CN-121980395-A - Bearing fault detection research method based on improved depth forest algorithm

Abstract

The invention designs a bearing fault detection method based on an improved depth forest model, and belongs to the technical field of fault detection. The method is characterized in that rolling bearing fault signals are used as features, the signals are converted into a frequency domain, an integrated learning framework is used for improving a depth forest algorithm to be used as a classifier for recognition on the basis of a Convolutional Neural Network (CNN) algorithm and a random forest model, the multi-granularity scanning structure of the convolutional neural network is used for replacing the depth forest to conduct feature extraction on the data model, cascading CatBoost (cascade catboost, casCatBoost) is used for replacing the cascading forest, diversity of the integrated learner is increased, when the signal to noise ratio is gradually reduced, the improved depth forest framework is lower in sensitivity degree to noise, the method can replace larger memory consumption of multi-granularity scanning in the original depth forest, and the problems of low single model prediction precision and low original structure operation efficiency are solved. Therefore, a great amount of signal processing work in bearing fault diagnosis is reduced, and the defects of excessive over-parameters, limited calculation facilities and the like are overcome.

Inventors

• ZHOU YUETING
• LIU XIEYU
• Xue Huitang
• CHENG XIANGDONG
• WANG CHONG
• YANG FUHUA

Assignees

• 江苏瀚微半导体科技有限责任公司

Dates

Publication Date

20260505

Application Date

20251204

Claims (4)

1. 1. The bearing fault detection research method based on the improved depth forest algorithm is characterized by comprising the following steps of: Step a, because noise exists in the collected signals, in order to reduce the influence of noise in the original signals on the detection result, kalman filtering processing is firstly adopted on the collected original signals, and the basic idea of the Kalman filtering is to correct the prediction estimation of the state variables by using the observation data so as to obtain the optimal estimation of the state variables; preprocessing the vibration signal of the rolling bearing by using the statistical measure of the time domain signal to obtain the frequency domain characteristic of the signal, wherein the spectrum obtained by Fast Fourier Transform (FFT) is divided into a plurality of wave bands, then carrying out iterative calculation on the frequency domain amplitude according to the formula (1), (1) Wherein: Representing the length of the one-dimensional time series, Representing the iteration characteristic of the current time t; step c, extracting Deep Iteration Features (DIF) by using a Convolutional Neural Network (CNN), namely setting the number of convolutional layers, the size of a convolutional kernel, the size of a pooling layer, the pooling step length and the number of filters of the convolutional neural network, taking the iteration features as the input of the convolutional neural network, performing feature analysis on time sequence signals by using the convolutional neural network, extracting local information, and mining the contained deep features, thereby obtaining the DIF; Constructing a cascade CatBoost (CasCatBoost) prediction model, namely constructing a first cascade layer by taking the extracted deep iteration characteristic as CasCatBoost input, wherein the cascade layer comprises four CatBoost, four CatBoost of the cascade layer respectively obtain a decision coefficient R2 as a prediction result, constructing a second cascade layer, continuously using four CatBoost as weak learners in the cascade layer, taking the output result and the original input of the first layer as the input of the weak learners, and respectively obtaining a decision coefficient R2 by the weak learners again; step e, inputting a test set into a model to detect whether faults occur, and obtaining a current signal time-frequency diagram of the rolling bearing, a current signal time-frequency diagram of the rolling bearing under real use and fitting degree between the current signal time-frequency diagram and the current signal time-frequency diagram of the rolling bearing by utilizing CasCatBoost prediction models as fault diagnosis indexes; Step f. Rolling bearing fault detection, firstly extracting DIF from rolling bearing original signals in a test set, then inputting the DIF into a trained CasCatBoost prediction model to obtain a performance gain index, fitting the DIF by using a linear function smoothing curve to obtain a fault detection percentage p value trend of each point in the future, when a threshold value p=1 is reached, recognizing that the p value trend is reached, extracting features from a feature map of a time sequence signal by using a convolutional neural network, inputting the extracted feature data into the CasCatBoost model to train, reflecting the quality of the fault detection performance of the model by using an error E displayed by an information gain index, And g, improving a determination coefficient R2 formula in the depth forest model as follows: 。
2. 2. The original signal is a measurement signal in a time domain obtained directly from a state monitoring system, the intelligent fault diagnosis method based on the improved depth forest consists of three main steps of data acquisition, feature extraction and fault type classification, starting from increasing the diversity of the ensemble learner, the disturbance on input attribute and algorithm parameters is achieved by adding CatBoost ensemble learners in a multi-granularity cascade forest (GCForest) framework, the diversity of the whole framework is increased, The method for detecting and researching bearing faults based on the improved depth forest algorithm as claimed in claim 1, wherein, The step e is that the convolution layer is the basic structure of the Convolution Neural Network (CNN), takes the convolution kernel as the basic unit, and in the cascade lamination, a series of deep features can be obtained by forming the convolution of the input one-dimensional time sequence by a plurality of convolution kernels and adding the offset and then passing through an activation function. The convolution process is shown in formula (1): (1) wherein: Is the j element of the first layer; A j-th convolution region that is a layer 1 feature; A weight matrix corresponding to the convolution kernel; is a bias term. To activate the function.
3. 3. The method for detecting and researching bearing faults based on the improved depth forest algorithm as claimed in claim 1, wherein, The calculation method has the advantages that although the calculation amount is large, the model training time and the test time are both increased in an exponential form, so that the real-time performance of an algorithm is affected, the model training time and the test time can be directly input into a full-connection layer for network calculation in principle, a pooling layer is added behind a convolution layer, and the network calculation amount is reduced by reducing the output characteristic diagram parameters. Assuming that the first layer is a pooling layer, the l-1 layer is a convolution layer, the pooling size is n, and the l-1 layer is obtained by pooling, and can be obtained by using the formula (2): (2) wherein: Representing a pooling operation, the method comprises the steps of, Weights representing the j-th feature of the first layer, The bias of the jth characteristic of the first layer is represented, and after pooling treatment, the dimension of the output characteristic is reduced to 1/n of the original dimension.
4. 4. The method for detecting and researching bearing faults based on the improved depth forest algorithm as claimed in claim 1, wherein, The step g specifically comprises that the rolling bearing fault detection belongs to the regression prediction problem, common regression evaluation indexes comprise error Square Sum (SSE), mean Square Error (MSE), mean absolute percentage error (MSE) and a determination coefficient R2, the calculation mode is shown as the formula (3-6), (3) (4) (5) (6) Wherein: to be a true value of the value, Is that The average value of the sum, In order to be able to predict the value, Under the same dataset, the smaller the value of the Sum of Squares Error (SSE) is, the smaller the error is, the better the prediction effect of the model is, the Mean Square Error (MSE) is the mean value of the sum of squares error, when the number of samples is increased, the value of the coefficient R2 is determined to be [0,1], the value is close to 1, the stronger the interpretation capability of the independent variable to the dependent variable is shown, and the better the fitting degree of the model to the data is proved, so that the determined coefficient R2 is selected as the performance gain index of CasCatBoost through experiments.

Description

Bearing fault detection research method based on improved depth forest algorithm Technical Field The invention belongs to the technical field of fault detection, and relates to a bearing fault detection research method based on an improved depth forest model. Background The rolling bearing is used as one of key components in mechanical equipment, is widely applied to manufacturing industry and production life, the health state of the rolling bearing directly influences the running performance of the whole equipment, bearing failure is one of main reasons for mechanical accidents, and in a real-time situation, the mechanical bearing is frequently worn and corroded for a long time due to environmental stress, manufacturing defects, natural aging and other reasons, and finally, catastrophic mechanical bearing failure is caused. Thus, to avoid the possibility of catastrophic bearing failure, the industry may perform early failure detection and diagnosis of bearing components that begin to degrade. The condition monitoring system continuously evaluates the individual components of the mechanical bearing over the entire service life by means of signals obtained in the time domain. With the advent of big data age, the running speed of the computer and the artificial intelligence algorithm are continuously updated and perfected, various factors promote people to try to use the relevant knowledge of deep learning, in recent years, the deep learning theory induces research waves in different fields, and in view of the strong data learning and analysis capability, the research theory of the deep learning method for mechanical bearing diagnosis is continuously explored and innovated, so that the bearing fault diagnosis result is more accurate than that of the traditional diagnosis method. The deep learning model has the capability of autonomously learning the bearing vibration signal characteristics, and can mine the deep essential characteristics of the bearing vibration signal. Mechanical bearing fault diagnosis relies on feature extraction to a great extent, and a great deal of prior knowledge such as signal processing technology, diagnosis expertise, mechanical expertise and the like is required. On the basis of Convolutional Neural Network (CNN) and random forest algorithm, the multi-granularity cascade function in the improved deep forest algorithm is a substitute for a deep learning framework, and can overcome the defects of deep learning such as the need of a large amount of training data, excessive super parameters, powerful computing facilities and the like. The improved deep forest consists of CNN deep feature extraction and CasCatBoost detection models, can realize end-to-end intelligent fault detection, does not need to preprocess non-numerical features, does not depend on prior knowledge of experts, can perform classification and identification on acquired time domain signals, and performs fault detection on signals converted into a frequency domain. The working principle of the convolutional neural network structure is as follows: The Convolutional Neural Network (CNN) is a typical algorithm in the deep neural network, and uses convolution operation to process spatial information among features, and by constructing a plurality of filters, the weight sharing greatly reduces the number of network parameters, reduces the over-fitting problem, and plays a great role in image and voice recognition. As shown in fig. 1, a typical Convolutional Neural Network (CNN) structure includes a convolutional layer, a pooling layer, and a full-link layer. The design adopts a convolutional neural network to extract the characteristics of one-dimensional vibration signals and digs out the deep characteristics of data. The method comprises the steps of extracting local features of a vibration signal by utilizing a one-dimensional convolution kernel, and then carrying out high-dimensional level comprehensive operation on the local features to obtain global information. (1) Convolutional layer The convolution layer is an infrastructure of a Convolution Neural Network (CNN) and takes a convolution kernel as a basic unit. In cascade lamination, a series of deep features can be obtained by convolving an input one-dimensional time sequence with a plurality of convolution kernels and adding an activation function after offset. The convolution process is shown in formula (1): (1) wherein: Is the j element of the first layer; A j-th convolution region that is a layer 1 feature; A weight matrix corresponding to the convolution kernel; is a bias term. To activate the function. The main function of the activation function is to map the linearly inseparable multidimensional feature in the original dimension to another spatial dimension, making it separable. After the neural network is subjected to multi-cascade addition, the expression capacity of the neural network is greatly improved by the activation function, and the neural netwo