A multi-source heterogeneous data driven large rotating machinery fault diagnosis method

Abstract

A multi-source heterogeneous data driven large rotating machinery fault diagnosis method, which first collects one-dimensional data composed of low-frequency vibration signals, medium-frequency vibration signals and temperature signals, and two-dimensional data composed of operation images at the same time, then fuses the one-dimensional data and two-dimensional data by Bayes theorem to obtain a diagnosis result, at the same time, obtains a diagnosis result B according to the low-frequency vibration signals and medium-frequency vibration signals, and obtains a diagnosis result C according to the temperature signals, then sums up the diagnosis results A, B and C according to the weighted summation method to obtain the final diagnosis result, if the final diagnosis result is greater than or equal to 0.5, it is judged that there is a fault, if it is less than 0.5, it is judged that there is no fault. The design not only monitors multiple signals, but also has good fault diagnosis effect.

Description

Technical Field

[0001] This invention relates to a fault diagnosis technology for large rotating machinery, belonging to the field of mechanical fault diagnosis, and particularly to a fault diagnosis method for large rotating machinery driven by multi-source heterogeneous data. Background Technology

[0002] With the advancement of technology, mechanical equipment is constantly developing towards higher speeds, heavier loads, and greater stability, resulting in increasingly complex mechanical structures and harsher working environments. During service, the health of mechanical equipment components gradually deteriorates over time. Therefore, research on structural health monitoring technologies for mechanical equipment is of great significance for ensuring safe operation and improving production efficiency.

[0003] Rotating machinery, in particular, plays an important role in engineering applications. The increasingly complex structure and operating conditions of rotating machinery make it more and more difficult to diagnose faults in harsh environments. Therefore, it is especially necessary to adopt effective fault diagnosis methods.

[0004] Given the complexity and harshness of the working environment of large rotating machinery, electrical sensors are not suitable. Therefore, fiber optic sensors have emerged in the prior art to monitor large rotating machinery and analyze its operating status, such as fault diagnosis. However, existing technologies mostly rely on a single type of sensor for feature extraction and classification, obtaining a single type of monitoring signal without taking into account the extremely complex operating conditions of large rotating machinery. When a fault occurs, it will generate multiple vibration signals, such as low-frequency and medium-frequency signals, as well as other signals. Therefore, it is difficult to determine the type of fault using only a single type of sensor, resulting in low accuracy and a high risk of missed or false diagnoses.

[0005] The information disclosed in this background section is intended only to enhance understanding of the overall background of this application and should not be construed as an admission or in any way implying that the information constitutes prior art known to those skilled in the art. Summary of the Invention

[0006] The purpose of this invention is to overcome the shortcomings and problems of existing technologies, such as single monitoring signals and poor diagnostic effects, and to provide a multi-source heterogeneous data-driven fault diagnosis method for large rotating machinery with multiple monitoring signals and better diagnostic effects.

[0007] To achieve the above objectives, the technical solution of the present invention is as follows: See Figure 1 and Figure 2 A fault diagnosis method for large rotating machinery driven by multi-source heterogeneous data, the fault diagnosis method comprising the following steps:

[0008] The first step, data acquisition: At the same time, low-frequency vibration signals and medium-frequency vibration signals of the large rotating machinery are acquired through low-frequency acceleration sensors and medium-frequency acceleration sensors; temperature signals of the rotor in the large rotating machinery are acquired through temperature sensors; and running images of the frame in the large rotating machinery are acquired through cameras. The low-frequency vibration signals, medium-frequency vibration signals, and temperature signals are classified as one-dimensional data, and the running images are classified as two-dimensional data.

[0009] The second step, the multi-source heterogeneous step: The diagnostic results from A, B, and C are summed using a weighted summation method to obtain the final diagnostic result. If the final diagnostic result is greater than or equal to 0.5, a fault is determined; if it is less than 0.5, no fault is determined. The formula for summation is:

[0010] Final diagnosis result = Weight of A * Diagnosis result of A + Weight of B * Diagnosis result of B + Weight of C * Diagnosis result of C;

[0011] The diagnostic result of A refers to: first, inputting one-dimensional data into the Stacking model to obtain a one-dimensional diagnostic result; at the same time, inputting two-dimensional data into the deep learning model to obtain a two-dimensional diagnostic result; and then fusing the one-dimensional and two-dimensional diagnostic results using Bayes' formula to obtain the diagnostic result of A. The diagnostic result is a value greater than or equal to 0.5 but less than 1.

[0012] The diagnostic result of B refers to the following: the low-frequency vibration signal and the medium-frequency vibration signal are substituted into the K-means algorithm optimized by the genetic algorithm to obtain the diagnostic result of B. If the diagnostic result is a fault, it is 1; otherwise, it is 0.

[0013] The diagnostic result of C refers to the result obtained by substituting the temperature signal into the base model. The result is 1 if the fault is detected and 0 if the fault is not detected. The base model can be any one of Random Forest, KNN, or LightGBM.

[0014] The value of weight A ranges from 0.5 to 0.7; the values ​​of weights B and C are the same, and both range from 0.25 to 0.35.

[0015] The Stacking model includes a base model in the first layer and a simple linear regression model in the second layer. The base model can be any one of Random Forest, KNN, or LightGBM.

[0016] The process of inputting one-dimensional data into a Stacking model to obtain one-dimensional diagnostic results refers to: first, passing the low-frequency vibration signal, medium-frequency vibration signal, and temperature signal in the one-dimensional data through a base model to obtain new low-frequency vibration signal, new medium-frequency vibration signal, and new temperature signal, and then inputting the new low-frequency vibration signal, new medium-frequency vibration signal, and new temperature signal into a simple linear regression model to obtain one-dimensional diagnostic results;

[0017] The one-dimensional diagnostic result includes faults and non-faults, and the two-dimensional diagnostic result includes faults and non-faults;

[0018] The base model, simple linear regression model, and Stacking model are all well-trained models, and the training dataset consists of one-dimensional data and its corresponding fault or non-fault data.

[0019] The deep learning model is a well-trained model, and the training dataset consists of two-dimensional data and its corresponding fault or non-fault data.

[0020] The process of fusing one-dimensional and two-dimensional diagnostic results using Bayesian formulas to obtain diagnostic result A refers to:

[0021] ,in,

[0022] This represents the posterior probability, i.e., the diagnosis result of patient A.

[0023] All are prior probabilities, and are known values;

[0024] B represents a one-dimensional diagnostic result, k is the result type of B, C represents a two-dimensional diagnostic result, l is the result type of C, A represents a fault of large rotating machinery, i is the result type of A; k, l, i being 0 indicates that the corresponding result is a non-fault, k, l, i being 1, 2, 3, ... indicates that the corresponding result is a different type of fault.

[0025] Both the low-frequency vibration signal and the medium-frequency vibration signal are wavelengths.

[0026] The process of substituting low-frequency and mid-frequency vibration signals into the K-means algorithm optimized by the genetic algorithm to obtain the diagnosis result for person B refers to:

[0027] First, the waveforms of the low-frequency vibration signal and the medium-frequency vibration signal are subjected to conformal interpolation resampling and wavelet denoising to obtain the preprocessed waveform. Then, the low-frequency descriptor and the medium-frequency descriptor are calculated from the time domain and frequency domain of the preprocessed waveform by fast Fourier transform. The low-frequency descriptor and the medium-frequency descriptor are then averaged to obtain the average descriptor. Finally, the average descriptor is substituted into the K-means algorithm optimized by the genetic algorithm to obtain the diagnosis result of B.

[0028] The K-means algorithm optimized by the genetic algorithm is a well-trained model, and its training dataset consists of descriptors and their corresponding faulty or non-faulty features.

[0029] The low-frequency descriptor and the mid-frequency descriptor are any one or any combination of time features and spectral features.

[0030] The low-frequency vibration signal and the medium-frequency vibration signal are in the form of wavelength, or displacement or acceleration calculated from the wavelength.

[0031] The low-frequency acceleration sensor and the medium-frequency acceleration sensor can collect data from any one or any combination of the upper frame, lower frame, top cover, and stator end of a large rotating machine.

[0032] The objects of the operation images are any one or any combination of the upper frame, lower frame, and top cover of large rotating machinery. If the objects shown in the operation images have minor deformations or cracks, the fault classification corresponding to the operation images is a fault.

[0033] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0034] 1. In this invention, a multi-source heterogeneous data-driven fault diagnosis method for large rotating machinery first acquires one-dimensional data consisting of low-frequency vibration signals, medium-frequency vibration signals, and temperature signals, as well as two-dimensional data consisting of running images, simultaneously. Then, based on the one-dimensional and two-dimensional data, a diagnosis result A is obtained; based on the low-frequency and medium-frequency vibration signals, a diagnosis result B is obtained; and based on the temperature signals, a diagnosis result C is obtained. Finally, the diagnosis results A, B, and C are summed using a weighted summation method to obtain the final diagnosis result. If the final diagnosis result is greater than or equal to 0.5, a fault is determined; otherwise, it is determined to be faulty. A value of 0.5 indicates no fault. This demonstrates that the entire diagnostic process relies on a variety of acquired signals, including low-frequency vibration signals, mid-frequency vibration signals, temperature signals, and operational image signals. Furthermore, the acquired signals are differentiated, each employing a different signal processing method. Diagnostic result A is the primary indicator, corresponding to all signal types, while results B and C are secondary indicators. Finally, a unique summation formula is constructed, assigning specific weight ranges to diagnostic results A, B, and C to ensure a more accurate diagnostic result. Therefore, this invention not only monitors multiple signals but also achieves good diagnostic performance.

[0035] 2. In the multi-source heterogeneous data-driven fault diagnosis method for large rotating machinery of this invention, when obtaining the diagnostic result A, it is necessary to fuse the one-dimensional and two-dimensional diagnostic results using Bayes' formula to obtain the posterior probability (i.e., the diagnostic result A). Therefore, in addition to fulfilling the basic function of the diagnostic result A, it can also pinpoint the specific type of fault, rather than simply determining whether it is a fault or not. This facilitates further optimization of the fault diagnosis method, improves diagnostic accuracy, and expands the application scope of the fault diagnosis results. Therefore, this invention can determine the specific type of fault with high accuracy.

[0036] 3. In the multi-source heterogeneous data-driven fault diagnosis method for large rotating machinery of the present invention, one-dimensional data is input into a Stacking model to obtain one-dimensional diagnostic results. The Stacking model is limited to a base model comprising a first layer and a simple linear regression model comprising a second layer. The base model can be any one of Random Forest, KNN, or LightGBM. In application, the low-frequency vibration signal, mid-frequency vibration signal, and temperature signal from the one-dimensional data are first processed through the base model to obtain new low-frequency vibration signal, new mid-frequency vibration signal, and new temperature signal (signals correspond one-to-one with fault types; signal updates do not change the corresponding fault type). Then, the new low-frequency vibration signal, new mid-frequency vibration signal, and new temperature signal are input into the simple linear regression model to obtain the one-dimensional diagnostic results. It can be seen that not only is the initially obtained one-dimensional data optimized through the first-layer calculation, but the second-layer calculation is also performed based on the optimized data to improve the overall accuracy of the one-dimensional diagnostic results, thereby improving the precision of the final diagnostic results. Therefore, the present invention not only has high accuracy in the one-dimensional diagnostic results but also high precision in the final diagnostic results. Attached Figure Description

[0037] Figure 1 This is a schematic diagram of the operation process of the present invention.

[0038] Figure 2 This is a schematic diagram of the monitoring of the large rotating machinery corresponding to this invention. Detailed Implementation

[0039] The present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0040] See Figure 1 and Figure 2 A fault diagnosis method for large rotating machinery driven by multi-source heterogeneous data, the fault diagnosis method comprising the following steps:

[0041] The first step, data acquisition: At the same time, low-frequency vibration signals and medium-frequency vibration signals of the large rotating machinery are acquired through low-frequency acceleration sensors and medium-frequency acceleration sensors; temperature signals of the rotor in the large rotating machinery are acquired through temperature sensors; and running images of the frame in the large rotating machinery are acquired through cameras. The low-frequency vibration signals, medium-frequency vibration signals, and temperature signals are classified as one-dimensional data, and the running images are classified as two-dimensional data.

[0042] The second step, the multi-source heterogeneous step: The diagnostic results from A, B, and C are summed using a weighted summation method to obtain the final diagnostic result. If the final diagnostic result is greater than or equal to 0.5, a fault is determined; if it is less than 0.5, no fault is determined. The formula for summation is:

[0043] Final diagnosis result = Weight of A * Diagnosis result of A + Weight of B * Diagnosis result of B + Weight of C * Diagnosis result of C;

[0044] The diagnostic result of A refers to: first, inputting one-dimensional data into the Stacking model to obtain a one-dimensional diagnostic result; at the same time, inputting two-dimensional data into the deep learning model to obtain a two-dimensional diagnostic result; and then fusing the one-dimensional and two-dimensional diagnostic results using Bayes' formula to obtain the diagnostic result of A. The diagnostic result is a value greater than or equal to 0.5 but less than 1.

[0045] The diagnostic result of B refers to the following: the low-frequency vibration signal and the medium-frequency vibration signal are substituted into the K-means algorithm optimized by the genetic algorithm to obtain the diagnostic result of B. If the diagnostic result is a fault, it is 1; otherwise, it is 0.

[0046] The diagnostic result of C refers to the result obtained by substituting the temperature signal into the base model. The result is 1 if the fault is detected and 0 if the fault is not detected. The base model can be any one of Random Forest, KNN, or LightGBM.

[0047] The value of weight A ranges from 0.5 to 0.7; the values ​​of weights B and C are the same, and both range from 0.25 to 0.35.

[0048] The Stacking model includes a base model in the first layer and a simple linear regression model in the second layer. The base model can be any one of Random Forest, KNN, or LightGBM.

[0049] The process of inputting one-dimensional data into a Stacking model to obtain one-dimensional diagnostic results refers to: first, passing the low-frequency vibration signal, medium-frequency vibration signal, and temperature signal in the one-dimensional data through a base model to obtain new low-frequency vibration signal, new medium-frequency vibration signal, and new temperature signal, and then inputting the new low-frequency vibration signal, new medium-frequency vibration signal, and new temperature signal into a simple linear regression model to obtain one-dimensional diagnostic results;

[0050] The one-dimensional diagnostic result includes faults and non-faults, and the two-dimensional diagnostic result includes faults and non-faults;

[0051] The base model, simple linear regression model, and Stacking model are all well-trained models, and the training dataset consists of one-dimensional data and its corresponding fault or non-fault data.

[0052] The deep learning model is a well-trained model, and the training dataset consists of two-dimensional data and its corresponding fault or non-fault data.

[0053] The process of fusing one-dimensional and two-dimensional diagnostic results using Bayesian formulas to obtain diagnostic result A refers to:

[0054] ,in,

[0055] This represents the posterior probability, i.e., the diagnosis result of patient A.

[0056] All are prior probabilities, and are known values;

[0057] B represents a one-dimensional diagnostic result, k is the result type of B, C represents a two-dimensional diagnostic result, l is the result type of C, A represents a fault of large rotating machinery, i is the result type of A; k, l, i being 0 indicates that the corresponding result is a non-fault, k, l, i being 1, 2, 3, ... indicates that the corresponding result is a different type of fault.

[0058] Both the low-frequency vibration signal and the medium-frequency vibration signal are wavelengths.

[0059] The process of substituting low-frequency and mid-frequency vibration signals into the K-means algorithm optimized by the genetic algorithm to obtain the diagnosis result for person B refers to:

[0060] First, the waveforms of the low-frequency vibration signal and the medium-frequency vibration signal are subjected to conformal interpolation resampling and wavelet denoising to obtain the preprocessed waveform. Then, the low-frequency descriptor and the medium-frequency descriptor are calculated from the time domain and frequency domain of the preprocessed waveform by fast Fourier transform. The low-frequency descriptor and the medium-frequency descriptor are then averaged to obtain the average descriptor. Finally, the average descriptor is substituted into the K-means algorithm optimized by the genetic algorithm to obtain the diagnosis result of B.

[0061] The K-means algorithm optimized by the genetic algorithm is a well-trained model, and its training dataset consists of descriptors and their corresponding faulty or non-faulty features.

[0062] The low-frequency descriptor and the mid-frequency descriptor are any one or any combination of time features and spectral features.

[0063] The low-frequency vibration signal and the medium-frequency vibration signal are in the form of wavelength, or displacement or acceleration calculated from the wavelength.

[0064] The low-frequency acceleration sensor and the medium-frequency acceleration sensor can collect data from any one or any combination of the upper frame, lower frame, top cover, and stator end of a large rotating machine.

[0065] The objects of the operation images are any one or any combination of the upper frame, lower frame, and top cover of large rotating machinery. If the objects shown in the operation images have minor deformations or cracks, the fault classification corresponding to the operation images is a fault.

[0066] The principle of this invention is explained as follows:

[0067] See Figure 2 In this invention, the vibration signal acquisition targets include the upper frame, lower frame, top cover, and stator end; the temperature signal (i.e., temperature value) acquisition targets the rotor; and the running image signal acquisition targets the upper frame, lower frame, and top cover. After data acquisition, the vibration and temperature signals are connected to a fiber optic demodulator, while the running images captured by the high-definition camera are transmitted to an image recognition module. Subsequently, a deep learning method is used to determine whether there are minute deformations and cracks in the frame structure in the image.

[0068] The fault classification in this invention is initially divided into two types: fault and non-fault. Further fault subdivisions include rolling bearing fault, gear fault, sliding bearing fault, rotor winding inter-turn short circuit, shaft bending, imbalance, shaft crack, eccentricity, etc., which are reflected in "k, l, i are 1, 2, 3..., which means that the corresponding result is various different fault types".

[0069] In this invention, if the diagnostic result of A is less than 0.5, manual intervention is required to modify and adjust the previous fault classification. Then, the model is run again until a diagnostic result of A that meets the requirements is obtained.

[0070] Example 1:

[0071] See Figure 1 and Figure 2 A fault diagnosis method for large rotating machinery driven by multi-source heterogeneous data, the fault diagnosis method comprising the following steps:

[0072] The first step, data acquisition: At the same time, low-frequency vibration signals and medium-frequency vibration signals of the large rotating machinery are acquired through low-frequency acceleration sensors and medium-frequency acceleration sensors; temperature signals of the rotor in the large rotating machinery are acquired through temperature sensors; and running images of the frame in the large rotating machinery are acquired through cameras. The low-frequency vibration signals, medium-frequency vibration signals, and temperature signals are classified as one-dimensional data, and the running images are classified as two-dimensional data.

[0073] The second step, the multi-source heterogeneous step: The diagnostic results from A, B, and C are summed using a weighted summation method to obtain the final diagnostic result. If the final diagnostic result is greater than or equal to 0.5, a fault is determined; if it is less than 0.5, no fault is determined. The formula for summation is:

[0074] Final diagnosis result = A weight (range 0.5-0.7) * A diagnosis result + B weight * B diagnosis result + C weight * C diagnosis result. The values ​​of B weight and C weight are the same, and the range of their values ​​is 0.25-0.35.

[0075] The diagnostic result of A refers to: first, inputting one-dimensional data into the Stacking model to obtain a one-dimensional diagnostic result; at the same time, inputting two-dimensional data into the deep learning model to obtain a two-dimensional diagnostic result; and then fusing the one-dimensional and two-dimensional diagnostic results using Bayes' formula to obtain the diagnostic result of A. The diagnostic result is a value greater than or equal to 0.5 but less than 1.

[0076] The diagnostic result of B refers to the following: the low-frequency vibration signal and the medium-frequency vibration signal are substituted into the K-means algorithm optimized by the genetic algorithm to obtain the diagnostic result of B. If the diagnostic result is a fault, it is 1; otherwise, it is 0.

[0077] The diagnostic result of C refers to the result obtained by substituting the temperature signal into the base model. The result is 1 if the fault is detected and 0 if the fault is not detected. The base model can be any one of Random Forest, KNN, or LightGBM.

[0078] Example 2:

[0079] The basic content is the same as in Example 1, except that:

[0080] The Stacking model comprises a base model in the first layer and a simple linear regression model in the second layer. The base model can be any one of Random Forest, KNN, or LightGBM. Inputting one-dimensional data into the Stacking model to obtain a one-dimensional diagnostic result means: first, passing the low-frequency vibration signal, mid-frequency vibration signal, and temperature signal from the one-dimensional data through the base model to obtain new low-frequency vibration signal, new mid-frequency vibration signal, and new temperature signal, respectively; then, inputting the new low-frequency vibration signal, new mid-frequency vibration signal, and new temperature signal into the simple linear regression model to obtain a one-dimensional diagnostic result. The base model, simple linear regression model, and Stacking model are all well-trained models, and the training dataset consists of one-dimensional data and its corresponding fault or non-fault data. The deep learning model is a well-trained model, and the training dataset consists of two-dimensional data and its corresponding fault or non-fault data.

[0081] The process of fusing one-dimensional and two-dimensional diagnostic results using Bayesian formulas to obtain diagnostic result A refers to:

[0082] ,in,

[0083] This represents the posterior probability, i.e., the diagnosis result of patient A.

[0084] All are prior probabilities, and are known values;

[0085] B represents a one-dimensional diagnostic result, k is the result type of B, C represents a two-dimensional diagnostic result, l is the result type of C, A represents a fault of large rotating machinery, i is the result type of A; k, l, i being 0 indicates that the corresponding result is a non-fault, k, l, i being 1, 2, 3, ... indicates that the corresponding result is a different type of fault.

[0086] In obtaining one-dimensional and two-dimensional diagnostic results, it is necessary to first train the relevant base models, simple linear regression models, and deep learning models. Only after the training is mature can the output results, i.e., the fault type (fault or non-fault), be obtained when one-dimensional and two-dimensional data are directly input. During training, the training set used includes not only one-dimensional and two-dimensional data but also their respective fault classifications.

[0087] Example 3:

[0088] The basic content is the same as in Example 1, except that:

[0089] The step of substituting the low-frequency and mid-frequency vibration signals into the genetic algorithm-optimized K-means algorithm to obtain the diagnosis result of B involves: first, performing conformal interpolation resampling and wavelet denoising on the waveforms of the low-frequency and mid-frequency vibration signals (both of which are wavelengths) to obtain a preprocessed waveform; then, calculating the low-frequency and mid-frequency descriptors from the time and frequency domains of the preprocessed waveform using fast Fourier transform; finally, averaging the low-frequency and mid-frequency descriptors to obtain an average descriptor; and finally, substituting the average descriptor into the genetic algorithm-optimized K-means algorithm to obtain the diagnosis result of B. The genetic algorithm-optimized K-means algorithm is a well-trained model, and its training dataset consists of descriptors and their corresponding fault or non-fault data.

[0090] Since the information collected by the low-frequency acceleration sensor and the medium-frequency acceleration sensor is complementary (collected at the same time, with the same fault classification, both being either faulty or non-faulty), the low-frequency and medium-frequency information can be merged. Therefore, the K-means algorithm optimized by the genetic algorithm can be used to fuse the information and obtain the fault classification output associated with the average descriptor.

[0091] The above description is only a preferred embodiment of the present invention. The scope of protection of the present invention is not limited to the above embodiments. Any equivalent modifications or changes made by those skilled in the art based on the content disclosed in the present invention should be included within the scope of protection set forth in the claims.

Claims

1. A fault diagnosis method for large rotating machinery driven by multi-source heterogeneous data, characterized in that: The fault diagnosis method Includes the following steps: The first step, data acquisition: At the same time, low-frequency vibration signals and medium-frequency vibration signals of the large rotating machinery are acquired through low-frequency acceleration sensors and medium-frequency acceleration sensors; temperature signals of the rotor in the large rotating machinery are acquired through temperature sensors; and running images of the frame in the large rotating machinery are acquired through cameras. The low-frequency vibration signals, medium-frequency vibration signals, and temperature signals are classified as one-dimensional data, and the running images are classified as two-dimensional data. The second step, the multi-source heterogeneous step: The diagnostic results from A, B, and C are summed using a weighted summation method to obtain the final diagnostic result. If the final diagnostic result is greater than or equal to 0.5, a fault is determined; if it is less than 0.5, no fault is determined. The formula for summation is: Final diagnosis result = Weight of A * Diagnosis result of A + Weight of B * Diagnosis result of B + Weight of C * Diagnosis result of C; The diagnostic result of A refers to: first, inputting one-dimensional data into the Stacking model to obtain a one-dimensional diagnostic result; at the same time, inputting two-dimensional data into the deep learning model to obtain a two-dimensional diagnostic result; and then fusing the one-dimensional and two-dimensional diagnostic results using Bayes' formula to obtain the diagnostic result of A. The diagnostic result is a value greater than or equal to 0.5 but less than 1. The diagnostic result of B refers to the following: the low-frequency vibration signal and the medium-frequency vibration signal are substituted into the K-means algorithm optimized by the genetic algorithm to obtain the diagnostic result of B. If the diagnostic result is a fault, it is 1; otherwise, it is 0. The diagnostic result of C refers to the result obtained by substituting the temperature signal into the base model. The result is 1 if the fault is detected and 0 if the fault is not detected. The base model can be any one of Random Forest, KNN, or LightGBM.

2. The method for fault diagnosis of large rotating machinery driven by multi-source heterogeneous data according to claim 1, characterized in that: The value of weight A ranges from 0.5 to 0.7; the values ​​of weights B and C are the same, and both range from 0.25 to 0.

35.

3. A method for fault diagnosis of large rotating machinery driven by multi-source heterogeneous data according to claim 1 or 2, characterized in that: The Stacking model includes a base model in the first layer and a simple linear regression model in the second layer. The base model can be any one of Random Forest, KNN, or LightGBM. The process of inputting one-dimensional data into a Stacking model to obtain one-dimensional diagnostic results refers to: first, passing the low-frequency vibration signal, medium-frequency vibration signal, and temperature signal in the one-dimensional data through a base model to obtain new low-frequency vibration signal, new medium-frequency vibration signal, and new temperature signal, and then inputting the new low-frequency vibration signal, new medium-frequency vibration signal, and new temperature signal into a simple linear regression model to obtain one-dimensional diagnostic results; The one-dimensional diagnostic results include faults and non-faults, and the two-dimensional diagnostic results include faults and non-faults.

4. The method for fault diagnosis of large rotating machinery driven by multi-source heterogeneous data according to claim 3, characterized in that: The base model, simple linear regression model, and Stacking model are all well-trained models, and the training dataset consists of one-dimensional data and its corresponding fault or non-fault data. The deep learning model is a well-trained model, and the training dataset consists of two-dimensional data and its corresponding fault or non-fault data.

5. The method for fault diagnosis of large rotating machinery driven by multi-source heterogeneous data according to claim 3, characterized in that: The process of fusing one-dimensional and two-dimensional diagnostic results using Bayesian formulas to obtain diagnostic result A refers to: ,in, This represents the posterior probability, i.e., the diagnosis result of patient A. All are prior probabilities, and are known values; B represents a one-dimensional diagnostic result, k is the result type of B, C represents a two-dimensional diagnostic result, l is the result type of C, A represents a fault of large rotating machinery, i is the result type of A; k, l, i being 0 indicates that the corresponding result is a non-fault, k, l, i being 1, 2, 3, ... indicates that the corresponding result is a different type of fault.

6. A method for fault diagnosis of large rotating machinery driven by multi-source heterogeneous data according to claim 1 or 2, characterized in that: Both the low-frequency vibration signal and the medium-frequency vibration signal are wavelengths. The process of substituting low-frequency and mid-frequency vibration signals into the K-means algorithm optimized by the genetic algorithm to obtain the diagnosis result for person B refers to: First, the waveforms of the low-frequency vibration signal and the medium-frequency vibration signal are subjected to conformal interpolation resampling and wavelet denoising to obtain the preprocessed waveform. Then, the low-frequency descriptor and the medium-frequency descriptor are calculated from the time domain and frequency domain of the preprocessed waveform by fast Fourier transform. The low-frequency descriptor and the medium-frequency descriptor are then averaged to obtain the average descriptor. Finally, the average descriptor is substituted into the K-means algorithm optimized by the genetic algorithm to obtain the diagnosis result of B. The K-means algorithm optimized by the genetic algorithm is a well-trained model, and its training dataset consists of descriptors and their corresponding faulty or non-faulty features.

7. The method for fault diagnosis of large rotating machinery driven by multi-source heterogeneous data according to claim 6, characterized in that: The low-frequency descriptor and the mid-frequency descriptor are any one or any combination of time features and spectral features.

8. A method for fault diagnosis of large rotating machinery driven by multi-source heterogeneous data according to claim 1 or 2, characterized in that: The low-frequency vibration signal and the medium-frequency vibration signal are in the form of wavelength, or displacement or acceleration calculated from the wavelength.

9. A method for fault diagnosis of large rotating machinery driven by multi-source heterogeneous data according to claim 1 or 2, characterized in that: The low-frequency acceleration sensor and the medium-frequency acceleration sensor can collect data from any one or any combination of the upper frame, lower frame, top cover, and stator end of a large rotating machine.

10. A method for fault diagnosis of large rotating machinery driven by multi-source heterogeneous data according to claim 1 or 2, characterized in that: The objects of the operation images are any one or any combination of the upper frame, lower frame, and top cover of large rotating machinery. If the objects shown in the operation images have minor deformations or cracks, the fault classification corresponding to the operation images is a fault.

Images