Equipment health monitoring method and system based on multivariate information fusion

Abstract

The invention relates to an equipment health monitoring method and system based on multivariate information fusion. The method comprises the following steps: firstly, collecting a multi-element vibration signal from an equipment sensor, and segmenting the multi-element vibration signal to obtain segmented signals; calculating a multi-dimensional characteristic value of each segment of signal in the segmented signals, and integrating the characteristic values to obtain a characteristic matrix; then, inputting the feature matrix into a sparse representation model, constructing a health index fitting function, performing recursive optimization of a residual error, and outputting to obtain a health index fitting result; and finally, performing equipment state evaluation by utilizing a health index fitting result, and monitoring and outputting an equipment abnormal point. Compared with the prior art, the method has the advantages of improving reliability and accuracy, enhancing tendency and monotonicity of indexes, comprehensively extracting multi-dimensional information in vibration signals, improving reliability and sensitivity of anomaly detection and the like.

Description

Technical Field

[0001] The present invention relates to the technical field of equipment monitoring, and in particular to an equipment health monitoring method and system based on multi-information fusion. Background Art

[0002] Mechanical equipment health monitoring is an important means to ensure production safety, improve equipment operation efficiency, and reduce maintenance costs. Since the equipment is subject to complex working conditions, vibration, friction, corrosion and other factors during operation, its condition will gradually deteriorate. If potential problems are not discovered in time, it may lead to sudden failures, which will not only affect production progress, but also may bring serious safety hazards and economic losses.

[0003] Fusion of multivariate information of equipment and construction of health indicators is an important part of equipment health monitoring. In terms of the construction of health indicators, the commonly used methods are dimensionality reduction methods, statistical methods based on single features, and data-driven methods. These methods have certain limitations in practical engineering applications. Dimensionality reduction-based methods such as principal component analysis (PCA) often assume that the main features of the data can be represented by linear combinations, which makes it difficult to capture nonlinear failure modes. At the same time, PCA is sensitive to noise and may affect the stability of the indicators. In addition, PCA cannot be updated dynamically, which leads to certain limitations in the application of PCA in real-time monitoring. When using statistical methods with single features such as mean, variance or peak to construct health indicators, although they can reflect certain characteristics of vibration signals, they lack the ability to characterize the comprehensive characteristics of multidimensional signals and are easily affected by single noise or outliers, resulting in insufficient robustness of health status assessment. Data-driven methods using neural networks often rely on a large amount of data for training and lack physical interpretation. Summary of the invention

[0004] The purpose of the present invention is to overcome the defects of the above-mentioned prior art and to provide a device health monitoring method and system based on multi-information fusion.

[0005] The purpose of the present invention can be achieved by the following technical solutions:

[0006] According to one aspect of the present invention, a device health monitoring method based on multivariate information fusion is provided, the method comprising the following steps:

[0007] S1, collecting a multivariate vibration signal from a device sensor and segmenting it to obtain a segmented signal;

[0008] S2, calculating the multi-dimensional eigenvalues ​​of each segment of the segmented signal, and integrating the eigenvalues ​​to obtain a feature matrix;

[0009] S3, input the feature matrix into the sparse representation model, construct the health index fitting function, recursively optimize the residual, and output the health index fitting result;

[0010] S4. Use the health indicator fitting results to evaluate the equipment status, monitor and output equipment abnormal points.

[0011] As a preferred technical solution, after the signal in S1 is segmented, its validity is judged by the mean and standard deviation of the segmented signal. If the mean and standard deviation of the signal meet the preset threshold requirements, the segmented signal is judged to be valid; otherwise, the segmented signal is discarded. The specific formula for judging the validity is:

[0012]

[0013] In the formula, μ x is the mean of the signal segment; σ x is the variance of the new number segment; mean_threshold is the preset signal mean threshold; std_threshold is the preset signal standard deviation threshold.

[0014] As a preferred technical solution, the multi-dimensionality in S2 is ten dimensions, including time domain features, frequency domain features and time-frequency domain features; among them, the time domain features include root mean square value, variance, peak value, skewness value and kurtosis value; the frequency domain features include peak-to-peak value, spectral skewness and spectral kurtosis; the time-frequency domain features include spectral power and wavelet packet energy.

[0015] As a preferred technical solution, after integrating the feature values ​​in S2 to obtain the feature matrix, the feature matrix is ​​standardized by column, and the specific processing formula is:

[0016]

[0017] In the formula, is the column vector of the characteristic matrix after normalization; is the mean of the i-th feature; is the variance of the i-th feature; t i is the column vector of the feature matrix.

[0018] As a preferred technical solution, the health index fitting function in S3 is:

[0019] H=T norm β

[0020] Where, T norm is the standardized feature matrix; β is the sparse coefficient.

[0021] As a preferred technical solution, the specific process of recursively optimizing the residual in S3 is: first initialize the parameters, then iterate, and when the iteration reaches a preset number of times, end the iteration and calculate the health indicator fitting result.

[0022] As a preferred technical solution, in each iteration, the parameters are first used to select the feature with the greatest correlation from the feature matrix, and then the parameters are calculated and updated; wherein the specific formula for selecting the feature with the greatest correlation and calculating the parameters is:

[0023]

[0024] S=S∪{i *}

[0025]

[0026] r k+1 =H true -T S β S

[0027] In the formula, i * is the residual r in the current iteration k The index of the most relevant feature; is the transpose of the i-th column of the feature matrix T; r k is the residual; S is the feature subset; β s is a sparse coefficient vector, including only the non-zero weights of the selected features; H true is the label value; T S is the currently selected feature subset.

[0028] As a preferred technical solution, when using the health index fitting results to evaluate the equipment status, the health index fitting results are firstly smoothed by sliding average, and then the monotonicity, trend and robustness are calculated based on them. The specific formula for smoothing the health index fitting results by sliding average is:

[0029]

[0030] Where i is the current index position; ω is the window length, H fit (j) is the fitting result of health indicators.

[0031] As a preferred technical solution, S4 monitors and outputs equipment abnormal points based on the 3σ principle of mean and variance. The specific formula is:

[0032] Anomaly={|i||H fit (i) -μ H |>3σ H}

[0033] In the formula, Anomaly is the abnormal point, H fit (i) is the health index fitting result; μ H is the mean value of health index; σ H is the variance of the health indicator.

[0034] According to another aspect of the present invention, there is provided an equipment health monitoring system based on multivariate information fusion, the system is operated by applying the equipment health monitoring method based on multivariate information fusion as described above, the system comprises an information acquisition module, a feature extraction module, a health index fitting module and an equipment status evaluation module;

[0035] Among them, the information acquisition module is used to collect multivariate vibration signals from equipment sensors and segment them to obtain segmented signals; the feature extraction module is used to calculate the multi-dimensional eigenvalues ​​of each signal in the segmented signal, and integrate the eigenvalues ​​to obtain the feature matrix; the health indicator fitting module is used to input the feature matrix into the sparse representation model, construct the health indicator fitting function, and recursively optimize the residual to output the health indicator fitting result; the equipment status evaluation module is used to use the health indicator fitting result to evaluate the equipment status, monitor and output equipment abnormal points.

[0036] Compared with the prior art, the present invention has the following beneficial effects:

[0037] 1. In the present invention, multivariate vibration signals are first collected from the equipment sensor and segmented to obtain segmented signals; then the multidimensional eigenvalues ​​of each segment of the segmented signals are calculated, and the eigenvalues ​​are integrated to obtain the feature matrix; then, the feature matrix is ​​input into the sparse representation model, a health index fitting function is constructed, and the residual is recursively optimized to output the health index fitting result; finally, the health index fitting result is used to evaluate the equipment status, monitor and output the equipment abnormal points. The present invention efficiently integrates multidimensional features, improves the interpretability of health indicators, and enhances the reliability and accuracy of health status monitoring.

[0038] 2. In the present invention, the feature matrix is ​​input into the sparse representation model, a health index fitting function is constructed, and the residual is recursively optimized. In each iteration, the parameters are first used to select the feature with the greatest correlation from the feature matrix, and then the parameters are calculated and updated. Finally, the iteration ends and the health index fitting result is output. The health index is fitted by sparse optimization, which enhances the trend and monotonicity of the index, accurately reflects the health status of the equipment, and improves the accuracy of equipment monitoring.

[0039] 3. The present invention calculates characteristic values ​​in multiple dimensions, and the multiple dimensions are ten dimensions, including time domain features, frequency domain features and time-frequency domain features; among them, the time domain features include root mean square value, variance, peak value, skewness value and kurtosis value; the frequency domain features include peak-to-peak value, spectral skewness and spectral kurtosis; the time-frequency domain features include spectral power and wavelet packet energy. By combining the time domain, frequency domain and time-frequency domain features to construct a sparse dictionary, the multidimensional information in the vibration signal can be fully extracted.

[0040] 4. The present invention determines whether a signal is valid based on whether the mean and standard deviation of the signal meet the preset threshold requirements, and continuously calculates and updates relevant parameters during the recursive optimization of the residual. The present invention combines dynamic thresholds and real-time update mechanisms, so that it can efficiently and accurately detect abnormal conditions in equipment operation, significantly improving the reliability and sensitivity of anomaly detection.

[0041] 5. The equipment health monitoring method based on multivariate information fusion in the present invention can complete data input only by collecting multivariate vibration signals from equipment sensors, and the rest of the calculation and monitoring process does not require a network, so that the method supports offline batch processing and online real-time expansion, and has good flexibility and practicality. BRIEF DESCRIPTION OF THE DRAWINGS

[0042] Figure 1 A schematic diagram of the steps of a multivariate information fusion method in the present invention;

[0043] Figure 2 is the original signal diagram in the embodiment;

[0044] Figure 3 This is a signal diagram after the preprocessing in the embodiment removes the non-powered data;

[0045] Figure 4a Schematic diagram of the signal root mean square value in the embodiment;

[0046] Figure 4b A schematic diagram of the signal spectrum skewness value in the embodiment;

[0047] Figure 5 A health indicator graph constructed by the method proposed for use in the embodiments;

[0048] Figure 6 Schematic diagram of the comparison results between the present method and other methods in the embodiment;

[0049] Figure 7 This is a schematic diagram of the results of abnormal monitoring using the method in the embodiment. DETAILED DESCRIPTION

[0050] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work should fall within the scope of protection of the present invention.

[0051] Mechanical equipment health monitoring is an important means to ensure production safety, improve equipment operation efficiency, and reduce maintenance costs. Due to the influence of complex working conditions, vibration, friction, corrosion and other factors during the operation of the equipment, its state will gradually deteriorate. If potential problems are not discovered in time, it may lead to sudden failures, which will not only affect the production progress, but also may bring serious safety hazards and economic losses. Fusion of multivariate information of equipment and then construction of health indicators is an important part of equipment health monitoring. In terms of the construction of health indicators, the commonly used methods currently include dimensionality reduction methods, statistical methods based on single features, and data-driven methods. These methods have certain limitations in practical engineering applications. Methods based on dimensionality reduction, such as principal component analysis (PCA), often assume that the main features of the data can be represented by linear combinations, which makes it difficult to capture nonlinear failure modes. At the same time, PCA is sensitive to noise, which may affect the stability of the indicators. In addition, PCA cannot be dynamically updated, which leads to certain limitations in the application of PCA in real-time monitoring. When constructing health indicators using statistical methods with single features such as mean, variance or peak, although it can reflect certain characteristics of vibration signals, it lacks the ability to characterize the comprehensive characteristics of multi-dimensional signals and is easily affected by single noise or outliers, resulting in insufficient robustness in health status assessment. Data-driven methods using neural networks often rely on large amounts of data for training and lack physical interpretability.

[0052] In response to the above problems, this application proposes a device health monitoring method based on multivariate information fusion. This method constructs a sparse dictionary by combining time domain, frequency domain and time-frequency domain features, and can comprehensively extract multidimensional information from vibration signals. The health indicators are fitted through sparse optimization, which enhances the trend and monotonicity of the indicators and accurately reflects the health status of the equipment. This method supports offline batch processing and online real-time expansion, and has good flexibility and practicality. This method also combines dynamic thresholds and real-time update mechanisms, which can efficiently and accurately detect abnormal conditions in equipment operation, significantly improving the reliability and sensitivity of anomaly detection.

[0053] Example 1

[0054] In this embodiment, a device health monitoring method based on multivariate information fusion is applied to perform device health monitoring. The method steps are as follows: Figure 1 As shown, including:

[0055] S1, collecting a multivariate vibration signal from a device sensor and segmenting it to obtain a segmented signal;

[0056] S2, calculating the multi-dimensional eigenvalues ​​of each segment of the segmented signal, and integrating the eigenvalues ​​to obtain a feature matrix;

[0057] S3, input the feature matrix into the sparse representation model, construct the health index fitting function, recursively optimize the residual, and output the health index fitting result;

[0058] S4. Use the health indicator fitting results to evaluate the equipment status, monitor and output equipment abnormal points.

[0059] In this embodiment, data collection and preprocessing are first performed. Data is collected from 11 fracturing machines, with machine numbers DA-001, ..., DA-009, DA-011, and DA-012. Each device has 5 effective collection channels, of which ch1, ch2, and ch3 are vibration data collection channels, ch6 is the outlet pressure data collection channel, and ch7 is the inlet pressure collection channel. The device sampling frequency is 125Hz, and there is no fixed sampling time. The sampling time depends on the actual operation of the device.

[0060] In this embodiment, the multivariate vibration signal is collected from the device sensor, denoted as x=[x1, x2,…, x N ], the signal is segmented, each segment length is 1024. The original signal of the data collected by channel ch3 of the fracturing machine No. DA-001 is shown in the figure Figure 2 shown.

[0061] In this embodiment, the vibration signals collected from the equipment sensor are stored in the form of CSV files, read in batches and processed in segments. The signal validity is judged by using the mean and standard deviation thresholds (mean>0.01, standard deviation>0.05) of the signal, invalid data is removed, and the collected data is preprocessed to remove the invalid data. Figure 3 shown.

[0062] In this embodiment, the mean values ​​μ of the signal segments are calculated respectively. x and variance σ x , the signal validity is determined by the following formula:

[0063]

[0064] In the formula, μ x is the mean of the signal segment; σ x is the variance of the new number segment; mean_threshold is the preset signal mean threshold; std_threshold is the preset signal standard deviation threshold.

[0065] In this embodiment, the preset signal mean threshold is 0.01; the preset signal standard deviation threshold is 0.05.

[0066] If the signal mean and standard deviation meet the above conditions, the signal is considered valid; otherwise, the signal is discarded.

[0067] In this embodiment, the 10-dimensional feature values ​​of each signal are calculated, including time domain features: root mean square value (RMS), variance (Var), peak value (Max), skewness value (Ske), and kurtosis value (Kur); frequency domain features: peak-to-peak value (P2P), spectral skewness (SpeSke), and spectral kurtosis (SpeKur); time-frequency domain features: spectral power (SpePow) and wavelet packet energy (Wave).

[0068]

[0069]

[0070] Max=max(x i )

[0071]

[0072] P2P=max(x i )-min(x i )

[0073]

[0074] In the formula, f i is the spectrum of the signal; S(f i ) is the power spectral density of the signal.

[0075] The RMS value of the signal is Figure 4a As shown, the spectrum skewness characteristics are as follows Figure 4b shown.

[0076] In this embodiment, the calculated signal eigenvalues ​​are combined into a characteristic matrix T = [t1, t2, ..., t i ], for each column t of the feature matrix T i Standardization is performed to eliminate scale differences between different features and provide consistent input for subsequent optimization calculations.

[0077] In this embodiment, after integrating the feature values ​​to obtain the feature matrix, the feature matrix is ​​standardized by column, and the specific processing formula is as follows:

[0078]

[0079] In the formula, is the column vector of the characteristic matrix after normalization; is the mean of the i-th feature; is the variance of the i-th feature; t i is the column vector of the feature matrix.

[0080] In this embodiment, a sparse representation model is used to select key features from the feature matrix to construct a health index fitting function. The accuracy of feature selection is ensured by recursively optimizing the residual, and the model complexity and fitting accuracy are balanced under the sparsity constraint.

[0081] In this embodiment, the health index fitting function is:

[0082] H=T norm β

[0083] Where, T norm is the standardized feature matrix; β is the sparse coefficient.

[0084] In this embodiment, the initialization parameters include: the initial residual is the label value, the initial subset is an empty set, and the initial iteration step is 0.

[0085] r0=H true

[0086] S = empty set

[0087] k=0

[0088] Among them: r0 is the initial residual, H true is the label value, S is the subset, and k is the number of iterations.

[0089] In this embodiment, in each iteration, the feature with the greatest correlation is first selected from the feature matrix using the parameters, and then the parameters are calculated and updated; wherein the specific formula for selecting the feature with the greatest correlation and calculating the parameters is:

[0090]

[0091] S=S∪{i *}

[0092]

[0093] r k+1 =H true -T S β S

[0094] In the formula, i * is the residual r in the current iteration k The index of the most relevant feature; is the transpose of the i-th column of the feature matrix T; r k is the residual; S is the feature subset; β sis a sparse coefficient vector, including only the non-zero weights of the selected features; H true is the label value; T S is the currently selected feature subset.

[0095] In this embodiment, the health index is calculated:

[0096] H fit =T S β S

[0097] In this embodiment, the state assessment is performed through the health indicator fitting results, and the health indicator is averaged to calculate its monotonicity, trend and robustness; based on the mean and standard deviation of the health indicator, abnormal points are detected and marked to support further fault prediction and alarm.

[0098] In this embodiment, the health index H fit Perform sliding average smoothing:

[0099]

[0100] Where i is the current index position; ω is the window length, H fit (j) is the fitting result of health indicators.

[0101] The health index after multivariate information fusion using this method is shown in the figure below. Figure 5 As shown in the figure, the comparison between the proposed method and the PCA method and the construction method with RMS as the single feature is shown in Figure 6 As shown. The blue solid line represents the original health index of this method, the blue dotted line represents the health index after smoothing by this method, the purple solid line represents the original health index using the PCA method, the purple dotted line represents the health index after smoothing by the PCA method, the green solid line represents the original health index using the RMS method, and the green dotted line represents the smoothed health index after smoothing by the RMS method. Figure 5 and Figure 6 It can be seen that the health index constructed by the proposed method can reflect the actual degradation trend of the equipment. Figure 6 It can be seen intuitively that compared with the PCA method and the construction method with RMS as a single feature, the proposed method has a better dynamic range and can more sensitively capture changes in equipment status. At the same time, although the curve fluctuation range of the other two methods is lower, it is still subject to greater noise interference. Although this method has large fluctuations, it can effectively remove noise after smoothing while maintaining the main trend, showing strong noise resistance and better capturing nonlinear fault modes.

[0102] In this embodiment, the specific formula for calculating the monotonicity of the health index is:

[0103]

[0104] Where D pos The number of health index derivative values ​​greater than 0 over time; D neg is the number of health indicator derivatives with time less than 0; h is the total number of sequences (observations) of health indicators.

[0105] In this embodiment, the specific formula for calculating the trend of the health index is:

[0106]

[0107] In the formula, x k is the health index value of observation period k; t k is the time value corresponding to the observation period k.

[0108] In this embodiment, the specific formula for calculating the robustness of the health indicator is:

[0109]

[0110] Where: m(k) is the kth value of the original health index; is the kth value of the smoothed health index; m j (k) is the kth value of the smoothed health index.

[0111] After calculation, the monotonicity, trend and robustness of the health index constructed by this method are 0.4927, 0.9312 and 0.9136 respectively, while the monotonicity, trend and robustness of the health index constructed by the PCA method are 0.3498, 0.8481 and 0.8118 respectively, and the monotonicity, trend and robustness of the health index constructed with RMS as the single feature are 0.3377, 0.8036 and 0.8727 respectively, which reflects the superiority of this method.

[0112] Finally, the 3σ principle based on mean and variance is used to monitor and output equipment abnormal points. The specific formula is:

[0113] Anomaly={|i||H fit (i) -μ H |>3σ H}

[0114] In the formula, Anomaly is the abnormal point, H fit (i) is the fitting result of health index; μ1 is the mean value of health index; σ H is the variance of the health indicator.

[0115] The abnormal monitoring results of this method are as follows Figure 7As shown in the figure, the red line is the 3σ principle anomaly detection line based on mean and variance. Figure 7 It can be seen that this method can effectively mark abnormal points and support further fault prediction and alarm for equipment.

[0116] Example 2

[0117] In this embodiment, a device health monitoring system based on multivariate information fusion is applied to perform device health monitoring, and the system includes an information collection module, a feature extraction module, a health index fitting module and an equipment status evaluation module;

[0118] Among them, the information acquisition module is used to collect multivariate vibration signals from equipment sensors and segment them to obtain segmented signals; the feature extraction module is used to calculate the multi-dimensional eigenvalues ​​of each signal in the segmented signal, and integrate the eigenvalues ​​to obtain the feature matrix; the health indicator fitting module is used to input the feature matrix into the sparse representation model, construct the health indicator fitting function, and recursively optimize the residual to output the health indicator fitting result; the equipment status evaluation module is used to use the health indicator fitting result to evaluate the equipment status, monitor and output equipment abnormal points.

[0119] In this embodiment, data collection and preprocessing are first performed. Data is collected from 11 fracturing machines, with machine numbers DA-001, ..., DA-009, DA-011, and DA-012. Each device has 5 effective collection channels, of which ch1, ch2, and ch3 are vibration data collection channels, ch6 is the outlet pressure data collection channel, and ch7 is the inlet pressure collection channel. The device sampling frequency is 125Hz, and there is no fixed sampling time. The sampling time depends on the actual operation of the device.

[0120] In this embodiment, the multivariate vibration signal is collected from the device sensor, denoted as x=[x1, x2,…, x N ], the signal is segmented, each segment length is 1024. The original signal of the data collected by channel ch3 of the fracturing machine No. DA-001 is shown in the figure Figure 2 shown.

[0121] In this embodiment, the vibration signals collected from the equipment sensor are stored in the form of CSV files, read in batches and processed in segments. The signal validity is judged by using the mean and standard deviation thresholds (mean>0.01, standard deviation>0.05) of the signal, invalid data is removed, and the collected data is preprocessed to remove the invalid data. Figure 3 shown.

[0122] In this embodiment, the mean values ​​μ of the signal segments are calculated respectively. x and variance σ x , the signal validity is determined by the following formula:

[0123]

[0124] In the formula, μ x is the mean of the signal segment; σ x is the variance of the new number segment; mean_threshold is the preset signal mean threshold; std_threshold is the preset signal standard deviation threshold.

[0125] In this embodiment, the preset signal mean threshold is 0.01; the preset signal standard deviation threshold is 0.05.

[0126] If the signal mean and standard deviation meet the above conditions, the signal is considered valid; otherwise, the signal is discarded.

[0127] In this embodiment, the 10-dimensional feature values ​​of each signal are calculated, including time domain features: root mean square value (RMS), variance (Var), peak value (Max), skewness value (Ske), and kurtosis value (Kur); frequency domain features: peak-to-peak value (P2P), spectral skewness (SpeSke), and spectral kurtosis (SpeKur); time-frequency domain features: spectral power (SpePow) and wavelet packet energy (Wave).

[0128]

[0129] Max=max(x i )

[0130]

[0131] P2P=max(x i )-min(x i )

[0132]

[0133] In the formula, f i is the spectrum of the signal; S(f i ) is the power spectral density of the signal.

[0134] The RMS value of the signal is Figure 4a As shown, the spectrum skewness characteristics are as follows Figure 4b shown.

[0135] In this embodiment, the calculated signal eigenvalues ​​are combined into a characteristic matrix T = [t1, t2, ..., t i ], for each column t of the feature matrix T i Standardization is performed to eliminate scale differences between different features and provide consistent input for subsequent optimization calculations.

[0136] In this embodiment, after integrating the feature values ​​to obtain the feature matrix, the feature matrix is ​​standardized by column, and the specific processing formula is as follows:

[0137]

[0138] In the formula, is the column vector of the characteristic matrix after normalization; is the mean of the i-th feature; is the variance of the i-th feature; t i is the column vector of the feature matrix.

[0139] In this embodiment, a sparse representation model is used to select key features from the feature matrix to construct a health index fitting function. The accuracy of feature selection is ensured by recursively optimizing the residual, and the model complexity and fitting accuracy are balanced under the sparsity constraint.

[0140] In this embodiment, the health index fitting function is:

[0141] H=T norm β

[0142] Where, T norm is the standardized feature matrix; β is the sparse coefficient.

[0143] In this embodiment, the initialization parameters include: the initial residual is the label value, the initial subset is an empty set, and the initial iteration step is 0.

[0144] r0=H true

[0145] S = empty set

[0146] k=0

[0147] Among them: r0 is the initial residual, H true is the label value, S is the subset, and k is the number of iterations.

[0148] In this embodiment, in each iteration, the feature with the greatest correlation is first selected from the feature matrix using the parameters, and then the parameters are calculated and updated; wherein the specific formula for selecting the feature with the greatest correlation and calculating the parameters is:

[0149]

[0150] S=S∪{i *}

[0151]

[0152] r k+1 =H true -T S β S

[0153] In the formula, i * is the residual r in the current iteration k The index of the most relevant feature; is the transpose of the i-th column of the feature matrix T; r k is the residual; S is the feature subset; β s is a sparse coefficient vector, including only the non-zero weights of the selected features; H true is the label value; T S is the currently selected feature subset.

[0154] In this embodiment, the health index is calculated:

[0155] H fit =T S β S

[0156] In this embodiment, the state assessment is performed through the health indicator fitting results, and the health indicator is averaged to calculate its monotonicity, trend and robustness; based on the mean and standard deviation of the health indicator, abnormal points are detected and marked to support further fault prediction and alarm.

[0157] In this embodiment, the health index H fit Perform sliding average smoothing:

[0158]

[0159] Where i is the current index position; ω is the window length, H fit (j) is the fitting result of health indicators.

[0160] The health indicators after multi-information fusion by this system are shown in the figure below. Figure 5 As shown in the figure, the comparison between the proposed system and the PCA system and the construction system with RMS as the single feature is shown in Figure 6 As shown. The blue solid line represents the original health index of this system, the blue dotted line represents the health index after smoothing, the purple solid line represents the original health index of the system using PCA, the purple dotted line represents the health index after smoothing using PCA, the green solid line represents the original health index of the system using RMS, and the green dotted line represents the smoothed health index after smoothing using RMS. Figure 5 and Figure 6 It can be seen that the health indicators constructed by the proposed system can reflect the actual degradation trend of the equipment. Figure 6It can be seen intuitively that compared with the PCA system and the system built with RMS as a single feature, the proposed system has a better dynamic range and can more sensitively capture changes in the equipment state. At the same time, although the curve fluctuation range of the other two systems is lower, they are still subject to greater noise interference. Although the system has large fluctuations, it can effectively remove noise after smoothing while maintaining the main trend, showing strong noise resistance and better capture of nonlinear fault modes.

[0161] In this embodiment, the specific formula for calculating the monotonicity of the health index is:

[0162]

[0163] Where D pos The number of health index derivative values ​​greater than 0 over time; D neg is the number of health indicator derivatives with time less than 0; h is the total number of sequences (observations) of health indicators.

[0164] In this embodiment, the specific formula for calculating the trend of the health index is:

[0165]

[0166] In the formula, x k is the health index value of observation period k; t k is the time value corresponding to the observation period k.

[0167] In this embodiment, the specific formula for calculating the robustness of the health indicator is:

[0168]

[0169] Where: m(k) is the kth value of the original health index; is the kth value of the smoothed health index; m j (k) is the kth value of the smoothed health index.

[0170] After calculation, the monotonicity, trend and robustness of the health index constructed by this system are 0.4927, 0.9312 and 0.9136 respectively, while the monotonicity, trend and robustness of the health index constructed by the PCA system are 0.3498, 0.8481 and 0.8118 respectively, and the monotonicity, trend and robustness of the health index constructed with RMS as the single feature are 0.3377, 0.8036 and 0.8727 respectively, which reflects the superiority of this system.

[0171] Finally, the 3σ principle based on mean and variance is used to monitor and output equipment abnormal points. The specific formula is:

[0172] Anomaly={|i||H fit (i) -μ H |>3σ H}

[0173] In the formula, Anomaly is the abnormal point, H fit (i) is the health index fitting result; μ H is the mean value of health index; σ H is the variance of the health indicator.

[0174] This system is suitable for offline analysis in batch mode and has the potential to be expanded to online data processing. It can efficiently integrate multiple information to reflect the current health status of the equipment, while also supporting further prediction of the remaining life of the equipment.

[0175] The system performs abnormal monitoring results such as Figure 7 As shown in the figure, the red line is the 3σ principle anomaly detection line based on mean and variance. Figure 7 It can be seen that this system can effectively mark abnormal points and support further fault prediction and alarm for equipment.

[0176] The above is only a specific embodiment of the present invention, but the protection scope of the present invention is not limited thereto. Any technician familiar with the technical field can easily think of various equivalent modifications or replacements within the technical scope disclosed by the present invention, and these modifications or replacements should be included in the protection scope of the present invention. Therefore, the protection scope of the present invention shall be based on the protection scope of the claims.

Claims

1. A device health monitoring method based on multivariate information fusion, characterized in that: The following steps are involved: S1, collecting a multivariate vibration signal from a device sensor and segmenting it to obtain a segmented signal; S2, calculating the multi-dimensional eigenvalues ​​of each segment of the segmented signal, and integrating the eigenvalues ​​to obtain a feature matrix; S3, input the feature matrix into the sparse representation model, construct the health index fitting function, recursively optimize the residual, and output the health index fitting result; S4. Use the health indicator fitting results to evaluate the equipment status, monitor and output equipment abnormal points.

2. The device health monitoring method based on multivariate information fusion according to claim 1 is characterized in that: After the signal in S1 is segmented, its validity is judged by the mean and standard deviation of the segmented signal. If the mean and standard deviation of the signal meet the preset threshold requirements, the segmented signal is judged to be valid; otherwise, the segmented signal is discarded. The specific formula for judging the validity is: In the formula, μ x is the mean of the signal segment; σ x is the variance of the new number segment; mean_threshold is the preset signal mean threshold; std_threshold is the preset signal standard deviation threshold.

3. The device health monitoring method based on multivariate information fusion according to claim 1 is characterized in that: The multi-dimensionality in S2 is ten dimensions, including time domain features, frequency domain features and time-frequency domain features; among which, the time domain features include root mean square value, variance, peak value, skewness value and kurtosis value; the frequency domain features include peak-to-peak value, spectral skewness and spectral kurtosis; the time-frequency domain features include spectral power and wavelet packet energy.

4. The device health monitoring method based on multivariate information fusion according to claim 1 is characterized in that: After integrating the feature values ​​in S2 to obtain the feature matrix, the feature matrix is ​​standardized by column. The specific processing formula is: In the formula, is the column vector of the characteristic matrix after normalization; is the mean of the i-th feature; is the variance of the i-th feature; t i is the column vector of the feature matrix.

5. The device health monitoring method based on multivariate information fusion according to claim 1 is characterized in that: The health index fitting function in S3 is: H=T norm b Where, T norm is the standardized feature matrix; β is the sparse coefficient.

6. The device health monitoring method based on multivariate information fusion according to claim 1 is characterized in that: The specific process of recursively optimizing the residual in S3 is: firstly, initializing the parameters, then iterating, and when the iteration reaches a preset number of times, ending the iteration, and calculating the health index fitting result.

7. The device health monitoring method based on multivariate information fusion according to claim 6 is characterized in that: In each iteration, the parameters are first used to select the feature with the greatest correlation from the feature matrix, and then the parameters are calculated and updated; wherein the specific formula for selecting the feature with the greatest correlation and calculating the parameters is: S=S∪{i * } r k+1 =H true -T S b S In the formula, i * is the residual r in the current iteration k The index of the most relevant feature; is the transpose of the i-th column of the feature matrix T; r k is the residual; S is the feature subset; β s is a sparse coefficient vector, including only the non-zero weights of the selected features; H true is the label value; T S is the currently selected feature subset.

8. The device health monitoring method based on multivariate information fusion according to claim 1 is characterized in that: When the health index fitting result is used to evaluate the equipment status, the health index fitting result is firstly subjected to sliding average smoothing, and then the monotonicity, trend and robustness are calculated based on the sliding average smoothing. The specific formula for the health index fitting result is: Where i is the current index position; ω is the window length, H fit (j) is the fitting result of health indicators.

9. The device health monitoring method based on multivariate information fusion according to claim 1 is characterized in that: In S4, the 3σ principle based on mean and variance is used to monitor and output equipment abnormal points. The specific formula is: Anomaly={|i||H fit (i)-μ H |>3σ H } In the formula, Anomaly is the abnormal point, H fit (i) is the health index fitting result; μ H is the mean value of health index; σ H is the variance of the health indicator.

10. An equipment health monitoring system based on multivariate information fusion, characterized in that: The system is applied to work by a device health monitoring method based on multivariate information fusion as described in claims 1-9, and the system includes an information acquisition module, a feature extraction module, a health index fitting module and an equipment status evaluation module; The information acquisition module is used to collect multivariate vibration signals from the equipment sensor and segment them to obtain segmented signals; The feature extraction module is used to calculate the multi-dimensional feature value of each segment signal in the segmented signal, and integrate the feature values ​​to obtain a feature matrix; The health index fitting module is used to input the feature matrix into the sparse representation model, construct a health index fitting function, recursively optimize the residual, and output the health index fitting result; The equipment status assessment module is used to evaluate the equipment status using the health index fitting results, and monitor and output equipment abnormal points.

Images