A Fault Diagnosis Method and System for a Hydraulic Motor with Integrated Signals

Abstract

The present invention belongs to the technical field of fault diagnosis of integrated signal hydraulic motors, and discloses a fault diagnosis method for integrated signal hydraulic motors. The fault diagnosis method for integrated signal hydraulic motors provided by the present invention collects the vibration and sound signals of the hydraulic motor, selects 10 characteristic values such as the root mean square and standard deviation with the greatest influence factors in the sound and vibration signals, and constructs a fusion feature vector with higher and more obvious fault information content by using the method of splicing similar tags. Combining the fusion feature vector and the LightGBM algorithm to construct a fault diagnosis model for the wear of the hydraulic motor plunger, realizing the rapid and accurate identification of three degrees of wear of the hydraulic motor plunger. Finally, the identification results of the wear of the hydraulic motor plunger between different algorithms and different signals are compared and analyzed, verifying the superiority of this method.

Description

Technical Field

[0001] The present invention belongs to the technical field of fault diagnosis of integrated signal hydraulic motors, and particularly relates to a fault diagnosis method for integrated signal hydraulic motors. Background Technique

[0002] Hydraulic steering gears are commonly used mechanical devices in fields such as aerospace, submarines, and ships, and play a crucial role in steering and maintaining stability. As an actuator of the steering gear hydraulic system, the working performance of the hydraulic motor determines the safe and stable operation of the entire steering gear system. Therefore, it is of great significance to adopt a simple and effective fault diagnosis method for the normal operation of the hydraulic system. Traditional hydraulic motor fault diagnosis methods rely strongly on empirical knowledge and have the problem of difficult extraction of fault features. Some researchers conduct fault diagnosis by collecting vibration signals or sound signals during the operation of hydraulic motors and combining machine learning. Vibration signals have a high signal-to-noise ratio and sensitivity. Sound sensors can collect non-contact signals, and sound signals have a higher bandwidth and a wider frequency range of collection, and are commonly used in fault diagnosis methods. Zhu, Jiang, etc. monitored the vibration signals of hydraulic piston pumps and used wavelet transform to decompose and reconstruct the vibration signals to extract fault features such as time domain and frequency domain, and combined with convolutional neural networks to achieve the classification of different faults of hydraulic piston pumps. Wang and Liu obtained the modal function components of vibration signals using empirical mode decomposition, then calculated the entropy of the time-frequency matrix, and finally used random forests to conduct fault diagnosis on centrifugal pumps. The sensitivity of acceleration sensors is a fixed value, which is not accurate enough for early fault identification, and vibration signals have strong nonlinearity and non-stationarity, which have a great impact on the signals collected by sensors. Sound sensors have the advantages of non-destructive installation and non-contact measurement, and higher sensitivity, and are applied by most scholars in the data collection of fault diagnosis. Tang, etc. reviewed mechanical fault diagnosis based on audio signal analysis, analyzed the advantages of using audio signals, and summarized the development prospects of fault diagnosis based on audio signals. Mian, etc. collected the sound signals of bearings in different fault states and extracted the sound quality features of the signals, and finally used the support vector machine algorithm to achieve bearing fault diagnosis. There is a lot of interference noise in the environment in the sound signal. To improve the signal-to-noise ratio of the sound signal, decomposition and noise reduction processing are required. If the number of decomposition layers is small, it is difficult to achieve the denoising effect. If the number of decomposition layers is large, it will reduce the useful information in the signal and affect the diagnostic accuracy.

[0003] Due to the uncertainties and concealments of faults in hydraulic systems, the analysis of fault points is relatively complex. It is difficult to accurately identify the faults in hydraulic systems only through a single signal source or fault feature. To ensure the stable operation of internal curve radial piston hydraulic motors and achieve timely and accurate fault monitoring, some scholars have proposed a fault diagnosis method based on the fusion features of multi-source sensors and achieved relatively ideal results. Li et al. proposed a multi-feature fusion fault diagnosis model based on Fourier descriptors and graphic features of indicator diagrams to detect the fault types of rod pumps, enhancing the robustness of features and verifying that the diagnostic accuracy of the multi-input feature fusion model is higher than that of the single-input feature model. Tang et al. proposed a fault diagnosis method for hydraulic piston pumps with an adaptive learning rate by using the vibration signals, pressure signals, and sound signals of hydraulic piston pumps and combining them with a convolutional neural network (CNN). The three original signals were converted into two-dimensional time-frequency images through continuous wavelet transform, and different fault types were identified using the CNN model. Karabacak et al. [9] processed the vibration, sound, and thermal image data of normal and faulty rotors to reduce the wear and fault risks of worm gearboxes. They used artificial neural networks (ANNs) and support vector machines (SVMs) to extract singular, dual, or triple forms of time-domain, frequency-domain, and thermal image features, realizing the wear fault diagnosis of worm gearboxes. Long et al. proposed a multi-sensor information-driven motor fault diagnosis method based on AdaBoost. The frequency-domain features of the current, magnetic, and vibration signals of the collected motors were extracted using Hilbert transform and Fourier transform, and the AdaBoost model was used for training and diagnosis. This fault diagnosis method has high robustness and generalization ability. Through the above literature analysis, it can be seen that using multi-signal fusion obviously has a higher diagnostic recognition accuracy than single signals. However, the processing of original signals is relatively complex, and the fault diagnosis methods combining traditional machine learning such as neural networks and AdaBoost algorithms have large memory occupancy and long operation time, and the fault recognition rate is not high.

[0004] Traditional machine learning algorithms such as GBDT and AdaBoost are implementations of the Boosting algorithm, suffering from the drawbacks of poor robustness and low diagnostic efficiency. The XGBoost and LightGBM algorithms are improvements based on the GBDT algorithm. The gradient boosting tree algorithm solves the problems of long training time and low fault recognition rate of traditional algorithms and is currently the most widely used fault diagnosis method. Xiang, Wang, etc. respectively processed the original data using the fast Fourier transform and support vector machine algorithms, and then combined with the XGBoost model to rank and diagnose feature importance, improving the accuracy of rolling bearing fault diagnosis. Wu, Zhang, etc. established an XGBoost fault recognition model to improve the accuracy and reliability of wind turbine fault diagnosis and compared the diagnostic results with those of the support vector machine and Adaboost algorithms. The results showed that the XGBoost algorithm had a higher classification accuracy. The research of the above scholars indicated that the XGBoost algorithm had a higher fault diagnosis accuracy than traditional machine learning algorithms and was widely applied to the fault diagnosis of various mechanical equipment. However, the way the XGBoost algorithm traverses data limits the operation speed and reduces the efficiency of fault diagnosis. The LightGBM algorithm can improve the data operation efficiency while ensuring the accuracy and is widely used in the fault diagnosis of gearboxes, rolling bearings, and hydraulic motors. Tang, etc. studied the application of the LightGBM algorithm in the fault diagnosis of wind turbine gearboxes for the problems of low efficiency and accuracy of traditional machine learning algorithms in wind turbine gearbox fault diagnosis. It further confirmed that the LightGBM algorithm had a higher detection accuracy than traditional algorithms and had lower false alarm and missed detection rates. Guo, etc. combined the wavelet denoising algorithm and the LightGBM model to diagnose the looseness fault of the hydraulic motor base bolts and compared it with traditional machine learning algorithms. The results showed that the LightGBM algorithm had a higher diagnostic efficiency. Xu, etc. proposed a rolling bearing fault diagnosis method combining a convolutional neural network and LightGBM to solve the problems of long training time and low diagnostic efficiency of traditional deep learning algorithm-based fault diagnosis models. And by constructing a data set, it was verified that the diagnostic efficiency and accuracy of the LightGBM model were superior to other models.

[0005] As can be seen from the above literature, the sensitivity of the vibration sensor is a fixed value, which is not accurate enough for early fault identification. Moreover, the vibration signal has strong non-linearity and non-stationarity, which has a great impact on the signal collected by the sensor. The sound signal contains a lot of interference noise in the environment. To improve the signal-to-noise ratio of the sound signal, decomposition and noise reduction processing are required. If the number of decomposition layers is small, it is difficult to achieve the noise reduction effect. If the number of decomposition layers is large, it will reduce the useful information in the signal and affect the diagnostic accuracy. Using a single signal for fault diagnosis has a low recognition rate, while using the signal fusion method can combine the advantages of the two sensors and utilize the fault-sensitive frequency bands of different sensors to improve the fault recognition rate. For the existing multi-source sensor fusion methods, some literature has confirmed that they have higher diagnostic effects than the single-signal feature model. However, the conversion of the original data and the extraction of features are relatively troublesome, and most of the fault recognition is based on traditional algorithms such as GBDT and AdaBoost. This method has a long training time for the fault diagnosis model and low diagnostic efficiency.

[0006] Through the above analysis, the problems and defects of the existing technology are as follows:

[0007] Only identifying the faults of the hydraulic motor through a single signal source or fault feature is subject to the uncertainty and concealment characteristics of the faults existing in the hydraulic motor system, and the accuracy of the fault recognition result is poor.

[0008] Using a single signal source combined with machine learning algorithms such as GBDT and AdaBoost to carry out fault identification of the hydraulic motor, both the robustness and efficiency of the diagnosis process are low. Summary of the Invention

[0009] Aiming at the problems existing in the existing technology, the present invention provides a method for diagnosing faults of a hydraulic motor with fused signals.

[0010] The present invention is implemented as follows. A method for diagnosing faults of a hydraulic motor with fused signals includes:

[0011] Step 1: Simulate the plunger fault through an in-line radial piston hydraulic motor test bench, and collect the vibration signal and sound signal of the motor when the plunger is normal and worn to three different degrees.

[0012] Step 2: Use the wavelet soft threshold algorithm for noise reduction processing, and extract time-domain and frequency-domain eigenvalue for feature fusion based on data correlation.

[0013] Step 3: Build a LightGBM motor plunger fault diagnosis model, and divide the fused features into a data set containing normal and 3 fault signal labels for training and diagnosis.

[0014] Step 4: Verify the superiority of this method through comparison between different algorithms and different signals.

[0015] Furthermore, the method for simulating plunger faults through an internal curve radial piston hydraulic motor test bench is as follows:

[0016] The structure of an internal curve radial piston hydraulic motor mainly consists of a stator, a rotor, an output shaft, pistons, and balls; the stator curve is composed of 6 uniformly distributed action arc segments, and 10 pistons are evenly distributed around the rotor; when the motor works, high-pressure oil enters the high-pressure chamber of the rotor through the oil distribution shaft, generating a large pressure to push the pistons to extend radially outward along the axis, causing the balls to contact the internal curve guide rail to generate mutual forces, among which the tangential component force drives the rotor to rotate continuously and output torque; at the same time, the hydraulic oil in the piston chamber in the low-pressure area is discharged through the low-pressure oil port of the oil distribution shaft, and the continuous work of the motor is realized by continuously introducing high-pressure oil and discharging low-pressure oil.

[0017] The vibration between the stator and the rotor is mainly caused by the piston assembly. From the piston motor structure diagram, it can be seen that 2 pistons are in the same position, and the typical vibration frequencies are mainly:

[0018]

[0019] In the formula, n is the rotational speed of the hydraulic motor, p is the number of internal curves, and q is the number of pistons; the frequency of one piston can be expressed as

[0020]

[0021] Furthermore, the method for extracting time-domain and frequency-domain eigenvalue feature fusion based on data correlation is as follows:

[0022] (1) Feature selection

[0023] The denoised signal is visually processed for data correlation in Python, and a total of 6 time-domain features such as mean value, root mean square, peak index, standard deviation, waveform index, and pulse index, and 6 frequency-domain features such as mean value of amplitude spectrum, standard deviation of amplitude spectrum, frequency centroid, spectral amplitude skewness, frequency standard deviation, and frequency skewness are selected from the vibration signal and sound signal data according to the feature importance.

[0024] (2) Model evaluation method

[0025] The confusion matrix is a commonly used indicator for evaluating model performance. It can reflect the relationship between the prediction result and the real situation. In model evaluation, 4 indicators such as accuracy, precision, recall, and f1-score are often selected for evaluation;

[0026] The expressions of each evaluation indicator are as follows:

[0027]

[0028]

[0029]

[0030]

[0031] PR curve and ROC curve are techniques and tools used to evaluate the performance of classifiers, which can calculate the accuracy and reliability of prediction results; in the PR curve, recall and precision are plotted on the same curve; the ROC curve can intuitively compare the performance of classifiers by plotting the true positive rate (TPR) and false positive rate (FPR) of the classifier on a curve; TPR and FPR can be expressed as:

[0032]

[0033]

[0034] (3) Diagnostic method process

[0035] Test bench setup and signal acquisition: Set up an inner curve radial fault diagnosis test bench, perform three different depths of wear on the plunger, and use vibration acceleration sensors and PCB sound sensors to collect the vibration and sound signals of the motor under different faults;

[0036] Signal denoising and feature extraction: Use the wavelet algorithm to denoise the vibration signal and sound signal respectively, and use mathematical methods to extract time-domain and frequency-domain features;

[0037] Signal feature fusion: Adopt a feature fusion method to fuse the feature values of the vibration signal and sound signal;

[0038] Model training and fault diagnosis: Divide the vibration signal data, sound signal data, and fused signal feature data sets into training sets and test sets respectively, input them into the fault diagnosis model for training and diagnosis, and set the maximum depth to prevent overfitting.

[0039] Furthermore, the method for collecting the vibration signals and sound signals of the motor when the plunger is normal and worn to three different degrees is as follows:

[0040] 1) Experimental method;

[0041] 2) Data processing;

[0042] 3) Feature selection.

[0043] Furthermore, the experimental method:

[0044] The present invention constructs a fault diagnosis experimental platform for an internal curve radial piston hydraulic motor, simulating fault types with different wear depths of the piston structure. The test bench consists of a fault simulation system and a data acquisition system:

[0045] The piston fault simulation test bench uses an internal curve radial piston hydraulic motor to simulate faults. The PCB sound sensor is installed 25 cm outside the hydraulic motor, and the vibration acceleration sensor is installed on the rear end cover of the hydraulic motor by magnetic adsorption. The X-axis is the same as the radial direction of the motor, and the speed sensor is installed 4 mm outside the hydraulic motor axially;

[0046] The data acquisition system. The PC terminal connects to the data acquisition card through Ethernet to receive and store data; the speed display can display the speed of the hydraulic motor in real time, ensuring the same speed for each experiment, eliminating the interference of irrelevant factors, and improving the accuracy of the experimental results; during the experiment, the servo control system is used to control the hydraulic oil source to supply a stable flow rate, ensuring that the speed of the hydraulic motor is stable at 300 r / min; the sampling frequency of the data acquisition system is set to 2000 Hz, and the vibration and sound signals are collected when the piston is normal and has three different degrees of wear faults. Each group of experimental data is collected for 6 minutes.

[0047] Furthermore, the data processing method:

[0048] The experimental data obtained through the vibration sensor and the sound sensor usually contains the interference of noise signals. Especially, the detection frequency of the sound sensor is low and it is sensitive to the sounds in the environment; therefore, according to the characteristics of the piston wear fault signal, the wavelet soft threshold algorithm is selected to decompose and reconstruct the vibration and sound signals to achieve the denoising effect of the original signal; the soft threshold function can be expressed as:

[0049]

[0050] The signal after wavelet soft threshold denoising is a time-domain signal. To better observe the characteristics of the signal and extract the frequency-domain features, the fast Fourier transform (FFT) is selected to transform the signal into a frequency-domain representation. The Fourier transform and its inverse transform formulas are respectively:

[0051]

[0052]

[0053] It can be seen from the frequency-domain comparison that the burrs in the signal are reduced after wavelet soft threshold denoising, which reflects that the cleanliness of the signal is improved and the denoising effect is achieved; the present invention adopts an internal curve radial piston hydraulic motor with 10 pistons, 6 internal curves in the housing, and an experimental speed of 300 r / min; from Figure 6As can be seen from the comparison, the vibration sensor is more sensitive to the half-frequency of the hydraulic motor speed, and the sound sensor is more sensitive to the first and second frequencies of the hydraulic motor speed, and has a better frequency response effect on the frequency of a single plunger. Considering that the vibration sensor and the sound sensor have different response frequencies to plunger faults, the present invention combines the respective advantages of the vibration and sound signals, fuses the characteristic values of the two signals, so as to improve the diagnostic accuracy of the hydraulic motor plunger faults.

[0054] Further, the feature selection method:

[0055] Use the inner curve radial piston hydraulic motor test bench to simulate the plunger fault, and collect the vibration and sound signals of the plunger under different states; the collection frequency is set to 2000Hz, each group of data is collected for 6 minutes, each group of data has 720,000 sampling points, and 4000 sampling points are selected in turn to generate a sample, with a total of 180 samples. The sample numbers of the training set and the test set are 80% and 20% respectively;

[0056] Time domain and frequency domain features are common indicators to reflect the operating state of the hydraulic motor; use vibration and sound sensors to collect signals of the plunger in normal and different wear conditions. After wavelet soft threshold denoising, use mathematical methods to extract 11 types of time domain and frequency domain characteristic values of each group of data respectively; generate vibration signal and sound signal characteristic value data sets, and conduct data set correlation analysis:

[0057] The data correlation heat map reflects the correlation information between data. The larger the positive and negative correlation values in the figure, the more it can reflect the operating state information of the hydraulic motor; 10 types of vibration signal characteristic values such as root mean square value, root amplitude, standard deviation, peak index, waveform index, spectral amplitude mean, spectral amplitude standard deviation, spectral amplitude skewness, spectral amplitude kurtosis, and frequency skewness are selected according to the heat map; 10 types of sound signal characteristic values such as root mean square value, root amplitude, peak value, standard deviation, peak index, pulse index, spectral amplitude mean, spectral amplitude standard deviation, frequency centroid, and frequency standard deviation are selected; finally, the normal and fault characteristic values of each group are horizontally fused according to the corresponding data labels, without changing the sample number, only increasing the data set length.

[0058] Further, the method of dividing the fused features into a data set containing normal and 3 types of fault signal labels for training and diagnosis:

[0059] (1) Data pre-classification and visualization;

[0060] (2) Model evaluation.

[0061] Further, the data pre-classification and visualization method:

[0062] To pre-observing the classification effect of the fusion signal data relative to the vibration and sound signals, the T-SNE algorithm is used to visualize the data; it can be seen from the figure that the classification of the vibration signal has a certain degree of aggregation, but there are too many aggregation points and the samples of the same category are not aggregated together; the classification of the sound signal has strong discreteness and different types of data points are fused together, resulting in a poor classification effect; for the classification of the fusion signal, it can be seen that different types of data points are almost completely separated, with few intersections between samples, and the classification effect is better.

[0063] Further, the model evaluation method:

[0064] The AdaBoost algorithm is an improvement based on the Boosting algorithm and has high effectiveness and practicability; the GBDT algorithm is a gradient boosting iterative decision tree algorithm, and the XGBoost and LightGBM algorithms are improvements based on GBDT and are also the most popular decision tree boosting algorithms with higher learning efficiency; in this invention, by comparing the classification accuracies of the four algorithms for the fusion signal, vibration signal, and sound signal respectively, the high efficiency of the LightGBM algorithm and its classification effect on the fusion signal data are verified;

[0065] For the diagnostic results of different models, the classification effect of the traditional AdaBoost algorithm is poor, and the diagnostic accuracies for the three data sets are all more than 10% lower than those of the other three algorithms; the classification effects of the GBDT, XGBoost, and LightGBM algorithms for the three data sets are not very different, but the diagnostic accuracy of the LightGBM algorithm is 0.17%-0.33% higher, and the diagnostic results of the three algorithms for the fusion data set have higher accuracies. The training time of the AdaBoost algorithm is the shortest, but its diagnostic accuracy is relatively low; the diagnostic accuracies of the GBDT, XGBoost, and LightGBM algorithms are not very different, but the training time of the LightGBM algorithm is shortened by 6 times and 3 times compared with the GBDT and XGBoost algorithms respectively. Therefore, the LightGBM algorithm has higher diagnostic accuracy and higher efficiency;

[0066] The vibration signal and sound signal are denoised using the wavelet soft threshold algorithm, and the eigenvalues are extracted and fused according to the data labels to generate a fused signal dataset. Each dataset contains normal and three fault types, but the number of samples remains unchanged. The LightGBM model is used to diagnose the faults of the fused signal, vibration signal, and sound signal, and accuracy, precision, recall, and f1-score are introduced as evaluation indicators for classification performance. The LightGBM model is used to train and diagnose the three datasets. By comparison, it can be seen that the diagnostic accuracy for the fused dataset is 4.86% and 14.59% higher than that of the vibration and sound signals respectively. The diagnostic results of precision, recall, and f1-score for the fused dataset are also improved by 4.5% - 13.25% compared with the vibration and sound datasets.

[0067] ROC and PR curves are used to evaluate the classification performance of a machine learning algorithm for a given dataset. Each dataset contains a fixed number of positive and negative samples. The larger the area under the curve, the better the classification performance of the machine learning algorithm for the dataset. FPR refers to the probability that a sample that is actually negative is predicted to be positive, and TPR refers to the probability that a sample that is actually positive is predicted to be positive. Precision refers to the proportion of samples predicted to be positive among positive samples, and Recall refers to the proportion of positive samples predicted to be positive among positive samples. Based on the classification effects of the LightGBM algorithm on the vibration, sound, and fused signal datasets, the classification effect of the LightGBM algorithm on the sound dataset is the worst, and the classification effect on the fused dataset is much higher than that of the vibration and sound datasets.

[0068] Another object of the present invention is to provide a fused signal hydraulic motor fault diagnosis system, including:

[0069] Fault simulation component: It includes an inner curve radial piston hydraulic motor test bench, which is used to simulate the normal state and different degrees of wear faults of the piston, and collect the corresponding vibration and sound signals;

[0070] Signal processing component module: The wavelet soft threshold algorithm is used to denoise the collected signals to improve the signal quality for subsequent analysis;

[0071] Feature extraction and fusion module: Based on data correlation analysis, eigenvalues are extracted from the processed time-domain and frequency-domain signals, and feature fusion is performed to form a comprehensive feature set for fault diagnosis;

[0072] Fault diagnosis model module: A motor piston fault diagnosis model based on the LightGBM algorithm is constructed, and the fused feature dataset containing normal and three fault signal labels is used for model training and diagnosis;

[0073] Performance verification component module: It includes tools for evaluating and verifying the performance of diagnostic models, such as confusion matrices, PR curves, and ROC curves, and conducts comparative analysis of different algorithms and signal types to verify the effectiveness and superiority of the proposed method.

[0074] Furthermore, it includes an inner curve radial piston hydraulic motor test bench for simulating normal and different degrees of wear faults of the piston, and is equipped with vibration and sound signal acquisition devices for collecting corresponding signals of the motor in various states.

[0075] Furthermore, it includes a signal processing module that uses the wavelet soft threshold algorithm to denoise the collected vibration and sound signals, improving the signal quality for more accurate fault diagnosis.

[0076] Furthermore, it includes a feature extraction and fusion module for extracting time-domain and frequency-domain eigenvalue from the denoised vibration and sound signals based on data correlation analysis and performing feature fusion to form a comprehensive feature set for fault diagnosis.

[0077] Furthermore, it includes a motor piston fault diagnosis model construction module based on the LightGBM algorithm for training and diagnosing a fused feature data set containing normal and three fault signal labels, thereby achieving accurate diagnosis of hydraulic motor piston faults.

[0078] Furthermore, it includes a model evaluation and verification module that uses evaluation tools such as confusion matrices, PR curves, and ROC curves to evaluate and verify the performance of the diagnostic model, including the calculation of indicators such as accuracy, precision, recall rate, and f1-score, and verifies the superiority of the diagnostic method through comparative analysis of different algorithms and signal types.

[0079] Combined with the above technical solutions and the technical problems solved, the advantages and positive effects of the technical solution to be protected by the present invention are:

[0080] First, a fault diagnosis method for a fused-signal hydraulic motor provided by the present invention. The plunger fault is simulated through an inner-curve radial piston hydraulic motor test bench, and the vibration signals and sound signals of the motor are collected when the plunger is normal and worn to three different degrees. The wavelet soft-threshold algorithm is used for noise reduction processing, and feature selection is carried out according to data correlation and feature importance. The time-domain and frequency-domain feature values are extracted for feature fusion. A LightGBM motor plunger fault diagnosis model is constructed, and the fused features are divided into a data set containing normal and three fault signal labels for training and diagnosis. Finally, the superiority of this method is verified through the comparison between different algorithms and different signals. The fault diagnosis method with fused features proposed by the present invention solves the problem of low diagnostic accuracy when using a single sensor to collect data. By using two different signals for diagnosis, the fault information of different sensors can be combined, and the fault features can be increased to improve the fault recognition rate. The LightGBM is used to establish a fault diagnosis model, which reduces the occupied running memory and improves the model training time while ensuring the classification accuracy, greatly improving the fault diagnosis efficiency.

[0081] The results show that: the diagnostic accuracy of the LightGBM algorithm is improved by 0.17%-0.33% compared with the AdaBoost, GBDT, and XGBoost algorithms; the diagnostic accuracy of the fused signal under the LightGBM model is improved by 4.86% and 14.59% respectively compared with the diagnosis of vibration and sound signals.

[0082] Second, a fault diagnosis method for a fused-signal hydraulic motor provided by the present invention. The plunger fault is simulated through an inner-curve radial piston hydraulic motor test bench, and the vibration signals and sound signals of the motor are collected when the plunger is normal and worn to three different degrees. The wavelet soft-threshold algorithm is used for noise reduction processing, and feature selection is carried out according to data correlation and feature importance. The time-domain and frequency-domain feature values are extracted for feature fusion. A LightGBM motor plunger fault diagnosis model is constructed, and the fused features are divided into a data set containing normal and three fault signal labels for training and diagnosis. Finally, the superiority of this method is verified through the comparison between different algorithms and different signals. The fault diagnosis method with fused features proposed by the present invention solves the problem of low diagnostic accuracy when using a single sensor to collect data. By using two different signals for diagnosis, the fault information of different sensors can be combined, and the fault features can be increased to improve the fault recognition rate. The LightGBM is used to establish a fault diagnosis model, which reduces the occupied running memory and improves the model training time while ensuring the classification accuracy, greatly improving the fault diagnosis efficiency.

[0083] The results show that: compared with the AdaBoost, GBDT, and XGBoost algorithms, the diagnostic accuracy of the LightGBM algorithm has increased by 0.17% - 0.33%; the diagnostic accuracy of the fused signal under the LightGBM model has increased by 4.86% and 14.59% respectively compared with the diagnosis of vibration and sound signals.

[0084] Third, the expected benefits and commercial value after the transformation of the technical solution of the present invention are as follows: a) Due to the uncertainty and concealment characteristics of the hydraulic motor system failure, it is difficult to accurately identify the faults existing in the hydraulic motor only through a single signal source or fault feature. The fused signal fault diagnosis method combines the respective advantages of sound and vibration sensors, contains more fault information, is simpler in feature extraction and fusion, and has more reliable diagnostic results; b) Select 10 eigenvalue features such as the root mean square and standard deviation that have the greatest influence factors in the sound and vibration signals, and use the splicing of the same type of tags for feature fusion to form a new feature vector, with higher and more obvious fault information content in the sample vector; c) Using the fused sample vector combined with the LightGBM algorithm model, multi-threaded parallel computing can be performed, saving the occupancy of computer memory, greatly reducing the diagnostic time, and improving the fault identification performance of the plunger motor.

[0085] The technical solution of the present invention fills the technical gaps at home and abroad in the industry: a) The present invention provides a diagnostic method for the plunger wear fault of an internal curve plunger type hydraulic motor based on the LightGBM algorithm to identify the characteristics of the sound-vibration fused signal, opening up a feasible technical route for the fault diagnosis of the plunger type hydraulic motor; b) The fault diagnosis method based on machine learning has been widely used in rotating machinery such as bearings and gears, but for the diagnosis of plunger motors, especially internal curve radial plunger type hydraulic motors, there is no application in the existing domestic and foreign literature.

[0086] Fourth, the significant technical progress of the fused signal hydraulic motor fault diagnosis method provided by the present invention can be summarized as follows:

[0087] 1. Precise acquisition of fault simulation and experimental data

[0088] Authenticity of fault simulation: By simulating the plunger fault on the internal curve radial plunger type hydraulic motor test bench, the fault conditions that occur in actual operation can be accurately simulated, so as to obtain test data closer to the actual situation.

[0089] Multi-dimensional data acquisition: Collect vibration signals and sound signals, combining two different signal types, improving the accuracy and reliability of fault diagnosis.

[0090] 2. Efficient signal processing and feature fusion

[0091] Application of Denoising Algorithm: The wavelet soft threshold algorithm is used for denoising, effectively removing the noise in the signal and improving the signal quality.

[0092] Innovation of Feature Fusion: The time-domain and frequency-domain eigenvalues are used for feature fusion, enhancing the diagnostic ability of the model and ensuring the comprehensiveness and depth of fault diagnosis.

[0093] 3. Utilizing Advanced Machine Learning Models

[0094] Application of LightGBM Model: A motor plunger fault diagnosis model based on LightGBM is constructed. This model features high efficiency and high precision and is suitable for processing large-scale data.

[0095] Optimization of Model Training and Diagnosis: The fused features are divided into a dataset containing normal and three types of fault signal labels for training, improving the generalization ability and diagnostic accuracy of the model.

[0096] 4. Comprehensive Verification of System Performance

[0097] Comprehensive Algorithm Comparison: By comparing different algorithms, the superiority of the proposed method is verified, ensuring the advancement and effectiveness of the method.

[0098] Comprehensive Analysis of Signal Types: The comparative analysis between different signals further verifies the importance of feature fusion in improving diagnostic accuracy.

[0099] The significant technical progress of the fused-signal hydraulic motor fault diagnosis method provided by the present invention is mainly reflected in improving the accuracy, reliability, and efficiency of fault diagnosis. By comprehensively utilizing various signals, advanced signal processing methods, and machine learning models, the performance of hydraulic motor fault diagnosis is significantly improved. Brief Description of the Drawings

[0100] Figure 1 is the flowchart of the fused-signal hydraulic motor fault diagnosis method provided by the embodiment of the present invention.

[0101] Figure 2 is the structural diagram of an inner curve radial plunger motor provided by the embodiment of the present invention.

[0102] Figure 3 is the schematic diagram of the LightGBM algorithm provided by the embodiment of the present invention.

[0103] Figure 4 is the flowchart of the method provided by the embodiment of the present invention.

[0104] Figure 5 is the experimental equipment diagram provided by the embodiment of the present invention.

[0105] Figure 6It is the plunger wear degree diagram provided by the embodiment of the present invention.

[0106] Figure 7 It is the signal denoising effect comparison diagram provided by the embodiment of the present invention.

[0107] Figure 8 It is the data correlation heat map provided by the embodiment of the present invention.

[0108] Figure 9 It is the T-SNE clustering diagram provided by the embodiment of the present invention. (a) Vibration signal (b) Sound signal (c) Fusion signal.

[0109] Figure 10 It is the confusion matrix diagram provided by the embodiment of the present invention. (a) Vibration signal (b) Sound signal (c) Fusion signal.

[0110] Figure 11 It is the diagnosis result diagram of different models provided by the embodiment of the present invention.

[0111] Figure 12 It is the training duration diagram of different models provided by the embodiment of the present invention.

[0112] Figure 13 It is the LightGBM prediction accuracy rate diagram provided by the embodiment of the present invention.

[0113] Figure 14 It is the classification effect diagram of the vibration, sound and fusion signal datasets based on the LightGBM algorithm provided by the embodiment of the present invention. (a) ROC curve; (b) PR curve.

[0114] Figure 2 Wherein: 1. Ball; 2. Plunger; 3. Rotor; 4. Output shaft; 5. Stator. Specific embodiments

[0115] In order to make the objectives, technical solutions and advantages of the present invention clearer, the present invention will be further described in detail below with reference to embodiments. It should be understood that the specific embodiments described herein are only used to explain the present invention and are not used to limit the present invention.

[0116] Based on the fusion signal hydraulic motor fault diagnosis method provided by the present invention, the following are two specific embodiments and their implementation schemes:

[0117] Embodiment 1: Diagnosis of slight wear fault

[0118] (1) Fault simulation: Simulate the plunger fault of slight wear on the test bench of an in-line radial piston hydraulic motor.

[0119] 2) Data acquisition: Use vibration acceleration sensors and PCB sound sensors to collect vibration signals and sound signals under normal conditions and slight wear conditions.

[0120] 3) Signal processing: Use the wavelet soft threshold algorithm to denoise the collected signals, and then extract eigenvalue from the time domain and frequency domain.

[0121] 4) Feature fusion and model construction: Fusion the time domain and frequency domain eigenvalues, and construct a motor plunger fault diagnosis model based on LightGBM.

[0122] 5) Model training and diagnosis: Use the dataset containing normal and slight wear signal labels to train and diagnose the model.

[0123] 6) Performance evaluation: Evaluate the accuracy and reliability of the model through confusion matrix, PR curve and ROC curve.

[0124] Example 2: Diagnosis of severe wear fault

[0125] 1) Fault simulation: Simulate the plunger fault of severe wear on the same test bench.

[0126] 2) Data acquisition: Also collect vibration and sound signals under normal conditions and severe wear conditions.

[0127] 3) Signal processing: Perform the same denoising process on the signals, and extract the corresponding time domain and frequency domain eigenvalues.

[0128] 4) Feature fusion and model optimization: Fusion the extracted features, and optimize the LightGBM model to adapt to the fault characteristics of severe wear.

[0129] 5) Model training and diagnosis: Train the model, and use the dataset containing severe wear signal labels for diagnosis.

[0130] 6) Performance verification: Use the model evaluation index to verify the performance of the model, and ensure that it also maintains high accuracy in the case of severe wear.

[0131] The two embodiments provided by the present invention show how to apply the fusion signal hydraulic motor fault diagnosis method for different degrees of wear faults (slight and severe), provide the actual operation process and steps, and ensure the high accuracy and effectiveness of fault diagnosis.

[0132] As Figure 1 shown, the present invention provides a fusion signal hydraulic motor fault diagnosis method, including the following steps:

[0133] S101. Simulate the plunger fault through an inner curve radial piston hydraulic motor test bench, and collect the vibration signals and sound signals of the motor when the plunger is normal and worn to three different degrees.

[0134] S102. Perform noise reduction processing using the wavelet soft threshold algorithm, and extract the time-domain and frequency-domain eigenvalues based on data correlation for feature fusion.

[0135] S103. Construct a LightGBM motor plunger fault diagnosis model, and divide the fused features into a data set containing normal and 3 types of fault signal labels for training and diagnosis.

[0136] S104. Verify the superiority of this method through the comparison between different algorithms and different signals.

[0137] The method for simulating the plunger fault through an inner curve radial piston hydraulic motor test bench provided by the present invention is as follows:

[0138] The working principle of the inner curve radial piston hydraulic motor is as Figure 2 shown. Its structure mainly consists of a stator, a rotor, an output shaft, a plunger, and a ball; the stator curve consists of 6 uniformly distributed action arc segments, and 10 plungers are evenly distributed around the rotor; when the motor works, high-pressure oil enters the high-pressure chamber of the rotor through the oil distribution shaft, generating a large pressure to push the plunger to extend radially outward along the axis, causing the ball to contact the inner curve guide rail to generate a mutual force, and the tangential component force drives the rotor to rotate continuously to output torque; at the same time, the hydraulic oil in the plunger chamber in the low-pressure area is discharged through the low-pressure oil port of the oil distribution shaft, and the continuous operation of the motor is realized by continuously introducing high-pressure oil and discharging low-pressure oil.

[0139] The vibration between the stator and the rotor is mainly caused by the plunger assembly. It can be seen from the plunger motor structure diagram that 2 plungers are in the same position, and the typical vibration frequencies are mainly:

[0140]

[0141] In the formula, n is the rotational speed of the hydraulic motor, p is the number of inner curves, and q is the number of plungers; the frequency of one plunger can be expressed as

[0142]

[0143] Principle of the LightGBM algorithm

[0144] Both the XGBoost and LightGBM algorithms are improvements based on the GBDT algorithm; both the GBDT and XGBoost models adopt the Level-wise growth strategy; by traversing the data once, the leaves of the same layer can be split simultaneously, reducing overfitting and making it easier to optimize for multiple threads; however, this method requires traversing the data layer by layer, which will generate many unnecessary searches or splits, thus consuming more running memory, increasing the computing time and reducing the efficiency; the LightGBM model incorporates the histogram algorithm, whose principle is to divide the original data of features into k discrete features and construct k bins for statistical features. When searching, there is no need to traverse the data, and only k bins need to be traversed to find the best split point, greatly shortening the computing time; the split method of the LightGBM model adopts the Leaf-wise growth strategy; this method searches for the leaf with the largest information gain in the current leaves. Under the same number of splits, the Leaf-wise growth strategy has lower error, less running memory occupation than the Level-wise growth strategy, and improves the accuracy and efficiency of the algorithm; the disadvantage of the LightGBM model is that it is prone to overfitting. Therefore, the maximum depth parameter is added to limit during model training;

[0145] The LightGBM model also incorporates two algorithms, Gradient-based One-Side Sampling (GOSS) and Exclusive Feature Bundling (EFB). The principle is as Figure 3 shown; GOSS samples the samples using gradient information. Samples with large gradients have a greater impact on information gain; when sampling, samples with large gradients are retained, and samples with small gradients are randomly sampled, which not only does not lose the model accuracy but also can reduce memory occupation and training time, greatly improving the diagnostic efficiency of the model; the feature spaces in the data are often mutually exclusive. The EFB algorithm starts from the perspective of reducing the feature dimension, bundles multiple mutually exclusive features together, reduces the number of features, and improves the running speed of the model.

[0146] The method for extracting time-domain and frequency-domain eigenvalue feature fusion based on data correlation provided by the present invention is as follows:

[0147] (1) Feature selection

[0148] The denoised signal is visualized for data correlation in Python. A total of 6 time-domain features, such as mean, root mean square, peak index, standard deviation, waveform index, and impulse index, and 6 frequency-domain features, such as mean of amplitude spectrum, standard deviation of amplitude spectrum, frequency centroid, spectral amplitude skewness, frequency standard deviation, and frequency skewness, are selected from the vibration signal and sound signal data according to the feature importance. These features can best reflect the working state of the inner curve piston motor and can reflect the fault information of the motor in a timely manner. The mathematical formulas of the time-domain and frequency-domain features are shown in Table 1:

[0149] Table 1 Time-domain and Frequency-domain Statistical Features

[0150]

[0151] In the table, x(n) (n = 1, 2,..., N) represents the original signal, N represents the number of data points; s(k) (k = 1, 2,..., M) represents the amplitude of the signal spectrum, M represents the number of spectral lines; f(k) represents the amplitude of the frequency of the k-th spectral line;

[0152] (2) Model Evaluation Method

[0153] The confusion matrix is a commonly used indicator to evaluate the performance of the model. As shown in Table 2, it can reflect the relationship between the prediction result and the real situation. In model evaluation, 4 indicators, such as accuracy, precision, recall, and f1-score, are often selected for evaluation;

[0154] Table 2 Confusion Matrix

[0155]

[0156] The expressions of each evaluation indicator are as follows:

[0157]

[0158]

[0159]

[0160]

[0161] The PR curve and ROC curve are techniques and tools used to evaluate the performance of classifiers. It can calculate the accuracy and reliability of the prediction results. In the PR curve, recall and precision are plotted on the same curve; the ROC curve can visually compare the performance of classifiers by plotting the true positive rate (TPR) and false positive rate (FPR) of the classifier on one curve. TPR and FPR can be expressed as:

[0162]

[0163]

[0164] (3) Diagnostic method process

[0165] Test bench construction and signal acquisition: Build an inner curve radial fault diagnosis test bench, conduct three different depths of wear on the plunger, and use vibration acceleration sensors and PCB sound sensors to collect the vibration and sound signals of the motor under different faults;

[0166] Signal denoising and feature extraction: Use the wavelet algorithm to denoise the vibration signal and the sound signal respectively, and use mathematical methods to extract time-domain and frequency-domain features;

[0167] Signal feature fusion: Adopt a feature fusion method to fuse the eigenvalues of the vibration signal and the sound signal;

[0168] Model training and fault diagnosis: Divide the vibration signal data, sound signal data, and fused signal feature data sets into training sets and test sets respectively, input them into the fault diagnosis model for training and diagnosis, and set the maximum depth to prevent overfitting.

[0169] The diagnostic method process is as Figure 4 shown.

[0170] The method for collecting the vibration signal and sound signal of the motor when the plunger is normal and worn to three different degrees provided by the present invention is as follows:

[0171] 1) Experimental method;

[0172] 2) Data processing;

[0173] 3) Feature selection.

[0174] The experimental method provided by the present invention:

[0175] The present invention builds an inner curve radial piston hydraulic motor fault diagnosis experimental platform, simulates the fault types of different wear depths of the plunger structure. This test bench consists of a fault simulation system and a data acquisition system. The installation positions of the experimental equipment and sensors are as Figure 5 shown:

[0176] The plunger fault simulation test bench is as Figure 5 (a) shown. Use an inner curve radial piston hydraulic motor to simulate the fault. The PCB sound sensor is installed 25 cm outside the hydraulic motor, and the vibration acceleration sensor is installed on the rear end cover of the hydraulic motor in a magnetic adsorption manner. The X-axis is the same as the radial direction of the motor, and the rotational speed sensor is installed 4 mm outside the axial direction of the hydraulic motor;

[0177] The data acquisition system is as Figure 5As shown in (b), the PC receives and stores data through an Ethernet connection to the data acquisition card; the rotational speed display can display the rotational speed of the hydraulic motor in real time, ensuring the same rotational speed for each experiment, eliminating the interference of irrelevant factors, and improving the accuracy of the experimental results; during the experiment, the servo control system is used to control the hydraulic oil source to supply a stable flow rate to ensure that the rotational speed of the hydraulic motor is stable at 300 r / min; the sampling frequency of the data acquisition system is set to 2000 Hz, and the vibration and sound signals of the plunger under normal conditions and three different degrees of wear faults are collected. Each group of experimental data is collected for 6 minutes; the models and parameters of each experimental equipment are shown in Table 3:

[0178] Table 3 Models and Parameters of Experimental Equipment

[0179] Device Name Model Parameter Vibration Sensor 1C302 <![CDATA[Frequency range: 20Hz - 12 KHz > Sound Sensor PCB378B02 <![CDATA[Frequency range: 3.75Hz - 20 KHz <!-- 13 -->]]> Plunger Hydraulic Motor 1001-0.1 Maximum Rotation Speed: 360 r / min Data Acquisition Card FK2012 16 Channels Rotation Speed Sensor ZSM12-CS-01 <![CDATA[Output current: 50 mA > Rotation Speed Display WR5135-FR-N Measurement Range: 0.2 Hz - 100 KHz

[0180] To make the experimental effect better, a lateral wear method is used to create faults along the axis of the plunger; according to the depth of wear, the plunger faults are divided into mild wear, moderate wear, and deep wear. The specific wear effects are as Figure 6 shown.

[0181] The data processing method provided by the present invention:

[0182] The experimental data obtained by the vibration sensor and the sound sensor usually contains the interference of noise signals. Especially, the detection frequency of the sound sensor is low and it is sensitive to the sounds in the environment; therefore, according to the characteristics of the plunger wear fault signal, the wavelet soft threshold algorithm is selected to decompose and reconstruct the vibration and sound signals to achieve the denoising effect on the original signal; the soft threshold function can be expressed as:

[0183]

[0184] The signal after wavelet soft threshold denoising is a time-domain signal. To better observe the characteristics of the signal and extract the frequency-domain features, the fast Fourier transform (FFT) is selected to transform the signal into the frequency-domain representation. The Fourier transform and its inverse transform formulas are respectively:

[0185]

[0186]

[0187] Figure 7 They are respectively the comparison diagrams of the denoising effects of the vibration signal and the sound signal; it can be seen from the frequency-domain comparison that after wavelet soft threshold denoising, the burrs in the signal are reduced, which reflects that the cleanliness of the signal has been improved and the denoising effect has been achieved; the present invention adopts an internal curve radial piston hydraulic motor, the number of pistons is 10, the internal curve of the housing is 6, and the experimental rotational speed is 300 r / min; from Figure 7It can be seen from the comparison that the vibration sensor is more sensitive to the half-frequency of the hydraulic motor speed, and the sound sensor is more sensitive to the first and second frequencies of the hydraulic motor speed, and the frequency response effect on a single plunger is better. Considering that the vibration sensor and the sound sensor have different response frequencies to plunger faults, the present invention combines the respective advantages of the vibration and sound signals, and fuses the characteristic values of the two signals to improve the diagnostic accuracy of the hydraulic motor plunger faults.

[0188] The feature selection method provided by the present invention:

[0189] Use the inner curve radial piston hydraulic motor test bench to simulate plunger faults, and collect vibration and sound signals of the plunger under different states; the collection frequency is set to 2000 Hz, each group of data is collected for 6 minutes, each group of data has 720,000 sampling points, and 4000 sampling points are sequentially selected to generate a sample, with a total of 180 samples. The sample numbers of the training set and the test set are 80% and 20% respectively; the plunger types and the sample numbers are shown in Table 4:

[0190] Table 4 Dataset information

[0191] Status Training Set / Test Set Label Normal 144 / 36 0 Fault A 144 / 36 1 Fault B 144 / 36 2 Fault C 144 / 36 3

[0192] Time domain and frequency domain features are common indicators reflecting the operating state of the hydraulic motor; use vibration and sound sensors to collect signals of the plunger under normal and different wear conditions. After wavelet soft threshold denoising, use mathematical methods to extract 11 types of time domain and frequency domain characteristic values for each group of data respectively; generate vibration signal and sound signal characteristic value datasets, and conduct correlation analysis of the datasets; the information correlation heat map between the datasets is as Figure 8 shown:

[0193] Figure 8 The data correlation heat map reflects the correlation information between the data. The larger the positive and negative correlation values in the figure, the more it can reflect the operating state information of the hydraulic motor; based on the heat map, 10 types of vibration signal characteristic values such as root mean square value, root amplitude, standard deviation, peak index, waveform index, spectral amplitude mean, spectral amplitude standard deviation, spectral amplitude skewness, spectral amplitude kurtosis, and frequency skewness are selected; 10 types of sound signal characteristic values such as root mean square value, root amplitude, peak value, standard deviation, peak index, pulse index, spectral amplitude mean, spectral amplitude standard deviation, frequency centroid, and frequency standard deviation are selected; finally, the normal and fault characteristic values of each group are horizontally spliced according to the corresponding data labels to form a new feature vector without changing the sample number.

[0194] The method for training and diagnosing the fusion features provided by the present invention by dividing them into a dataset including normal and 3 types of fault signal labels:

[0195] (1) Data pre-classification and visualization;

[0196] (2) Model evaluation.

[0197] The data pre-classification and visualization method provided by the present invention:

[0198] In order to pre-observing the classification effect of the fusion signal data relative to the vibration and sound signals, the T-SNE algorithm is used to visualize the data, and the T-SNE clustering map is as Figure 9 shown; it can be seen from the figure that the classification of vibration signals has a certain degree of aggregation, but there are too many aggregation points and the samples of the same category are not aggregated together; the classification of sound signals has strong discreteness, and different types of data points are fused together, and the classification effect is poor; for the classification of fusion signals, it can be seen that different types of data points are almost completely separated, and there are few intersection points between samples, and the classification effect is better;

[0199] Figure 10 is the four-class confusion matrix of the vibration signal, sound signal and fusion signal data sets, which is used to visualize the classification effects of the three data sets; the numbers of samples with labels 0-3 in the test set are 28, 32, 44 and 40 respectively; from Figure 10 the confusion matrix of the vibration signal in (a), it can be seen that there are 3 samples that are actually normal but predicted as faulty, and there are 9 misclassifications between the three types of faults; from Figure 10 the confusion matrix of the sound signal in (b), it can be seen that all samples that are actually normal are predicted correctly, but there are 23 misclassifications between the three types of faults; from Figure 10 the confusion matrix of the fusion signal in (c), it can be seen that all samples that are actually normal are predicted correctly, and there are only 5 misclassifications between fault B and fault C among the three types of faults; the classification accuracy rate of the fusion signal data set is much higher than that of the vibration signal and sound signal.

[0200] The model evaluation method provided by the present invention:

[0201] The AdaBoost algorithm is an improvement based on the Boosting algorithm, with high effectiveness and practicability; the GBDT algorithm is a gradient boosting iterative decision tree algorithm, and the XGBoost and LightGBM algorithms are improvements based on GBDT and are also the most popular decision tree boosting algorithms, with higher learning efficiency; the present invention verifies the high efficiency of the LightGBM algorithm and the classification effect on the fusion signal data by comparing the classification accuracies of the four algorithms for the fusion signal, vibration signal and sound signal respectively;

[0202] The diagnostic results of different models are as Figure 11As shown, the traditional AdaBoost algorithm has a relatively poor effect, and the diagnostic accuracy for the three datasets is more than 10% lower than that of the other three algorithms; the classification effects of the GBDT, XGBoost, and LightGBM algorithms on the three datasets are not very different, but the diagnostic accuracy of the LightGBM algorithm is 0.17% - 0.33% higher, and the diagnostic results of the three algorithms for the fused dataset are more accurate; combined with Figure 12 the training durations of different models in Figure 12 , it can be seen that the AdaBoost algorithm has the shortest training time, but its diagnostic accuracy is relatively low; the diagnostic accuracies of the GBDT, XGBoost, and LightGBM algorithms are not very different, but the training time of the LightGBM algorithm is 6 times and 3 times shorter than that of the GBDT and XGBoost algorithms respectively. Therefore, the LightGBM algorithm has higher diagnostic accuracy and higher efficiency;

[0203] The wavelet soft threshold algorithm is used to denoise the vibration signal and sound signal, extract the eigenvalues and fuse them according to the data labels to generate a fused signal dataset. Each dataset contains normal and three fault types, but the sample quantity remains unchanged; Figure 13 The fault diagnosis results of the LightGBM model for the fused signal, vibration signal, and sound signal are shown. The accuracy, precision, recall, and f1-score are introduced as evaluation indicators for classification performance; the LightGBM model is used to train and diagnose the three datasets. By comparison, it can be seen that the diagnostic accuracy for the fused dataset is 4.86% and 14.59% higher than that of the vibration and sound signals respectively; the diagnostic results of precision, recall, and f1-score for the fused dataset are also improved by 4.5% - 13.25% compared with the vibration and sound datasets;

[0204] ROC and PR curves are used to evaluate the classification performance of machine learning algorithms for a given dataset. Each dataset contains a fixed number of positive and negative samples. The larger the area under the curve, the better the classification performance of the machine learning algorithm for the dataset; FPR refers to the probability that a sample that is actually negative is predicted as positive, and TPR refers to the probability that a sample that is actually positive is predicted as negative; Precision refers to the proportion of samples predicted as positive among positive samples, and Recall refers to the proportion of positive samples predicted as positive among positive samples; the classification effects of the LightGBM algorithm on the vibration, sound, and fused signal datasets are as Figure 14 shown; it can be seen from the figure that the LightGBM algorithm has the worst classification effect on the sound dataset, and the classification effect on the fused dataset is much higher than that of the vibration and sound datasets.

[0205] A fault diagnosis method for a combined-signal hydraulic motor provided by the present invention. The plunger fault is simulated through an inner-curve radial piston hydraulic motor test bench, and the vibration signals and sound signals of the motor are collected when the plunger is in normal condition and worn to three different degrees. The wavelet soft-threshold algorithm is used for noise reduction processing, and feature selection is carried out based on data correlation and feature importance. The time-domain and frequency-domain feature values are extracted for feature fusion. A LightGBM motor plunger fault diagnosis model is constructed, and the fused features are divided into a data set containing normal and three fault signal labels for training and diagnosis. Finally, the superiority of this method is verified through comparison between different algorithms and different signals. The fault diagnosis method with fused features proposed by the present invention solves the problem of low diagnostic accuracy when using a single sensor to collect data. By using two different signals for diagnosis, the fault information of different sensors can be combined to increase fault features and improve the fault recognition rate. The LightGBM is used to establish a fault diagnosis model, which reduces the occupied running memory and improves the model training time while ensuring the classification accuracy, greatly improving the fault diagnosis efficiency.

[0206] The combined-signal hydraulic motor fault diagnosis system provided by the present invention includes an inner-curve radial piston hydraulic motor test bench for simulating the normal condition and different degrees of wear faults of the plunger, and is equipped with vibration and sound signal acquisition devices for collecting the corresponding signals of the motor in various states. It includes a signal processing module that uses the wavelet soft-threshold algorithm to perform noise reduction processing on the collected vibration and sound signals to improve the signal quality for more accurate fault diagnosis. It includes a feature extraction and fusion module for extracting the time-domain and frequency-domain feature values from the noise-reduced vibration and sound signals based on data correlation analysis and performing feature fusion to form a comprehensive feature set for fault diagnosis. It includes a motor plunger fault diagnosis model construction module based on the LightGBM algorithm for training and diagnosing the fused feature data set containing normal and three fault signal labels, so as to achieve accurate diagnosis of the hydraulic motor plunger fault. It includes a model evaluation and verification module that uses evaluation tools such as confusion matrix, PR curve, and ROC curve to evaluate and verify the performance of the diagnosis model, including the calculation of indicators such as accuracy, precision, recall rate, and f1-score, and verifies the superiority of the diagnosis method through comparative analysis of different algorithms and signal types.

[0207] The connection relationship and working principle of the combined-signal hydraulic motor fault diagnosis system provided by the present invention are as follows:

[0208] 1) Fault simulation component: This component includes an inner-curve radial piston hydraulic motor test bench for simulating the normal condition and different degrees of wear faults of the plunger. It will simultaneously collect the corresponding vibration and sound signals of the motor in various states.

[0209] 2) Signal Processing Component Module: This module receives vibration and sound signals from the fault simulation component. Then, it uses the wavelet soft threshold algorithm to denoise these signals to improve the signal quality and make it more suitable for subsequent fault diagnosis.

[0210] 3) Feature Extraction and Fusion Module: This module receives the processed signals from the signal processing component module. Based on data correlation analysis, it extracts eigenvalue from the processed time-domain and frequency-domain signals and performs feature fusion to form a comprehensive feature set for fault diagnosis.

[0211] 4) Fault Diagnosis Model Module: This module receives the comprehensive feature set from the feature extraction and fusion module. Then, it constructs a motor plunger fault diagnosis model based on the LightGBM algorithm and uses the fusion feature data set containing normal and three types of fault signal labels for model training and diagnosis.

[0212] 5) Performance Verification Component Module: This module is used to evaluate and verify the performance of the diagnosis model. It uses tools such as confusion matrix, PR curve, and ROC curve to calculate the accuracy, precision, recall rate, and f1-score metrics of the diagnosis model. In addition, it also verifies the superiority of the diagnosis method through comparative analysis of different algorithms and signal types.

[0213] The working process of the whole system is as follows: First, the fault simulation component simulates and collects the vibration and sound signals of the hydraulic motor. Then, the signal processing component module denoises these signals. Next, the feature extraction and fusion module extracts and fuses features from the processed signals. After that, the fault diagnosis model module uses these features to train and diagnose the model. Finally, the performance verification component module evaluates and verifies the performance of the diagnosis model and verifies the superiority of the diagnosis method. It should be noted that the embodiments of the present invention can be implemented by hardware, software, or a combination of software and hardware. The hardware part can be implemented using dedicated logic; the software part can be stored in a memory and executed by an appropriate instruction execution system, such as a microprocessor or dedicated design hardware. Those of ordinary skill in the art can understand that the above devices and methods can be implemented using computer-executable instructions and / or included in processor control code, such as provided on a carrier medium such as a disk, CD, or DVD-ROM, a programmable memory such as read-only memory (firmware), or a data carrier such as an optical or electronic signal carrier. The devices and modules of the present invention can be implemented by hardware circuits of programmable hardware devices such as very large scale integrated circuits or gate arrays, semiconductors such as logic chips, transistors, etc., or field programmable gate arrays, programmable logic devices, etc., can also be implemented by software executed by various types of processors, or can be implemented by a combination of the above hardware circuits and software such as firmware.

[0214] To solve the problem of low accuracy in diagnosing faults using single-signal data, improve the diagnostic efficiency of plunger wear faults in an internal curve radial piston hydraulic motor, and ensure the stable operation of the hydraulic system, a method for diagnosing plunger wear faults in a hydraulic motor based on identifying the fusion features of vibration and sound signals using LightGBM is proposed. By collecting the vibration signals and sound signals of the motor when the plunger is normal and worn to three different degrees on a fault diagnosis test bench, feature selection is carried out based on data correlation and feature importance, and time-domain and frequency-domain feature values are extracted for feature fusion. A LightGBM motor plunger fault diagnosis model is constructed, and the fusion features are divided into a data set containing normal and three fault signal labels for training and diagnosis. Finally, the superiority of this method is verified through comparisons between different algorithms and different data sets. The specific conclusions are as follows:

[0215] Build a plunger fault test bench for an internal curve radial piston hydraulic motor, and collect the vibration signals and sound signals of the plunger when the motor is running normally and in three different degrees of wear states.

[0216] Use the wavelet algorithm to denoise the two signals, and extract the time-domain and frequency-domain feature values of the signals for data fusion of the same category to generate a fusion data set.

[0217] Build a LightGBM fault diagnosis model and compare it with the AdaBoost, GBDT, and XGBoost models. The results show that the fault diagnosis accuracy of the LightGBM model is 0.17%-0.33% higher, and the training time is shortened by 3-5 times, verifying that the LightGBM model has higher diagnostic efficiency.

[0218] Through comparison, it shows that the diagnostic accuracy of the LightGBM model for the fusion signal data set is 4.86% and 14.59% higher than that of the vibration signal and the sound signal respectively, and evaluation indicators such as precision, recall, and f1-score are also 4.5%-13.25% higher than those of the vibration and sound data sets.

[0219] Through the analysis of the ROC and PR curves, it can be seen that the classification effect of the fusion data set using the LightGBM algorithm is better than that of the vibration and sound data.

[0220] The method proposed in this invention has achieved the expected results in the fault diagnosis of plunger hydraulic motors, but there are still some deficiencies. Future research should also consider the following issues: 1) When collecting sound signals, the sound sensor is directly exposed to the environment. A soundproof cover can be added to reduce the interference of environmental noise. 2) In the experiment, each group of signals was collected for 6 minutes, and the number of data points is not large. The number of samples can be increased to improve the fault diagnosis accuracy.

[0221] A fault diagnosis method for a fused signal hydraulic motor provided by the present invention. The plunger fault is simulated through an inner curve radial piston hydraulic motor test bench, and the vibration signals and sound signals of the motor are collected when the plunger is normal and worn to three different degrees. The wavelet soft threshold algorithm is used for noise reduction processing, and feature selection is carried out based on data correlation and feature importance. The time-domain and frequency-domain feature values are extracted for feature fusion. A LightGBM motor plunger fault diagnosis model is constructed, and the fused features are divided into a data set containing normal and three fault signal labels for training and diagnosis. Finally, the superiority of this method is verified by the comparison between different algorithms and different signals. The fault diagnosis method with fused features proposed by the present invention solves the problem of low diagnostic accuracy when using a single sensor to collect data. By using two different signals for diagnosis, the fault information of different sensors can be combined to increase fault features and improve the fault recognition rate. The LightGBM is used to establish a fault diagnosis model, which reduces the occupied running memory and improves the model training time while ensuring the classification accuracy, greatly improving the fault diagnosis efficiency.

[0222] As described above, the above are only specific embodiments of the present invention, but the protection scope of the present invention is not limited thereto. Any person skilled in the art within the technical scope disclosed by the present invention, any modifications, equivalent replacements, and improvements made within the spirit and principle of the present invention shall be covered by the protection scope of the present invention.

Claims

1. A fault diagnosis method for a combined-signal hydraulic motor, characterized in that First, simulate the plunger fault through an inner curve radial piston hydraulic motor test bench, and collect the vibration and sound signals of the motor in the normal state and under different degrees of wear. Secondly, use the wavelet soft threshold algorithm to denoise these signals, and then based on data correlation analysis, extract the eigenvalues from the time domain and frequency domain and perform feature fusion. Then, construct a motor plunger fault diagnosis model based on LightGBM, and use a data set containing normal and three types of fault signal labels for model training and diagnosis. Among them, the LightGBM model integrates the histogram algorithm, divides the original data of the features into k discrete features, constructs k bins for statistical features, and does not need to traverse the data during the search, only needs to traverse k bins to find the best split point. And the split method of the LightGBM model adopts the Leaf-wise growth strategy, and searches for the leaf with the largest information gain in the current leaf. This LightGBM model also integrates the gradient-based one-sided sampling algorithm and the exclusive feature bundling algorithm. The one-sided sampling algorithm uses gradient information to sample the samples, and the exclusive feature bundling algorithm starts from the perspective of reducing the feature dimension and bundles multiple exclusive features together. Among them, use vibration sensors and sound sensors to collect signals when the plunger is normal and under different degrees of wear. After wavelet soft threshold denoising, use mathematical methods to extract the time domain and frequency domain eigenvalues of each group of data respectively. Generate the vibration signal and sound signal eigenvalue data sets, and perform data set correlation analysis. Select 10 types of vibration signal eigenvalues including root mean square value, root amplitude, standard deviation, peak index, waveform index, spectral amplitude mean, spectral amplitude standard deviation, spectral amplitude skewness, spectral amplitude kurtosis, and frequency skewness. Select 10 types of sound signal eigenvalues including root mean square value, root amplitude, peak value, standard deviation, peak index, pulse index, spectral amplitude mean, spectral amplitude standard deviation, frequency centroid, and frequency standard deviation. Finally, horizontally splice the normal and fault eigenvalues of each group according to the corresponding data labels to form a new feature vector. To pre-observing the classification effect of the fused signal data relative to the vibration and sound signals, use the T-SNE algorithm to visualize the data.

2. A fusion signal hydraulic motor fault diagnosis system based on the method described in claim 1, characterized in that, Including: Fault simulation component: including an inner curve radial piston hydraulic motor test bench, which is used to simulate the normal state and different degrees of wear faults of the plunger, and collect the corresponding vibration and sound signals. Signal processing component module: use the wavelet soft threshold algorithm to denoise the collected signals to improve the signal quality for subsequent analysis. Feature extraction and fusion module: based on data correlation analysis, extract eigenvalues from the processed time domain and frequency domain signals, and perform feature fusion to form a comprehensive feature set for fault diagnosis. Fault diagnosis model module: construct a motor plunger fault diagnosis model based on the LightGBM algorithm, and use the fused feature data set containing normal and three types of fault signal labels for model training and diagnosis. Performance verification component module: It includes tools for evaluating and verifying the performance of diagnostic models, such as confusion matrix, PR curve, and ROC curve, and conducts comparative analysis of different algorithms and signal types to verify the effectiveness and superiority of the proposed method.

3. A fusion signal hydraulic motor fault diagnosis system based on the method described in claim 1, characterized in that, Including: An inner curve radial piston hydraulic motor test bench, which is used to simulate normal and different degrees of wear faults of the piston, and is equipped with vibration and sound signal acquisition devices to collect corresponding signals of the motor in various states; A signal processing module, which uses the wavelet soft threshold algorithm to denoise the collected vibration and sound signals to improve the signal quality for more accurate fault diagnosis; A feature extraction and fusion module, which is used to extract time-domain and frequency-domain feature values from the denoised vibration and sound signals based on data correlation analysis, and conduct feature fusion to form a comprehensive feature set for fault diagnosis; A motor piston fault diagnosis model construction module based on the LightGBM algorithm, which is used to train and diagnose a fusion feature data set containing normal and three fault signal labels, so as to achieve accurate diagnosis of hydraulic motor piston faults; A model evaluation and verification module, which uses evaluation tools such as confusion matrix, PR curve, and ROC curve to evaluate and verify the performance of the diagnostic model, including the calculation of accuracy, precision, recall rate, and f-score indicators, and verifies the superiority of the diagnostic method through comparative analysis of different algorithms and signal types.

Images