Early Anomaly Detection Method for Rotating Machinery Based on Gram and Gray Matrix Co-generation

Abstract

This invention discloses an early anomaly detection method for rotating machinery based on Gram and grayscale matrix co-occurrence. It transforms a one-dimensional time-series signal into a two-dimensional image signal using Gram angle field analysis, preserving the signal's time dependence while enhancing its feature representation capability. Multi-dimensional texture features are extracted from the reconstructed image signal through grayscale co-occurrence matrix analysis to meticulously characterize abnormal patterns within the signal. The system integrates a dynamic threshold calculation and continuous alarm mechanism that adaptively adjusts based on real-time extracted feature values. An alarm is triggered immediately when the feature value exceeds the calculated threshold, enabling highly sensitive monitoring of equipment status. Furthermore, the system is equipped with a data preprocessing module and a feature fusion module, responsible for denoising and filtering the vibration signal and performing principal component analysis for dimensionality reduction, respectively, ensuring signal quality and feature validity. This invention demonstrates excellent anomaly detection accuracy and reliability in high-noise and heavy-load environments.

Description

Technical Field

[0001] This invention belongs to the field of mechanical equipment maintenance, and in particular relates to an early anomaly detection method for rotating machinery based on the co-generation of Gram and grayscale matrix. Background Technology

[0002] With the increasing complexity of industrial equipment, early fault detection of heavy-duty rotating machinery has become particularly critical in industrial applications. This type of equipment typically operates in high-load and high-intensity environments, such as power plants, petrochemical plants, mining industries, and marine engineering projects. Failures in these environments often result in severe economic losses and safety risks. Therefore, early fault detection of heavy-duty rotating machinery has become a key research area for industrial maintenance and safety assurance. Current fault detection methods can be mainly categorized into three types: knowledge-driven, model-driven, and data-driven. However, these methods have limitations in their effectiveness under complex operating conditions and variable environments.

[0003] Knowledge-driven approaches rely on expert knowledge and pre-defined fault rule bases to determine known fault modes of equipment based on expert experience. While these methods can be effective under simple operating conditions, the coverage and effectiveness of expert knowledge are limited for the complex structures and diverse fault modes of heavy rotating machinery. Especially with new equipment, experts' insufficient understanding of its internal working principles and potential fault modes poses challenges to the accuracy and reliability of knowledge-driven approaches. Furthermore, the reliance on manually defined rules and knowledge base construction makes rule updates and maintenance particularly difficult under complex operating conditions, hindering adaptation to dynamic changes in equipment status.

[0004] Model-driven methods are typically based on physical models and attempt to detect faults by describing the normal and abnormal behavior of equipment. However, for heavy rotating machinery, the establishment of physical models has significant limitations. Since such equipment usually operates under unpredictable load fluctuations and environmental changes, a single physical model can hardly comprehensively describe the normal and abnormal states of the equipment. Furthermore, the complex model calculations make real-time detection difficult to achieve in actual operation. Therefore, model-driven methods have poor adaptability and real-time performance under complex operating conditions, limiting their widespread application in heavy rotating machinery.

[0005] Data-driven approaches rely on historical equipment data for fault detection through signal analysis and pattern recognition, without depending on expert knowledge or physical models. While these methods show promise for complex systems, they face challenges in practical applications, such as missing or incomplete data, especially since monitoring data for heavy rotating machinery often exhibits non-uniformity under various operating conditions. Furthermore, data-driven methods depend on large amounts of labeled data, which is often lacking in industrial settings for novel fault modes, limiting the model's generalization ability and making it difficult to address new types of faults.

[0006] The application of signal processing technology in fault detection of heavy rotating machinery is gaining increasing attention. Traditional time-domain, frequency-domain, and time-frequency-domain analysis methods, such as Short-Time Fourier Transform (STFT), wavelet transform, and envelope analysis, can extract fault features from vibration signals. However, with the increasing complexity of equipment operating environments, traditional signal processing methods are insufficient when facing variable operating conditions, especially in capturing hidden early fault features. Therefore, there is an urgent need for an early anomaly detection method for rotating machinery to achieve more accurate fault warnings. Summary of the Invention

[0007] Purpose of the Invention: The purpose of this invention is to provide an early anomaly detection method for rotating machinery based on the co-generation of Gram and grayscale matrices. This method can adapt to complex operating conditions and achieve efficient and accurate fault early warning. Based on advanced signal processing technology, this invention overcomes the limitations of existing methods through multi-step feature extraction, feature fusion, dynamic threshold calculation, and an adaptive alarm mechanism, achieving highly sensitive and reliable fault detection.

[0008] Technical solution: The present invention provides an early anomaly detection method for rotating machinery based on the co-occurrence of Gram and grayscale matrices, comprising the following steps:

[0009] Step S1: Collect the original vibration signal of the target machine and perform segmented aggregation approximation processing to obtain a one-dimensional time series vibration signal. Use Gram angle field (GAF) to reconstruct the one-dimensional time series vibration signal to convert the one-dimensional signal into a two-dimensional image signal and obtain a GAF image.

[0010] Step S2: Perform texture analysis on the GAF image using the gray-level co-occurrence matrix to extract texture features, including contrast, second angular moment, entropy, inverse difference moment and correlation features, in order to capture subtle texture changes in the image and achieve multi-dimensional feature representation. The texture features are sorted by monotonicity using the Spearman correlation coefficient and dimensionality is reduced using PCA to obtain the processed texture features.

[0011] Step S3: Perform dynamic alarm threshold calculation and alarm triggering mechanism design. Based on the real-time texture features extracted in step S2, use statistical characteristics to calculate the dynamic threshold for anomaly detection, improve the real-time performance of alarm response, and achieve early warning of target mechanical failure.

[0012] Furthermore, in step S1, the segmented aggregation approximation process specifically involves: dividing the original signal into segments and calculating the average value of each segment through segmented aggregation approximation to remove signal redundancy and improve the computational efficiency of subsequent feature extraction. The calculation formula for the segmented aggregation approximation process is as follows:

[0013]

[0014] in, Where N is the segment length and N is the time series length. The original signal value, This is a one-dimensional time series signal obtained after segmentation and aggregation approximation.

[0015] Furthermore, in step S1, the reconstruction of the one-dimensional spatial sequence vibration signal using Gram angle field (GAF) specifically involves: for a one-dimensional time sequence signal... The formula for calculating the Gram's angle field (GAF) is as follows:

[0016]

[0017] in, and These are the results of standardizing the data points at times i and j in the time series and mapping them to polar coordinates, calculated using the following formulas:

[0018]

[0019] The generated two-dimensional image signal retains the time dependence of the original signal and enhances its feature representation ability.

[0020] Furthermore, in step S2, the step of performing texture analysis on the GAF image using the gray-level co-occurrence matrix to extract texture features specifically involves: performing texture analysis on the GAF image using the gray-level co-occurrence matrix to extract key features such as contrast, second angular moment, entropy, inverse difference moment, and correlation, in order to capture subtle texture changes in the image and achieve multi-dimensional feature representation. The specific calculation formula is as follows:

[0021] Contrast ratio is defined as follows:

[0022]

[0023] The second moment of an angle is defined as follows:

[0024]

[0025] Entropy is defined as follows:

[0026]

[0027] The inverse moment is defined as follows:

[0028]

[0029] The correlation is defined as follows:

[0030]

[0031] in, These are the element values ​​of the gray-level co-occurrence matrix. and These represent the mean and standard deviation of the corresponding rows, respectively. Texture features describe the local symmetry, detail, and smoothness information of the image, and identify abnormal patterns in mechanical vibration signals.

[0032] Furthermore, the calculation of the gray-level co-occurrence matrix involves the analysis of multi-directional texture features, specifically including the extraction of primary information in the 135° direction and the extraction of secondary information in the 0° or 90° direction, so as to achieve a comprehensive analysis of multi-directional texture features and thereby improve the reliability of abnormal pattern recognition.

[0033] Furthermore, in step S3, the formula for calculating the dynamic alarm threshold is as follows:

[0034]

[0035] in, and Let represent the mean and standard deviation of the real-time feature, respectively. This represents the adjustment coefficient, and the alarm mechanism activates when it detects a feature value exceeding a dynamic threshold. It can trigger an alarm signal in time to achieve early warning of mechanical failure.

[0036] Furthermore, in step S3, the alarm triggering mechanism design includes the following steps:

[0037] Step S3.1: Smooth the extracted feature values ​​through multiple rounds of first-order exponential smoothing to reduce the impact of signal noise on the feature values ​​and improve the stability of feature extraction. The calculation formula is as follows:

[0038]

[0039] in, The current smoothing value, For the observed values, For smoothing coefficients;

[0040] Step S3.2: Perform Z-score normalization on the smoothed features. The normalization formula is as follows:

[0041]

[0042] Where x is the eigenvalue, The mean of the eigenvalues, Z-score normalization, which uses the standard deviation of the features, ensures comparability between different features;

[0043] Step S3.3: When the normalized fusion feature value exceeds the preset warning threshold, an abnormal alarm signal is triggered, indicating a potential equipment failure risk.

[0044] The present invention also discloses a system for implementing an early anomaly detection method for rotating machinery, comprising a signal reconstruction module, a texture analysis module, a feature fusion module, and an alarm module;

[0045] The signal reconstruction module is used to receive mechanical vibration signals and convert the time series signals into two-dimensional images through GAF to form image signals that can represent vibration characteristics.

[0046] The texture analysis module is used to perform gray-level co-occurrence matrix analysis to extract texture features of image contrast, second angular moment, entropy, inverse difference moment and correlation, thereby identifying abnormal signal features;

[0047] The feature fusion module sorts the texture features by monotonicity using the Spearman correlation coefficient and uses PCA for dimensionality reduction to enhance the representativeness of the features and reduce redundant information.

[0048] The alarm module is used to calculate the real-time vibration threshold based on the fused features and generate an alarm signal based on the normalized vibration sequence. The alarm module has the function of analyzing the main and secondary features to ensure the real-time triggering of the alarm signal. The alarm module has the ability to adjust the threshold automatically according to different working conditions of the equipment to improve the sensitivity and accuracy of the system's anomaly detection, and reduce the probability of false alarms and missed alarms through the adaptive threshold mechanism.

[0049] Furthermore, the system further includes a data preprocessing module, which is used to filter and denoise the input vibration signal to ensure signal quality and provide optimized input for feature extraction.

[0050] Furthermore, the system is applied to the early anomaly detection of high-speed, heavy-duty mechanical equipment, and is especially suitable for fault early warning of complex mechanical components such as rolling bearings and gearboxes. The system maintains detection accuracy under high noise and high load conditions to prevent potential mechanical failures.

[0051] Beneficial effects: Compared with the prior art, the present invention has the following significant advantages:

[0052] 1. Enhanced feature representation capability: Through GAF reconstruction, one-dimensional vibration signals are transformed into two-dimensional images, effectively preserving the time dependence of the signals, making early fault features more prominent in the images, and improving the ability to represent fault features.

[0053] 2. Multi-dimensional texture feature extraction: Based on the gray-level co-occurrence matrix, various texture features are extracted, including contrast, second angular moment, entropy, inverse difference moment and correlation, which can comprehensively characterize the small anomalies in the vibration signal and significantly improve the sensitivity and accuracy of detection.

[0054] 3. Efficient feature fusion and dimensionality reduction: Features are screened and dimensionality reduced by using Spearman correlation coefficient and principal component analysis (PCA), which effectively removes redundant information, simplifies calculation, improves detection efficiency, and ensures the preservation of key fault features.

[0055] 4. Adaptive dynamic threshold calculation: The adaptive dynamic threshold calculation mechanism designed in this invention can adjust the threshold according to changes in real-time characteristics, thereby adapting to changing working conditions and improving the accuracy and real-time performance of fault detection in complex environments.

[0056] 5. Noise suppression and signal stability improvement: Through data preprocessing and first-order exponential smoothing, noise interference in the signal is effectively reduced, and Z-score standardization ensures the comparability between different features, further improving the stability of fault detection.

[0057] Overall, this invention enables accurate detection and rapid response to early faults in heavy rotating machinery in complex industrial environments with high noise and variable operating conditions, significantly improving the reliability and applicability of the system and providing strong technical support for the intelligent maintenance of heavy rotating machinery. Attached Figure Description

[0058] Figure 1 This is the test bench for the heavy rotating machinery - large roller press - in this embodiment;

[0059] Figure 2 This is a graph showing the change in operating parameters verified by experiments in this embodiment of the invention;

[0060] Figure 3 This is a flowchart of the process of the present invention;

[0061] Figure 4 This is a conceptual diagram of the workflow of the present invention;

[0062] Figure 5 This is a GAF reconstruction fault image of heavy rotating machinery in this invention;

[0063] Figure 6 This is the pseudocode for the continuous early warning mechanism in this invention;

[0064] Figure 7 This is a diagram showing the results of early abnormality detection of a large rolling mill in the example. Detailed Implementation

[0065] The technical solution of the present invention will be further described below with reference to the accompanying drawings.

[0066] The test data used in this example comes from the bearing life cycle test completed by Nanjing University of Aeronautics and Astronautics and Nanjing High Speed ​​Gear Co., Ltd. on a large rolling mill test bench. Figure 1 The diagram shows the site layout of the heavy-duty rotating machinery—large roller press test bench—used in the experiment, and some parameters are marked in detail. Figure 2 The figure shows the curves of the operating parameters as verified by the experiment.

[0067] like Figure 3 As shown, this invention presents a real-time detection method for early anomalies in heavy rotating machinery based on the co-generation of Gram and grayscale matrices. A more illustrative conceptual diagram is shown below. Figure 4 As shown. Specifically, it includes the following steps:

[0068] Step S1: Acquire vibration data of heavy rotating machinery and complete signal reconstruction;

[0069] First, the acquired one-dimensional time series signal of mechanical vibration was analyzed. Standardize the data to ensure it falls within the range [-1, 1]. The standardization formula is as follows:

[0070]

[0071] Then, the standardized signal Mapped to polar coordinates angle The Gram angle field is used to generate two-dimensional image signals, and the angle... The calculation formula is as follows:

[0072]

[0073] The two-dimensional image signal generated through the above steps retains the temporal dependence of the original signal and enhances its feature representation ability, laying the foundation for subsequent texture feature analysis. The reconstructed GAF ​​fault image is shown below. Figure 5 As shown.

[0074] The time series signal undergoes segmented aggregation approximation processing before reconstruction. This processing divides the original signal into segments and calculates the average value of each segment to effectively remove signal redundancy and improve the computational efficiency of subsequent feature extraction. The calculation formula for the segmented aggregation approximation processing is as follows:

[0075]

[0076] in, Where N is the segment length and N is the time series length. The original signal value, This is the new signal obtained after segmented aggregation approximation processing.

[0077] Step S2: Perform texture analysis on the GAF image using the gray-level co-occurrence matrix to extract key features such as contrast, second angular moment, entropy, inverse difference moment, and correlation, in order to capture subtle texture changes in the image and achieve multi-dimensional feature representation. The specific calculation formula is as follows:

[0078] Contrast ratio is defined as follows:

[0079]

[0080] The second moment of an angle is defined as follows:

[0081]

[0082] Entropy is defined as follows:

[0083]

[0084] The inverse moment is defined as follows:

[0085]

[0086] The correlation is defined as follows:

[0087]

[0088] in, These are the element values ​​of the gray-level co-occurrence matrix. and These represent the mean and standard deviation of the corresponding rows. The texture features describe information such as the local symmetry, detail, and smoothness of the image, effectively identifying abnormal patterns in mechanical vibration signals. The calculation of the gray-level co-occurrence matrix involves the analysis of multi-directional texture features, specifically including the extraction of primary information in the 135° direction and the extraction of secondary information in the 0° or 90° direction, to achieve a comprehensive analysis of multi-directional texture features, thereby improving the reliability of abnormal pattern recognition.

[0089] Step S3: Calculate the dynamic alarm threshold and design the alarm triggering mechanism. Based on the real-time texture features extracted in Step S2, calculate the dynamic threshold for anomaly detection using statistical properties to improve the real-time performance of the alarm response. The dynamic threshold calculation formula is as follows:

[0090]

[0091] in, and Let represent the mean and standard deviation of the real-time feature, respectively. This represents the adjustment coefficient. The alarm mechanism detects that the feature value exceeds the dynamic threshold. An alarm signal is triggered in a timely manner to provide early warning of mechanical failures. The pseudocode of the early warning mechanism algorithm is as follows: Figure 6 As shown.

[0092] Step S4: Smooth the extracted feature values ​​through multiple rounds of first-order exponential smoothing to reduce the impact of signal noise on the feature values ​​and improve the stability of feature extraction. The calculation formula is as follows:

[0093]

[0094] in, The current smoothing value, For the observed values, This is the smoothing coefficient.

[0095] Step S5: Perform Z-score normalization on the smoothed features. The normalization formula is as follows:

[0096]

[0097] Where x is the eigenvalue, The mean of the eigenvalues, Z-score normalization, where Z is the standard deviation of the feature, ensures comparability between different features.

[0098] Step S6: When the normalized fusion feature value exceeds the preset warning threshold, an abnormal alarm signal is triggered, indicating a potential equipment failure risk.

[0099] Step S7: Signal reconstruction module establishment: used to receive mechanical vibration signals and convert time series signals into two-dimensional images through GAF to form image signals that can represent vibration characteristics.

[0100] Step S8, Texture Analysis Module: Used to perform gray-level co-occurrence matrix analysis to extract texture features such as contrast, second moment of angle, entropy, inverse difference moment and correlation of the image, thereby identifying abnormal signal features.

[0101] Step S9: To optimize the feature space, the feature fusion module performs monotonicity ranking of texture features using the Spearman correlation coefficient and dimensionality reduction using PCA to enhance feature representativeness and reduce redundant information. The specific formula for calculating the Spearman correlation coefficient is as follows:

[0102]

[0103] Step S10, Alarm Module: This module calculates the real-time vibration threshold based on the fused features and generates an alarm signal based on the normalized vibration sequence. The alarm module has the ability to analyze both primary and secondary features to ensure real-time triggering of the alarm signal.

[0104] The system further includes a data preprocessing module, which filters and denoises the input vibration signal to ensure signal quality and provide optimized input for feature extraction. The final output is as follows: Figure 7 As shown.

[0105] In summary, this invention provides a method and system for early fault detection in heavy rotating machinery. Through signal reconstruction, texture feature extraction, feature fusion and dimensionality reduction, adaptive threshold calculation, and an alarm mechanism, it achieves multi-dimensional analysis of vibration signals, overcoming the limitations of traditional detection methods under complex operating conditions. This invention can effectively detect early faults in high-noise and variable environments, improving detection accuracy and real-time performance, and providing reliable technical support for the intelligent maintenance of heavy rotating machinery.

Claims

1. A method for early anomaly detection in rotating machinery based on the co-generation of Gram and grayscale matrices, characterized in that, Includes the following steps: Step S1: Collect the original vibration signal of the target machine and perform segmented aggregation approximation processing to obtain a one-dimensional time series vibration signal. Use Gram angle field (GAF) to reconstruct the one-dimensional time series vibration signal to convert the one-dimensional signal into a two-dimensional image signal and obtain a GAF image. Step S2: Perform texture analysis on the GAF image using the gray-level co-occurrence matrix to extract texture features, including contrast, second angular moment, entropy, inverse difference moment and correlation features. Sort the texture features by monotonicity using Spearman correlation coefficient and perform dimensionality reduction using PCA to obtain the processed texture features. Step S3: Perform dynamic alarm threshold calculation and alarm triggering mechanism design. Based on the texture features processed in step S2, use statistical characteristics to calculate the dynamic threshold for anomaly detection, improve the real-time performance of alarm response, and achieve early warning of target mechanical failure. In step S3, the alarm triggering mechanism design includes the following steps: Step S3.1: Smooth the extracted feature values ​​through multiple rounds of first-order exponential smoothing to reduce the impact of signal noise on the feature values ​​and improve the stability of feature extraction. The calculation formula is as follows: ； in, The current smoothing value, For the observed values, For smoothing coefficients; Step S3.2: Perform Z-score normalization on the smoothed features. The normalization formula is as follows: ； Where x is the eigenvalue, The mean of the eigenvalues, Z-score normalization, which uses the standard deviation of the features, ensures comparability between different features; Step S3.3: When the normalized fusion feature value exceeds the preset warning threshold, an abnormal alarm signal is triggered, indicating a potential equipment failure risk.

2. The method for early anomaly detection of rotating machinery based on Gram and grayscale matrix co-occurrence as described in claim 1, characterized in that, In step S1, the segmented aggregation approximation process specifically involves: dividing the original signal into segments and calculating the average value of each segment to remove signal redundancy and improve the computational efficiency of subsequent feature extraction. The calculation formula for the segmented aggregation approximation process is as follows: ； in, Where N is the segment length and N is the time series length. The original signal value, This is a one-dimensional time series signal obtained after segmentation and aggregation approximation.

3. The method for early anomaly detection of rotating machinery based on Gram and grayscale matrix co-occurrence as described in claim 1, characterized in that, In step S1, the reconstruction of the one-dimensional spatial sequence vibration signal using Gram angle field (GAF) specifically involves: for a one-dimensional time series signal... The formula for calculating the Gram's angle field (GAF) is as follows: ； in, and These are the results of standardizing the data points at times i and j in the time series and mapping them to polar coordinates, calculated using the following formulas: ； The generated two-dimensional image signal retains the time dependence of the original signal and enhances its feature representation ability.

4. The method for early anomaly detection of rotating machinery based on Gram and grayscale matrix co-occurrence as described in claim 1, characterized in that, In step S2, the step of performing texture analysis on the GAF image using the gray-level co-occurrence matrix and extracting texture features specifically involves: performing texture analysis on the GAF image using the gray-level co-occurrence matrix to extract key features such as contrast, second angular moment, entropy, inverse difference moment, and correlation, in order to capture subtle texture changes in the image and achieve multi-dimensional feature representation. The specific calculation formula is as follows: Contrast ratio is defined as follows: ； The second moment of an angle is defined as follows: ； Entropy is defined as follows: ； The inverse moment is defined as follows: ； The correlation is defined as follows: ； in, These are the element values ​​of the gray-level co-occurrence matrix. and These represent the mean and standard deviation of the corresponding rows, respectively. Texture features describe the local symmetry, detail, and smoothness information of the image, and identify abnormal patterns in mechanical vibration signals.

5. A method for early anomaly detection of rotating machinery based on Gram and grayscale matrix co-occurrence, as described in claim 1 or 4, characterized in that, The calculation of the gray-level co-occurrence matrix involves the analysis of multi-directional texture features, specifically including the extraction of primary information in the 135° direction and the extraction of secondary information in the 0° or 90° direction.

6. A system for implementing the early anomaly detection method for rotating machinery as described in claim 1, characterized in that, It includes a signal reconstruction module, a texture analysis module, a feature fusion module, and an alarm module; The signal reconstruction module is used to receive mechanical vibration signals and convert the time series signals into two-dimensional images through GAF to form image signals that can represent vibration characteristics. The texture analysis module is used to perform gray-level co-occurrence matrix analysis to extract texture features of image contrast, second angular moment, entropy, inverse difference moment and correlation, thereby identifying abnormal signal features; The feature fusion module sorts the texture features by monotonicity using the Spearman correlation coefficient and uses PCA for dimensionality reduction to enhance the representativeness of the features and reduce redundant information. The alarm module is used to calculate the real-time vibration threshold based on the fused features and generate an alarm signal based on the normalized vibration sequence. The alarm module has the function of analyzing major and minor features to ensure the real-time triggering of the alarm signal. The alarm module has the ability to adaptively adjust the threshold, which can automatically adjust the alarm threshold according to different operating conditions of the equipment to improve the sensitivity and accuracy of the system's anomaly detection, and reduce the probability of false alarms and missed alarms through the adaptive threshold mechanism. The system further includes a data preprocessing module, which is used to filter and denoise the input vibration signal to ensure the signal quality and provide optimized input for feature extraction.

7. The system according to claim 6, characterized in that, The system is used for early anomaly detection in high-speed, heavy-duty mechanical equipment.

8. The system according to claim 7, characterized in that, The system is particularly suitable for fault early warning of complex mechanical components such as rolling bearings and gearboxes. The system maintains detection accuracy under high noise and high load conditions to prevent potential mechanical failures.

Images