning on multi-domain fusion features; the deep learning AE model includes an encoder and a decoder. When training the deep learning AE model, a GAN adversarial training mechanism is introduced to construct a discriminator for distinguishing between original and reconstructed features.

[0075] An autoencoder (AE) model based on deep learning is constructed to perform deep feature learning and compression on multi-domain fusion features. The AE model consists of two parts: an encoder and a decoder. The encoder maps the high-dimensional original feature vector to a low-dimensional hidden layer representation, and learns the potential feature pattern of the data through a multi-layer fully connected neural network and nonlinear activation functions (such as Relu, Sigmoid, etc.); the decoder restores the hidden layer representation back to the original feature space, and trains the model by minimizing the reconstruction error, so that the model can automatically extract the most representative features.

[0076] In the training process of the AE model, an adversarial training mechanism based on the generative adversarial network (GAN) is introduced to build a discriminator network for adversarial training with the autoencoder. The task of the discriminator is to distinguish between the real original features and the reconstructed features generated by the autoencoder, while the autoencoder tries to deceive the discriminator. Through this adversarial training, the autoencoder can learn a more powerful and discriminative feature representation, improving the model's feature extraction ability and generalization performance.

[0077] In feature correlation analysis and screening, a method based on the combination of maximum information coefficient (MIC) and Pearson correlation coefficient (PCC) is used for feature correlation analysis. MIC can measure the nonlinear correlation between two variables, while PCC is mainly used to evaluate linear correlation. By combining these two methods, the complex correlation relationship between features can be comprehensively analyzed.

[0078] In the ranking and selection of feature importance, feature importance evaluation methods based on ensemble learning algorithms such as random forest (RF) and XGBoost are used. By performing multiple random sampling and model training on a large amount of sample data, the contribution of each feature in the model decision process, i.e., the feature importance score, is calculated.

[0079] According to the feature importance score, the features are sorted in descending order, and the first k features with importance scores higher than a certain threshold or the greatest contribution to model performance improvement are selected using threshold-based selection strategies or stepwise regression methods such as forward selection and backward selection (k is determined by cross-validation and other methods based on actual conditions) to form the final high-quality feature subset F selected , which is used for subsequent model training and analysis to ensure that the selected features can reflect the operating status and fault information of the equipment to the greatest extent.

[0080] In the feature validity verification and optimization of feature extraction and selection, it should be noted that the validity of the selected feature subset is verified using a method based on leave-one-out (LOO) or K-fold cross-validation. The data set is divided into a training set and a test set, and the training and testing process is repeated many times. Different samples are selected as test sets each time, and the performance indicators of the model on different test sets are calculated, such as accuracy, recall, F1 value, root mean square error (RMSE), etc., to evaluate the impact of feature subsets on model performance. If it is found that the performance of the feature subset does not meet the expected goal, such as the accuracy is lower than 90% or the RMSE is greater than 0.5 (the specific threshold is determined according to the actual application requirements), the feature extraction and selection method is optimized and adjusted.

[0081] In the establishment of the feature adaptive update mechanism, with the operation of water conservancy equipment and the continuous accumulation of data, a feature adaptive update mechanism is established. Newly collected data is analyzed regularly (such as monthly or quarterly), the importance and relevance of features are re-evaluated, and new valid features are incorporated into the feature subset using online learning and incremental learning methods, while features that may become no longer important as the equipment ages or operating conditions change are eliminated.

[0082] In step S13, it includes: hybrid life prediction model architecture design, model parameter initialization and pre-training, and model training based on improved particle swarm optimization.

[0083] In the design of the hybrid life prediction model architecture, it should be noted that a hybrid life prediction model architecture is designed that integrates convolutional neural network (CNN), recurrent neural network (RNN) and its variant long short-term memory network (LSTM) and gated recurrent unit (GRU) and support vector regression (SVR). Construct a hybrid model that integrates CNN, RNN (taking LSTM as an example), GRU and SVR. For the input data set X = {x1, x2, …, x n}(x i is the i-th sample feature vector, with dimension d), and the l-th convolutional layer of the CNN part outputs the feature map H l The following improved formula is used for calculation:

[0084]

[0085] Among them, K l-1 is the number of convolution kernels in the l-1th layer, is the kth convolution kernel weight matrix of the lth layer, H l-1 is the output feature map of the previous layer (H 0 =X), b l is the bias vector, * is the convolution operation, f(·) is the activation function (such as ReLU), represents element-wise multiplication, It is a learnable mask matrix introduced to adaptively adjust the weight distribution of feature maps and enhance the model's ability to extract key features. Its initial value is randomly generated in the interval [0,1] and updated through back propagation during training. The pooling layer uses an improved adaptive pooling method and outputs P l The calculation of is as follows:

[0086]

[0087] Among them, R ij is the area corresponding to the feature map of the pooling window in the lth layer, and β is a learnable parameter used to control the importance weight of features at different positions. The LSTM unit processes the time series features of the CNN output, and its formula is improved as follows:

[0088] i t =σ(W xi x t +W hi h t-1 +U ci c t-1 +b i )

[0089] f t =σ(W xf x t +W hf h t-1 +U cf c t-1 +b f )

[0090] o t =σ(W xo x t +W ho h t-1 +U co c t-1 +b o )

[0091]

[0092]

[0093] Among them, the newly added U ci ,U cf ,U co It is the weight matrix between the input and memory units, which is used to more flexibly control the flow of information and the updating of memory, and enhance LSTM's ability to process long sequence data and capture complex time dependencies.

[0094] In the attention mechanism, let the LSTM hidden state sequence H = {h1,h2,…,h T}, calculate the weight α of each hidden state t The formula is improved to:

[0095] e t =v T tanh(W h h t +U h (h t -h t-1 )+b h )

[0096]

[0097] Among them, U h is a weight matrix used to consider the change of hidden state, so that the attention mechanism can better pay attention to the dynamic change information in the sequence. The final attention weighted hidden state hatt is:

[0098]

[0099] Among them, W att and b att It is a parameter used to further transform the weighted hidden state to enhance the expressiveness of the feature.

[0100] The SVR part uses an improved radial basis function (RBF) kernel function:

[0101] K(x i ,x j )=exp(-γx i -x j +Δx ij 2 )

[0102] Where Δx ij It is a correction vector calculated based on the local density information of the data, which is used to adjust the distance metric between samples so that the kernel function can better adapt to the distribution characteristics of the data. The SVR model is trained by minimizing the following objective function:

[0103]

[0104] Among them, λ is the balance term coefficient, θ ij It is a weight coefficient calculated based on the similarity and local structure information of the data. It is used to constrain the smoothness of the model in the feature space and improve the generalization ability of the model.

[0105] In the model parameter initialization and pre-training, the parameters of each sub-model in the hybrid model are initialized reasonably according to the historical data and operation characteristics of the water conservancy equipment. For the CNN part, the model parameters pre-trained on large-scale image data sets or other data with similar characteristics are used as the initial values, and fine-tuned according to the characteristics of the water conservancy equipment data, such as adjusting the size, number and step size of the convolution kernel; for the RNN and its variants, the number of neurons in the hidden layer, the weight matrix, and the bias parameters of the forget gate, input gate and output gate are initialized according to the time series length and feature dimension of the data: for the SVR part, according to the prior knowledge and preliminary statistical analysis of the data, the initial parameter range of its kernel function (such as radial basis function RBF kernel, polynomial kernel, etc.) is determined, such as the gamma value of the kernel parameter, the polynomial degree, etc., and the parameters are initialized within a reasonable range by random sampling, which provides a good starting point for subsequent model training, accelerates the convergence speed of the model and improves the performance stability of the model.

[0106] A pre-training strategy based on transfer learning is adopted to pre-train the hybrid model using a large amount of data from other similar water conservancy equipment or related fields so that the model can learn some common feature representations and prediction patterns. During the pre-training process, feature alignment and domain adaptation technology of pre-training data and target water conservancy equipment data are used to reduce the impact of data distribution differences on model performance, improve the generalization ability and adaptability of the model, and enable the model to quickly converge to a better performance state when facing a small amount of target equipment data.

[0107] In the model training based on improved particle swarm optimization, the hybrid life prediction model is iteratively trained using feature subsets, and the parameters of the hybrid life prediction model are optimized using a parameter optimization method based on the improved particle swarm optimization algorithm; the improved particle swarm optimization algorithm, on the basis of the particle swarm optimization algorithm, introduces an adaptive inertia weight adjustment strategy and a dynamic neighborhood search mechanism, and dynamically adjusts the inertia weight according to the distance between the current position of the particle and the historical optimal position and the distribution of the entire particle swarm; in the iterative training process, a weighted combined loss function that minimizes the mean square error and the mean absolute error between the predicted value and the true value is used as the objective function.

[0108] In one embodiment, the dataset F after feature selection is selected Divide into training set F train , validation set F val and the test set F test , the division ratio can be set to 70:15:15. Using the training set F trainThe hybrid life prediction model is trained using a parameter optimization method based on an improved particle swarm optimization (PSO) algorithm. Based on the traditional PSO algorithm, an adaptive inertia weight adjustment strategy and a dynamic neighborhood search mechanism are introduced to dynamically adjust the inertia weight according to the distance between the current position of the particle and the historical optimal position and the distribution of the entire particle swarm, so that the particles can better balance the global search and local search capabilities in the search space; at the same time, through the dynamic neighborhood search mechanism, particles can communicate information and conduct collaborative searches with other particles in the neighborhood to avoid falling into the local optimal solution.

[0109] During the training process, the weighted combination loss function of minimizing the mean square error (MSE) and mean absolute error (MAE) between the predicted value and the true value is used as the objective function. The parameters of each sub-model in the hybrid model are optimized and adjusted through the improved PSO algorithm, so that the model can better fit the training data and improve the accuracy and stability of the prediction. After multiple iterations of training, until the performance indicators of the model on the validation set reach a certain convergence condition or no longer have a significant improvement, the initially trained hybrid life prediction model M is obtained. hybrid .

[0110] In step S14, it includes: multivariate Gaussian mixture model (GMM) construction and parameter estimation, dynamic fault threshold determination and adaptive adjustment, and fault feature extraction and classification based on deep belief network (DBN).

[0111] In the construction and parameter estimation of multivariate Gaussian mixture model (GMM), it should be noted that the improved EM algorithm based on hierarchical clustering and variational inference is used to estimate the GMM model parameters.

[0112] For the normal operation status data set X = {x1, x2, ..., x n}(x i is the i-th sample feature vector with a dimension of d). Assuming that it is generated by a GMM model consisting of K Gaussian components, its probability density function is:

[0113]

[0114] Among them, π k is the kth Gaussian component mixing coefficient The mean is μ k , the covariance matrix is ​​∑ k Gaussian distribution probability density function:

[0115]

[0116] The GMM model parameters are estimated using an improved EM algorithm based on hierarchical clustering and variational inference. In the E step, the data is first pre-grouped by hierarchical clustering to cluster similar data points into multiple subsets, and then the posterior probability of the data point belonging to each Gaussian component is calculated in each subset:

[0117]

[0118] In the M step, the model parameters are updated:

[0119]

[0120] Among them, η k ,δ k ,Σ k is a small correction term introduced to prevent overfitting and numerical instability during parameter update, and I is the unit matrix. Through multiple iterations of E and M steps until the model parameters converge to a better value, this improved algorithm can better handle the complex distribution and noise of data and improve the accuracy of GMM modeling of the normal operation status of the equipment.

[0121] In the determination and adaptive adjustment of dynamic fault thresholds, the Mahalanobis distance (MD) threshold and probability density threshold of each feature dimension are calculated according to the parameter estimation results of the GMM model for fault detection. The Mahalanobis distance threshold is determined by statistical analysis of normal data, for example, the Mahalanobis distance critical value under a certain confidence level (such as 95% or 99%) is selected; the probability density threshold is calculated according to the probability density function of each Gaussian component, and the probability value below the threshold is regarded as abnormal. The threshold adaptive adjustment mechanism based on sliding window and online learning is introduced. With the operation of the equipment and the continuous update of data, the sliding window is used to monitor and analyze the newly collected data in real time. When the statistical characteristics of the data are detected to have obvious changes (such as significant shifts in the mean and variance), the GMM model is updated through the online learning algorithm (such as incremental, EM algorithm), and the fault threshold is adjusted accordingly, so that it can adapt to the dynamic changes of the equipment operation status, effectively reduce the false alarm rate and missed alarm rate, and improve the timeliness and accuracy of fault monitoring.

[0122] In the fault feature extraction and classification based on deep belief network (DBN), a deep belief network (DBN) model is constructed to perform deep fault feature extraction and classification on equipment operation data. DBN is composed of multiple restricted Boltzmann machines (RBMs) stacked together and trained by combining unsupervised layer-by-layer pre-training and supervised reverse fine-tuning. First, each RBM is pre-trained using the contrastive divergence (CD) algorithm so that it can automatically learn the multi-level abstract feature representation of the data; then, based on the pre-training, the entire DBN is fine-tuned in a supervised manner using the back propagation algorithm so that it can accurately classify the fault type based on the extracted features.

[0123] In order to improve the learning ability and generalization performance of the DBN model for small sample fault data, a pre-training strategy based on transfer learning is introduced. The DBN is pre-trained using a large amount of fault data from other related fields or similar equipment, and then fine-tuned on a small amount of fault data of the target water conservancy equipment, so that the model can quickly adapt to the fault characteristic mode of the target equipment, effectively solving the problem of model training difficulties caused by the relatively small amount of fault data of water conservancy equipment, and improving the accuracy and reliability of fault classification.

[0124] In step S15, it includes: real-time data collection and synchronous processing, real-time data cleaning and dynamic normalization update, real-time feature extraction and adaptive selection, real-time update and prediction of life prediction model, and utilization of fault monitoring model.

[0125] In real-time data collection and synchronous processing, according to the predetermined intelligent collection frequency f, through the optimized data collection network and communication protocol, it is ensured that the operation data of the water conservancy equipment can be transmitted to the data processing center in real time and accurately. During the collection process, a synchronization mechanism based on timestamp and data verification is adopted to ensure the consistency and accuracy of the data collected by different sensors in time, and avoid analysis errors caused by data synchronization problems.

[0126] The collected real-time data is quickly and preliminarily processed, including data format conversion, outlier marking (but not temporarily eliminated to avoid losing possible fault information), and other operations. The processed data is stored in the real-time data buffer, awaiting further cleaning, normalization and feature extraction processing to ensure the efficiency and continuity of the data processing process, and provide timely and reliable data support for subsequent real-time analysis and prediction.

[0127] In real-time data cleaning and dynamic normalization update, the deep learning-based anomaly detection model trained on historical data (such as the model in step S11) is used to perform real-time outlier detection and cleaning on the data in the real-time data buffer. At the same time, according to the statistical characteristics and distribution of the newly collected data, the parameters of data normalization (such as mean, standard deviation, etc.) are dynamically updated, and a dynamic normalization method based on a sliding window is adopted to ensure that the normalized real-time data can accurately reflect the change trend of the current operating status of the equipment, maintain good comparability and consistency with historical data, and provide a stable data foundation for subsequent feature extraction and model prediction.

[0128] In real-time feature extraction and adaptive selection, a feature extraction model trained on historical data (such as the deep feature learning and automatic encoding model in step S12) is used to extract features from the cleaned and normalized real-time data to obtain a real-time feature vector. At the same time, based on the characteristics of the real-time data and the feedback information of the model, the algorithm and parameters of feature extraction, as well as the selected feature subset, are dynamically adjusted using online learning and incremental learning methods. For example, when a significant change in the operating conditions of the equipment is detected, a new feature extraction method is automatically introduced or the threshold of feature selection is adjusted so that the extracted features can better adapt to the current state of the equipment, improve the effectiveness and adaptability of the features, and ensure that the model can capture subtle changes in the operating status of the equipment and potential fault information in a timely manner.

[0129] In the real-time update and prediction of the life prediction model, the real-time feature vector is input into the trained life prediction model (such as the hybrid life prediction model in step S13) to obtain the real-time life prediction value of the equipment. At the same time, according to the new operation data, regularly (for example, every certain period of time T update ) The life prediction model is updated online, and a small sample incremental learning method is adopted to incorporate new data samples into the model training process at a lower computational cost, and the model parameters are fine-tuned so that the model can continuously adapt to changes in the equipment's operating status, thereby improving the accuracy and timeliness of life prediction and providing timely and accurate basis for equipment maintenance decisions.

[0130] In utilizing the fault monitoring model, the real-time feature vector is input into a trained life prediction model (such as the fault monitoring model in step S14), and the fault monitoring model can be used to perform real-time fault monitoring on the water conservancy equipment.

[0131] Compared with the prior art, the present invention has the following beneficial effects:

[0132] Multi-source heterogeneous data collection planning uses modeling and simulation technology to accurately locate monitoring points, dynamically adjusts collection frequency according to working conditions, and ensures that the data is comprehensive and targeted; real-time adaptive data cleaning combines deep neural networks and advanced mechanisms to efficiently process massive data and greatly improve data quality; dynamic intelligent data normalization is dynamically adjusted according to data distribution to keep data stable and comparable, laying a solid foundation for subsequent analysis and ensuring data validity from the source.

[0133] The multi-domain fusion feature extraction framework integrates multiple time-frequency domain methods to capture equipment fault characteristics in all aspects; deep feature learning and automatic encoding use deep learning models and adversarial training to enhance feature discrimination; feature correlation analysis and screening comprehensively consider multiple coefficients to ensure that the selected features are closely related to the equipment status; operations based on feature importance screen key features, and after effectiveness verification and adaptive updates, continuously optimize feature subsets to improve model input quality.

[0134] The life prediction model integrates multiple neural networks and improved algorithms, and optimizes and innovates various components, such as CNN mask matrix and LSTM unit improvements. After parameter initialization, transfer learning and efficient training, it can accurately predict the equipment life. In the fault monitoring model, GMM improved parameter estimation, combined with dynamic threshold adjustment and DBN transfer learning training, effectively monitors faults, reduces misjudgment, provides strong decision-making support for equipment maintenance, ensures the reliable operation of water conservancy equipment, and reduces failure losses.

[0135] According to a second aspect of an embodiment of the present invention, there is provided a hydraulic equipment life prediction and fault monitoring device, comprising:

[0136] A main controller, and a memory connected to the main controller;

[0137] a memory in which program instructions are stored;

[0138] The main controller is used to execute program instructions stored in the memory to perform any of the above methods.

[0139] According to a third aspect of an embodiment of the present invention, a computer-readable storage medium is provided, storing a computer program, and when the computer program is executed by a processor, the method described in any one of the above is implemented.

[0140] It can be understood that the same or similar parts of the above embodiments can be referenced to each other, and the contents not described in detail in some embodiments can refer to the same or similar contents in other embodiments.

[0141] It should be noted that, in the description of the present invention, the terms "first", "second", etc. are only used for descriptive purposes and cannot be understood as indicating or implying relative importance. In addition, in the description of the present invention, unless otherwise specified, the meaning of "plurality" refers to at least two.

[0142] Any process or method description in a flowchart or otherwise described herein may be understood to represent a module, segment or portion of code that includes one or more executable instructions for implementing the steps of a specific logical function or process, and the scope of the preferred embodiments of the present invention includes alternative implementations in which functions may not be performed in the order shown or discussed, including performing functions in a substantially simultaneous manner or in the reverse order depending on the functions involved, which should be understood by those skilled in the art to which the embodiments of the present invention belong.

[0143] It should be understood that the various parts of the present invention can be implemented by hardware, software, firmware or a combination thereof. In the above-mentioned embodiments, a plurality of steps or methods can be implemented by software or firmware stored in a memory and executed by a suitable instruction execution system. For example, if implemented by hardware, as in another embodiment, it can be implemented by any one of the following technologies known in the art or their combination: a discrete logic circuit having a logic gate circuit for implementing a logic function for a data signal, a dedicated integrated circuit having a suitable combination of logic gate circuits, a programmable gate array (PGA), a field programmable gate array (FPGA), etc.

[0144] A person skilled in the art may understand that all or part of the steps in the method for implementing the above-mentioned embodiment may be completed by instructing related hardware through a program, and the program may be stored in a computer-readable storage medium, which, when executed, includes one or a combination of the steps of the method embodiment.

[0145] In addition, each functional unit in each embodiment of the present invention may be integrated into a processing module, or each unit may exist physically separately, or two or more units may be integrated into one module. The above-mentioned integrated module may be implemented in the form of hardware or in the form of a software functional module. If the integrated module is implemented in the form of a software functional module and sold or used as an independent product, it may also be stored in a computer-readable storage medium.

[0146] The storage medium mentioned above can be a read-only memory, a magnetic disk or an optical disk, etc.

[0147] In the description of this specification, the description with reference to the terms "one embodiment", "some embodiments", "examples", "specific examples", or "some examples" means that the specific features, structures, materials or characteristics described in conjunction with the embodiment or example are included in at least one embodiment or example of the present invention. In this specification, the schematic representation of the above terms does not necessarily refer to the same embodiment or example. Moreover, the specific features, structures, materials or characteristics described may be combined in any one or more embodiments or examples in a suitable manner.

[0148] Although the embodiments of the present invention have been shown and described above, it is to be understood that the above embodiments are exemplary and are not to be construed as limitations of the present invention. A person skilled in the art may change, modify, replace and vary the above embodiments within the scope of the present invention.

Claims

1. A method for predicting the life of hydraulic equipment and monitoring faults, characterized in that: include: S11. Carry out comprehensive digital modeling of the water conservancy equipment to be tested, determine the fault-sensitive areas and key monitoring points according to the equipment model, and generate a sensor deployment plan; and performing simulation according to the constructed device model and sensor deployment scheme to obtain simulated sensor data, and preprocessing the sensor data; S12, extracting multi-domain fusion features from the sensor data to obtain multi-domain fusion features, performing deep feature learning and feature correlation analysis on the multi-domain fusion features in sequence, and placing features whose feature importance is greater than a threshold into a feature subset; S13, constructing a hybrid life prediction model architecture, initializing and pre-training parameters of the hybrid life prediction model; using a feature subset to formally train the pre-trained hybrid life prediction model to obtain a trained hybrid life prediction model; S14, construct a multivariate Gaussian mixture model, and perform parameter estimation on the multivariate Gaussian mixture model according to a preset data set of normal operation status of the equipment; calculate the Mahalanobis distance threshold and probability density threshold of each feature dimension in the feature subset according to the parameter estimation result; construct a deep belief network for fault feature extraction and classification; and obtain a fault monitoring model based on the multivariate Gaussian mixture model and the deep belief network; S15, acquiring sensor data in real time, wherein the sensor is deployed according to the sensor deployment plan; preprocessing the sensor data, extracting features from the preprocessed sensor data using step S12 to obtain a real-time feature vector; inputting the real-time feature vector into a hybrid life prediction model to perform life prediction, and inputting the real-time feature vector into a fault monitoring model to perform fault monitoring.

2. The method for predicting the life of hydraulic equipment and monitoring faults according to claim 1, characterized in that: In step S11, it also includes: constructing an acquisition frequency adaptive adjustment model using a machine learning algorithm according to a dynamic acquisition strategy of the working condition and environment of the water conservancy equipment to be tested; In step S15, when the sensor data is acquired in real time, the method further includes: inputting the sensor data acquired in real time into the acquisition frequency adaptive adjustment model to obtain the sensor acquisition frequency, and controlling the acquisition frequency of each sensor according to the sensor acquisition frequency.

3. The method for predicting the life of hydraulic equipment and monitoring faults according to claim 1, characterized in that: In step S12, it also includes: Construct a multi-domain fusion feature extraction framework, and use the multi-domain fusion feature extraction framework to extract multi-domain fusion features from the sensor data to obtain multi-domain fusion features; the multi-domain fusion feature extraction framework uses a combination of VMD and HHT in the time domain, a WPT and FFT joint analysis method in the frequency domain, and SST and GST tools in the time-frequency domain; A deep learning AE model is constructed to perform deep feature learning on multi-domain fusion features; the deep learning AE model includes an encoder and a decoder. When training the deep learning AE model, a GAN adversarial training mechanism is introduced to construct a discriminator for distinguishing original and reconstructed features; Feature correlation analysis was performed based on a method combining the maximum information coefficient and the Pearson correlation coefficient; The feature importance evaluation method is used to obtain the importance score of each feature, and the features with feature importance scores greater than the threshold are placed in the feature subset.

4. The method for predicting the life of hydraulic equipment and monitoring faults according to claim 3 is characterized in that: In step S12, it also includes: The validity of the selected feature subset is verified using a leave-one-out method or a K-fold cross-validation method. If the verification fails, step S12 is optimized and adjusted.

5. The method for predicting the life of hydraulic equipment and monitoring faults according to claim 1, characterized in that: In step S13, it also includes: The hybrid life prediction model architecture is composed of a convolutional neural network, a recurrent neural network and its variant, a long short-term memory network, a gated recurrent unit, and a support vector regression; During formal training: the hybrid life prediction model is iteratively trained using feature subsets, and the parameters of the hybrid life prediction model are optimized using a parameter optimization method based on an improved particle swarm optimization algorithm; the improved particle swarm optimization algorithm, based on the particle swarm optimization algorithm, introduces an adaptive inertia weight adjustment strategy and a dynamic neighborhood search mechanism, and dynamically adjusts the inertia weight according to the distance between the current position of the particle and the historical optimal position and the distribution of the entire particle swarm; During the iterative training process, the objective function is the weighted combination loss function that minimizes the mean square error and the mean absolute error between the predicted value and the true value.

6. The method for predicting the life of hydraulic equipment and monitoring faults according to claim 1, characterized in that: In step S14, Parameter estimation is performed on the multivariate Gaussian mixture model, including: estimating GMM model parameters using an improved EM algorithm based on hierarchical clustering and variational inference.

7. The method for predicting the life of hydraulic equipment and monitoring faults according to claim 1, characterized in that: Preprocessing the sensor data includes: Build a deep neural network anomaly detection model, introduce sliding windows and dynamic threshold mechanisms, segment new sensor data into fixed-length windows, calculate statistical features and compare them with dynamic thresholds, and mark potential abnormal data; Cluster analysis is performed on the abnormal data to remove abnormal values ​​that meet the abnormal conditions to obtain a cleaned data set.

8. The method for predicting the life of hydraulic equipment and monitoring faults according to claim 7, characterized in that: After getting the cleaned data set, it also includes: Perform real-time probability distribution estimation on the cleaned data set to determine the distribution type and parameters of each dimension of data; According to the data distribution type, the corresponding normalization transformation method is selected to normalize the data set to obtain a normalized data set; during the normalization process, the time series trend of the data set is modeled and predicted, and when it is detected that the data has a preset trend change, the normalization parameters are dynamically adjusted.

9. A hydraulic equipment life prediction and fault monitoring device, characterized in that: include: A main controller, and a memory connected to the main controller; a memory in which program instructions are stored; The main controller is used to execute program instructions stored in the memory and perform the method according to any one of claims 1 to 8.

10. A computer-readable storage medium storing a computer program, characterized in that: When the computer program is executed by a processor, the method according to any one of claims 1 to 8 is implemented.

Images