A rotating machinery fault diagnosis method and system based on multi-feature fusion machine learning

Abstract

This invention provides a multi-feature fusion machine learning-based method and system for fault diagnosis of rotating machinery, comprising the following steps: acquiring vibration, rotational speed, load, and eccentricity signals of the rotating machinery; preprocessing the signals, including missing value imputation, outlier handling, and standardization; extracting multiple mathematical statistical features from the signals, including time-domain and frequency-domain statistical features; evaluating and filtering the extracted features, selecting a subset of preferred features with high discriminative power; then training multiple machine learning models based on the filtered features, and autonomously selecting the optimal diagnostic model by comparing the diagnostic performance of each model; finally, intelligently calling the optimal diagnostic model to predict and diagnose the vibration state of the rotating machinery. This invention achieves intelligent feature selection and model optimization, improving the accuracy and adaptability of diagnosis, and can be widely applied to vibration monitoring and fault prediction and diagnosis of various types of rotating machinery.

Description

Technical Field

[0001] This invention relates to the technical field of rotating machinery fault diagnosis and machine learning algorithms, specifically to a rotating machinery fault diagnosis method and system based on multi-feature fusion machine learning. Background Technology

[0002] Rotating machinery (such as motors, fans, pumps, compressors, and various bearings) are key components in modern industrial equipment, and their operational reliability and safety are of paramount importance. Vibration signal analysis is one of the most effective and commonly used techniques for monitoring the condition and diagnosing faults in rotating machinery. Traditional vibration fault diagnosis methods mainly rely on field experts to manually observe and judge the characteristics of the spectrum and time-domain waveforms. This approach is not only inefficient and highly dependent on human experience, but also difficult to achieve real-time, online, large-scale equipment monitoring.

[0003] With the development of machine learning technology, data-driven intelligent diagnostic methods have gradually become mainstream. These methods typically preprocess and extract features from the collected vibration signals before using classification models for fault identification. However, existing methods generally suffer from the following limitations: First, feature extraction often relies on a few fixed or manually selected indicators, lacking a systematic evaluation and automatic screening mechanism for the effectiveness of features under different operating conditions and fault types, which may lead to redundant or irrelevant features interfering with model performance. Second, the selection of diagnostic models is often based on trial and error or a single model preference, failing to fully utilize the characteristics of different models and select the optimal solution through systematic comparison, thus affecting the accuracy and generalization ability of the diagnosis. Furthermore, when facing complex scenarios such as imbalanced sample classes, cross-operating conditions, or cross-equipment situations, the adaptability and robustness of existing methods are often insufficient. Summary of the Invention

[0004] This invention is made to solve the above-mentioned problems, and aims to provide a method and system for diagnosing rotating machinery faults by multi-feature fusion machine learning.

[0005] This invention provides a multi-feature fusion machine learning-based method for diagnosing rotating machinery faults, comprising the following steps: S1: Signal acquisition and preprocessing step, acquiring signals from the rotating machinery and preprocessing the signals; S2: Multi-dimensional feature extraction step, extracting time-domain statistical features and frequency-domain statistical features from the preprocessed signals to form an initial feature set; S3: Feature evaluation and screening step, evaluating the effectiveness of each feature in the initial feature set, and selecting a preferred feature subset for fault classification based on the evaluation results; S4: Multi-model training and comparison step, using the preferred feature subset, training multiple different types of machine learning classification models in parallel, and comparing the diagnostic performance of each machine learning classification model based on a preset evaluation system to obtain comparison results; S5: Optimal model deployment step, based on the comparison results, selecting the machine learning classification model with the best diagnostic performance as the final diagnostic model; S6: Fault prediction and diagnosis step, after processing the signal to be diagnosed in S1 and S2, extracting the feature values ​​corresponding to the preferred feature subset, inputting them into the final diagnostic model, and outputting the fault category diagnosis result.

[0006] The rotating machinery fault diagnosis method based on multi-feature fusion machine learning provided by this invention may also have the following features: In S1, the signals include vibration, rotational speed, load, and eccentricity. The preprocessing method is as follows: fill in the missing values ​​of the signals. For each missing value, fill it with the average or adjacent value of the adjacent values ​​according to its position, and calculate the mean and standard deviation of the filled signal. Values ​​exceeding a preset multiple of the standard deviation are regarded as outliers and replaced with the average or adjacent value of the adjacent values. Then, standardize all signal data using the StandardScaler method.

[0007] The rotating machinery fault diagnosis method based on multi-feature fusion machine learning provided by this invention may also have the following features: wherein, in S2, the time-domain statistical features include:

[0008]

[0009] in, It is the number of signal sampling points. It is the first The signal value at the sampling point, The first correction factor is determined based on the rotational speed of the rotating machinery.

[0010]

[0011] in, It is the average value of the signal. This is the second correction factor, determined based on the rotational speed of the rotating machinery.

[0012]

[0013] in, It is the standard deviation of the signal.

[0014]

[0015]

[0016]

[0017]

[0018] Frequency domain statistical characteristics include:

[0019]

[0020] in, is the number of positive frequencies, and k is the summation index. It is the signal at frequency The result of the fast Fourier transform at the given location. It is the amplitude value. This is the third correction factor, determined based on the distribution range of the fault characteristic frequencies.

[0021]

[0022] in, It is the first frequency domain eigenvalue. This is the fourth correction factor, determined based on the dispersion of the fault characteristic frequency distribution.

[0023]

[0024] in, This is the fifth correction factor, used to emphasize the fault characteristic frequency band.

[0025] The rotating machinery fault diagnosis method based on multi-feature fusion machine learning provided by this invention may also have the following feature: wherein, the first correction coefficient Calculated using the following formula:

[0026]

[0027] in, The rotational speed of the rotating machinery.

[0028] Second correction factor Calculated using the following formula:

[0029]

[0030] Third correction factor Calculated using the following formula:

[0031]

[0032] in, The centroid frequency of the spectrum, The highest analysis frequency, The sensitivity coefficient,

[0033] Fourth correction factor Calculated using the following formula:

[0034]

[0035] in, The standard deviation of the spectrum,

[0036] Fifth correction factor Calculated using the following formula:

[0037]

[0038] in, For high-frequency energy, Low-frequency energy, This is an adjustment factor, with a value range of 0.01 to 0.05.

[0039] In the rotating machinery fault diagnosis method based on multi-feature fusion machine learning provided by the present invention, it may also have the following feature: wherein, in S3, the method for selecting the preferred feature subset is: calculating the variance analysis F value of each feature relative to different fault categories to evaluate the ability of each feature to distinguish fault categories, sorting them from high to low according to the F value, and selecting a preset number of features with the highest ranking to form the preferred feature subset.

[0040] The rotating machinery fault diagnosis method based on multi-feature fusion machine learning provided by this invention may also have the following features: various types of machine learning classification models include: random forest, support vector machine, gradient boosting tree family model, and multilayer perceptron neural network. S4 specifically includes the following sub-steps: S4-1: Based on the preferred feature subset, extract corresponding feature data from the labeled signal feature dataset to form an effective dataset for model training. The labeled signal feature dataset is generated based on the initial feature set and historical signals of known fault categories. S4-2: Divide the effective dataset into a training set, a validation set, and an independent test set. S4-3: Use the training set to train various types of machine learning classification models in parallel, and use the validation set to perform hyperparameter tuning and performance evaluation on each machine learning classification model. S4-4: Based on a preset evaluation system, compare the diagnostic performance of each machine learning classification model to obtain comparison results.

[0041] The rotating machinery fault diagnosis method based on multi-feature fusion machine learning provided by this invention may also have the following features: In S4, the preset evaluation system is executed according to the following rules: Safety filtering: A minimum average recall threshold is set as a safety requirement. If the average recall of any machine learning classification model is lower than this threshold, it will be excluded; Key fault constraints: For one or more preset key fault categories, their independent minimum performance thresholds are set respectively; For any machine learning classification model, if its evaluation index on the key fault category is lower than the corresponding minimum performance threshold, it will be directly excluded or downgraded in the comprehensive performance ranking; Comprehensive performance ranking: For machine learning classification models that meet the safety requirements, comprehensive performance ranking is performed according to preset evaluation indicators, including: accuracy, precision, recall, and F1 score; Engineering optimization: Based on the comprehensive performance ranking results, the machine learning classification model with the best comprehensive performance ranking is selected. If the comprehensive performance rankings of multiple machine learning classification models are the same or the differences are within the preset allowable error range, the machine learning classification model with a simpler model structure, faster inference speed, or less computational resources is given priority.

[0042] This invention also provides a multi-feature fusion machine learning-based rotating machinery fault diagnosis system, comprising: a signal acquisition and preprocessing module for acquiring and preprocessing signals from the rotating machinery; a multi-dimensional feature extraction module for extracting time-domain and frequency-domain statistical features from the preprocessed signals to form an initial feature set; a feature evaluation and screening module for evaluating the effectiveness of each feature in the initial feature set and selecting a preferred feature subset for fault classification based on the evaluation results; a multi-model training and comparison module for training multiple different types of machine learning classification models in parallel using the preferred feature subset and comparing the diagnostic performance of each machine learning classification model based on a preset evaluation system to obtain comparison results; an optimal model deployment module for selecting the machine learning classification model with the best diagnostic performance as the final diagnostic model based on the comparison results; and a fault prediction and diagnosis module for extracting feature values ​​corresponding to the preferred feature subset from the signal to be diagnosed after processing by the signal acquisition and preprocessing module and the multi-dimensional feature extraction module, inputting them into the final diagnostic model, and outputting the fault category diagnosis result.

[0043] Compared with the prior art, the present invention has the following advantages:

[0044] This invention employs a data-driven technical solution. Through an automated feature selection mechanism, it objectively selects a subset of highly discriminative features from high-dimensional features. Combined with a systematic model comparison strategy, it autonomously selects the diagnostic model with the best overall performance from multiple candidate models. This achieves intelligent feature selection and model optimization, improving the accuracy and adaptability of diagnosis. Compared with existing technologies, this invention effectively reduces reliance on expert experience, automatically adapts to different equipment and operating conditions, and enhances the accuracy and robustness of fault diagnosis while increasing the method's engineering applicability and promotion potential. It is suitable for vibration monitoring and intelligent fault early warning of various rotating machinery. Attached Figure Description

[0045] Figure 1 This is a flowchart of a rotating machinery fault diagnosis method based on multi-feature fusion machine learning, as described in an embodiment of the present invention.

[0046] Figure 2 This is a flowchart illustrating the multi-feature fusion machine learning-based rotating machinery fault diagnosis method in an embodiment of the present invention. Detailed Implementation

[0047] To make the technical means, creative features, objectives and effects of this invention easy to understand, the following embodiments, in conjunction with the accompanying drawings, specifically illustrate the rotating machinery fault diagnosis method and system based on multi-feature fusion machine learning of this invention.

[0048] The multi-feature fusion machine learning-based method for diagnosing rotating machinery faults provided in this embodiment includes the following steps:

[0049] Figure 1 This is a flowchart of a rotating machinery fault diagnosis method based on multi-feature fusion machine learning, as described in an embodiment of the present invention. Figure 2 This is a flowchart illustrating the multi-feature fusion machine learning-based rotating machinery fault diagnosis method in an embodiment of the present invention.

[0050] like Figures 1-2 As shown, step S1 is the signal acquisition and preprocessing step, which involves acquiring signals from the rotating machinery and preprocessing the signals, specifically as follows:

[0051] The system collects signals such as vibration, rotational speed, load, and eccentricity of rotating machinery and preprocesses these signals.

[0052] The main steps of preprocessing are: imputing missing values ​​in the signal; for each missing value, filling it with the average or adjacent values ​​of the adjacent values ​​according to its location; calculating the mean and standard deviation of the imputed signal; considering values ​​exceeding 4 times the standard deviation as outliers and replacing them with the average or adjacent values ​​of the adjacent values; and then standardizing all signal data using the StandardScaler method.

[0053] StandardScaler is a data standardization technique based on Z-score. Its core idea is to center and scale the original features so that the processed data conforms to a standard normal distribution with a mean of 0 and a standard deviation of 1. Its mathematical expression is:

[0054]

[0055] in, The mean of the features on the training set, denoted as the standard deviation of the feature on the training set.

[0056] In rotating machinery fault diagnosis, the collected vibration, rotational speed, and load signals have different dimensions and numerical ranges (e.g., vibration amplitude is typically 0–10 m / s², while rotational speed ranges from hundreds to thousands of rpm). If the raw data is used directly for modeling, features with larger numerical values ​​may dominate model training, leading to a decline in model performance. StandardScaler standardization can eliminate the influence of dimensions, allowing all features to participate in subsequent feature evaluation and model training on the same scale. This not only improves the model's convergence speed but also enhances the stability and generalization ability of the diagnostic model.

[0057] Step S2 is a multi-dimensional feature extraction step, in which time-domain statistical features and frequency-domain statistical features are extracted from the preprocessed signal to form an initial feature set.

[0058] Specifically, time-domain statistical features include, but are not limited to:

[0059]

[0060] in, It is the number of signal sampling points. It is the first The signal value at the sampling point, This is the first correction factor.

[0061]

[0062] in, It is the average value of the signal. This is the second correction factor.

[0063]

[0064] in, It is the standard deviation of the signal.

[0065]

[0066]

[0067]

[0068]

[0069] Frequency domain statistical characteristics include:

[0070]

[0071] in, is the number of positive frequencies, and k is the summation index. It is the signal at frequency The result of the fast Fourier transform at the given location. It is the amplitude value. This is the third correction factor.

[0072]

[0073] in, It is the first frequency domain eigenvalue. This is the fourth correction factor.

[0074]

[0075] in, This is the fifth correction factor.

[0076] It should be noted that in this embodiment, five adjustable correction coefficients are introduced during the time-domain and frequency-domain feature extraction process: the first correction coefficient α, the second correction coefficient β, the third correction coefficient γ, the fourth correction coefficient δ, and the fifth correction coefficient ε. These correction coefficients are not fixed but are flexibly adjusted according to the specific type of rotating machinery, operating conditions, and fault characteristics, and are determined according to the following rules:

[0077] First correction factor Determined based on the rotational speed of the rotating machinery, and calculated using the following formula:

[0078]

[0079] in, Rotational speed of the rotating machinery (unit: rpm).

[0080] This formula is designed based on the empirical relationship between rotational speed and vibration signal characteristics. Lower rotational speeds result in more pronounced impact fault characteristics, necessitating increased sensitivity to the root mean square (RMS) value. Take a larger value; the higher the rotational speed, the greater the high-frequency noise interference, so the root mean square value needs to be appropriately suppressed to avoid false alarms. Take the smaller value.

[0081] In practical applications, if the rotational speed exceeds the range of 0–6000 rpm, it can be limited. The value of the formula can be expanded by taking boundary values ​​or adjusting the coefficients.

[0082] Second correction factor Determined based on the rotational speed of the rotating machinery, and calculated using the following formula:

[0083]

[0084] in, Rotational speed of rotating machinery (unit: rpm).

[0085] Used to adjust the standard deviation feature, its value selection rules are the same as Similarly, but with a focus on reflecting the volatility of the signal. The formula is in... rpm = 1.13, in rpm = 0.85, decreasing linearly.

[0086] Third correction factor Determined based on the main distribution range of fault characteristic frequencies, and calculated using the following formula:

[0087]

[0088] in, The centroid frequency of the spectrum, The highest analysis frequency, This is the sensitivity coefficient.

[0089] Spectral centroid frequency The calculation formula is:

[0090]

[0091] Used to adjust the sensitivity of frequency domain centroid characteristics to fault frequency distribution. When the fault is dominated by low-frequency components (such as imbalance, misalignment), , < 1, thus highlighting low-frequency energy; when the fault is dominated by high-frequency components (such as early bearing damage), , > 1. Enhance high-frequency characteristics.

[0092] In practical applications, take When, the range of values ​​for γ varies. It varies from 0.2 to 0.8, ranging from 0.97 to 1.03.

[0093] Fourth correction factor The degree of dispersion of fault characteristic frequencies is reflected by the following formula:

[0094]

[0095] in, The standard deviation of the spectrum reflects the degree of dispersion of frequency components, and is calculated using the following formula:

[0096]

[0097] Used to adjust the frequency domain standard deviation characteristic to adapt to the fault bandwidth. For narrowband faults (such as single frequency components). Smaller (e.g., 0.1) 1. Emphasize the main frequency; for broadband faults (such as impacts, friction). Larger (e.g., 0.5) 1. Enhanced spectral impact; for hybrid faults, The value is centered.

[0098] Fifth correction factor High-frequency or low-frequency components used to enhance fault characteristics are represented as follows:

[0099]

[0100] in, This refers to high-frequency energy (e.g., the sum of spectral energy above 500 Hz). This refers to low-frequency energy (e.g., the total spectral energy below 500 Hz). This is an adjustment factor, with a value ranging from 0.01 to 0.05, preferably... .

[0101] Used to adjust the emphasis of the frequency domain squared centroid feature on high and low frequency energy. When it is necessary to emphasize high-frequency faults (such as pitting), , > 1; When it is necessary to emphasize low-frequency faults (such as imbalance), , <1.

[0102] Pick When the energy ratio varies within the typical range of 0.2 to 5.0, The corresponding value range is 0.984 to 1.08. If the actual energy ratio exceeds this range, it can be adjusted by limiting the boundary value or by adjusting... Keep Within a reasonable range of values ​​for the project (0.80 to 1.10).

[0103] The above value selection rules are based on vibration signal analysis and fault diagnosis simulation experiments of various rotating machinery (including motors, fans, pumps, compressors, bearings, etc.). In actual operation, the selection can be made according to the equipment model and operating parameters, referring to the above rules. Alternatively, optimization and calibration can be performed using a small amount of historical data (such as normal samples), without the need for extensive trial and error.

[0104] The introduction of these five correction coefficients enables the feature extraction process to adapt to different types of mechanical equipment, operating conditions, and signal characteristics. For example, for a high-speed rotating centrifugal compressor, its vibration signal has high frequency components and nonlinear characteristics. Appropriately adjusting the correction coefficients can enhance the sensitivity of the frequency domain features to high-frequency components. For a low-speed, heavy-load bearing system, adjusting the correction coefficients helps to highlight the feature components in the time domain waveform related to faults such as impact and friction.

[0105] Therefore, these five correction coefficients are not only key parameters in the feature extraction process, but also one of the important reasons why the method of this invention can be widely adapted to various rotating machinery (such as motors, fans, pumps, compressors, bearings, etc.). They enable this invention to achieve flexible adaptation to different equipment types, different fault modes, and different operating conditions by adjusting a few parameters without changing the core algorithm framework, significantly improving the versatility, robustness, and engineering promotion value of the method.

[0106] Step S3 is the feature evaluation and selection step, which evaluates the effectiveness of each feature in the initial feature set and, based on the evaluation results, automatically selects a preferred subset of features for fault classification. Specifically:

[0107] Calculate the F-value of each feature relative to different fault categories using analysis of variance, and assess the significance of the difference in the mean of the feature under different fault categories, so as to measure the ability of each feature to distinguish fault categories.

[0108] The ANOVA F-value is a statistical indicator used to assess the ability of a feature variable to distinguish a target variable for classification. In rotating machinery fault diagnosis, it measures whether the mean difference of a feature (such as the root mean square value of a vibration signal, center of gravity of frequency, etc.) is significant across different fault categories (such as imbalance, misalignment, gear tooth breakage, etc.). A larger F-value indicates a more significant difference in the distribution of that feature under different fault states, and a greater contribution to fault classification. The specific calculation formula is as follows:

[0109]

[0110] in, It is the total number of fault categories. It is the first Number of samples of this type of fault Representing the The mean of this type of fault on this feature, It is the population mean of all samples on this feature. No. The first type of fault One sample, It represents the total number of samples.

[0111] Sort the features from high to low according to their F-values, and select a preset number of features that rank highly to form an optimal feature subset.

[0112] Step S4 is the multi-model training and comparison step. Using the preferred feature subset, multiple different types of machine learning classification models are trained in parallel, and the diagnostic performance of each machine learning classification model is compared based on the preset evaluation system to obtain the comparison results.

[0113] Specifically, various types of machine learning classification models include, but are not limited to: random forests, support vector machines, gradient boosting tree family models, and multilayer perceptron neural networks.

[0114] Step S4 specifically includes the following sub-steps:

[0115] S4-1: Based on the optimized feature subset, extract corresponding feature data from the labeled signal feature dataset to form an effective dataset for model training. The labeled signal feature dataset is generated by calculating historical signals of known fault categories based on the initial feature set.

[0116] In this embodiment, corresponding feature data is extracted from the labeled vibration signal feature dataset to form an effective dataset for model training.

[0117] S4-2: Divide the effective dataset into a training set, a validation set, and an independent test set, and apply the SMOTE oversampling technique to the training set to balance the number of samples for different fault categories.

[0118] SMOTE (Synthetic Minority Over-sampling Technique) is an effective method for addressing class imbalance in classification tasks. Its basic principle is not to simply copy minority class samples, but rather to artificially synthesize new samples based on the feature space distribution of the minority class samples through linear interpolation. The specific implementation process is as follows:

[0119] For each minority class sample First, find the k nearest neighbors (usually k=5) in the same type of sample, and then randomly select a nearest neighbor. , and in and A new sample is randomly generated on the connection line:

[0120]

[0121] In this way, SMOTE effectively expands the number of minority class samples, improving the class distribution of the dataset. In rotating machinery fault diagnosis, data from normal operation is usually far more abundant than data from fault conditions, while samples of certain critical faults (such as broken gear teeth or broken bearing cages) are particularly scarce. If the original imbalanced data is used directly to train the model, the model often biases towards the majority class (normal state), leading to a higher false negative rate for minority class faults. After applying the SMOTE oversampling technique, the number of fault samples for each class tends to be balanced, which can significantly improve the model's recall and F1 score for minority class faults, thereby enhancing the robustness and reliability of the diagnostic method in real industrial data environments.

[0122] S4-3: Use the training set to train multiple different types of machine learning classification models in parallel, and use the validation set to perform hyperparameter tuning and performance evaluation on each machine learning classification model.

[0123] Hyperparameter tuning refers to the process of optimizing and selecting structural parameters of a machine learning model (such as the kernel type of a support vector machine, the number of trees in a random forest, and the number of layers in a neural network) using systematic search methods (such as grid search, random search, or Bayesian optimization) before training the model. Its purpose is to obtain optimal model performance on the validation set, avoiding underfitting or overfitting. In rotating machinery fault diagnosis, hyperparameter tuning helps improve the model's adaptability to different operating conditions and fault types, enhancing the stability and reliability of diagnostic results.

[0124] S4-4: Based on the preset evaluation system, compare the diagnostic performance of each machine learning classification model and obtain the comparison results.

[0125] The pre-set evaluation system will be implemented according to the following rules:

[0126] Security filtering: Set a minimum average recall threshold (e.g., 90%, which can be set and adjusted based on the application scenario) as a security requirement. If the average recall of any machine learning classification model is lower than this threshold, it will be excluded first to ensure that critical faults are not missed.

[0127] Critical Fault Constraints: For one or more preset critical fault categories, separate minimum performance thresholds are set for each. For any machine learning classification model, if its evaluation metric for a critical fault category is lower than the corresponding minimum performance threshold, it is directly excluded or downgraded in the overall performance ranking. Specifically:

[0128] Particularly critical or high-consequence fault categories can be defined based on actual circumstances. These categories refer to faults that, if missed, could lead to serious equipment damage, unplanned downtime, safety accidents, or significant economic losses. Examples include, but are not limited to, severe rotor imbalance, broken gear teeth, and broken fan blades. Comprehensive performance ranking: For machine learning classification models that meet safety requirements, a comprehensive performance ranking is performed based on preset evaluation metrics.

[0129] Evaluation metrics include: accuracy, precision, recall, and F1 score.

[0130] Accuracy is one of the most commonly used performance evaluation metrics for machine learning classification models (hereinafter referred to as models). It is defined as the proportion of samples correctly predicted by the model out of the total number of samples. In rotating machinery fault diagnosis, accuracy reflects the model's overall ability to identify all fault categories (including normal states). However, in cases of imbalanced sample categories (such as a much larger number of normal samples than fault samples), accuracy may be distorted. Therefore, it is necessary to combine it with other metrics for comprehensive evaluation.

[0131] Precision refers to the proportion of samples that a model predicts as belonging to a certain fault category, but which actually belong to that category; in other words, "how many samples predicted as positive are actually positive." In fault diagnosis, a higher precision indicates a lower false alarm rate for that type of fault, making it suitable for scenarios where false alarm warnings are critical (such as avoiding unnecessary downtime for inspection).

[0132] Recall refers to the proportion of samples correctly identified by a model within a specific fault category out of the total number of actual samples in that category; in other words, "how many true positive samples are correctly identified." In rotating machinery fault diagnosis, recall reflects the model's ability to detect a particular type of fault. Especially for critical faults with serious consequences (such as broken teeth or broken blades), a high recall rate is an important indicator for ensuring the safe operation of equipment.

[0133] The F1 score is the harmonic mean of precision and recall, comprehensively reflecting the model's ability to identify a particular class. Its calculation formula is:

[0134]

[0135] In rotating machinery fault diagnosis, the F1 score is often used for model comparison in imbalanced scenarios, especially in key fault categories, where it can more comprehensively measure the balance between the model's diagnostic accuracy and the risk of missed detections.

[0136] In this embodiment, the F1 score is used as an evaluation metric for illustration:

[0137] For certain critical or serious fault categories, separate minimum F1 score requirements are set, and machine learning classification models that fail to meet the requirements will be downgraded or excluded.

[0138] For machine learning classification models that meet security requirements, the overall performance ranking is mainly based on the macro average F1 score.

[0139] Engineering Optimization: Based on the above ranking results, select the machine learning classification model with the best overall performance. If multiple machine learning classification models have the same overall performance ranking (e.g., the same macro average F1 score) or the difference is within the preset allowable error range, then the machine learning classification model with a simpler model structure, faster inference speed, or less computational resources is preferred to facilitate subsequent engineering deployment and real-time diagnosis.

[0140] Step S5 is the optimal model deployment step. Based on the comparison results, the machine learning classification model with the best diagnostic performance is selected as the final diagnostic model, and its performance is independently verified using a test set.

[0141] Step S6 is the fault prediction and diagnosis step. After the signal of the rotating machinery to be diagnosed is processed by S1 and S2, the feature values ​​corresponding to the preferred feature subset are extracted and input into the final diagnosis model, and the fault category diagnosis result is output.

[0142] This embodiment also provides a multi-feature fusion machine learning-based rotating machinery fault diagnosis system, including:

[0143] The signal acquisition and preprocessing module is used to implement step S1, namely: acquiring the signal of the rotating machinery and preprocessing the signal.

[0144] The multidimensional feature extraction module is used to implement step S2, namely: extracting time-domain statistical features and frequency-domain statistical features from the preprocessed signal to form an initial feature set.

[0145] The feature evaluation and screening module is used to implement step S3, namely: to evaluate the effectiveness of each feature in the initial feature set, and to screen out the preferred feature subset for fault classification based on the evaluation results.

[0146] The multi-model training and comparison module is used to implement step S4, namely: using a preferred feature subset, multiple different types of machine learning classification models are trained in parallel, and the diagnostic performance of each machine learning classification model is compared based on a preset evaluation system to obtain comparison results.

[0147] The optimal model deployment module is used to implement step S5, namely: based on the comparison results, select the machine learning classification model with the best diagnostic performance as the final diagnostic model.

[0148] The fault prediction and diagnosis module is used to implement step S6, namely: after the signal to be diagnosed is processed by the signal acquisition and preprocessing module and the multi-dimensional feature extraction module, the feature values ​​corresponding to the preferred feature subset are extracted and input into the final diagnosis model, and the fault category diagnosis result is output.

[0149] The role and effect of the embodiments

[0150] The multi-feature fusion machine learning-based rotating machinery fault diagnosis method and system according to the present invention have the following beneficial effects:

[0151] This invention introduces five adjustable correction coefficients during the time-domain and frequency-domain feature extraction process: a first correction coefficient α, a second correction coefficient β, a third correction coefficient γ, a fourth correction coefficient δ, and a fifth correction coefficient ε. These correction coefficients are not fixed but are flexibly adjusted according to the specific type and operating characteristics of the rotating machinery. The introduction of these five correction coefficients enables the feature extraction process to adapt to different types of machinery, operating conditions, and signal characteristics. These five correction coefficients are not only key parameters in the feature extraction process but also one of the important reasons why the method of this invention can be widely adapted to various types of rotating machinery (such as motors, fans, pumps, compressors, bearings, etc.). They allow this invention to achieve flexible adaptation to different equipment types, different fault modes, and different operating conditions through a small number of parameter adjustments without changing the core algorithm framework, significantly improving the versatility, robustness, and engineering application value of the method.

[0152] This invention quantifies the distinguishing ability of features for different fault categories by calculating the F-value of each feature relative to different fault categories through analysis of variance (ANOVA), and autonomously selects a preferred feature subset based on the ranking of F-values. This mechanism eliminates the need for manual feature selection based on expert experience, avoids interference from subjective bias and redundant features, and effectively improves the discriminative power and modeling efficiency of the preferred feature subset.

[0153] This invention constructs a progressive evaluation system of "security assessment—comprehensive performance ranking—critical faults—engineering optimization," balancing security and diagnostic accuracy: by setting a minimum average recall threshold, models with high false negative risk are prioritized for exclusion, ensuring no false negatives are missed; by setting independent single performance thresholds for critical fault categories, stricter access control is implemented for fault types with severe consequences; based on meeting security requirements, comprehensive performance is ranked according to preset evaluation indicators, and the model with the best diagnostic accuracy is selected; when performance is similar, machine learning classification models with simple structure, fast inference speed, and low resource consumption are prioritized, achieving synergistic optimization of diagnostic performance and engineering deployment feasibility. This evaluation system has good configurability and generalization ability, and can adapt to the security levels and performance requirements of different industrial scenarios.

[0154] This invention applies oversampling technology to the training set before training the machine learning classification model, effectively balancing the number of samples of different fault categories. This avoids the machine learning classification model from ignoring minority faults due to the dominance of majority class samples, improving the recall and F1 score for rare but critical faults, and enhancing the robustness of this invention in real industrial data environments.

[0155] This invention achieves data-driven autonomous decision-making throughout the entire process, from feature extraction and feature selection to model selection and deployment, without the need for manual intervention or repeated trial and error. The method of this invention is not dependent on specific equipment, operating conditions, or fault types, and can be easily transferred and applied to vibration monitoring and fault diagnosis tasks of various rotating machinery (such as motors, fans, pumps, compressors, bearings, etc.), demonstrating good versatility and promotional value.

[0156] In summary, because this invention adopts the aforementioned data-driven technical solution, it objectively selects a subset of highly discriminative features from high-dimensional features through an automated feature selection mechanism, and autonomously selects the diagnostic model with the best overall performance from multiple candidate models using a systematic model comparison strategy. Compared with existing technologies, this invention effectively reduces reliance on expert experience, automatically adapts to different equipment and operating conditions, improves the accuracy and robustness of fault diagnosis, and enhances the method's engineering applicability and promotion potential. It is suitable for vibration monitoring and intelligent fault early warning of various rotating machinery.

[0157] Those skilled in the art should understand that this invention is not limited to the above embodiments. The embodiments and descriptions in the specification are merely illustrative of the principles of the invention. Various changes and modifications can be made to this invention without departing from its spirit and scope, and all such changes and modifications fall within the scope of the invention as claimed. The scope of protection of this invention is defined by the appended claims and their equivalents.

Claims

1. A multi-feature fusion machine learning method for rotating machinery fault diagnosis, characterized in that, Includes the following steps: S1: Signal acquisition and preprocessing step, acquiring signals from the rotating machinery and preprocessing the signals; S2: Multidimensional feature extraction step, extracting time-domain statistical features and frequency-domain statistical features from the preprocessed signal to form an initial feature set; S3: Feature evaluation and screening step, evaluating the effectiveness of each feature in the initial feature set, and screening out the preferred feature subset for fault classification based on the evaluation results; S4: Multi-model training and comparison step: Using the preferred feature subset, multiple different types of machine learning classification models are trained in parallel, and the diagnostic performance of each machine learning classification model is compared based on a preset evaluation system to obtain comparison results; S5: Optimal model deployment steps: Based on the comparison results, select the machine learning classification model with the best diagnostic performance as the final diagnostic model; S6: Fault prediction and diagnosis step, after the signal to be diagnosed is processed by S1 and S2, the feature values ​​corresponding to the preferred feature subset are extracted and input into the final diagnosis model, and the fault category diagnosis result is output.

2. The rotating machinery fault diagnosis method based on multi-feature fusion machine learning according to claim 1, characterized in that: in, In S1, the signals include: vibration, rotational speed, load, and eccentricity. The preprocessing method is as follows: fill in missing values ​​in the signal. For each missing value, fill it with the average or adjacent value of the adjacent values ​​according to its position, and calculate the mean and standard deviation of the filled signal. Values ​​exceeding a preset multiple of the standard deviation are regarded as outliers and replaced with the average or adjacent value of the adjacent values. Then, the StandardScaler method is used to standardize all signal data.

3. The rotating machinery fault diagnosis method based on multi-feature fusion machine learning according to claim 1, characterized in that: in, In S2, the time-domain statistical features include: ， in, It is the number of signal sampling points. It is the first The signal value at the sampling point, The first correction factor is determined based on the rotational speed of the rotating machinery. ， in, It is the average value of the signal. This is the second correction factor, determined based on the rotational speed of the rotating machinery. ， in, It is the standard deviation of the signal. ， ， ， ， Frequency domain statistical characteristics include: ， in, is the number of positive frequencies, and k is the summation index. It is the signal at frequency The result of the fast Fourier transform at the given location. It is the amplitude value. This is the third correction factor, determined based on the distribution range of the fault characteristic frequencies. ， in, It is the first frequency domain eigenvalue. The fourth correction factor is determined based on the dispersion of the fault characteristic frequencies. ， in, This is the fifth correction factor, used to emphasize the fault characteristic frequency band.

4. The rotating machinery fault diagnosis method based on multi-feature fusion machine learning according to claim 3, characterized in that: in, First correction coefficient Calculated using the following formula: ， in, The rotational speed of the rotating machinery. Second correction coefficient Calculated using the following formula: ， The third correction coefficient Calculated using the following formula: ， in, The centroid frequency of the spectrum, The highest analysis frequency, The sensitivity coefficient, The fourth correction coefficient Calculated using the following formula: ， in, The standard deviation of the spectrum, The fifth correction coefficient Calculated using the following formula: ， in, For high-frequency energy, Low-frequency energy, This is an adjustment factor, with a value range of 0.01 to 0.

05.

5. The rotating machinery fault diagnosis method based on multi-feature fusion machine learning according to claim 1, characterized in that: in, In S3, the method for selecting the preferred feature subset is as follows: calculate the variance analysis F value of each feature relative to different fault categories to evaluate the ability of each feature to distinguish fault categories, sort them from high to low according to the F value, and select a preset number of features with the highest ranking to form the preferred feature subset.

6. The rotating machinery fault diagnosis method based on multi-feature fusion machine learning according to claim 1, characterized in that: in, In S4, various types of machine learning classification models include: random forest, support vector machine, gradient boosting tree family model, and multilayer perceptron neural network. S4 specifically includes the following sub-steps: S4-1: Based on the preferred feature subset, extract the corresponding feature data from the labeled signal feature dataset to form an effective dataset for model training. The labeled signal feature dataset is generated by calculating historical signals of known fault categories based on the initial feature set. S4-2: Divide the effective dataset into a training set, a validation set, and an independent test set; S4-3: Use the training set to train multiple different types of machine learning classification models in parallel, and use the validation set to perform hyperparameter tuning and performance evaluation on each machine learning classification model; S4-4: Based on the preset evaluation system, compare the diagnostic performance of each machine learning classification model and obtain the comparison results.

7. The rotating machinery fault diagnosis method based on multi-feature fusion machine learning according to claim 1, Its features are: In step S4, the preset evaluation system is implemented according to the following rules: Safety filtering: A minimum average recall threshold is set as a safety requirement. Any machine learning classification model whose average recall is lower than this threshold will be excluded. Critical Fault Constraints: For one or more preset critical fault categories, set their own independent minimum performance thresholds; for any machine learning classification model, if its evaluation index on the critical fault category is lower than the corresponding minimum performance threshold, it will be directly excluded or downgraded in the overall performance ranking. Overall performance ranking: For machine learning classification models that meet the safety requirements, overall performance ranking is performed based on preset evaluation metrics, including: accuracy, precision, recall, and F1 score. Engineering Optimization: Based on the overall performance ranking results, select the machine learning classification model with the best overall performance ranking. If the overall performance rankings of multiple machine learning classification models are the same or the differences are within the preset allowable error range, then the machine learning classification model with a simpler model structure, faster inference speed, or less computational resources required will be given priority.

8. A rotating machinery fault diagnosis system based on multi-feature fusion machine learning, characterized in that, include: The signal acquisition and preprocessing module acquires signals from the rotating machinery and preprocesses the signals. The multi-dimensional feature extraction module extracts time-domain statistical features and frequency-domain statistical features from the preprocessed signal to form an initial feature set. The feature evaluation and screening module evaluates the effectiveness of each feature in the initial feature set and selects a preferred feature subset for fault classification based on the evaluation results. The multi-model training and comparison module uses the preferred feature subset to train multiple different types of machine learning classification models in parallel, and compares the diagnostic performance of each machine learning classification model based on a preset evaluation system to obtain comparison results. The optimal model deployment module selects the machine learning classification model with the best diagnostic performance as the final diagnostic model based on the comparison results. The fault prediction and diagnosis module processes the signal to be diagnosed through the signal acquisition and preprocessing module and the multi-dimensional feature extraction module, extracts the feature values ​​corresponding to the preferred feature subset, inputs them into the final diagnosis model, and outputs the fault category diagnosis result.

Images