Multi-source data fusion fault diagnosis method and system based on DS evidence theory

Abstract

The present invention belongs to the technical field of signal processing, and particularly relates to a multi-source data fusion fault diagnosis method and system based on the DS evidence theory. The method of the present invention comprises: first, collecting multi-source information of a mechanical device, wherein the multi-source information includes vibration data, image data and sound data; respectively calculating corresponding feature indicators; then, using a PCA algorithm to perform dimensionality reduction processing on the indicators; using a Bayesian fault diagnosis model to perform fault diagnosis, so as to obtain a diagnosis result from each sensor; and then, using the DS evidence theory to perform fault fusion diagnosis. The present invention can comprehensively consider the characteristics of various types of data, extract more effective fault indicators, achieve more comprehensive fault identification, and fuse diagnosis results from a plurality of sensors, thereby ensuring the accuracy of final diagnosis. The present invention is suitable for fault diagnosis of various complex mechanical systems.

Description

A multi-source data fusion fault diagnosis method and system based on DS evidence theory TECHNICAL FIELD

[0001] The present application belongs to the technical field of signal processing, and specifically relates to a multi-source data fusion fault diagnosis method and system based on DS evidence theory. BACKGROUND

[0002] Mechanical fault diagnosis algorithms are crucial in modern industry. Through real-time monitoring, fault diagnosis algorithms can promptly identify potential hazards, ensuring the safety of employees and assets. They also extend equipment life, reduce wear and tear, and identify factors affecting device performance through analysis of operational data, providing maintenance recommendations.

[0003] Mechanical equipment operational data comes from a wide range of sources, including vibration data, sound data, image data, etc. The diversity of these data poses challenges such as inconsistent information and missing information. Through multi-source data fusion, the strengths of different data sources can be fully utilized, improving the accuracy and robustness of fault diagnosis. Fusion algorithms can consider the characteristics of various types of data, enabling more comprehensive fault identification. Multi-source data fusion alarm diagnosis algorithms based on Dempster-Shafer (DS) theory are gradually becoming an important research direction in the field of mechanical fault diagnosis due to their advantages in handling uncertainty and information fusion.

[0004] Existing methods or research have one or more of the following defects and deficiencies:

[0005] 1. Many existing algorithms are still based on traditional methods such as Kalman filtering and fuzzy logic, lacking innovative algorithms for new fault patterns.

[0006] 2. Although research on multi-sensor data fusion has increased, data processing techniques for different types of sensors (such as vibration, temperature, and acoustics) are still lacking.

[0007] 3. Although deep learning has achieved significant results in image processing and natural language processing, its application in fault diagnosis is still in its early stages, with fewer related patents.

[0008] 4. Existing patents focus on single algorithms or techniques, lacking comprehensive solutions, and are unable to effectively address complex mechanical system failures.

[0009] 5. In high-speed, dynamic environments, existing technologies have limitations in data processing speed and real-time performance, making it difficult to meet the needs of industrial sites.

[0010] 6. There are many characteristic indicators of data collected by sensors, making it difficult to extract effective indicators, resulting in deviations in alarms and diagnoses. SUMMARY

[0011] The present application aims to provide a multi-source data fusion fault diagnosis method and system based on DS evidence theory, to solve the technical problems of inconsistent information, insufficient fusion technology, lack of comprehensive solution, poor real-time performance and difficulty in extracting effective feature indicators when existing mechanical fault diagnosis algorithms process multi-source data.

[0012] The present application achieves the above-mentioned purposes through the following technical solutions:

[0013] In a first aspect, the present application provides a multi-source data fusion fault diagnosis method based on DS evidence theory, which comprises:

[0014] Collecting multi-source information of the target device, including vibration data, image data and sound data;

[0015] Determining and extracting feature indicators corresponding to the vibration data, the image data and the sound data;

[0016] Using principal component analysis algorithm to reduce dimension of the feature indicators;

[0017] Training a pre-constructed fault diagnosis model based on time-frequency domain features and naive Bayes according to the reduced dimension feature indicators, and determining a first diagnosis result of the target device fault based on the fault diagnosis model;

[0018] Fusing the first diagnosis result using DS evidence theory to determine a second diagnosis result. Further, the determining and extracting feature indicators corresponding to the vibration data, the image data and the sound data comprises:

[0019] Determining and extracting feature indicators of vibration data of multiple target devices, including acceleration effective value, acceleration peak-to-peak value, acceleration pulse, acceleration margin, acceleration kurtosis, acceleration skewness, speed effective value, speed peak-to-peak value, speed pulse, speed margin, speed kurtosis and speed skewness, to form a first feature set;

[0020] Determining and extracting feature indicators of sound data of multiple target devices, including effective value, loudness, voiceprint roughness, frequency barycenter, sharpness, high-low frequency band energy ratio, peak-to-peak value, pulse, margin, kurtosis and skewness, to form a second feature set;

[0021] Determining and extracting feature indicators of image data of multiple target devices, including texture features, edge features, temperature distribution features and crack distribution features, to form a third feature set.

[0022] Further, the using principal component analysis algorithm to reduce dimension of the feature indicators comprises:

[0023] calculating a covariance matrix of each of the normalized feature sets;

[0024] calculating eigenvalues and eigenvectors of each of the covariance matrices based on a numerical algorithm;

[0025] determining target eigenvectors according to the sizes of the eigenvalues, and constructing a principal component space based on the target eigenvectors;

[0026] projecting the normalized covariance matrix into the principal component space for matrix multiplication to obtain a reduced dimension feature index.

[0027] Further, the training of the pre-constructed fault diagnosis model based on the time-frequency domain features and the Naive Bayes based on the reduced dimension feature index to determine the first diagnosis result of the target equipment fault includes:

[0028] determining a target equipment fault type according to a fault mode or a historical fault record of the target equipment, taking the fault type as a class label, and forming a training set including the reduced dimension feature index and the corresponding fault class label;

[0029] calculating a prior probability P(C k ) of each of the fault types in the training set;

[0030] calculating a likelihood probability P(x i ∣C k ) of each feature index under each fault type;

[0031] calculating a posterior probability P(C k ∣X) of the reduced dimension feature index X of the target equipment belonging to each fault type by Bayes' theorem, and selecting a fault type with the maximum posterior probability as the first diagnosis result, expressed as:

[0032]

[0033]

[0034]

[0035] wherein C k represents a fault type; P(C k ) is a prior probability of a fault class C k ; P(X∣C k ) represents a conditional probability of X under the class C k ; and P(C k ∣X) represents a fault class C k posterior probability; P(X) represents a normalization constant of all fault types; arg max is to select C k a fault category that maximizes the posterior probability.

[0036] Further, the adopting the DS evidence theory to fuse the first diagnosis result to determine a second diagnosis result comprises:

[0037] converting the first diagnosis result into evidence bodies in the DS evidence theory, each of the evidence bodies representing a degree of trust of a signal source to a fault type, the signal source including a vibration sensor, an image sensor and a sound sensor;

[0038] constructing a trust degree function to assign a trust degree function to each fault type for each signal source;

[0039] fusing the trust degree functions of the multiple evidence bodies to obtain a merged trust degree function;

[0040] selecting, according to the merged trust degree function, a fault type with a maximum trust degree value as the second diagnosis result.

[0041] Further, the selecting, according to the merged trust degree function, a fault type with a maximum trust degree value as the second diagnosis result comprises:

[0042] updating the merged trust degree function based on a set conflict management strategy, and selecting, from the updated trust degree function, a fault type corresponding to a maximum trust degree value as the second diagnosis result;

[0043] The set conflict management strategy comprises:

[0044] (1) quantifying a conflict degree between different signal sources based on a conflict, and when the conflict degree exceeds a preset threshold, selecting to exclude conflict information from evidence synthesis to update the merged trust degree function; the conflict information is one or more of inconsistent diagnosis results, inconsistent probability distribution, inconsistent feature indicators, inconsistent time or space;

[0045] (2) adjusting a weight of conflict evidence based on reliability or historical data of an evidence source to update the merged trust degree function.

[0046] In a second aspect, the present application provides a multi-source data fusion fault diagnosis system based on the DS evidence theory, which is applied to the multi-source data fusion fault diagnosis method based on the DS evidence theory described in any of the above aspects, and comprises:

[0047] a data acquisition module configured to acquire multi-source information of a target device, the multi-source information including vibration data, image data and sound data;

[0048] a feature extraction module configured to extract feature indicators corresponding to the vibration data, the image data, and the sound data;

[0049] a principal component analysis module configured to perform dimensionality reduction processing on the feature indicators by using a principal component analysis algorithm;

[0050] a first diagnosis module configured to train a pre-constructed fault diagnosis model based on time-frequency domain features and a naive Bayes algorithm according to the dimensionally reduced feature indicators, and determine a first diagnosis result of a fault of the to-be-tested equipment based on the fault diagnosis model;

[0051] a second diagnosis module configured to fuse the first diagnosis result by using a DS evidence theory to determine a second diagnosis result.

[0052] The present application has the following beneficial effects:

[0053] The present application can comprehensively collect multi-source information such as vibration, image, sound, and the like of a mechanical equipment, and calculate corresponding feature indicators, so that the equipment state can be more comprehensively reflected, and the accuracy of fault recognition can be improved. The PCA algorithm is used to perform dimensionality reduction processing on the feature indicators, so that the features with the highest contribution degree can be effectively extracted, and the accuracy of alarm and diagnosis can be further improved. Meanwhile, the combination of the Bayes fault diagnosis model and the DS evidence theory for fault fusion diagnosis can not only enhance the comprehensive analysis capability of the system on various types of data, but also significantly improve the anti-interference capability of the system on noise and external interference. Even in the case of partial sensor or model failure, the present application can ensure that the overall system maintains a high diagnosis level, and provides a strong guarantee for the stable operation of the mechanical equipment. BRIEF DESCRIPTION OF DRAWINGS

[0054] FIG. 1 is a flowchart of a multi-source data fusion fault diagnosis method based on a DS evidence theory according to an embodiment of the present application;

[0055] FIG. 2 is another flowchart of a multi-source data fusion fault diagnosis method based on a DS evidence theory according to an embodiment of the present application;

[0056] FIG. 3 is a structural diagram of a multi-source data fusion fault diagnosis system based on a DS evidence theory according to an embodiment of the present application;

[0057] FIG. 4 is a vibration data diagram of a case in the specific embodiment of the present application;

[0058] FIG. 5 is a sound data diagram of a case in the specific embodiment of the present application;

[0059] FIG. 6 is an image data diagram of a case in the specific embodiment of the present application;

[0060] Figure 7 is a scatter plot of the dimension-reduced vibration data of the specific embodiment case after PCA processing. DETAILED DESCRIPTION

[0061] The application will be described in further detail below with reference to the drawings. It is necessary to point out here that the following detailed description is only used to further illustrate the application and cannot be understood as limiting the scope of protection of the application. Those skilled in the art can make some non-essential improvements and adjustments to the application according to the above application content.

[0062] Embodiment 1

[0063] As shown in Figures 1-2, the embodiment proposes a multi-source data fusion fault diagnosis method based on DS evidence theory, characterized in that the method comprises the following steps:

[0064] S1. Collecting multi-source information of the target device, the multi-source information including vibration data, image data and sound data;

[0065] In specific implementation, the multi-source information is collected by sensors installed on the mechanical equipment, the vibration sensor is used to collect the vibration data of the equipment, the sound sensor is used to collect the sound data of the equipment, and the image sensor is used to collect the image data of the equipment. These sensors will collect the real-time data for subsequent processing.

[0066] S2. Determining and extracting the characteristic indexes corresponding to the vibration data, image data and sound data;

[0067] As a preferred, determining and extracting the characteristic indexes corresponding to the vibration data, image data and sound data, comprising:

[0068] Determining and extracting the characteristic indexes of the vibration data of the plurality of target devices, including acceleration effective value, acceleration peak-to-peak value, acceleration pulse, acceleration margin, acceleration kurtosis, acceleration skewness, velocity effective value, velocity peak-to-peak value, velocity pulse, velocity margin, velocity kurtosis, velocity skewness, forming a first characteristic set;

[0069] Determining and extracting the characteristic indexes of the sound data of the plurality of target devices, including effective value, loudness, voiceprint roughness, frequency barycenter, sharpness, high-to-low frequency band energy ratio, peak-to-peak value, pulse, margin, kurtosis and skewness, forming a second characteristic set;

[0070] Determining and extracting the characteristic indexes of the image data of the plurality of target devices, including texture features, edge features, temperature distribution features and crack distribution features, forming a third characteristic set.

[0071] In step S2, the vibration sensor data feature extraction includes extracting time domain and frequency domain feature data of the vibration data, and the specific indexes are as follows:

[0072] (1) Vibration acceleration amplitude, vibration amplitude refers to the maximum deviation value of the vibration signal, which is usually used to represent the intensity of vibration.

[0073] (2) Acceleration effective value, which represents the intensity of vibration, is often used to measure the energy or power of vibration.

[0074] (3) Margin index, margin is the ratio of peak absolute value to square root amplitude, which reflects the impact characteristics of the signal. If this value is too large, it means that the signal has strong impact and is not sensitive to kurtosis.

[0075] (4) Kurtosis index, kurtosis reflects the impact characteristics of the vibration signal. Kurtosis is sensitive to impact, and the kurtosis value should be around 3 in general case, because the kurtosis of normal distribution is equal to 3. If it deviates from 3 too much, it means that the mechanical equipment has certain impact vibration and may have some hidden trouble. Kurtosis is more sensitive than margin, but the stability of the index is poor.

[0076] (5) Skewness index, skewness reflects the asymmetry of the vibration signal. In general case, the vibration signal is symmetric about the x-axis, so the skewness should tend to 0. If the friction or collision in a certain direction of the device is large, it will cause the asymmetry of the vibration and make the skewness larger.

[0077] (6) Peak-to-peak value, which has similar meaning as peak value, represents the change range of maximum and minimum values of vibration waveform.

[0078] (7) Pulse reflects the overall fluctuation degree of vibration data, and the calculation formula is the maximum value of data divided by the absolute average value.

[0079] In step S2, the sound sensor data feature extraction, the specific indicators are as follows:

[0080] (1) Short-time zero-crossing rate (ST-ZCR): it can reflect the fluctuation degree of sound signal, and simply reflect the frequency domain information from time domain. The original sound signal needs to pass through an anti-aliasing band-pass filter first, with a low-end cutoff frequency higher than 50 Hz, to filter out environmental noise and power interference.

[0081] (2) Short-time energy (STE): used to distinguish the boundary of sound.

[0082] (3) Linear prediction coefficient (LPC): the basic idea is that the human vocal system is a linear time-invariant system, and the sound sampling data close in time have strong correlation.

[0083] (4) Linear Predictive Cepstral Coefficients (LPCC): computed from LPC, the last C stands for cepstral domain, and LPC alike, it models linear systems. Compared to LPC, LPCC uses less data to express more effective information, and it is easy to implement.

[0084] (5) Mel Frequency Cepstral Coefficients (MFSC): Mel frequency converts the non-uniform distribution of frequency into a uniform distribution of Mel frequency (good when the signal-to-noise ratio is low). The main process of sound MFSC extraction algorithm is: framing, windowing, fast Fourier transform, modulus, Mel filtering, logarithm (mfccs = librosa.feature.mfcc(y=y, sr=sr, n_mfcc=40)).

[0085] (6) Mel Frequency Cepstral Coefficients (MFCC): MFSC is subjected to discrete cosine transform, compared to MFSC, MFCC outputs less data, lower dimension, weaker correlation, and less information, so generally different orders of difference coefficients are combined to use. (preprocessing, Hamming windowing, FFT change, through Mel triangular filter bank, mean and variance, transpose, FFT change, logarithmic operation through filter bank, DCT change.)

[0086] (7) GFCC coefficients: Gamma-tone filter bank is used to filter the sound.

[0087] (8) Peak value and corresponding frequency of power spectrum: according to the size of spectral energy, the first 8 peak values and corresponding frequencies (energy) are taken as feature vectors.

[0088] (9) Loudness: the loudness of sound is a subjective index to describe the intensity of sound perceived by human ear. Loudness is used to reflect the attribute of auditory judgment of sound intensity, which is related not only to sound pressure, but also to frequency, waveform and other factors.

[0089] (10) Sharpness: sharpness index describes the timbre characteristics in sound quality evaluation. The sharpness of sound describes the feeling related to the frequency component of sound but unrelated to the loudness of sound. It reflects the degree of feeling of sharpness, high frequency and discomfort, which is the contrast between high frequency energy and total energy in sound.

[0090] In step S2, the feature extraction of image data, specific indicators include:

[0091] (1) Texture features Texture features can characterize the state of the equipment surface through the gray value, edge, directionality, roughness, etc. When the equipment fails, different texture features often appear on the surface, which can reflect wear, cracks, corrosion, etc.

[0092] (2) Edge Features Edge features are used to identify abrupt changes, defects, and variations in the surface or internal structure of a device. Faults often result in significant changes at the edges, such as cracks, wear, or the presence of foreign objects.

[0093] (3) Temperature Distribution Map (Thermal Imaging Features) In thermal imaging images, faults are usually manifested as uneven heat distribution. For example, faults such as motor overheating and bearing wear may cause local temperature increases, and these anomalies can be detected by temperature distribution maps.

[0094] (4) Crack Detection: Image processing technology can effectively detect cracks on the surface or inside of equipment, especially in structural components, welds, pipes, etc. The occurrence of cracks is usually a precursor to failure.

[0095] S3. Principal component analysis algorithm is used to reduce the dimensionality of the feature indicators.

[0096] It should be noted that before applying the Principal Component Analysis (PCA) algorithm, the data needs to be preprocessed to ensure data quality and applicability.

[0097] The steps include:

[0098] (1) Data cleaning: Delete or correct missing values, outliers and erroneous data.

[0099] (2) Standardization (Normalization): Since PCA depends on the covariance matrix, and the magnitude of the covariance is affected by the scale of each feature, it is necessary to standardize or normalize the data to ensure that all features have the same scale. The mean of the standardized data is 0 and the standard deviation is 1.

[0100] Preferably, principal component analysis (PCA) is used to reduce the dimensionality of the feature indices, including:

[0101] Calculate the covariance matrix of each standardized feature set;

[0102] The eigenvalues ​​and eigenvectors of each covariance matrix are calculated using numerical algorithms (such as the QR algorithm, the Jacobi method, etc.).

[0103] According to the size of the eigenvalue, the target feature vector is determined, and a principal component space is constructed therefrom; specifically, according to the size of the eigenvalue, the first several principal components whose cumulative explained variance reaches a predetermined threshold (such as 90% or 95%) are selected to form a new feature space;

[0104] The standardized covariance matrix is projected into the principal component space for matrix multiplication to obtain the dimension-reduced feature index.

[0105] S4. According to the dimension-reduced feature index, a pre-constructed fault diagnosis model based on time-frequency domain features and Naive Bayes is trained, and a first diagnosis result of the fault of the to-be-tested equipment is determined based on the fault diagnosis model;

[0106] As preferred, according to the dimension-reduced feature index, a pre-constructed fault diagnosis model based on time-frequency domain features and Naive Bayes is trained, and a first diagnosis result of the fault of the to-be-tested equipment is determined based on the fault diagnosis model, comprising:

[0107] According to the fault mode or historical fault record of the target equipment, the fault type of the target equipment is determined, the fault type is taken as a class label, and a training set including the dimension-reduced feature index and the corresponding fault class label is formed;

[0108] The prior probability P(Ck) of each fault type in the training set is calculated;

[0109] The likelihood probability P(xi|Ck) of each feature index under each fault type is calculated;

[0110] The posterior probability P(Ck|X) of the dimension-reduced feature index X of the to-be-tested equipment belonging to each fault type is calculated through Bayes' theorem, and the fault type with the maximum posterior probability is selected as the first diagnosis result, and the expression is:

[0111]

[0112]

[0113]

[0114] wherein, C k represents the fault type; P(C k ) is the prior probability of the fault class C k ; P(X|C k ) represents the conditional probability of X under the class C k ; P(C k |X) represents the posterior probability of the fault class C k given the feature X; P(X) represents a normalization constant of all fault types; and arg max is the fault class C k that maximizes the posterior probability.

[0115] S5. Fusing the first diagnostic results by DS evidence theory to determine the second diagnostic result.

[0116] As preferred, fusing the first diagnostic results by DS evidence theory to determine the second diagnostic result comprises:

[0117] converting the first diagnostic results into evidence bodies in DS evidence theory, each evidence body representing a degree of belief of a signal source to a fault type, the signal sources including vibration sensors, image sensors and sound sensors;

[0118] constructing a belief function for each fault type in each signal source;

[0119] fusing the belief functions of the multiple evidence bodies to obtain a merged belief function;

[0120] selecting a fault type with a maximum belief value from the merged belief function as the second diagnostic result.

[0121] In some optional embodiments, selecting a fault type with a maximum belief value from the merged belief function as the second diagnostic result comprises:

[0122] updating the merged belief function based on a set conflict management strategy, and selecting a fault type corresponding to a maximum belief value from the updated belief function as the second diagnostic result;

[0123] wherein the set conflict management strategy comprises:

[0124] (1) quantifying a conflict degree between different signal sources based on a conflict, and when the conflict degree exceeds a preset threshold, selecting to exclude the conflict information from evidence synthesis to update the merged belief function, the conflict information being one or more of inconsistent diagnostic results, inconsistent probability distribution, inconsistent feature indicators, inconsistent time or space;

[0125] (2) adjusting a weight of the conflict evidence based on a reliability or historical data of the evidence source to update the merged belief function.

[0126] More specifically, regarding the definition of the above conflict information:

[0127] inconsistent diagnostic results: different signal sources give different diagnostic results for the same fault type.

[0128] inconsistent probability distribution: in a Bayesian fault diagnosis model, there is a significant difference in the probability distribution of the same fault type by different signal sources.

[0129] Inconsistency in features: After PCA dimensionality reduction, the feature indicators extracted by each signal source have obvious differences in numerical value or trend.

[0130] Inconsistency in time or space: Different signal sources do not match in the time point or spatial position of fault occurrence, for example, the vibration sensor detects the fault earlier than the image sensor.

[0131] In specific implementation, after obtaining the first diagnostic results of each sensor, DS evidence theory is used for fault fusion diagnosis. The specific steps are as follows:

[0132] (1) Convert the first diagnostic results into evidence bodies in DS evidence theory, and each evidence body represents the confidence degree of a signal source (such as vibration sensor, image sensor and sound sensor) to the fault type.

[0133] (2) Construct a belief function (Basic Probability Assignment, BPA) for each fault type in each signal source, and assign a belief function value to each fault type, indicating the credibility of the fault type.

[0134] (3) Fuse the belief functions of multiple evidence bodies to obtain the merged belief function. In the fusion process, the conflict management strategy set above is used to handle the conflict and uncertainty between evidences.

[0135] (4) According to the merged belief function, select the fault type with the maximum belief value as the second diagnostic result. This fault type is the most likely fault type, which is used to guide the subsequent maintenance and fault handling.

[0136] It should be further explained that in the present embodiment, the updated merged belief function based on the set conflict management strategy in actual application, different evidence sources may give conflicting information. DS theory measures these conflicts through conflict degree (K value). The greater the conflict degree, the lower the credibility of the final diagnostic result.

[0137] Specific steps:

[0138] In the synthesis process, when conflicts are encountered, conflicts need to be handled. There are usually two ways to handle them:

[0139] (1) Ignore the conflict information: exclude the conflict information from the synthesis.

[0140] (2) Adjust BPA: adjust the final BPA by reducing the weight of the conflicting evidence.

[0141] After completing the update of the fusion belief function, subsequent diagnostic reasoning is performed according to the synthesis result.

[0142] In the embodiment, according to the combined confidence function, the fault type with the maximum confidence value is selected as the second diagnostic result, and in the specific implementation, the following steps are included:

[0143] From the fused BPA, the fault type with the maximum BPA value is selected as the final diagnostic result. This fault type is the most likely fault type.

[0144] If the BPA values of certain fault types are very close, further diagnosis can be performed to confirm the final result.

[0145] According to the above embodiments of the present application, the present application can comprehensively consider the vibration, image and sound data of the mechanical equipment, extract more effective fault indicators, and realize more comprehensive fault identification. At the same time, through PCA dimension reduction processing and construction and training of the Bayesian fault diagnosis model, the complexity and cost of data processing are reduced. Finally, through DS evidence theory, the fault fusion diagnosis is performed, and the accuracy and reliability of the fault diagnosis are improved. This method can significantly improve the operation efficiency and safety of the mechanical equipment, and reduce the maintenance cost and production loss in practical application.

[0146] In combination with FIG. 3, in another embodiment of the present application, a multi-source data fusion fault diagnosis system based on DS evidence theory is proposed, which is applied to the multi-source data fusion fault diagnosis method based on DS evidence theory described above, and the system includes:

[0147] A data acquisition module is configured to acquire multi-source information of a target device, and the multi-source information includes vibration data, image data and sound data. Correspondingly, the data acquisition module includes a vibration sensor, an image sensor and a sound sensor.

[0148] A feature extraction module is configured to extract feature indicators corresponding to the vibration data, the image data and the sound data.

[0149] A principal component analysis module is configured to perform dimension reduction processing on the feature indicators by using a principal component analysis algorithm.

[0150] A first diagnostic module is configured to train a pre-constructed fault diagnosis model based on time-frequency domain features and naive Bayes according to the dimension-reduced feature indicators, and determine a first diagnostic result of a fault of the target device based on the fault diagnosis model.

[0151] A second diagnostic module is configured to fuse the first diagnostic result by using DS evidence theory to determine a second diagnostic result.

[0152] It should be noted that the modules in the above fault diagnosis system correspond to steps S1 to S5 in the implementation of the above fault diagnosis method, and the instances and application scenarios realized by the modules and the corresponding steps are the same, but are not limited to the content disclosed in the above embodiment 1.

[0153] In order to make the present application and its advantages more clearly understood, the following further explains the fault diagnosis method provided by the present application in conjunction with specific cases and relevant parts of the drawings.

[0154] 1. Data acquisition

[0155] The sensor collects the vibration data, sound, and image data of the actual bearing on site and inputs the server. The server can include the feature extraction module, the principal component analysis module, the first diagnosis module, and the second diagnosis module of the fault diagnosis system in the above embodiments. The corresponding vibration data is shown in FIG. 4 (in FIG. 4, the horizontal axis represents the corresponding time, and the vertical axis represents the vibration amplitude), the sound data is shown in FIG. 5 (in FIG. 5, the horizontal axis represents the corresponding time, and the vertical axis represents the sound waveform amplitude), the equipment image data is shown in FIG. 6, and the PCA graph after dimension reduction is shown in FIG. 7 (in FIG. 7, each coordinate system axis corresponds to a principal component, the first principal component (PC1) is the direction with the largest variance in the data, and the second principal component (PC2) is the direction orthogonal to the first principal component).

[0156] 2. Data feature extraction

[0157] 2.1. Extracting vibration data feature indexes, calculating the corresponding vibration indexes according to the acceleration vibration waveform data collected in FIG. 4, and extracting 12 indexes including the acceleration effective value, the acceleration peak-to-peak value, the acceleration pulse, the acceleration margin, the acceleration kurtosis, the acceleration skewness, the speed effective value, the speed peak-to-peak value, the speed pulse, the speed margin, the speed kurtosis, and the speed skewness. The vibration feature index array is calculated as follows:

[0158] a = [12, 106, 5.47, 10.48, 3.05, -0.048, 21, 72.89, 1.52, 0.01, 19.34, 1.99]

[0159] Selecting multiple devices including healthy and faulty device data, measuring multiple groups of data to establish a vibration data set A:

[0160] A = [a1, a2, a3,... an]

[0161] 2.2. Extracting sound data feature indexes, calculating the corresponding sound indexes according to the sound waveform data collected in FIG. 5, and extracting indexes including the effective value, the loudness, the voiceprint roughness, the frequency gravity center, the sharpness, the high-to-low frequency band energy ratio, the peak-to-peak value, the pulse, the margin, the kurtosis, and the skewness. The calculation indexes are shown as follows:

[0162] s = [435, 51, 976, 890, 0.264, 0.359, 3670, 8.042, 6.459, 3.69, -0.018]

[0163] Select multiple devices, including health and failure device data, measure multiple sets of data to establish a vibration data set S:

[0164] S = [s1,s2,s3,...sn]

[0165] 2.3, extract image data feature index, calculate the corresponding image data feature index according to the device picture collected according to figure 6, the extracted index is gray level co-occurrence matrix (GLCM) describing the spatial relationship between pixel gray in image, usually used to detect cracks, wear and other defects, commonly used GLCM features include contrast, homogeneity, roughness, complexity. The calculation index is as follows:

[0166] P = [65.15, 0.4594, 0.0284, 7.96144]

[0167] 3, PCA dimension reduction processing

[0168] Taking the vibration data as an example, the vibration data is reduced in dimension, and the following data is obtained:

[0169] The variance ratio of each principal component is [0.67284757, 0.10768493];

[0170] The eigenvalue of each principal component is [8.65089732, 1.3845205];

[0171] Cumulative variance ratio: [0.67284 0.7805325];

[0172] The scatter plot after dimension reduction is shown in figure 7, and the sound, image is also processed by the above method, and the corresponding principal component is obtained.

[0173] 4, construction of fault diagnosis model based on time-frequency domain features and naive Bayes

[0174] 4.1, sort the data set after dimension reduction, including corresponding features and labels.

[0175] 4.2, calculate the prior probability P(Ck) of each category, that is, the frequency of each category in all samples:

[0176]

[0177] 4.3, For each feature, calculate the probability of the feature given the class, P(Xi|Ck). If the feature is continuous data, it is usually assumed to follow a Gaussian distribution and the probability is calculated by the mean and variance. If it is discrete data, the conditional probability is calculated using frequencies.

[0178] 4.4, For a new sample X, calculate its posterior probability of belonging to each class according to Bayes' theorem:

[0179]

[0180] where P(X) is the normalization constant for all classes, which can be ignored because it is the same for each class.

[0181] 4.5, Select the maximum posterior probability: Calculate the posterior probability of all classes and select the class corresponding to the maximum value as the prediction result:

[0182]

[0183] 5, Fault diagnosis fusion

[0184] Based on the above embodiment, the diagnostic results of the sound, image and vibration three sensor data are fused by the DS evidence theory to obtain a final conclusion.

[0185] It should be noted that the algorithm mentioned in the present application can be realized in the host computer software in the form of signal processing, or can be realized in other forms such as digital chips, hardware circuits, etc. to realize the operation process.

[0186] Those skilled in the art can realize that the units and algorithm steps of each example described in combination with the embodiments disclosed herein can be realized in electronic hardware or a combination of computer software and electronic hardware. Whether the functions are realized in hardware or software depends on the specific application and design constraints of the technical solution. Professional technicians can use different methods to realize the described functions for each specific application, but such implementation should not be considered beyond the scope of the present application.

[0187] In addition, each functional module in each embodiment of the present application can be integrated in one processing unit, or each unit can exist physically alone, or two or more units can be integrated in one unit.

[0188] The above examples are only used to illustrate the technical solutions of the present application, and not to limit them; although the present application has been described in detail with reference to the foregoing examples, those skilled in the art will understand that the technical solutions recorded in the foregoing examples can still be modified, or some technical features therein can be replaced by equivalents; and these modifications or replacements do not drive the essence of the corresponding technical solutions out of the spirit and scope of the technical solutions of the embodiments of the present application.

Claims

1. A multi-source data fusion fault diagnosis method based on DS evidence theory, characterized in that, The method comprises: Collecting multi-source information of a target device, the multi-source information comprising vibration data, image data and sound data; Determining and extracting feature indicators corresponding to the vibration data, the image data and the sound data; Performing dimension reduction processing on the feature indicators by using a principal component analysis algorithm; Training a pre-constructed fault diagnosis model based on time-frequency domain features and a naive Bayes algorithm according to the dimension-reduced feature indicators, and determining a first diagnosis result of a fault of a device under test based on the fault diagnosis model; Fusing the first diagnosis result by using a DS evidence theory to determine a second diagnosis result.

2. The multi-source data fusion fault diagnosis method based on DS evidence theory according to claim 1, characterized in that, The determining and extracting of the feature indicators corresponding to the vibration data, the image data and the sound data comprises: Determining and extracting feature indicators of vibration data of a plurality of target devices, including acceleration effective value, acceleration peak-to-peak value, acceleration pulse, acceleration margin, acceleration kurtosis, acceleration skewness, velocity effective value, velocity peak-to-peak value, velocity pulse, velocity margin, velocity kurtosis and velocity skewness, to form a first feature set; Determining and extracting feature indicators of sound data of a plurality of target devices, including effective value, loudness, voiceprint roughness, frequency barycenter, sharpness, high-to-low frequency band energy ratio, peak-to-peak value, pulse, margin, kurtosis and skewness, to form a second feature set; Determining and extracting feature indicators of image data of a plurality of target devices, including texture feature, edge feature, temperature distribution feature and crack distribution feature, to form a third feature set.

3. The multi-source data fusion fault diagnosis method based on DS evidence theory according to claim 2, characterized in that, The dimension reduction processing on the feature indicators by using the principal component analysis algorithm comprises: Calculating a covariance matrix of each of the standardized feature sets; Calculating eigenvalues and eigenvectors of each of the covariance matrices based on a numerical algorithm; Determining target eigenvectors according to the sizes of the eigenvalues, and constructing a principal component space based on the target eigenvectors; Projecting the standardized covariance matrices into the principal component space to perform matrix multiplication, to obtain the dimension-reduced feature indicators.

4. The multi-source data fusion fault diagnosis method based on DS evidence theory according to claim 3, characterized in that, The training of the pre-constructed fault diagnosis model based on time-frequency domain features and the naive Bayes algorithm according to the dimension-reduced feature indicators, and the determination of the first diagnosis result of the fault of the device under test based on the fault diagnosis model, comprises: Determining a fault type of a target device according to a fault mode or a historical fault record of the target device, taking the fault type as a class label, and forming a training set comprising the dimension-reduced feature indicators and corresponding fault class labels; calculating a prior probability P(Ci) of each of the fault types in the training set k ); calculating a likelihood probability P(x i ∣C k ) for each feature indicator under each failure type The posterior probability P(C) of the dimension-reduced feature index X of the device under test belonging to each of the aforementioned fault types is calculated using Bayes' theorem. k |X), select the fault type with the highest posterior probability as the first diagnostic result, and the expression is: Among them, C k Indicates the fault type; P(C k ) is fault category C k The prior probability; P(X|C) k ) indicates that X belongs to category C k The conditional probability under the following conditions; P(C k |X) represents the fault category C given feature X. k The posterior probability; P(X) represents the normalization constant for all fault types; arg max is the choice C k The fault category that maximizes the posterior probability.

5. The multi-source data fusion fault diagnosis method based on DS evidence theory according to claim 4, characterized in that, The fusing of the first diagnosis result by using the DS evidence theory to determine the second diagnosis result, comprises: Converting the first diagnosis result into evidence bodies in the DS evidence theory, each of the evidence bodies representing a trust degree of a signal source to a fault type, the signal sources including vibration sensors, image sensors and sound sensors; Constructing a trust degree function, and assigning a trust degree function to each of the fault types in each of the signal sources; Fusing the trust degree functions of the evidence bodies to obtain a combined trust degree function; Selecting a fault type with a maximum trust degree value as the second diagnosis result according to the combined trust degree function.

6. The multi-source data fusion fault diagnosis method based on DS evidence theory according to claim 5, characterized in that, The selecting of the fault type with the maximum trust degree value as the second diagnosis result according to the combined trust degree function, comprises: updating the combined belief function based on a set conflict management strategy, and selecting a fault type corresponding to a maximum value of the updated belief function as the second diagnosis result; wherein the set conflict management strategy comprises: (1) quantifying the conflict degree between different signal sources based on the conflict, and when the conflict degree exceeds a preset threshold, selecting to exclude the conflict information from evidence synthesis to update the combined belief function; the conflict information is one or more of inconsistent diagnosis results, inconsistent probability distribution, inconsistent feature indicators, inconsistent time or space; (2) adjusting the weight of the conflict evidence based on the reliability or historical data of the evidence source to update the combined belief function.

7. A DS evidence theory based multi-source data fusion fault diagnosis system, applied to execute the DS evidence theory based multi-source data fusion fault diagnosis method of any one of claims 1-6, characterized in that, The system comprises: a data acquisition module for acquiring multi-source information of a target device, the multi-source information comprising vibration data, image data and sound data; a feature extraction module for extracting feature indicators corresponding to the vibration data, the image data and the sound data; a principal component analysis module for performing dimensionality reduction processing on the feature indicators using a principal component analysis algorithm; a first diagnosis module for training a pre-constructed fault diagnosis model based on time-frequency domain features and Naive Bayes based on the dimensionally reduced feature indicators, and determining a first diagnosis result of a fault of the device under test based on the fault diagnosis model; a second diagnosis module for fusing the first diagnosis result using DS evidence theory to determine a second diagnosis result.

Images