A fault diagnosis method, product and equipment based on a continuous homology fluctuation sequence

Abstract

This invention discloses a fault diagnosis method, product, and device based on persistent homophonic wave sequences, belonging to the field of fault diagnosis technology. The method includes: extracting multi-dimensional features, including time-domain features, frequency-domain features, and topological features; the topological features include phase space reconstruction and persistent homophonic wave sequences; fusing the extracted multi-dimensional features; and inputting the fused feature data into a machine learning classification model for model training or classification. This invention can improve the model's ability to represent complex data by capturing subtle patterns in signals through topological features, without relying on large-scale datasets.

Description

Technical Field

[0001] This invention belongs to the field of fault diagnosis technology, specifically relating to a fault diagnosis method, product, and equipment based on a continuous homophonic fluctuation sequence. Background Technology

[0002] Equipment safety plays a crucial role in industrial applications. However, due to equipment aging and harsh operating environments, equipment failures occur frequently, potentially leading to severe damage and significant economic losses. Therefore, developing technologies capable of timely detection of these failures is essential.

[0003] Fault diagnosis methods can be divided into two categories: model-based methods and data-based methods. Model-based methods utilize mathematical methods, such as partial differential equations and stochastic processes, to analyze faults by establishing a system analysis model related to the equipment. Model-based methods heavily rely on precise information about the equipment-related system, which limits their application in other scenarios. With technological advancements, data-based methods have been applied to fault diagnosis in many scenarios. Data-based methods use artificial intelligence techniques to accomplish fault diagnosis tasks. They do not require analysis of the coupling relationships within the equipment-related system, thus making them suitable for diverse scenarios.

[0004] For data-driven methods, fault diagnosis often employs approaches based on machine learning and deep learning.

[0005] In machine learning-based methods, Guo Chuangxin et al. established a transformer fault diagnosis model based on a multi-class, multi-kernel trained support vector machine (SVM) and verified its effectiveness. Wu Chenxi et al. used wavelet transform and support vector machine to identify rolling bearing faults. Gonzalez et al. used envelope analysis and machine learning to detect bearing faults. Zhou et al. proposed a signal processing method for bearing fault diagnosis using fast Fourier transform (FFT) and machine learning. Goyal et al. used Hilbert transform and support vector machine to analyze laser beam signals to measure machine vibration. In deep learning-based methods, Huang Ting et al. proposed a hierarchical deep learning algorithm based on long short-term memory (LSTM) neural networks for transformer fault diagnosis. Deng Linfeng et al. combined short-time Fourier transform and convolutional neural networks (CNN) for bearing fault diagnosis. Tang Aimin et al. proposed a wavelet kernel CNN-BiLSTM neural network for drilling pumps.

[0006] Data-driven methods do not require precise information about the system, and therefore achieve excellent results. However, these methods have two limitations:

[0007] 1) Most methods require large-scale training data to obtain satisfactory results, and their performance is affected when only a small training dataset is available.

[0008] 2) Most methods do not consider the impact of low signal-to-noise ratio (SNR) on model performance, which limits their practical applications.

[0009] Most previous studies have not taken these two limitations into account. Therefore, it is important to research new fault diagnosis methods.

[0010] Topology is a long-established field of mathematics. With the emergence of Topological Data Analysis (TDA) technology, topology has gradually demonstrated its practical value. TDA is a relatively new branch of data analysis that utilizes techniques derived from topology to analyze data. TDA has already seen successful applications in several fields, such as biomolecular chemistry, drug design, image analysis, time series data analysis, and network analysis. Notably, the winners of the D3R challenge have incorporated TDA into their algorithmic approaches. Furthermore, TDA can be synergistically integrated with machine learning techniques, including deep learning and statistical methods. Despite its applications in many fields, TDA's use in the electrical and smart grid sectors remains very limited. Summary of the Invention

[0011] The present invention aims to at least partially solve one of the technical problems in the aforementioned related technologies.

[0012] Therefore, the purpose of this invention is to provide a fault diagnosis method, product, and device based on a continuous homophonic wave sequence, which can capture subtle patterns in signals through topological features without relying on large-scale datasets, thereby improving the model's ability to represent complex data.

[0013] To solve the above-mentioned technical problems, the present invention is implemented as follows:

[0014] This invention provides a fault diagnosis method based on a continuous homophonic fluctuation sequence. The method combines feature extraction and topological data analysis to construct a fault diagnosis framework suitable for small data volumes and low signal-to-noise ratios.

[0015] By capturing subtle patterns in signals through topological features, the model's ability to represent complex data can be improved.

[0016] In addition, the fault diagnosis method based on a continuous homogeneous fluctuation sequence according to the present invention may also have the following additional technical features:

[0017] In some embodiments, the method includes:

[0018] Multi-dimensional feature extraction is performed, including time-domain features, frequency-domain features, and topological features;

[0019] The extracted multi-dimensional features are fused;

[0020] The fused feature data is then input into a machine learning classification model for model training or classification.

[0021] In some of these implementations, the time-series characteristics include absolute energy, mean, standard deviation, skewness, kurtosis, absolute mean variation, mean second difference center, and / or median.

[0022] In some implementations, the frequency domain features are extracted by converting the original signal to the frequency domain using a fast Fourier transform and capturing the frequency component distribution as the extracted frequency domain features.

[0023] In some of these implementations, the topological features include phase space reconstruction and persistent homophonic wave sequences.

[0024] In some implementations, the phase space reconstruction includes converting a one-dimensional time series into high-dimensional point cloud data based on Takens' embedding theorem, while preserving the signal dynamics characteristics.

[0025] In some implementations, the extraction of the persistent cohomology wave sequence is performed by: performing sliding window processing on the phase space reconstruction time series, calculating the Wassenberg distance between the persistent cohomology maps of adjacent windows, generating 0-dimensional and 1-dimensional persistent cohomology wave sequences, characterizing the dynamic changes in the signal topology, and using these as the extracted persistent cohomology wave sequences.

[0026] In some of these implementations, the specific content of fusing the extracted multi-dimensional features includes: concatenating time-domain features, frequency-domain features, and topological features, and using the concatenated features as the fused feature data.

[0027] In some of these implementations, the machine learning classification model is a support vector machine, a random forest, or a neural network.

[0028] This invention also provides a computer program product, including a computer program that, when executed by a processor, implements the fault diagnosis method based on a continuous homophonic fluctuation sequence as described in any of the preceding embodiments.

[0029] This invention also provides a computer device, including: a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the fault diagnosis method based on a continuous coherent fluctuation sequence as described in any of the preceding embodiments.

[0030] Compared with the prior art, the present invention has at least the following beneficial effects:

[0031] In this embodiment of the invention, the fault diagnosis method based on continuous homophonic fluctuation sequence provides a way to effectively capture subtle signal patterns that are difficult to characterize using traditional methods by quantifying the differences in the topology of adjacent windows using Wassenberg distance.

[0032] In this embodiment of the invention, the fault diagnosis method based on continuous homophonic fluctuation sequence provided has strong noise resistance: under low SNR (e.g., SNR=4) conditions, the overall F1-score of CWRU-4 and CWRU-10 is improved by 3%-4% respectively compared with the baseline model.

[0033] In this embodiment of the invention, the fault diagnosis method based on continuous homophonic fluctuation sequence provided has high data efficiency: with only 2.5% of the training data, the accuracy can be improved by at least 20% compared with traditional methods, significantly reducing the dependence on large-scale labeled data;

[0034] In this embodiment of the invention, the topological features of the fault diagnosis method based on continuous homophonic wave sequence provided are effective: after incorporating topological features, the fault identification accuracy is improved by at least 14%, which verifies the ability of TDA to characterize complex signal structures.

[0035] In this embodiment of the invention, the fault diagnosis method based on continuous homogeneous fluctuation sequence has excellent generalization ability: it is stable under different fault types (inner circle, sphere, outer circle faults) and data granularity (coarse granularity, fine granularity), and is suitable for diverse industrial scenarios.

[0036] Additional aspects and advantages of the invention will be set forth in part in the description which follows, and in part will be obvious from the description, or may be learned by practice of the invention. Attached Figure Description

[0037] Figure 1 This is a flowchart of a fault diagnosis method based on a continuous homophonic wave sequence disclosed in an embodiment of the present invention;

[0038] Figure 2 These are two examples of continuous homophonic wave sequence extraction disclosed in one embodiment of the present invention. Detailed Implementation

[0039] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of the present invention. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0040] The embodiments of the present invention will be described in detail below with reference to the accompanying drawings and specific examples and application scenarios.

[0041] With continuous technological advancements, some deep learning algorithms have been used to improve traditional feature extraction methods. However, the features extracted by these methods lack clear mathematical meaning and consume considerable computation time. The main contributions of this invention are as follows:

[0042] 1) A novel feature extraction method for equipment fault diagnosis based on topology data analysis is proposed. By converting the original equipment operation signal time series into a new time series, the TDA method can be used to effectively reveal subtle patterns in the original signal.

[0043] 2) The proposed feature extraction method, combined with machine learning, provides a new framework for equipment fault diagnosis based on time series data.

[0044] 3) The method was tested on two public datasets with different signal-to-noise ratios and training datasets of different sizes. Experimental results show that it exhibits superior performance even with low signal-to-noise ratios or small training datasets.

[0045] Please see Figure 1 As shown, in some embodiments of the present invention, a fault diagnosis method based on a continuous homogeneous fluctuation sequence is provided for equipment fault diagnosis. This method is implemented based on topological data analysis. The method includes:

[0046] (1) Temporal characteristics of time series

[0047] This section mainly includes the following features:

[0048] Absolute energy: the sum of squares of all components in a time series.

[0049] Average value: the mean of the time series.

[0050] Standard deviation:

[0051] Skewness: the third standardized moment of a time series. It is a measure of the direction and degree of skewness in the distribution of statistical data.

[0052] Kurtosis: the fourth-order normalized moment of a time series. It is a characteristic quantity that describes the peak value of the probability density distribution curve at the mean.

[0053] Absolute average change: the mean of the absolute values ​​of continuous changes in a time series.

[0054] Mean second difference center: the mean of the second-order variation of the time series.

[0055] Median: The median of each value in a time series arranged from smallest to largest.

[0056] (2) Frequency domain characteristics of time series

[0057] This section primarily utilizes the Fourier transform to extract the frequency domain features of the time series. Specifically, given a discrete signal x[n] of length N, its Fourier transform is defined as:

[0058]

[0059] The Fourier transform can be used to transform the original signal into a new signal of equal length to describe its frequency domain information.

[0060] (3) Time series topological features

[0061] To analyze the topological properties of a time series, it is necessary to first convert the time series into a point cloud representation. The method used in this invention is phase space reconstruction, and its theoretical basis comes from the following theorem:

[0062] (Takens embedding theorem

[28] ) if For an n-dimensional compact manifold, φ: R×M→M and f: M→R are almost everywhere smooth mappings. Given a delay parameter τ, for a point x in M, define x as... i =φ(i·τ,x), then g: M→R 2n+1 ,x→(f(x),f(x1),...,f(x 2n )) is an embedding mapping between manifolds.

[0063] Specifically, in the scenario studied in this invention, if X = {x0, ..., x...} n-1 If} is a one-dimensional time series (where d∈N), then given a delay parameter τ, the Takens phase space reconstruction point at time t is defined as:

[0064] Z t :=(x t x t+τ , ..., x t+(d-1)τ )∈R d ;

[0065] Given parameters (d, T), phase space reconstruction transforms the time series into point cloud data.

[0066] (4) Wassenstan distance

[0067] Given continuous graphs D1 and D2, their Wassenstein distance is defined as:

[0068]

[0069] (5) Continuous Homogeneous Oscillation Sequence

[0070] To analyze time series, this invention proposes the concept of persistent homology fluctuation sequences to transform time series. Specifically, given a time series and a homology dimension, a sliding window is applied. For the subsequence corresponding to each window, it is transformed into point cloud data through Takens embedding and its persistent homology is calculated. The difference in persistent homology between two consecutive windows is then calculated using the Wassenberg distance.

[0071] In some embodiments of the present invention, the implementation steps of the method include:

[0072] a. Input the original fault time series s;

[0073] b. Following the time-domain feature extraction method and the frequency-domain feature extraction method of Fast Fourier Transform described above, extract the time-domain and frequency-domain features of the time series s, the 0-dimensional sequence s0, and the 1-dimensional sequence s1. Specifically, s0 and s1 are calculated as follows: First, the original sequence is windowed. Given the homology dimension (0 and 1), for a given window, calculate the Wassenberg distance between its persistent homology graph after Takens embedding and the persistent homology graph of the previous window after Takens embedding. Connecting all windows with their corresponding previous windows yields the persistent homology fluctuation sequences s0 and s1. Two examples are shown below. Figure 2 As shown (from left to right, the original signal, s0, s1);

[0074] c. Consolidate all extracted features; the concatenation method can be to concatenate different features column by column.

[0075] d. Input the feature space into the support vector machine model for training to obtain the classification results of the fault.

[0076] The research material for this invention uses the publicly available CWRU dataset for experiments. The CWRU dataset is a publicly available dataset of gear bearing vibration signals, widely used for fault diagnosis. This dataset contains 4 classes of fault data at a coarse-grained level and 10 classes at a fine-grained level, as follows:

[0077] Coarse-grained (CWRU-4): No fault, inner ring fault, ball fault, and outer ring fault;

[0078] Fine-grained (CWRU-1O): No fault, 7mil inner ring fault, 14mil inner ring fault, 21mil inner ring fault, 7mil sphere fault, 14mil sphere fault, 21mil sphere fault, 7mil outer ring fault, 14mil outer ring fault, and 21mil outer ring fault.

[0079] This invention specifically tested the performance under different signal-to-noise conditions and different training data volumes, and selected domain adaptive convolutional neural network, multi-scale inner product convolutional neural network, and multi-scale residual neural network as baseline models. The results are shown in Table 1 and Table 2, where Table 1 shows the results of CWRU-4 and Table 2 shows the results of CWRU-10.

[0080] Table 1

[0081]

[0082] Table 2

[0083]

[0084]

[0085] In the tables of this invention, IR007 represents a 7-mil inner ring defect, IR014 represents a 14-mil inner ring defect, IR021 represents a 21-mil inner ring defect, B007 represents a 7-mil rolling element defect, B014 represents a 14-mil rolling element defect, B021 represents a 21-mil rolling element defect, OR007 represents a 7-mil outer ring defect, OR014 represents a 14-mil outer ring defect, and OR021 represents a 21-mil outer ring defect.

[0086] The two tables above show the results obtained by training with 80% of the data under different signal-to-noise ratio conditions. It can be seen that the model of the present invention has improved the accuracy by 3%-4% compared with other methods.

[0087] Tables 3 and 4 show the improvement effect of topology features on fault diagnosis. Table 3 presents the results for CWRU-4, and Table 4 presents the results for CWRU-10. It is clear that incorporating topology features improves the accuracy of fault identification by at least 14%.

[0088] Table 3

[0089]

[0090]

[0091] Table 4

[0092]

[0093]

[0094] Tables 5 and 6 show the results of training with only 2.5% of the data, with Table 5 showing the results for CWRU-4 and Table 6 showing the results for CWRU-10. The results demonstrate that the method of this invention achieves at least 20% higher accuracy than other methods.

[0095] Table 5

[0096]

[0097] Table 6

[0098]

[0099]

[0100] All parts of this invention not described in detail herein can be referred to in the prior art or are known to those skilled in the art. This embodiment does not limit these aspects and will not describe them in detail here.

[0101] The embodiments of the present invention have been described above with reference to the accompanying drawings. However, the present invention is not limited to the specific embodiments described above. The specific embodiments described above are merely illustrative and not restrictive. Those skilled in the art can make many other forms under the guidance of the present invention without departing from the spirit and scope of the claims, and all of these forms are within the protection scope of the present invention.

Claims

1. A fault diagnosis method based on a persistent homophonic fluctuation sequence, characterized in that, The method combines feature extraction and topological data analysis to construct a fault diagnosis framework suitable for small data volumes and low signal-to-noise ratios. By capturing subtle patterns in signals through topological features, the model's ability to represent complex data is improved. The method includes the following: Multi-dimensional feature extraction is performed, including time-domain features, frequency-domain features, and topological features; the topological features are the time-domain and frequency-domain features of the 0-dimensional sequence s0 and the 1-dimensional sequence s1. The extracted multi-dimensional features are fused; The fused feature data is then input into a machine learning classification model for model training or classification. The topological features include phase space reconstruction and persistent homophonic wave sequences; The phase space reconstruction includes: based on Takens embedding theorem, converting a one-dimensional time series into high-dimensional point cloud data while preserving the signal dynamics characteristics; The extraction method of the persistent homology wave sequence is as follows: the original sequence is windowed, and given the homology dimension 0 and 1, for a given window, the Wassenberg distance between the persistent homology graph after Takens embedding and the persistent homology graph after Takens embedding of the previous window is calculated; all windows are connected to their corresponding previous windows to obtain the persistent homology wave sequences s0 and s1. The specific process of fusing the extracted multi-dimensional features includes: concatenating time-domain features, frequency-domain features, and topological features, and using the concatenated features as the fused feature data.

2. The fault diagnosis method based on a persistent homophonic fluctuation sequence according to claim 1, characterized in that, The time-domain features include absolute energy, mean, standard deviation, skewness, kurtosis, absolute mean variation, mean second difference center, and / or median.

3. The fault diagnosis method based on a persistent homophonic fluctuation sequence according to claim 1, characterized in that, The frequency domain features are extracted by converting the original signal to the frequency domain using a fast Fourier transform and capturing the frequency component distribution as the extracted frequency domain features.

4. A computer program product, comprising a computer program, characterized in that, When executed by a processor, the computer program implements the fault diagnosis method based on a continuous homophonic fluctuation sequence as described in any one of claims 1-3.

5. A computer device, comprising: A memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that the processor executes the computer program to implement the fault diagnosis method based on a continuous coherent fluctuation sequence as described in any one of claims 1-3.

Images