Adaptive rolling bearing early fault real-time detection method based on data fusion

Abstract

The invention belongs to the technical field of fault diagnosis, and discloses a self-adaptive rolling bearing early fault real-time detection method based on data fusion, and the method comprises the steps: determining the installation positions of a vibration sensor and a rotation speed sensor, and setting detection parameters; obtaining one-dimensional synchronous vibration data X; preprocessing the one-dimensional synchronous vibration data X; calculating the mean square value of the preprocessed data, the cosine similarity of the synchronous data of two adjacent dimensions and the DTW distance; constructing a feature matrix; calculating an F statistic and a statistic value pv of the feature matrix; and judging whether the bearing has an early fault or not and realizing self-adaptive data acquisition. According to the method, a plurality of fault features of the bearing are extracted, and fault information of the plurality of fault features is fused by adopting Hotelling T2, so that a standardized fault index is obtained, the reliability, the universality and the actual operability of bearing fault detection are improved, the method can be applied to small sample data and high-dimensional data, the detection deviation caused by dimension change is avoided, and the detection accuracy is improved. Therefore, the method adapts to the number change caused by adaptive sampling. Besides, the number of synchronous sampling data points in each rotation period of the bearing is determined in a self-adaptive mode, the data acquisition amount is reduced, and therefore the data acquisition, processing, transmission and storage cost is reduced.

Description

Technical Field

[0001] The present invention belongs to the technical field of rotating machinery fault detection and diagnosis, and relates to a real-time bearing fault detection method, in particular to a real-time adaptive rolling bearing early fault detection method based on data fusion. Background Art

[0002] Rotating machinery is a type of mechanical equipment that uses the motion of rotating parts to achieve functions such as energy conversion, material transportation, and machining. Rotating machinery plays a vital role in energy generation, chemical production, and transportation. For example, in energy production, a turbine, as a type of rotating machinery, uses the energy of water flow to propel the rotor. In transportation, rotating parts such as the crankshaft in an automobile engine are crucial to the vehicle's power output and performance. Rolling bearings, as key components of rotating machinery, are susceptible to damage due to long-term high-load operation. Bearing failures often cause a chain reaction, leading to gear wear, shaft breakage, and even complete machine damage. Therefore, real-time fault detection and regular maintenance of bearings are essential to ensure the stable operation of rotating machinery.

[0003] Currently, most real-time detection technologies for early-stage bearing faults utilize either fixed or variable detection intervals. Fixed detection intervals involve testing the bearing at fixed intervals during the inspection process. Variable detection intervals dynamically adjust the detection interval based on the bearing's degradation status. When bearing degradation is mild, a longer detection interval is used, while when bearing degradation is severe, a shorter detection interval is used. Both fixed and variable detection intervals suffer from a high missed detection rate. Bearing failures can occur at any time during the bearing's use, not just during fault detection. Therefore, neither detection method can overcome this high missed detection rate.

[0004] Furthermore, real-time detection of early bearing faults primarily focuses on fault feature extraction. Common early bearing fault features include standard deviation, correlation coefficient, and various entropy types. A Chinese invention patent (Patent No. 2023112773207) discloses a rolling bearing fault diagnosis method that extracts the correlation coefficient as a fault feature. However, extracting a single fault feature can only reflect a specific aspect of a rolling bearing fault and cannot fully describe the fault. Furthermore, a single fault feature can sometimes be poorly adaptable to noise and dynamic changes.

[0005] In order to overcome the shortcomings of the above fault detection methods, the present invention discloses a real-time detection method for early faults of adaptive rolling bearings based on data fusion. 2This method fuses fault information from multiple fault signatures to comprehensively characterize various bearing faults. Furthermore, based on the fault indicators derived from the fused data and considering the bearing's degradation patterns, it adaptively determines the number of sampling data points per shaft revolution. Specifically, during periods of mild bearing degradation, a larger sampling interval is used, while during periods of severe degradation, a smaller sampling interval is used. This effectively reduces data collection, transmission, and processing costs while ensuring continuous detection, making it more suitable for industrial applications requiring high safety. Summary of the Invention

[0006] Purpose of the invention: The purpose of the present invention is to provide a real-time detection method for early failure of an adaptive rolling bearing based on data fusion. The method extracts multiple fault features of the bearing and adopts Hotelling T 2 By fusing fault information from multiple fault characteristics, standardized fault indicators are obtained, which improves the reliability, universality and practical operability of bearing fault detection, enabling it to be applied to small sample data and high-dimensional data, avoiding inspection bias caused by dimensionality changes, and thus adapting to quantity changes caused by adaptive sampling; the sampling interval is adaptively adjusted according to the fault indicator, which reduces the amount of data collected and effectively reduces the cost of data collection, transmission and processing, making it more suitable for the industrial field.

[0007] Technical solution: The present invention discloses a real-time detection method for early-stage failure of adaptive rolling bearings based on data fusion, which specifically includes the following steps:

[0008] Step s1: Determine the installation positions of the vibration sensor and the speed sensor, and set the detection parameters;

[0009] Step s2: Acquire one-dimensional synchronous vibration data X;

[0010] Step s3: preprocessing the one-dimensional synchronous vibration data X;

[0011] Step s4: Calculate the mean square value of the preprocessed data, the cosine similarity and DTW distance of the synchronized data of two adjacent dimensions;

[0012] Step s5: construct feature matrix;

[0013] Step s6: Calculate the F statistic and statistical value p of the feature matrix v ;

[0014] Step s7: Determine whether the bearing has an early fault and implement adaptive data collection.

[0015] Furthermore, in step s1, the vibration sensor is installed near the bearing to be tested to measure the bearing vibration data; the speed sensor is installed near the shaft to measure the shaft speed data; the detection parameters include: the number of shaft revolutions b, the initial number a0 of the amount of synchronous sampling data a per shaft revolution, the minimum number a min , the maximum number a max , threshold T. When the bearing being tested is a brand new bearing, set a0 to the minimum number a min ; When the health status of the detected bearing is unknown, set Among them, round() is a rounding function, and the result is an integer; the threshold T is obtained based on statistical experience.

[0016] Furthermore, in step s2, according to the preset number of synchronous sampling points a per rotation of the shaft and the number of rotations b of the shaft, the rotation speed signal of the shaft is used as a reference signal to obtain synchronous vibration data during the working process of the bearing. The obtained one-dimensional synchronous vibration data can be expressed as X = {x(1), x(2), x(3)...x(ab)};

[0017] Furthermore, the preprocessing process of step s3 is as follows:

[0018] The preprocessing of one-dimensional synchronous vibration data X includes three parts: standardization processing, data expansion processing, and zero-centering decomposition processing.

[0019] First, the one-dimensional synchronous vibration data X is normalized using the Z-score normalization method. The normalized one-dimensional synchronous vibration data X′ is shown as follows:

[0020]

[0021] in, Represents the mean value of one-dimensional synchronous vibration data X, σ X represents the standard deviation of the one-dimensional synchronous vibration data X. Then, according to the number of synchronous sampling points a per shaft revolution and the number of shaft revolutions b, the standardized synchronous vibration data X′ is expanded into a-dimensional synchronous data Y. The expanded a-dimensional synchronous data Y is shown as follows:

[0022]

[0023] make Then the a-dimensional synchronous data Y can be expressed as follows:

[0024]

[0025] Finally, the zero-centering technology is used to decompose the multidimensional synchronous data Y to obtain the multidimensional synchronous data Y1 and Y2, as shown below:

[0026]

[0027] in, Represents the mean of the i-th dimension data in the multidimensional data Y, and can also be seen as the mean of the i-th row data in the multidimensional data Y.

[0028]

[0029] make Then the a-dimensional synchronous data Y1 can be expressed as follows:

[0030]

[0031] Furthermore, in step s4: the mean square value data S1 of each dimension of the multi-dimensional synchronous data Y1 is calculated respectively: [s1(1) s1(2) s1(3) … s1(i) … s1(a)] T , where the mean square value of the i-th dimension synchronization data of the multi-dimensional synchronization data Y1 is:

[0032]

[0033] Furthermore, in step s4, the cosine similarities of adjacent two-dimensional synchronous data in the multi-dimensional synchronous data Y1 are calculated respectively to obtain cosine similarity data R:

[0034] R=[R(1) R(2) R(3) … R(i) … R(a-1)] T ,

[0035] Where R(i) represents the cosine similarity between the i-th dimension and the i+1-th dimension of the multidimensional synchronous data Y1.

[0036]

[0037] Furthermore, in step s4, the DTW algorithm is used to calculate the DTW distances of adjacent two-dimensional synchronization data in the multi-dimensional synchronization data Y1 to obtain DTW distance data D:

[0038] D=[d(1) d(2) d(3) … d(i) … d(a-1)] T ,

[0039] Wherein, d(i) represents the DTW distance between the i-th dimension and the (i+1)-th dimension synchronization data in the multi-dimensional synchronization data Y1.

[0040] Furthermore, in step s4, the mean square value data S2 of each dimension of the multi-dimensional synchronous data Y2 is calculated respectively: [s2(1) s2(2) s2(3) … s2(i) … s2(a)] T, where the mean square value of the i-th dimension synchronous data in the multi-dimensional synchronous data Y2 is:

[0041]

[0042] Furthermore, the implementation process of step s5 is as follows: the first a-1 mean square values of the two mean square value data S1 and S2, the cosine similarity data R and the DTW distance data D are respectively taken to construct the feature matrix C:

[0043]

[0044] Furthermore, the implementation process of step s6 is as follows:

[0045] First, calculate the mean vector of the fault characteristics in, are the mean values of data S1, R, D, and S2 respectively; then the fault information of the feature matrix C is accumulated to obtain Hotelling T 2 value:

[0046] T 2 =(a-1)·(Z-μ0) T ∑ -1 (Z-μ0),

[0047] Among them, a-1 represents the number of samples of each feature in the feature matrix C,

[0048] μ0 represents the mean vector of the four data S1, R, D, and S2 under the healthy state of the bearing,

[0049] ∑ -1 is the inverse matrix of the covariance matrix of the feature matrix C;

[0050] Furthermore, in step s6, Hotelling T 2 The process of converting statistics into F statistics is as follows:

[0051]

[0052] Where p represents the number of features in the feature matrix C;

[0053] Furthermore, in step s6, the statistical value p is calculated based on the F statistic. v , the formula is:

[0054] p v =1-CDF F(F，a-1，(a-1)-p) ,

[0055] Among them, CDF F is the cumulative distribution function of the F distribution,

[0056] F is the F statistic value.

[0057] Furthermore, the implementation process of step s7 of determining whether the bearing is faulty and implementing adaptive data collection is as follows:

[0058] The statistical value p v As a fault indicator; According to the hypothesis testing principle, it is assumed that the bearing is in a healthy state, that is, the null hypothesis H0, and the bearing is in a fault state as the alternative hypothesis H1; If the fault indicator p v If it is greater than T, the bearing failure is judged to be a low-probability event. At this time, the number of synchronous sampling data points a per rotation period of the shaft is set to Among them, round() is a rounding function, the result is an integer, and returns to step s2 to continue the bearing early fault detection in the next time window; if the fault indicator p v If the value is less than or equal to the threshold value T, the alternative hypothesis H1 is accepted, and the rolling bearing is judged to be faulty, and an alarm is issued, and then the number of synchronous sampling data points per rotation period of the shaft is adaptively adjusted to a max , and return to step 2 to continue bearing early fault detection in the next time window until engineering personnel participate in resolving the bearing fault.

[0059] Beneficial effects:

[0060] (1) The adaptive rolling bearing early fault real-time detection method based on data fusion disclosed in the present invention overcomes the disadvantage that a single fault feature cannot comprehensively evaluate various bearing faults by calculating four fault features, namely, the mean square value of the multi-dimensional synchronous data Y1, the cosine similarity and DTW distance of the two adjacent dimensional synchronous data, and the mean square value of the multi-dimensional synchronous data Y2. It improves the characterization ability of bearing faults, reduces the missed alarm rate and false alarm rate of bearing early fault detection, and enhances the robustness of the fault detection method.

[0061] (2) The adaptive rolling bearing early fault real-time detection method based on data fusion disclosed by the present invention overcomes the disadvantage of low reliability of existing methods and uses the statistical method Hotelling T 2 Fault information that integrates multiple fault characteristics can comprehensively characterize various bearing faults and reliably detect early-stage bearing faults.

[0062] (3) The adaptive real-time detection method for early-stage rolling bearing faults based on data fusion disclosed in the present invention reduces the amount of data collected by adaptively determining the number of synchronously sampled data points within each rotation period of the bearing, thereby reducing the cost of data collection, processing, transmission and storage;

[0063] (4) The present invention discloses a real-time detection method for early failure of adaptive rolling bearings based on data fusion, using Hotelling T 2 Statistics and F statistics are used to obtain standardized fault indicators, which improve the reliability, universality and practical operability of bearing fault detection. They can then be applied to small sample data and high-dimensional data, avoiding inspection bias caused by dimensional changes, and adapting to changes in sampling quantity caused by adaptive sampling. BRIEF DESCRIPTION OF THE DRAWINGS

[0064] Figure 1 This is a flow chart of the real-time detection method for early-stage faults of adaptive rolling bearings based on data fusion disclosed by the present invention;

[0065] Figure 2 This is a diagram showing the test results when the bearing is fault-free according to an embodiment of the present invention;

[0066] Figure 3 This is a schematic diagram of the number of synchronously sampled data points in each rotation period of the shaft when the bearing is fault-free according to an embodiment of the present invention;

[0067] Figure 4 This is a schematic diagram of the detection results of an early bearing failure according to an embodiment of the present invention;

[0068] Figure 5 This is a schematic diagram of the number of synchronously sampled data points per rotation period of the rotating shaft when a bearing fails in its early stage according to an embodiment of the present invention. DETAILED DESCRIPTION

[0069] The implementation process of the method of the present invention is described below with reference to the accompanying drawings.

[0070] As attached Figure 1 As shown, the present invention provides a method for real-time detection of early-stage rolling bearing faults based on adaptive data fusion, comprising the following steps:

[0071] Step s1: Determine the installation positions of the vibration sensor and the speed sensor, and set the detection parameters;

[0072] The vibration sensor is installed near the bearing to be tested to measure the bearing vibration data; the speed sensor is installed near the shaft to measure the shaft speed data, such as the installation position and method disclosed in the invention patent (application number: CN 110186510B); the detection parameters include: the number of shaft revolutions b, the initial number a0 of synchronous sampling data per shaft revolution period, the minimum number a min , the maximum number a max , threshold T; when the detected bearing is a fault-free bearing, set a0 to a min ; When the health status of the detected bearing is unknown, set Among them, round() is a rounding function, and the result is an integer.

[0073] In this example, the initial number of synchronous sampling points per shaft revolution is a0=51, the number of shaft revolutions is b=10, and the minimum number is a min =2, maximum number a max =100, the threshold T is selected as 0.05; according to statistical experience, T = 0.05 means that the probability of a second type error in bearing detection is 5%;

[0074] Step s2: Acquire one-dimensional synchronous vibration data X;

[0075] According to the preset initial number of synchronous sampling points per shaft revolution a0=51 and the shaft revolution number b=10, the shaft speed signal is used as the reference signal to obtain the synchronous vibration data of the bearing during operation. The obtained one-dimensional synchronous vibration data can be expressed as X={x(1),x(2),x(3)...x(510)}.

[0076] Step s3: preprocessing the one-dimensional synchronous vibration data X;

[0077] The preprocessing of one-dimensional synchronous vibration data X includes three parts: standardization processing, data expansion processing, and zero-centering decomposition processing.

[0078] First, the one-dimensional synchronous vibration data X is normalized using the Z-score normalization method. The normalized one-dimensional synchronous vibration data X′ is shown as follows:

[0079]

[0080] in, Represents the mean value of one-dimensional synchronous vibration data X, σ X represents the standard deviation of the one-dimensional synchronous vibration data X; then, based on the number of synchronous sampling points a = 51 per shaft revolution and the number of shaft revolutions b = 10, the standardized synchronous vibration data X′ is expanded into a-dimensional synchronous data Y. The expanded a-dimensional synchronous data Y is shown below:

[0081]

[0082] make Then the a-dimensional synchronous data Y can be expressed as follows:

[0083]

[0084] Finally, the zero-centering technology is used to decompose the multidimensional synchronous data Y to obtain the multidimensional synchronous data Y1 and Y2, as shown below:

[0085]

[0086] in, Represents the mean of the i-th dimension data in the multidimensional data Y, and can also be seen as the mean of the i-th dimension data in the multidimensional data Y.

[0087]

[0088] make Then the a-dimensional synchronous data Y1 can be expressed as follows:

[0089]

[0090] Step s4: Calculate the mean square value of the preprocessed data, the cosine similarity and DTW distance of the synchronized data of two adjacent dimensions;

[0091] Calculate the mean square value S1 of each dimension of the multidimensional synchronous data Y1 respectively: [s1(1) s1(2) s1(3) …s1(i) …s1(51)] T , where the mean square value of the i-th dimension synchronous data is:

[0092]

[0093] Calculate the cosine similarity of adjacent two-dimensional synchronous data in the multi-dimensional synchronous data Y1 respectively to obtain the cosine similarity data R:

[0094] R=[R(1) R(2) R(3) … R(i) … R(50)] T ,

[0095] Where R(i) represents the cosine similarity between the i-th dimension and the i+1-th dimension synchronization data in the multidimensional synchronization data Y1.

[0096]

[0097] The DTW algorithm is used to calculate the DTW distance of two adjacent dimensional synchronous data in the multi-dimensional synchronous data Y1, and the DTW distance data D is obtained:

[0098] D=[d(1)d(2)d(3)…d(i)…d(50)] T ,

[0099] Wherein, d(i) represents the DTW distance between the i-th dimension and the i+1-th dimension synchronization data in the multidimensional synchronization data Y1;

[0100] Calculate the mean square value data S2 of each dimension of the multidimensional synchronous data Y2 respectively = [s2(1)s2(2)s2(3)…s2(i)…s2(51)] T , where the mean square value of the i-th dimension synchronous data in the multi-dimensional synchronous data Y2 is:

[0101]

[0102] Step s5: construct feature matrix;

[0103] For the detection of the first time window, the first 50 mean square values of the two mean square value data S1 and S2, the cosine similarity data R and the DTW distance data D are taken to construct the fault feature matrix C:

[0104]

[0105] Step s6: Calculate the F statistic and statistical value p of the feature matrix v ;

[0106] The Hotelling T of the calculation fault feature matrix C 2 The statistical process is as follows:

[0107] First calculate the mean vector of the fault feature matrix C in, are the mean values of data S1, R, D, and S2 respectively; then the fault information of the feature matrix C is accumulated to obtain Hotelling T 2 value:

[0108] T 2 =(a-1)·(Z-μ0) T ∑ -1 (Z-μ0),

[0109] Among them, a-1 represents the number of samples of each feature in the feature matrix C,

[0110] μ0 represents the mean vector of the four data S1, R, D, and S2 under the healthy state of the bearing,

[0111] ∑ -1 is the inverse matrix of the covariance matrix ∑ of the feature matrix C;

[0112] In this example, μ0 is calculated based on a large amount of sample data under the healthy state of the bearing, μ0 = [0.9-0.942.410.09] T .

[0113] Hotelling T 2 The process of converting statistics into F statistics is as follows:

[0114]

[0115] Where p represents the number of features in the feature matrix C;

[0116] According to the F statistic, calculate the statistical value pv , the formula is:

[0117] p v =1-CDF F(F，a-1，(a-1)-p) ,

[0118] Among them, CDF F is the cumulative distribution function of the F distribution

[0119] F is the F statistic value.

[0120] Step s7: Determine whether the bearing has an early fault and implement adaptive data collection.

[0121] The statistical value p v As a fault indicator; According to the hypothesis testing principle, it is assumed that the bearing is in a healthy state, that is, the null hypothesis H0, and the bearing is in a fault state as the alternative hypothesis H1; If the fault indicator p v If it is greater than T, the bearing failure is judged to be a low-probability event. At this time, the number of synchronous sampling data points a per rotation period of the shaft is set to Among them, round() is a rounding function, the result is an integer, and returns to step s2 to continue the bearing early fault detection in the next time window; if the fault indicator p v If the value is less than or equal to the threshold value T, the alternative hypothesis H1 is accepted, and the rolling bearing is judged to be faulty, and an alarm is issued, and then the number of synchronous sampling data points per rotation period of the shaft is adaptively adjusted to a max , and return to step 2 to continue bearing early fault detection in the next time window until engineering personnel participate in resolving the bearing fault.

[0122] The present invention discloses a method for real-time detection of early-stage rolling bearing faults based on adaptive data fusion. First, the one-dimensional synchronous vibration data X is preprocessed, including standardization, data expansion, and zero-centering, which can improve the performance of the model and the accuracy of detection. Then, a feature matrix is constructed and the statistical method Hotelling T is used to calculate the early-stage rolling bearing faults. 2 Get T 2 value, overcome the shortcoming that a single fault feature cannot fully evaluate the bearing fault, improve the fault characterization ability, reduce the missed alarm rate and false alarm rate of early bearing fault detection, and enhance the robustness of the fault detection method; finally, T 2 The value is converted into F statistics to obtain the standardized fault index p v , after improving the reliability, universality and practical operability of bearing fault detection, it can be applied to small sample data and high-dimensional data, avoiding the inspection bias caused by dimensionality change, and thus adapting to the change in sampling quantity caused by adaptive sampling. In addition, the patent of this invention is based on the fault indicator p vAdaptively adjusting the number of samples taken per bearing revolution can reduce data acquisition, processing, transmission, and storage costs.

[0123] Figure 2 This is the detection result when the bearing is fault-free. As can be seen from the figure, if the fault indicators are all greater than the threshold T, the bearing is judged to be in a fault-free state, and the number of synchronous sampling data points a per rotation period of the shaft detected in the next time window is adaptively adjusted to And return to step 2 to continue bearing early fault detection in the next time window.

[0124] Figure 3 The number of synchronously sampled data points, a, per shaft revolution is shown when the bearing is fault-free. During each fault detection, the number of synchronously sampled data points per shaft revolution varies with the fault indicator. Because the bearing is fault-free, the number of synchronously sampled data points, a, per shaft revolution is always less than the maximum value of 100 during each detection. This reduces the amount of data collected and, in turn, lowers data transmission, processing, and storage costs.

[0125] Figure 4 The figure shows the detection results of the bearing failure. As can be seen from the figure, when the fault detection index is less than the threshold T in the 2nd, 5th, 9th, and 10th time windows, it is considered that the bearing has an early fault.

[0126] Figure 5 is the number of synchronously sampled data points per shaft revolution when a bearing early-stage fault occurs. During the second, fifth, ninth, and tenth time windows, an early-stage bearing fault is detected. Therefore, the number of synchronously sampled data points per shaft revolution in the next detection time window is increased to its maximum value.

[0127] The above embodiments are intended only to illustrate the technical concepts and features of the present invention. Their purpose is to enable those skilled in the art to understand the contents of the present invention and implement them accordingly. They are not intended to limit the scope of protection of the present invention. Any equivalent changes or modifications made in accordance with the spirit of the present invention are intended to be covered by the scope of protection of the present invention.

Claims

1. The present invention discloses a real-time detection method for early-stage failure of an adaptive rolling bearing based on data fusion, the method comprising the following steps: Step s1: Determine the installation positions of the vibration sensor and the speed sensor, and set the detection parameters; Step s2: Acquire one-dimensional synchronous vibration data X; Step s3: preprocessing the one-dimensional synchronous vibration data X; Step s4: Calculate the mean square value of the preprocessed data, the cosine similarity and DTW distance of the synchronized data of two adjacent dimensions; Step s5: construct feature matrix; Step s6: Calculate the F statistic and statistical value p of the feature matrix v ; Step s7: Determine whether the bearing has an early fault and implement adaptive data collection.

2. The method for real-time detection of early-stage rolling bearing faults based on adaptive data fusion according to claim 1 is characterized in that: The vibration sensor in step s1 is installed near the bearing to be tested to measure the bearing vibration data; the speed sensor is installed near the shaft to measure the shaft speed data; the detection parameters include: the shaft speed b, the initial number a0 of the synchronous sampling data volume a per shaft rotation period, the minimum number a min , the maximum number a max , threshold T; when the bearing being tested is a new bearing, set a0 to the minimum number a min ; When the health status of the detected bearing is unknown, set Among them, round() is a rounding function, and the result is an integer; the threshold T is obtained based on statistical experience.

3. The method for real-time detection of early-stage rolling bearing faults based on adaptive data fusion according to claim 1 is characterized in that: In step s2, based on the preset number of synchronous sampling points a per shaft revolution and the shaft revolution number b, the shaft speed signal is used as a reference signal to obtain synchronous vibration data during the bearing operation process. The obtained one-dimensional synchronous vibration data can be expressed as X = {x(1), x(2), x(3)...x(ab)}.

4. The method for real-time detection of early-stage rolling bearing faults based on adaptive data fusion according to claim 1 is characterized in that: The preprocessing process described in step s3 is as follows: the preprocessing of the one-dimensional synchronous vibration data X includes three parts: standardization processing, data expansion processing, and zero-centering decomposition processing. First, the one-dimensional synchronous vibration data X is normalized using the Z-score normalization method. The normalized one-dimensional synchronous vibration data X′ is shown as follows: in, Represents the mean value of one-dimensional synchronous vibration data X, σ X represents the standard deviation of the one-dimensional synchronous vibration data X. Then, according to the number of synchronous sampling points a per shaft revolution and the number of shaft revolutions b, the standardized synchronous vibration data X′ is expanded into a-dimensional synchronous data Y. The expanded a-dimensional synchronous data Y is shown as follows: make Then the a-dimensional synchronous data Y can be expressed as follows: Finally, the zero-centering technology is used to decompose the multidimensional synchronous data Y to obtain the multidimensional synchronous data Y1 and Y2, as shown below: in, Represents the mean of the i-th dimension data in the multidimensional data Y, and can also be seen as the mean of the i-th row data in the multidimensional data Y. make Then the a-dimensional synchronous data Y1 can be expressed as follows:

5. The method for real-time detection of early-stage rolling bearing faults based on adaptive data fusion according to claim 1 is characterized in that: Step s4 calculates the mean square value of each dimension of the multi-dimensional synchronous data Y1: S1 = [s1(1)s1(2)s1(3) ... s1(i) ... s1(a)] T , where the mean square value of the i-th dimension synchronization data of the multi-dimensional synchronization data Y1 is: Calculate the cosine similarity of adjacent two-dimensional synchronous data in the multi-dimensional synchronous data Y1 respectively to obtain the cosine similarity data R: R=[R(1) R(2) R(3) … R(i) … R(a-1)] T , Where R(i) represents the cosine similarity between the i-th dimension and the i+1-th dimension of the multidimensional synchronous data Y1. The DTW algorithm is used to calculate the DTW distance of two adjacent dimensional synchronous data in the multi-dimensional synchronous data Y1, and the DTW distance data D is obtained: D=[d(1) d(2) d(3) … d(i) … d(a-1)] T , Wherein, d(i) represents the DTW distance between the i-th dimension and the i+1-th dimension synchronization data in the multidimensional synchronization data Y1; Calculate the mean square value data S2 of each dimension of the multidimensional synchronous data Y2 respectively: [s2(1) s2(2) s2(3) …s2(i) …s2(a)] T , where the mean square value of the i-th dimension synchronous data in the multi-dimensional synchronous data Y2 is:

6. The method for real-time detection of early-stage rolling bearing faults based on adaptive data fusion according to claim 1, characterized in that: In step s5, the first a-1 mean square values of the two mean square value data S1 and S2, the cosine similarity data R and the DTW distance data D are respectively taken to construct the feature matrix C:

7. The method for real-time detection of early-stage rolling bearing faults based on adaptive data fusion according to claim 1, characterized in that: The implementation process of step s6 is as follows: First, calculate the mean vector of the fault characteristics in; are the mean values of data S1, R, D, and S2 respectively; then the fault information of the feature matrix C is accumulated to obtain Hotelling T 2 value: T 2 =(a-1)·(Z-μ0) T ∑ -1 (Z-μ0), Among them, a-1 represents the number of samples of each feature in the feature matrix C, μ0 represents the mean vector of the four data S1, R, D, and S2 under the healthy state of the bearing, ∑ -1 is the inverse matrix of the covariance matrix of the feature matrix C; Hotelling T 2 The process of converting statistics into F statistics is as follows: Where p represents the number of features in the feature matrix C; According to the F statistic, calculate the statistical value p v , the formula is: p v =1-CDF F(F，a-1，(a-1)-p) , Among them, CDF F is the cumulative distribution function of the F distribution, F is the F statistic value.

8. The method for real-time detection of early-stage rolling bearing faults based on adaptive data fusion according to claim 1, characterized in that: The process of determining whether the bearing is faulty and implementing adaptive data acquisition in step s7 is as follows: The statistical value p v As a fault indicator; According to the hypothesis testing principle, it is assumed that the bearing is in a healthy state, that is, the null hypothesis H0, and the bearing is in a fault state as the alternative hypothesis H1; If the fault indicator p v If it is greater than T, the bearing failure is judged to be a low-probability event. At this time, the number of synchronous sampling data points a per rotation period of the shaft is set to Among them, round() is a rounding function, the result is an integer, and returns to step s2 to continue the bearing early fault detection in the next time window; if the fault indicator p v If the value is less than or equal to the threshold value T, the alternative hypothesis H1 is accepted, and the rolling bearing is judged to be faulty, and an alarm is issued, and then the number of synchronous sampling data points per rotation period of the shaft is adaptively adjusted to a max , and return to step 2 to continue bearing early fault detection in the next time window until engineering personnel participate in resolving the bearing fault.

Images