A vacuum dry pump multi-fault prediction method and system based on improved C4.5

Abstract

The application discloses a kind of vacuum dry pump multiple fault prediction method and system based on improved C4.5, method includes: using improved C4.5 algorithm to construct decision tree model and using the full life vibration signal data of vacuum dry pump to train, obtain trained multiple fault prediction model;Obtain the vibration signal data of real-time operation monitoring of vacuum dry pump, and the data is preprocessed;The vibration signal data after pre-processing is input into trained multiple fault prediction model, to predict whether fault appears in vibration signal data, and output fault prediction result, the prediction result includes fault type and fault occurrence time.The application solves the deficiency of prior art through transparent decision process, real-time updating, fine feature extraction and powerful multiple fault classification capability, can effectively predict multiple overlapping faults of vacuum dry pump, provides timely decision support for equipment maintenance.

Description

Technical Field

[0001] This invention relates to equipment health monitoring, and more particularly to a method and system for predicting multiple faults in a vacuum dry pump based on an improved C4.5. Background Technology

[0002] During the operation of a vacuum dry pump, dust accumulates at various locations, including the inlet and the upper and lower pump sections. As operating time increases, the residual dust gradually increases. When it reaches a certain threshold, the squeezing action of the dust can cause friction between internal mechanical components. This friction not only accelerates equipment wear but can also lead to serious malfunctions, causing the vacuum dry pump to shut down. Furthermore, numerous external and internal disturbances, including dust jamming, can cause faults at different locations on the pump, resulting in multiple fault characteristics in vibration signal data, making fault diagnosis exceptionally complex. The cumulative effect of multiple faults further complicates fault identification and prediction, posing challenges to the normal operation and maintenance of the equipment.

[0003] Currently, vacuum dry pump fault prediction methods are mainly divided into physical models and data-driven methods. Physical models predict equipment status by establishing complex mathematical relationships. Although they have high accuracy under specific conditions, they usually rely on a large amount of expert knowledge and are difficult to adapt to complex operating environments. Data-driven methods, such as deep learning, clustering, Transformer, and XGBoost, have improved fault detection capabilities to some extent, but they still face problems such as insufficient real-time performance, low accuracy, high false alarms and false negatives, weak fault classification capabilities, and excessive reliance on historical data. In particular, the identification capabilities of existing methods are still insufficient when dealing with specific fault modes. Summary of the Invention

[0004] The purpose of this invention is to solve the problems in the prior art.

[0005] The technical solution adopted by this invention to solve its technical problem is: to provide a multi-fault prediction method for vacuum dry pumps based on improved C4.5, comprising:

[0006] A decision tree model was constructed using the improved C4.5 algorithm and trained using the full-life vibration signal data of the vacuum dry pump to obtain a well-trained multi-fault prediction model.

[0007] Acquire vibration signal data from real-time operation monitoring of the vacuum dry pump, and preprocess the data;

[0008] The preprocessed vibration signal data is input into the trained multi-fault prediction model to predict whether a fault occurs in the vibration signal data, and the fault prediction result is output, which includes the fault type and the fault occurrence time.

[0009] Preferably, the step of constructing a decision tree model using the improved C4.5 algorithm and training it with full-life vibration signal data of the vacuum dry pump includes the following steps:

[0010] A decision tree model was constructed using the improved C4.5 algorithm;

[0011] The vibration signal data of the vacuum dry pump throughout its entire lifespan is acquired and preprocessed. The preprocessed data is then labeled to obtain a dataset including sample data and corresponding labels.

[0012] Feature extraction is performed on the dataset to obtain a feature set; the features of the feature set include time-domain features and frequency-domain features.

[0013] The decision tree model is trained using a feature set.

[0014] Preferably, the step of constructing a decision tree model using the improved C4.5 algorithm includes the following steps:

[0015] An improved C4.5 algorithm is obtained by introducing an attribute weight mechanism into the original C4.5 algorithm.

[0016] The improved C4.5 algorithm selects split points using information gain ratio to construct a decision tree model, as follows:

[0017] M = C4.5(F train Y train );

[0018] Among them, F train Y represents the training features. trai Represents the training features F train The label; the information gain ratio of the improved C4.5 algorithm is expressed as:

[0019] G ′ (A) = FII(A,t)·G(A);

[0020] FII(A,t)=Var(A,t)×Exp(A)×λ(t);

[0021]

[0022] Where FII(A,t) is the weight of attribute A, G(A) is the information gain ratio of attribute A calculated by the original C4.5 algorithm, and G... ′(A) is the information gain ratio of attribute A calculated by the improved C4.5 algorithm; Var(A,t) represents the variance of attribute A within the time window t, reflecting its degree of change during that time period; Exp(A) represents the empirical importance index of attribute A as assessed by experts, based on expert knowledge; λ(t) represents the time adjustment coefficient, used to dynamically adjust feature weights according to time changes; tk is the starting point of the time window, and Var(A,t-1) is the variance of the previous time window. This adjustment coefficient reflects the change in the importance of the attribute within different time windows.

[0023] Preferably, the acquisition of vibration signal data from real-time operation monitoring of the vacuum dry pump specifically involves: installing multiple sensors at different locations on each vacuum dry pump, with each sensor independently acquiring vibration signals at that location; each sensor represents a channel, thereby acquiring vibration data from multiple channels to form a multi-dimensional time-series dataset.

[0024] Preferably, the preprocessing includes: data dimensionality reduction, handling missing and outlier values, and normalizing the data to a uniform range.

[0025] Preferably, the labeling of the preprocessed data specifically involves: performing a sliding window on the preprocessed life-cycle vibration signal data, first performing a progressively increasing real-time normalization on the window and the data before it, then extracting the data from the window to construct a multi-class dataset containing sample data from multiple windows and corresponding labels.

[0026] Preferably, the features of the feature set further include:

[0027] The zero-crossing rate is expressed as:

[0028]

[0029] Where, N ZCR N is the number of points that cross the zero point between adjacent points. win It represents the total number of data points within the window;

[0030] Stability is expressed as:

[0031]

[0032] Where, N win x is the total number of data points within the window. i+1 x is the (i+1)th data point in the signal. i It is the i-th data point in the signal;

[0033] The percentage of time with unstable data is represented as follows:

[0034]

[0035] Where, N unstable N is the number of adjacent points whose difference exceeds a set threshold. win It represents the total number of data points within the window;

[0036] The proportion of high-frequency components is expressed as:

[0037]

[0038] Where f represents frequency, f c Represents the high-frequency threshold, |X(f)| 2 |X(D)| represents the energy of the high-frequency components. 2 This represents the total energy of the signal;

[0039] Hausdorff distance, expressed as:

[0040]

[0041] Among them, X a X represents a sample of a certain type of fault. b This indicates a sample selected from other types of fault or normal data.

[0042] Preferably, inputting the preprocessed vibration signal data into the trained multi-fault prediction model includes the following steps:

[0043] For the preprocessed vibration signal data, by setting the window size W and the sliding step size S, the data in the window and before it are first normalized by gradual growth, and then the sliding window is extracted to form window data.

[0044] Feature extraction is performed on the data from each window to obtain a feature matrix;

[0045] The characteristic matrix F real The input is fed into a pre-trained multi-fault prediction model for classification, and the model outputs the class label C. pred Represented as:

[0046] C pred =C4.5(F real );

[0047] If the model identifies a fault, it will output a fault report, represented as:

[0048] Fault report = {C pred T pred};

[0049] Among them, C pred Indicates the fault type, T pred Indicates the time of occurrence.

[0050] Preferably, the feature matrix includes important features selected from the feature set, and the selection process includes the following steps:

[0051] Standardize the data in the feature set;

[0052] Calculate the covariance matrix C based on the standardized data:

[0053] Perform eigenvalue decomposition on the covariance matrix C to obtain the eigenvalues ​​λ. i and the corresponding feature vector v i , represented as:

[0054] Cv i =λ i v i ;

[0055] Where, λ i For v i The corresponding eigenvalue, v i Let i be the i-th eigenvector of the feature set;

[0056] The eigenvalue λ i Sort by value from largest to smallest, and select the features corresponding to the top few feature values ​​as important features.

[0057] This invention also provides a multi-fault prediction system for vacuum dry pumps based on an improved C4.5, characterized in that it includes:

[0058] The training module uses the improved C4.5 algorithm to build a decision tree model and uses the full-life vibration signal data of the vacuum dry pump for training to obtain a trained multi-fault prediction model.

[0059] The acquisition module acquires vibration signal data from the real-time operation monitoring of the vacuum dry pump and performs data preprocessing on the data.

[0060] The identification module inputs the preprocessed vibration signal data into the trained multi-fault prediction model to predict whether a fault occurs in the vibration signal data and outputs the fault prediction result, which includes the fault type and the fault occurrence time.

[0061] This invention offers the following advantages: By acquiring real-time vibration signal data during the operation of a vacuum dry pump and performing data preprocessing, it can accurately predict potential faults and output specific fault types and occurrence times, thus improving the accuracy of fault detection. Furthermore, the method employs sliding window technology and progressively increasing real-time normalization, enabling the model to quickly adapt to new operating conditions and fault modes, achieving real-time updates and significantly improving the real-time performance and accuracy of predictions. During the feature engineering stage, not only time-domain and frequency-domain features are calculated, but also zero-crossing rate, stationarity, proportion of non-stationary data time, proportion of high-frequency components, Hausdorff distance, etc., effectively distinguishing different types of faults and enhancing the model's ability to identify multiple fault modes. The improved C4.5 algorithm is suitable for multi-classification tasks, possesses good interpretability, and by combining the information gain rate of attribute weights to select the optimal partitioning attribute, it can accurately classify multiple fault types. This method can effectively identify multiple fault types caused by dust jamming and other reasons, providing accurate fault classification and helping maintenance personnel take timely measures to reduce equipment downtime and maintenance costs.

[0062] The present invention will be further described in detail below with reference to the accompanying drawings and embodiments, but the present invention is not limited to the embodiments. Attached Figure Description

[0063] Figure 1 This is a diagram illustrating the method steps of an embodiment of the present invention;

[0064] Figure 2 This is a diagram illustrating the model training steps in an embodiment of the present invention.

[0065] Figure 3 This is a flowchart illustrating the model training process according to an embodiment of the present invention.

[0066] Figure 4 This is a flowchart of the model testing process according to an embodiment of the present invention;

[0067] Figure 5 This is a system structure diagram of an embodiment of the present invention. Detailed Implementation

[0068] See Figure 1 The diagram shows the method steps of an embodiment of the present invention, including the following steps:

[0069] S101, a decision tree model was constructed using the improved C4.5 algorithm and trained using the full-life vibration signal data of the vacuum dry pump to obtain a well-trained multi-fault prediction model.

[0070] S102, acquire vibration signal data from real-time operation monitoring of the vacuum dry pump, and preprocess the data;

[0071] S103, input the preprocessed vibration signal data into the trained multi-fault prediction model to predict whether a fault occurs in the vibration signal data and output the fault prediction result, which includes the fault type and the fault occurrence time.

[0072] For details, see Figure 2 and Figure 3 The diagram shown is a flowchart and diagram of the model training steps according to an embodiment of the present invention. Step S101 includes the following steps:

[0073] A1. Acquire vibration signal data of the vacuum dry pump throughout its entire lifespan, and perform data preprocessing, including data dimensionality reduction, handling missing and outlier values, and normalizing the data to a uniform range. Specifically, this includes steps A11 to A14.

[0074] A11 acquired time-series vibration signal data of a vacuum dry pump throughout its entire lifespan. A portion of the lifespan data was used to extract windowed sample data to construct a multi-classification dataset for model training and testing, while the other portion was used to simulate real-time fault prediction, with a 7:3 ratio. Multiple sensors were installed on each pump, distributed in different locations to ensure comprehensive monitoring of the equipment's operating status. Each sensor represents a channel, independently acquiring vibration signals at that location. This allows each pump to acquire vibration data from multiple channels, forming a multi-dimensional time-series dataset. The acquired lifespan data covers the entire process of the vacuum dry pump from start-up to shutdown. This data records the equipment's operation at different stages, including startup, normal operation, fault occurrence, and final shutdown. Different types of fault data may exhibit specific vibration patterns and may appear in different forms within the same channel. Some faults may occur individually, while others may overlap, resulting in complex signal characteristics.

[0075] A12, In the data dimensionality reduction stage, a sliding window method is used to process continuous vibration data points. The window size is k and the step size is also k. The input vibration signal data is D = [d0, d1, ..., d...]. n-1 ], where n represents the total number of data points in the original dataset, d i This represents the i-th data point.

[0076] First, the number of data points after dimensionality reduction can be determined using the following formula:

[0077]

[0078] Where m represents the total number of data points after dimensionality reduction. This indicates rounding up to the nearest integer.

[0079] By taking the average of every k data points, the reduced-dimensional data R is calculated as follows:

[0080]

[0081] where R[j] represents the j-th data point after dimensionality reduction.

[0082] If the number of data points in the last window is less than k, assuming the starting index of this window is p, then calculate the average of all data points in the range from p to n - 1:

[0083] where p = (m - 1)k and p < n

[0084] A13. When dealing with outliers and missing values, the values in the dataset that deviate significantly from other data points are regarded as outliers, usually manifested as particularly large or particularly small numerical values. The data point R[i] is defined as an outlier if it satisfies the following condition:

[0085] R[i] > μ + 3σ or R[i] < μ - 3σ

[0086] where μ represents the mean of the dataset. σ represents the standard deviation of the dataset.

[0087] A common method for dealing with outliers and missing values is mean imputation. The formula for mean imputation is as follows:

[0088] (where R j is a non-missing value)

[0089] where, R i is the result after substituting the missing value, and N is the number of non-missing values.

[0090] A14. Finally, in the data normalization stage, the min-max normalization method is adopted, and its calculation formula is as follows:

[0091]

[0092] where r represents the normalized value, R is the original data, R min and R max represent the minimum and maximum values in the dataset respectively.

[0093] After dimensionality reduction, the specification size of the output data R is m, that is, the total number of data points after dimensionality reduction. Usually, m < n. The reduced-dimensional data can be represented as a vector:

[0094] R = [r0, r1,..., r m-1

[0095] A2 involves setting the sliding window size and step size to perform a sliding window operation on the full-lifetime vibration signal data. First, a progressively increasing normalization is applied to the data within the window and before it. Then, the data within the window is extracted, resulting in multiple window sample data. Expert annotation is used to label the samples, and each sample is classified into fault categories, constructing a multi-class dataset containing multiple window sample data and corresponding labels. Finally, the dataset is randomly divided into training and testing sets. This process includes steps A21 to A25.

[0096] A21, in the processed vibration signal data R=[r0, r1, ..., r m-1 In this context, window parameters are set to extract samples. The window size W is the number of consecutive data points contained in each sample, and the sliding step S is the number of points the window slides forward each time a new sample is extracted. The number of windows M that can be extracted from a vibration signal data set is calculated using the following formula:

[0097]

[0098] A22 employs a progressively increasing normalization method to better meet real-time requirements. Before extracting window data, for the current window position j, the window data and all previous data, i.e., R′=[r0, r1, ..., r...], are processed. j·s+w-1 Normalization is performed using the following formula:

[0099]

[0100] A23, for data R′, extract samples using the set window parameters W and S. Specifically, at each position j in data R, extract a window X. j As a new sample:

[0101] X j =[R ′ [j·S],R ′ [j·S+1],…,R ′ [j·S+W-1], (j=0, 1,..., M-1)

[0102] A24, Sample X per window j The labels will be manually assigned based on expert experience, generating the corresponding label Y. j Experts evaluated each window sample by analyzing historical fault records and equipment operating status, determining its corresponding fault type or normal state, and recording it as label Y. j .

[0103] For vibration signal data R of different vacuum dry pumps, each window sample X j and its corresponding label Y jStored in a data structure, forming sample pairs (X) j Y j By using window sample X j and corresponding label Y j Combine and construct a dataset V containing sample data and labels:

[0104] V = {(X j Y j )∣j=0,1,2,…,L-1}

[0105] Where L is the number of samples in the dataset.

[0106] A25, randomly divide the dataset V into training set V. train and test set V test The division ratio is set to p = 0.7, and the specific division method is as follows:

[0107] p·L samples are randomly selected from the dataset V as the training set V. train The remaining samples are used as the test set V. test .

[0108] V train =(X train Y train )

[0109] V test =(X test Y test )

[0110] In feature engineering (A3), multiple features are extracted and a feature set is constructed for each sample data. Besides calculating some commonly used time-domain and frequency-domain features, new features are also calculated, including zero-crossing rate, stationarity, proportion of non-stationary data time, and proportion of high-frequency components. Simultaneously, Hausdorff distance is calculated to assess the differences between a certain type of fault and other types of fault data, as well as normal data. This specifically includes steps A31 to A34.

[0111] A31, In the time-domain feature extraction stage, the input data is X, X i Let X be the i-th data point of the signal, and N be the total number of data points of signal X.

[0112] The maximum value is the highest level of the signal, reflecting the signal strength, and is expressed by the formula: X max =max(X)

[0113] The minimum value signal represents the lowest level of the signal, indicating the lower limit of the signal. The formula is: X min =min(X)

[0114] Peak-to-peak value represents the distance of a signal from its maximum to its minimum value, and the formula is: Xpp =X max -X min

[0115] The mean represents the average level of a signal, and the formula is:

[0116] The variance quantization formula is used to determine the degree of fluctuation in a signal:

[0117] The formula for the square root of the standard deviation and variance is:

[0118] The mean square value reflects the energy of the signal, and the formula is:

[0119] The root mean square (RMS) value is used to assess signal strength, and the formula is:

[0120] Kurtosis reflects the influence of outliers, and the formula is:

[0121]

[0122] Skewness represents the symmetry of the signal distribution, and the formula is:

[0123] The waveform factor reflects the sharpness of a signal, and the formula is:

[0124] The ratio of the maximum value to the mean value of the peak factor signal is given by the following formula:

[0125] The ratio of the maximum value to the mean square value of the impulse factor signal is given by the following formula:

[0126] The ratio of the maximum to the minimum value of the margin factor signal is given by the following formula:

[0127] A32, in the frequency domain feature extraction stage, Fourier transform is used to analyze the frequency components of the signal. A Fourier transform is performed on the signal X(t) to obtain the frequency domain representation X(f), where f represents the frequency. high and f low Indicates the upper and lower limits of the frequency range.

[0128] The energy distribution of a signal is analyzed by power spectral density, expressed as: P(f)=|X(f)| 2

[0129] The centroid frequency describes the frequencies at which signal energy is concentrated; the formula is:

[0130] Mean square frequency represents the frequency distribution characteristics of a signal, and the formula is:

[0131] The root mean square frequency is used to measure the intensity of a frequency, and the formula is:

[0132] Frequency variance quantifies the degree of dispersion of the frequency distribution, and the formula is:

[0133] The standard deviation of frequency represents the degree of dispersion of the frequency distribution, and the formula is:

[0134] Bandwidth is the frequency range occupied by a signal in the frequency domain, and the formula is: BW = f high -f low

[0135] The dominant frequency is usually the frequency with the highest energy in the frequency spectrum, and the formula is:

[0136] Spectral kurtosis describes the sharpness of a spectrum, and its formula is:

[0137] Spectral skewness represents the symmetry of the spectral distribution, and the formula is:

[0138] Total energy represents the total energy of the signal, and the formula is: E total =∑ t |X(t)| 2

[0139] Spectral entropy is used to measure the uncertainty of the spectrum, and the formula is:

[0140] Where, E = ∑ f P(f).

[0141] Relative energy represents the proportion of energy in a certain frequency band to the total energy, and the formula is:

[0142] In extracting other features, A33 utilizes some uncommon metrics designed to analyze the characteristics of vibration signals more deeply and better distinguish between different types of faults, especially when different faults overlap. Specifically, the zero-crossing rate measures the number of times the signal crosses zero, differentiating between faults with rapidly changing characteristics and relatively stable normal data; a high zero-crossing rate usually indicates drastic signal fluctuations. Stationarity reflects signal stability, helping to identify faults with drastic changes, focusing on measuring the overall trend of signal change rather than instantaneous fluctuations. The proportion of non-stationary data time calculates the duration of the fault occurrence, distinguishing between trend changes and abrupt changes caused by the fault, helping to identify gradual and sudden faults. The proportion of high-frequency components distinguishes between faults caused by high-frequency vibrations and faults dominated by low-frequency vibrations. Hausdorff distance calculates the similarity of the curve trends between samples; a large distance indicates significant sample differences and different fault types, while a small distance indicates sample similarity, suggesting they are similar in characteristic change trends and have the same fault type.

[0143] The zero-crossing rate is the ratio of data points in a data window X that cross a zero. When the product of adjacent data point values ​​is less than zero, it indicates that the signal has crossed a zero at that point. The zero-crossing rate can capture frequently changing features in a signal, and is particularly suitable for detecting high-frequency noise or abrupt changes in a signal. Its calculation formula is:

[0144]

[0145] Where, N ZCR N is the number of points that cross the zero point between adjacent points. win This represents the total number of data points within the window. The condition for adjacent points passing through zero is:

[0146] x i ·x i+1 <0

[0147] Stationarity is assessed by calculating the average difference between all adjacent points in a window of data X. Stationarity reflects the temporal consistency of a signal and can identify unstable components within it. Its calculation formula is:

[0148]

[0149] The proportion of non-stationary data time measures the ratio of the duration of non-stationary segments in a window of data X to the total time. This is achieved by calculating the number of adjacent data points whose differences exceed a certain threshold. The formula is as follows:

[0150]

[0151] Where, N unstable It is the number of adjacent points whose differences exceed a set threshold.

[0152] The proportion of high-frequency components refers to frequencies greater than a certain threshold f. c The proportion of signal energy to total frequency energy is calculated using the following formula:

[0153]

[0154] Where f represents frequency, f c Represents the high-frequency threshold, |X(f)| 2 |X(D)| represents the energy of the high-frequency components. 2 This represents the total energy of the signal.

[0155] The Hausdorff distance is used to calculate the difference between a certain type of fault data and other types of fault data, as well as normal data. The formula is:

[0156]

[0157] Among them, X a X represents a sample of a certain type of fault. b This means selecting one sample from other types of fault or normal data.

[0158] A34, Finally, the features extracted from the j-th sample will be combined into a feature vector F. j The specific form of a vector can be represented as:

[0159] F j ={X max X min X pp X mean X var X std X ms X rms ,K,S,WF,PF,IF,MF,f c f rms Var(f), f std ZCR, K, P unstable HF Ratio, d H}

[0160] Using the above comprehensive feature extraction methods, the extracted feature vector F j With the corresponding label Y j Form a complete dataset V s , can be represented as:

[0161] V s ={(F j Y j )∣j=0,1,2,…,L-1}

[0162] The training set and the test set are as follows:

[0163]

[0164] A4 uses an improved C4.5 algorithm to construct a decision tree model. The model is trained by inputting the feature set of training set samples and their corresponding labels, and a decision tree is built on the training set. Subsequently, test set samples are input into the model for classification prediction, and model performance is evaluated using metrics such as confusion matrix, accuracy, recall, and F1 score. See also... Figure 4 As shown, the specific steps include A41 to A44.

[0165] The A41 and C4.5 algorithms primarily select split points and construct decision trees using information gain ratio. The C4.5 algorithm, a classic decision tree construction method, generates a decision tree by selecting the feature that best divides the dataset for splitting. An improved C4.5 algorithm is obtained by introducing a dynamically adjusted Feature Importance Index (FII) mechanism during the calculation of information gain and information gain ratio for each attribute. In machine learning and decision tree algorithms, attributes are used to describe and distinguish data samples, typically representing a specific physical quantity or category. Attributes can refer to equipment status indicators after feature extraction. The FII method dynamically updates the importance weight of attributes based on their performance under different failure modes. In multidimensional feature time-series data, for each attribute A, its dynamic Feature Importance Index (FII) is calculated over different time periods. This index comprehensively considers three factors: variance, expert experience, and time adjustment coefficient. Variance reflects the fluctuation of features within different time windows; features with larger fluctuations are usually more revealing of failures. Expert experience assigns higher weights to certain features based on experts' understanding of equipment failure modes. The importance of dynamically adjusting the time adjustment factor as equipment operating time changes, in order to address the impact of equipment aging or wear. The formula is as follows:

[0166] FII(A,t)=Var(A,t)×Exp(A)×λ(t)

[0167] Where FII(A,t) represents the feature importance index of attribute A within the time window t; Var(A,t) represents the variance of attribute A within the time window t, reflecting its degree of change during that time period; Exp(A) represents the expert-assessed empirical importance index of attribute A, set based on expert knowledge; and λ(t) represents the time adjustment coefficient, used to dynamically adjust feature weights according to time changes. The attribute weights are dynamically adjusted by calculating the ratio of the variance of the current time window to the variance of the previous time window, as shown in the following formula:

[0168]

[0169] Where tk is the starting point of the time window, and Var(A,t-1) is the variance of the previous time window. This adjustment coefficient reflects the change in the importance of the attribute within different time windows.

[0170] Taking attributes A and B as examples, assume that the variance of attribute A in the current time window is 0.064, while the variance of the previous time window was 0.05; the variance of attribute B in the current time window is 0.02, while the variance of the previous time window was 0.018. Furthermore, based on expert experience, we assign empirical weights of 0.7 and 0.5 to attributes A and B, respectively.

[0171] To dynamically adjust the attribute weights, the variance ratio between the current time window and the previous time window is first calculated. For attribute A, the variance ratio is 1.28, i.e. For attribute B, the variance ratio is 1.11, that is... This ratio reflects how the attribute changes over different time periods and is used to adjust the weights.

[0172] By introducing the variance ratio, the formula for calculating attribute weights becomes:

[0173] W(A,t)=Var(A,t)×Exp(A)×λ(t)

[0174] W(B,t)=Var(B,t)×Exp(B)×λ(t)

[0175] After substituting the known data, the final weights of attributes A and B are obtained, and the weights are:

[0176] FII(A,t)=0.064×0.7×1.28=0.057344

[0177] FII(B,t)=0.02×0.5×1.11=0.0111

[0178] The improved C4.5 algorithm selects split points using information gain ratio to construct a decision tree model, as follows:

[0179] M = C4.5(F train Y train );

[0180] F train Y represents the training features. train Represents the training features F train The label; the information gain ratio of the improved C4.5 algorithm is expressed as:

[0181] G ′ (A) = FII(A,t)·G(A);

[0182] Where FII(A,t) is the weight of attribute A, G(A) is the information gain ratio of attribute A calculated by the original C4.5 algorithm, and G... ′ (A) is the information gain ratio of attribute A calculated by the improved C4.5 algorithm.

[0183] This method significantly differs from traditional C4.5 algorithms and decision tree methods such as random forests. Its core innovation lies in the introduction of a dynamic feature importance index (FII), which greatly enhances the model's adaptability and real-time performance. Traditional C4.5 and random forest algorithms rely on static information gain ratios during feature selection, making it difficult to maintain efficiency and accuracy when dealing with dynamic changes in equipment operating states and failure modes. In contrast, this invention, by introducing the FII mechanism, combines the variance of each feature within different time windows, expert experience, and time adjustment coefficients to dynamically adjust the weights of features. This allows the model to flexibly adjust the importance of each feature based on the long-term trends and instantaneous fluctuations of the equipment.

[0184] FII (Fixed Indicator) not only considers the performance of features across different time periods but also integrates changes in the actual working environment and operating status of the equipment, enabling the model to respond to fluctuations under various operating conditions in real time. This dynamic adjustment mechanism, particularly through the introduction of variance ratio optimization in the feature selection process, significantly improves the accuracy and flexibility of fault prediction compared to traditional static methods. Unlike traditional methods such as the C4.5 algorithm and random forests, this invention can dynamically adjust the importance and selection process of features based on the real-time status of the equipment, thereby more accurately capturing early signs of equipment failure and updating fault prediction results in real time. This method is particularly suitable for complex industrial equipment fault monitoring systems, capable of responding to changes in equipment operation in real time and improving fault diagnosis and prediction capabilities.

[0185] A43, in constructing the decision tree, the improved C4.5 algorithm selects the optimal feature to partition the nodes, thus forming the decision tree. The training data is then partitioned according to the selected feature, forming child nodes, and the above process is repeated for each child node until the stopping condition is met. At this point, model M will be represented by the following formula:

[0186] M = C4.5(F train Y train )

[0187] After the model training is complete, the test set features F can be used. test The input is fed into a trained decision tree model M for classification and prediction, generating a prediction result Y. pred :

[0188] Y pred =M(F test )

[0189] A44. To ensure the effectiveness and reliability of the model, metrics such as confusion matrix, accuracy, recall, and F1 score are used to evaluate the model's performance. The confusion matrix is ​​a summary of the model's prediction results, including true positives (TP), false positives (FP), true negatives (TN), and false negatives (FN).

[0190] Among them, True Positive Class (TP) is the number of samples correctly predicted as positive, False Positive Class (FP) is the number of negative class samples incorrectly predicted as positive, False Negative Class (FN) is the number of positive class samples incorrectly predicted as negative, and True Negative Class (TN) is the number of samples correctly predicted as negative.

[0191] The formulas for calculating the relevant indicators are as follows, and accuracy can be expressed as:

[0192]

[0193] The recall rate is:

[0194]

[0195] The F1 score is the harmonic mean of precision and recall, calculated as follows:

[0196]

[0197] A5 involves constructing a feature subset through feature dimensionality reduction analysis. First, the covariance matrix among the features in the dataset is calculated, and eigenvectors and eigenvalues ​​are extracted to determine the principal components. The top k features are selected based on the magnitude of their eigenvalues ​​to form a new feature subset. This includes steps A51 to A55.

[0198] A51, the entire dataset can be represented as a matrix. Where L is the number of samples and K is the number of features.

[0199] To eliminate the influence of different dimensions, the data is first standardized. The formula for calculating the standardized data matrix Z is:

[0200]

[0201] Among them, F ij μ represents the value of the i-th sample on the j-th feature. j and σ j Let Z be the mean and standard deviation of the j-th feature, respectively. A52, based on the standardized data Z, calculate the covariance matrix C:

[0202]

[0203] The elements C of the covariance matrix C pq It is defined as the covariance between the p-th feature and the q-th feature.

[0204] A53, perform eigenvalue decomposition on the covariance matrix C to obtain the eigenvalues ​​λ. i and the corresponding feature vector v i It satisfies the following relationship:

[0205] Cv i =λ i v i

[0206] Where, λ i For the i-th eigenvalue, v i This is the corresponding feature vector.

[0207] A54, sort the eigenvalues ​​in descending order to obtain the sorted eigenvalue array λ. (1) , λ (2) , …, λ (K) And arrange the eigenvectors accordingly. Select the first k eigenvalues ​​(λ). (1) , λ (2) , …, λ (k) ) and its corresponding feature vector (v (1) v (2) , ..., v (k) ), to form a new feature subset.

[0208] Construct a matrix from the selected eigenvectors. New feature matrix F s It can be calculated using the following formula:

[0209]

[0210] Among them, F s The feature matrix after dimensionality reduction contains the values ​​of L samples in the new k feature dimensions.

[0211] A55 will use a subset of features to train and test the model, evaluate the model's predictive performance, and determine whether these key components are effective.

[0212] A6 performs sliding window and real-time fault prediction on vibration signal data throughout the entire lifespan. The same window size and step size as in A2 are set on the vibration signal data, and progressively increasing normalization and sliding window extraction are performed. Features are calculated for each window's data, input into the C4.5 model for classification, determining whether the data in that window belongs to a faulty or normal data category. If a fault is predicted, it is promptly reported, and the corresponding fault type and occurrence time are marked, thus achieving effective prediction of multiple fault types in the vacuum dry pump. Specifically, steps A61 to A65 are included.

[0213] A61, Input vibration signal data for the entire lifespan D real After preprocessing, the data preprocessing process is the same as that described in claim 3. Then, a window size W and sliding step S, consistent with claim 4, are set on the vibration signal data. First, the data within and before this window are gradually normalized, and then sliding window extraction is performed to form window data X. real X real Represented as:

[0214] X real ={x i x i+1 , ..., x i+w-1}

[0215] Where i is the starting index of the window.

[0216] A62, for each window data X real Calculate the characteristic matrix F real The feature matrix has a dimension of 1×k, where k is the number of selected features. The feature matrix can be represented as follows:

[0217] F real = [f1, f2, ..., f k ]

[0218] A63, the characteristic matrix F real The input is fed into the improved C4.5 model for classification to determine whether the window belongs to a certain type of fault characteristic. The model outputs the category label C. pred It can be represented as:

[0219] C pred =C4.5(F real )

[0220] A64, If the model identifies a fault, the system will promptly report the fault and record the corresponding fault type C. pred and the time of occurrence T pred , can be represented as:

[0221] Fault report = {C pred T pred}

[0222] A65, finally evaluate the accuracy of the fault type prediction. If the fault type C output by the model... pred Compared with the actual fault type C true If the match is consistent, the prediction is correct; otherwise, it is incorrect. The evaluation formula is:

[0223]

[0224] The accuracy of time interval prediction is determined by calculating the predicted failure occurrence time T using the model. pred Compared with the actual time T true The evaluation is based on the differences between them. The evaluation formula is:

[0225]

[0226] Where ΔT is the allowable time error range.

[0227] When both the fault type and time prediction are correct, it is recorded as a successful prediction. The total number of correct predictions is N. correct The calculation formula is:

[0228]

[0229] Where N represents the actual number of failures.

[0230] To verify the effectiveness of the proposed model and its solution, data from vacuum dry pumps with full-lifetime vibration signal data under a specific process condition at a certain company were selected as the research object. The data includes full-lifetime vibration signal data from 49 vacuum dry pumps. Data from 34 vacuum dry pumps were selected to construct a multi-classification dataset, and data from 15 vacuum dry pumps were selected to simulate online real-time fault prediction.

[0231] For vibration data from 34 vacuum dry pumps, a sliding window was created on the full-life vibration signal data by setting the sliding window size and step size. First, the data within and before the window was progressively normalized, and then the data within the window was extracted, thus extracting multiple window sample data. Expert annotation was used to label the samples, and each sample was classified into fault categories, constructing a multi-class dataset containing multiple window sample data and corresponding labels. Finally, the dataset was randomly divided into training and testing sets. The experimental parameters of this invention are shown in Table 1 below:

[0232] Table 1. Experimental parameter settings

[0233] Parameter Name Setting Value Sliding Window Size 128 data points Step Size 16 data points C4.5 Max Depth 15 C4.5 Min Sample Split 5 C4.5 Gain Threshold 0.01 Model Training Rounds 100 Number of Predicted Fault Types 4 classes (3 fault + 1 normal) Real-time Prediction Update Frequency Every 5 seconds

[0234] The training and testing model for multi-fault prediction of vacuum dry pumps based on the improved C4.5 of this invention was used to observe its performance. The confusion matrix of the test set results in the multi-class dataset is shown in Table 2.

[0235] Table 2. Confusion Matrix

[0236] Predicted Class 1 Predicted Class 2 Predicted Class 3 Predicted Normal Actual Class 1 140 31 15 14 Actual Class 2 19 138 22 21 Actual Class 3 5 12 79 5 Actual Normal 27 15 8 350

[0237] Based on the confusion matrix, the accuracy of multi-class prediction is calculated to be 72.66%.

[0238] The recall and F1 score for multi-class prediction are shown in Table 3.

[0239] Table 3. Recall and F1 score results

[0240] Class Recall F1 Score Class 1 68.48％ 70.17％ Class 2 70.85％ 69.58％ Class 3 79.45％ 74.75％ Normal 87.50％ 87.50％

[0241] The accuracy of real-time fault prediction based on vibration data of 15 vacuum dry pumps is shown in Table 4.

[0242] Table 4. Real-time Fault Prediction Accuracy

[0243] Fault Type Actual Occurrences Predicted Faults Accuracy Fault Type 1 147 104 70.73％ Fault Type 2 95 62 65.26％ Fault Type 3 49 37 75.51％ Overall Accuracy 291 203 69.74％

[0244] For details, see Figure 5 The diagram shown is a system structure diagram according to an embodiment of the present invention, including:

[0245] Training module 501 uses the improved C4.5 algorithm to construct a decision tree model and uses the full-life vibration signal data of the vacuum dry pump for training to obtain a trained multi-fault prediction model.

[0246] The acquisition module 502 acquires vibration signal data from the real-time operation monitoring of the vacuum dry pump and performs data preprocessing on the data.

[0247] The identification module 503 inputs the preprocessed vibration signal data into the trained multi-fault prediction model to predict whether a fault occurs in the vibration signal data and outputs the fault prediction result, which includes the fault type and the fault occurrence time.

[0248] In summary, compared with existing technologies, the decision tree structure generated by the C4.5 algorithm in this invention is clear, easy to understand and interpret, enabling engineers to intuitively analyze prediction results and possessing high operability. Other black-box models, such as deep learning, typically lack transparency and make it difficult to trace the decision-making basis. Secondly, this invention combines sliding window technology and progressive normalization to effectively capture dynamic changes in vibration signals, ensuring the model can be updated in real time and adapt to new operating conditions. Compared to clustering and Transformer methods, which may neglect local details, this method has stronger adaptability. In terms of feature engineering, this method significantly improves the accuracy of fault classification by introducing time-frequency domain features such as stationarity, high-frequency component proportion, and Hausdorff distance, as well as other features. Algorithms such as Transformer, XGBoost, and Catboost typically rely on automatic feature selection, which may not adequately handle specific local features. The improved C4.5 algorithm can also update the prediction model in real time, adapting to new data and maintaining high accuracy, while methods such as Transformer typically require retraining the entire model, increasing computational cost and time overhead. Furthermore, this method is specifically designed for complex faults caused by multiple factors, such as dust jamming, and can effectively distinguish different types of faults, reducing false alarms and false negatives. Compared with other methods, such as fault prediction methods based on Transformer, XGBoost, and Catboost, it has stronger recognition capabilities when handling complex faults. Finally, by combining information gain ratio based on attribute weights, the contribution of each feature in the prediction is ensured, thereby improving prediction accuracy. In contrast, methods such as Transformer and XGBoost may rely on automatic feature selection strategies, failing to fully consider the importance of each feature in different fault modes. Overall, this invention addresses the shortcomings of existing technologies through a transparent decision-making process, real-time updates, refined feature extraction, and powerful multi-fault classification capabilities, possessing higher practical value.

[0249] The above are merely preferred embodiments of the present invention and are not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A multi-fault prediction method for vacuum dry pumps based on improved C4.5, characterized in that, include: A decision tree model was constructed using the improved C4.5 algorithm and trained using the full-life vibration signal data of the vacuum dry pump to obtain a well-trained multi-fault prediction model. Acquire vibration signal data from real-time operation monitoring of the vacuum dry pump, and preprocess the data; The preprocessed vibration signal data is input into the trained multi-fault prediction model to predict whether a fault occurs in the vibration signal data, and the fault prediction result is output, which includes the fault type and the fault occurrence time. The method of constructing a decision tree model using the improved C4.5 algorithm includes the following steps: An improved C4.5 algorithm is obtained by introducing an attribute weight mechanism into the original C4.5 algorithm. The improved C4.5 algorithm selects split points using information gain ratio to construct a decision tree model, as follows: ； in, Representing training features, Representing training features The label; the information gain ratio of the improved C4.5 algorithm is expressed as: ； ； ； in, It is an attribute The weight, It is the information gain ratio of attribute A calculated by the original C4.5 algorithm. It is the information gain ratio of attribute A calculated by the improved C4.5 algorithm; This represents the variance of attribute A within a time window t, reflecting the degree of its change during that time window. The experiential importance index for attribute A as assessed by experts is based on expert knowledge. This represents the time adjustment coefficient, used to dynamically adjust feature weights according to changes over time. The starting point of the time window. This represents the variance of the previous time window, reflecting the change in the importance of the attribute within different time windows.

2. The method for predicting multiple faults in a vacuum dry pump based on improved C4.5 according to claim 1, characterized in that, The process of constructing a decision tree model using the improved C4.5 algorithm and training it with full-life vibration signal data of a vacuum dry pump includes the following steps: A decision tree model was constructed using the improved C4.5 algorithm; The vibration signal data of the vacuum dry pump throughout its entire lifespan is acquired and preprocessed. The preprocessed data is then labeled to obtain a dataset including sample data and corresponding labels. Feature extraction is performed on the dataset to obtain a feature set; the features of the feature set include time-domain features and frequency-domain features. The decision tree model is trained using a feature set.

3. The method for predicting multiple faults in a vacuum dry pump based on improved C4.5 according to claim 1, characterized in that, The acquisition of vibration signal data for real-time operation monitoring of vacuum dry pumps specifically involves: installing multiple sensors at different locations on each vacuum dry pump, with each sensor independently acquiring vibration signals at that location; each sensor represents a channel, thereby acquiring vibration data from multiple channels to form a multi-dimensional time-series dataset.

4. The method for predicting multiple faults in a vacuum dry pump based on improved C4.5 according to claim 1 or 2, characterized in that, The preprocessing includes: data dimensionality reduction, handling missing and outlier values, and normalizing the data to a uniform range.

5. A multi-fault prediction method for vacuum dry pumps based on improved C4.5 according to claim 2, characterized in that, The labeling of the preprocessed data specifically involves: performing a sliding window on the preprocessed whole-life vibration signal data; first, performing a progressively increasing real-time normalization on the window and the data before it; then, extracting the data from the window to construct a multi-class dataset containing sample data from multiple windows and corresponding labels.

6. The method for predicting multiple faults in a vacuum dry pump based on an improved C4.5 according to claim 2, characterized in that, The features of the feature set also include: The zero-crossing rate is expressed as: ； in, It is the number of times that adjacent points cross the zero point. It represents the total number of data points within the window; Stability is expressed as: ； in, It is the (i+1)th data point in the signal. It is the i-th data point in the signal; The percentage of time with unstable data is represented as follows: ； in, It is the number of adjacent points whose differences exceed a set threshold; The proportion of high-frequency components is expressed as: ； in, Indicates frequency, Indicates the high-frequency threshold. This represents the energy of the high-frequency components. This represents the total energy of the signal; Hausdorff distance, expressed as: ； in, A sample representing a certain type of fault. This indicates a sample selected from other types of fault or normal data.

7. A multi-fault prediction method for vacuum dry pumps based on improved C4.5 according to claim 6, characterized in that, The process of inputting the preprocessed vibration signal data into the trained multi-fault prediction model includes the following steps: For the preprocessed vibration signal data, the window size W and sliding step size are set. First, the data in the window and the data before it are gradually normalized, and then the sliding window is used to extract the data to form the window data. Feature extraction is performed on the data from each window to obtain a feature matrix; feature matrix The input is fed into a pre-trained multi-fault prediction model for classification, and the model outputs the category labels. Represented as: ； If the model identifies a fault, it will output a fault report, represented as: ； in, Indicates the time of occurrence.

8. The method for predicting multiple faults in a vacuum dry pump based on an improved C4.5 according to claim 7, characterized in that, The feature matrix includes important features selected from the feature set, and the selection process includes the following steps: Standardize the data in the feature set; Calculate the covariance matrix based on the standardized data. : For covariance matrix Perform eigenvalue decomposition to obtain eigenvalues. and the corresponding feature vector , represented as: ； in, for The corresponding eigenvalues, Let i be the i-th eigenvector of the feature set; eigenvalues Sort by value from largest to smallest, and select the features corresponding to the top few feature values ​​as important features.

9. A multi-fault prediction system for vacuum dry pumps based on an improved C4.5, characterized in that, include: The training module uses the improved C4.5 algorithm to build a decision tree model and uses the full-life vibration signal data of the vacuum dry pump for training to obtain a trained multi-fault prediction model. The acquisition module acquires vibration signal data from the real-time operation monitoring of the vacuum dry pump and performs data preprocessing on the data. The identification module inputs the preprocessed vibration signal data into the trained multi-fault prediction model to predict whether a fault occurs in the vibration signal data and outputs the fault prediction result, which includes the fault type and the fault occurrence time. The method of constructing a decision tree model using the improved C4.5 algorithm includes the following steps: An improved C4.5 algorithm is obtained by introducing an attribute weight mechanism into the original C4.5 algorithm. The improved C4.5 algorithm selects split points using information gain ratio to construct a decision tree model, as follows: ； in, Representing training features, Representing training features The label; the information gain ratio of the improved C4.5 algorithm is expressed as: ； ； ； in, It is an attribute The weight, It is the information gain ratio of attribute A calculated by the original C4.5 algorithm. It is the information gain ratio of attribute A calculated by the improved C4.5 algorithm; This represents the variance of attribute A within a time window t, reflecting the degree of its change during that time window. The experiential importance index for attribute A as assessed by experts is based on expert knowledge. This represents the time adjustment coefficient, used to dynamically adjust feature weights according to changes over time. The starting point of the time window. This represents the variance of the previous time window, reflecting the change in the importance of the attribute within different time windows.

Images