Weak fault diagnosis method for ball screw pair based on wireless follow-up vibration sensor

Abstract

The invention discloses a ball screw pair weak fault diagnosis method based on a wireless follow-up vibration sensor, and relates to the technical field of mechanical fault detection. A ball screw pair weak fault test platform is built, various weak fault simulation conditions and normal operation states are checked, one-dimensional vibration signals in different weak fault states are collected in real time through the ball screw pair weak fault test platform, and wavelet noise reduction and scale unified processing are carried out on the one-dimensional vibration signals; converting the processed one-dimensional vibration signal into a two-dimensional image by using GADF coding to form an initial data set; expanding the initial data set, constructing a weak fault classification model, dividing the expanded initial data set into a test set and a training set, and respectively inputting the test set and the training set into the constructed weak fault classification model for training and verification; by adopting the ball screw pair weak fault diagnosis method based on the wireless follow-up rotation vibration sensor, the feature information of industrial equipment can be automatically extracted, and the intelligence and the accuracy of the equipment are improved.

Description

Technical Field

[0001] This invention relates to the field of mechanical fault detection technology, and in particular to a method for diagnosing minor faults in ball screw pairs based on wireless rotational vibration sensors. Background Technology

[0002] As a core component of mechanical transmission systems, the operating status of ball screw pairs directly affects the accuracy, efficiency, and reliability of equipment. Once faults such as ball wear, screw cracking, or nut loosening occur, it may lead to a decrease in equipment processing accuracy, abnormal operation, or even shutdown, resulting in significant economic losses. With the development of industrial automation, the stability requirements of equipment transmission systems are constantly increasing. Therefore, efficient fault diagnosis of ball screw pairs has become a key technology to ensure the safe operation of equipment. Fault diagnosis technology mainly consists of three stages: signal acquisition, feature extraction, and fault classification.

[0003] The accuracy and robustness of fault classification depend on the effectiveness of feature extraction and the adaptability of the model. Traditional machine learning methods rely on manual feature engineering, have limited ability to process high-dimensional data, and are prone to overfitting or parameter sensitivity problems. Although a single deep learning model can automatically extract features, it is difficult to combine the spatial features and time series information of dynamic signals, and it does not fully capture the multi-dimensional features of complex fault data. Although most methods can identify faults to a certain extent, they are still insufficient in fault classification, diagnosis of weak faults, and model generalization ability, making it difficult to meet the needs of accurate diagnosis of early weak faults in industrial scenarios. Summary of the Invention

[0004] The purpose of this invention is to provide a method for diagnosing weak faults in ball screw pairs based on wireless rotational vibration sensors. By converting one-dimensional vibration signals into two-dimensional images and combining them with a multi-modal feature fusion model, this method achieves efficient and accurate classification of weak faults in ball screw pairs. It improves upon the difficulties in extracting nonlinear features of weak faults and the insufficient fusion of multi-modal information in traditional fault diagnosis. Furthermore, the model has strong generalization ability and can accurately identify weak faults under complex working conditions without affecting the normal operation of the ball screw pair. This provides an efficient solution for early warning and intelligent maintenance of weak faults in industrial equipment, significantly improving the accuracy and generalization ability of weak fault diagnosis.

[0005] To achieve the above objectives, this invention provides a method for diagnosing minor faults in ball screw pairs based on a wireless rotational vibration sensor, comprising the following steps: S1: Build a ball screw pair weak fault test platform and verify the simulation of various weak faults and normal operation status. S2: The one-dimensional vibration signal under different weak fault conditions is collected in real time by the wireless rotation vibration sensor in the S1 test platform, and the one-dimensional vibration signal is processed by wavelet noise reduction and the data scale is unified. S3: The processed one-dimensional vibration signal obtained in S2 is converted into a 64×64 pixel two-dimensional image by Gram angle difference domain image GADF encoding to form the initial dataset; S4: Expand the initial dataset generated in S3, and divide the expanded dataset into a training set and a test set according to a ratio of 70% and 30%, respectively. The proportion of weak fault samples in the test set shall not be less than 50%. S5: Construct a CNN-GRU-Attention weak fault classification model, including a CNN module, a GRU module, and an attention mechanism; the CNN module extracts image spatial features, the GRU module captures temporal dependencies, the attention mechanism optimizes the weight allocation of key features, performs structural optimization on weak fault features, and outputs the identification results of four fault categories; S6: Use the Adam optimizer to train the CNN-GRU-Attention weak fault classification model in S5 until the weak fault classification model converges on the training set obtained in S4. S7: Evaluate the performance of the CNN-GRU-Attention model using the test set in S4. Complete the classification and identification of four fault states using the confusion matrix and accuracy metrics. If the accuracy is not up to standard, return to S5 for retraining. If the accuracy is up to standard, complete the classification of fault signals.

[0006] Preferably, in step S2, the wireless rotational vibration sensor is installed at the end of the ball screw assembly via a sensor connecting rod to directly capture the one-dimensional vibration signal of the rotating component. The specific steps are as follows: S21: External testing equipment is used to test whether the wireless rotation vibration sensor's signal acquisition function is normal; if the signal acquisition function is abnormal, replace the wireless rotation vibration sensor until the signal acquisition function is normal, then proceed to the next step. S23: A connecting sleeve is set on the wireless rotation vibration sensor. The two ends of the connecting sleeve are connected to the sensor connecting rod and the 3D printed cylindrical shell, respectively. The 3D printed cylindrical shell houses the sensing unit, the processing unit and the battery. S24: The wireless rotational vibration sensor rotates synchronously with the ball screw pair to capture the one-dimensional vibration signal of the ball screw pair.

[0007] Preferably, the specific process of expanding the initial dataset in S4 is as follows: S41: Define the core parameters of the initial dataset, use a weighted augmentation strategy to assign higher augmentation weights to weakly faulty samples and lower augmentation weights to normal samples, and calculate the proportion of weakly faulty samples. S42: Add ±5% Gaussian noise to the one-dimensional vibration signal to simulate the characteristics of weak fault signals under different noise environments; S43: Increase the diversity of the 64×64 pixel 2D image by random cropping to complete the initial dataset expansion.

[0008] Preferably, the calculation logic for strengthening weak fault samples through a weighted augmentation strategy in S41 is as follows: S411: Set weighting coefficients: The augmentation weighting coefficients for weak fault samples are as follows: The augmentation weighting coefficient for normal state samples is ,satisfy Set the weight ratio: ,in k This is the weighting amplification factor. k =1.5~2.0; S412: Calculate the number of augmented samples per class: The target number of augmented samples for the normal state is... , ,in The number of samples in the normal state; the number of target augmentations for each type in the weak fault samples is... The calculation method is as follows: ,in, This indicates the initial number of minor ball bearing fault samples; the initial number of minor lead screw fault samples. Initial number of samples of minor nut faults Initial number of ball bearing minor fault samples same; S413: Calculate the proportion of weak fault samples after expansion.

[0009] Preferably, the specific processing procedure for adding ±5% Gaussian noise in S42 is as follows: S421: Set the one-dimensional vibration signal as... , Given the number of signal sampling points, calculate the peak value of the one-dimensional vibration signal. : ; S422: Set the noise intensity range to within ±5%, and the standard deviation of the noise. The relationship with the peak value is: Based on the Gaussian distribution characteristics, 99.7% of the noise values ​​fall within... Within the range, the peak value of the one-dimensional vibration signal is ; The generated follows a mean of 0 and a standard deviation of . Gaussian noise sequence , ; S423: The one-dimensional vibration signal in S421 Compared with the Gaussian noise sequence obtained in S422 The augmented one-dimensional vibration signal is obtained by superposition. : ; The amplitude range is set at 10. -3 ~10 1 m / s 2 .

[0010] Preferably, S43 performs the following operations: S431. Input a 64×64 pixel 2D image, and set the pixel matrix of the 2D image as follows. , For the row index of the original image, =1~64, For column indexes of the original image, =1~64, the range of pixel grayscale values ​​is set to [0, 255]; S432: Perform a standard crop, and the cropped image size is 70%~90% of the original size; S433: Calculate the clipping coordinates; S434: Pixel matrix of the cropped image Represented as: ; in, This is the row index of the cropped image. =1~56, For the column index of the cropped image, =1~56.

[0011] Preferably, the CNN module in S5 contains two convolutional blocks, and the identification of four fault categories specifically includes the following steps: S51: Extracting spatial features: The first convolutional block uses a 5×5 convolutional kernel with 8 channels. Combined with the same padding method, after batch normalization and sigmoid activation, it is reduced in dimensionality by 2×2 max pooling. The second convolutional block uses a 5×5 convolutional kernel with 16 channels. It performs the same batch normalization, activation, and pooling dimensionality reduction operations as the first convolutional block. The weak fault spatial features are extracted by multi-scale convolutional kernel deflection. S52: Strengthen the weight allocation of key features of weak faults: After the extracted spatial features are flattened by the dropout layer, they are input into the GRU layer to capture the temporal dependence in the weak fault signal. The number of hidden layer neurons in the GRU layer is increased to 128. The weight allocation of key features of weak faults is strengthened by the dynamic weight allocation algorithm in the attention mechanism. S53: After being processed by 128 neurons in the fully connected layer and the sigmoid activation function, the softmax classification layer outputs the identification results of four weak fault categories.

[0012] Preferably, the initial learning rate is set to 1e-3 in S6.

[0013] Preferably, the dropout rate in S52 is set to 0.5.

[0014] Therefore, the ball screw pair weak fault diagnosis method based on a wireless rotational vibration sensor described above has the following advantages compared with the prior art: 1. This application uses a wireless vibration sensor to accurately collect weak fault signals. When the ball screw pair is running, it can accurately collect weak fault signals with an amplitude of less than 0.01mm, which improves the problem that traditional wired vibration sensors are not good at capturing weak fault signals and are easily interfered with. 2. In this invention, GADF encodes the time series into a two-dimensional matrix using trigonometric functions and amplifies the minute fluctuations of weak faults through angle sine and cosine transformation, thereby enhancing the spatial representation capability of weak faults while preserving the temporal characteristics of the signal.

[0015] 3. This invention combines GADF with the CNN-GRU-Attention model, which solves the shortcomings of traditional methods in the nonlinear feature extraction and multimodal fusion of weak faults, and realizes high-precision diagnosis of weak faults in ball screw pairs, providing reliable technical support for condition monitoring and intelligent maintenance of industrial equipment.

[0016] The technical solution of the present invention will be further described in detail below with reference to the accompanying drawings and embodiments. Attached Figure Description

[0017] Figure 1 This is an overall flowchart of the method for diagnosing weak faults in ball screw pairs based on a wireless rotational vibration sensor, as described in this invention. Figure 2 This is an experimental installation diagram of the wireless rotational vibration sensor in an embodiment of the present invention; Figure 3 This is a diagram of triaxial acceleration signals collected by the wireless rotational vibration sensor in this embodiment of the invention; Figure 4 This is a low-frequency spectrum diagram of the triaxial acceleration signal collected by the wireless rotation vibration sensor in this embodiment of the invention; Figure 5 This is a broadband spectrum diagram of the triaxial acceleration signal collected by the wireless rotation vibration sensor in this embodiment of the invention; Figure 6This is the Gracaoum difference domain diagram obtained by the Gracaoum transform of the fault signal in this embodiment of the invention; Figure 7 This is a structural diagram of the CNN-GRU-Attention model in an embodiment of the present invention; Figure 8 This is a diagram showing the classification results of fault signals by different models in the comparative examples of this invention; Figure Labels 1. Ball screw assembly; 2. Wireless vibration sensor; 3. Controller. Detailed Implementation

[0018] In the description of this invention, it should be noted that the terms "upper," "lower," "inner," "outer," etc., indicate the orientation or positional relationship based on the orientation or positional relationship shown in the accompanying drawings, or the orientation or positional relationship in which the product of this invention is usually placed when in use. They are only for the convenience of describing this invention and simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation. Therefore, they should not be construed as limiting this invention.

[0019] Example like Figure 1 As shown, the present invention provides a method for diagnosing minor faults in ball screw pairs based on wireless rotational vibration sensors. Includes the following steps: S1: Build a ball screw pair weak fault testing platform and verify the simulation of various weak faults and normal operating conditions; such as... Figure 2 As shown, the wireless rotation vibration sensor 2 is installed at the end of the ball screw assembly 1 via the sensor connecting rod, and the wireless rotation vibration sensor 2 is connected to the controller 3.

[0020] S2: One-dimensional vibration signals under different weak fault conditions are collected in real time using the wireless rotational vibration sensor in the S1 test platform. Wavelet noise reduction processing is then applied to the one-dimensional vibration signals, and the data scale is standardized. For example... Figures 3-5 As shown, the wireless rotational vibration sensor directly captures the one-dimensional vibration signal of the rotating component. The collected one-dimensional vibration signal is then transmitted to a computer for storage via a data acquisition device. The specific steps are as follows: S21: External testing equipment is used to test whether the wireless rotation vibration sensor's signal acquisition function is normal; if the signal acquisition function is abnormal, replace the wireless rotation vibration sensor until the signal acquisition function is normal, then proceed to the next step. S23: The wireless rotational vibration sensor ORS is equipped with a connecting sleeve. The total weight of ORS is only 87g. It adopts an axially symmetrical installation method. The two ends of the connecting sleeve are connected to the sensor connecting rod and the 3D printed cylindrical shell, respectively. The 3D printed cylindrical shell houses the sensing unit, processing unit and battery. The sensing unit is an ADXL345 triaxial MEMS accelerometer, and the processing unit is an nRF52840 processor. The sensing unit is located at the bottom of the 3D printed cylindrical shell, which is separate from the processing unit to reduce vibration interference.

[0021] The 3D-printed cylindrical shell is bolted to the connecting base plate. The connecting base plate and the connecting sleeve are connected to avoid dynamic imbalance, further improving assembly accuracy and stability, and ensuring coaxial rotation of the ORS and the workpiece. The connecting sleeve does not require modification of the lathe's tool post, spindle, or other original components. The installation of the sensing system can be completed simply by detaching the connecting sleeve to the end of the ball screw pair, avoiding interference with the cutting process. S24: The wireless vibration sensor rotates synchronously with the ball screw pair, capturing the one-dimensional vibration signal of the ball screw pair. This "source sensing" method can directly capture the original vibration signal of the rotor system, avoiding energy loss and noise interference during vibration transmission, achieving high-sensitivity acquisition, and providing high-quality raw data for subsequent feature extraction.

[0022] S3: As Figure 6 As shown, the processed one-dimensional vibration signal obtained in S2 is converted into a 64×64 pixel two-dimensional image by Gram angle difference domain image GADF encoding. The signal difference of weak faults is amplified by angle mapping, the temporal features of the signal are preserved and the spatial representation ability of weak faults is enhanced, forming an initial dataset containing 800 samples / classes, with a total of 3200 samples. S4: Expand the initial dataset generated in S3, and divide the expanded dataset into a training set and a test set according to a ratio of 70% and 30%, respectively. The proportion of weak fault samples in the test set shall not be less than 50%. The specific process of expanding the initial dataset in S4 is as follows: S41: Define the core parameters of the initial dataset, use a weighted augmentation strategy to assign higher augmentation weights to weakly faulty samples and lower augmentation weights to normal samples, and calculate the proportion of weakly faulty samples. The initial dataset includes four classes of samples: Normal state sample set: Sample size =800; Ball bearing minor fault sample set: Sample size =800; Minor fault sample set of leadscrew: Sample size =800; Minor nut fault sample set: Sample size =800; Total number of minor fault samples: = + + =2400; Total number of samples in the initial dataset: = + =3200; The calculation logic for strengthening weak fault samples through a weighted augmentation strategy in S41 is as follows: S411: Set weighting coefficients: The augmentation weighting coefficients for weak fault samples are as follows: The augmentation weighting coefficient for normal state samples is ,satisfy Set the weight ratio: ,in k This is the weighting amplification factor. k =1.5~2.0; S412: Calculate the number of augmented samples per class: The target number of augmented samples for the normal state is... , ,in The number of samples in the normal state; the number of target augmentations for each type in the weak fault samples is... The calculation method is as follows: ,in, This indicates the initial number of minor ball bearing fault samples; the initial number of minor lead screw fault samples. Initial number of samples of minor nut faults Initial number of ball bearing minor fault samples same; S413: Calculate the proportion of weak fault samples after expansion; set up k =2.0, , =800, =800, obtained through the calculation formulas in S411 and S412: Target augmentation number of normal state samples =800; Weighting coefficient for weak fault samples =2.0; Number of target augmentations for each type in the weak fault samples =2.0 ; pass The number of normal samples after augmentation was calculated. =800+800=1600; pass = + The number of ball bearing weak fault samples after expansion was calculated. = ; pass = + The number of minor fault samples of the leadscrew was calculated after expansion. = ; pass = + The number of weak nut fault samples after expansion was calculated. = ; Total number of weak fault samples after expansion = + + =7200; Total number of samples in the expanded dataset = + =8800; Calculate the proportion of weak fault samples after expansion. The proportion of samples with minor faults is greater than 50%, which meets the core objective of enhancing the proportion of samples with minor faults.

[0023] S42: Add ±5% Gaussian noise to the one-dimensional vibration signal to simulate the characteristics of weak fault signals under different noise environments; The specific process for adding ±5% Gaussian noise to S42 is as follows: S421: Set the one-dimensional vibration signal as... , Given the number of signal sampling points, calculate the peak value of the one-dimensional vibration signal. : ; S422: Set the noise intensity range to within ±5%, and the standard deviation of the noise. The relationship with the peak value is: Based on the Gaussian distribution characteristics, 99.7% of the noise values ​​fall within... Within the range, the peak value of the one-dimensional vibration signal is ; The generated follows a mean of 0 and a standard deviation of . Gaussian noise sequence , ; S423: The one-dimensional vibration signal in S421 Compared with the Gaussian noise sequence obtained in S422 The augmented one-dimensional vibration signal is obtained by superposition. : ; The amplitude range is set at 10. -3 ~10 1 m / s 2 .

[0024] S43: Increase the diversity of the 64×64 pixel two-dimensional image by random cropping to complete the initial dataset expansion; S43 performs the following operations: S431. Input a 64×64 pixel 2D image, and set the pixel matrix of the 2D image as follows. , For the row index of the original image, =1~64, For column indexes of the original image, =1~64, the range of pixel grayscale values ​​is set to [0, 255]; S432: Perform a standard crop, and the cropped image size is 70%~90% of the original size; S433: Calculate the clipping coordinates; Calculate the starting coordinates of the top left corner of the cropping region. The cropped area is entirely within the original image. The calculation process is as follows: x-axis Scope: , H The height of the original image. This represents the height of the cropped image. Randomly select an integer between 1 and 9; ordinate Scope: , W The original image width, This represents the width of the cropped image. Randomly select an integer between 1 and 9; Generated by a uniform random function , , This indicates that an integer is randomly selected.

[0025] S434: Pixel matrix of the cropped image Represented as: ; in, This is the row index of the cropped image. =1~56, For the column index of the cropped image, =1~56.

[0026] S5: As Figure 7 As shown, a CNN-GRU-Attention weak fault classification model is constructed, including a CNN module, a GRU module, and an attention mechanism. The CNN module extracts spatial features of the image, the GRU module captures temporal dependencies, and the attention mechanism optimizes the weight allocation of key features. The weak fault features are structurally optimized, and the recognition results of four fault categories are output. The CNN module in S5 contains two convolutional blocks, and the identification of four fault categories specifically includes the following steps: S51: Extracting spatial features: The first convolutional block uses a 5×5 convolutional kernel with 8 channels. Combined with the same padding method, after batch normalization and sigmoid activation, it is reduced in dimensionality by 2×2 max pooling. The second convolutional block uses a 5×5 convolutional kernel with 16 channels. It performs the same batch normalization, activation, and pooling dimensionality reduction operations as the first convolutional block. The extraction of weak fault spatial features is completed through multi-scale convolutional kernels. S52: Enhancing the weight allocation of key features in weak faults: The extracted spatial features are flattened by a dropout layer with a dropout rate of 0.5. The input GRU layer captures the temporal dependencies in the weak fault signals. The number of hidden neurons in the GRU layer is increased to 128, and the weight allocation of key features in weak faults is enhanced through a dynamic weight allocation algorithm in the attention mechanism. S53: After processing by 128 neurons in the fully connected layer and the sigmoid activation function, the softmax classification layer outputs the identification results of four weak fault categories. The four weak faults include normal operation, ball bearing weak fault, lead screw weak fault, and nut weak fault.

[0027] S6: Train the CNN-GRU-Attention weak fault classification model from S5 using the Adam optimizer, with an initial learning rate of 1e-3. Continue until the weak fault classification model converges on the training set obtained in S4; S7: Evaluate the performance of the CNN-GRU-Attention model using the test set in S4. Complete the classification and identification of four fault states using the confusion matrix and accuracy metrics. If the accuracy is not up to standard, return to S5 for retraining. If the accuracy is up to standard, complete the classification of fault signals.

[0028] Comparative Example Table 1 shows a comparison between the wireless rotation-following vibration sensor and the traditional wired vibration sensor. The wireless rotation-following vibration sensor has the advantages of high signal-to-noise ratio, high sensitivity, and all-dimensional monitoring in the process of acquiring signals. Table 1

[0029] As can be seen from Table 1, in this scheme, the wireless rotational vibration sensor is directly attached to the rotating ball screw pair spindle through the connecting sleeve, which can more sensitively capture the dynamic vibration of the ball screw pair spindle, especially responding more directly to the radial and tangential acceleration changes caused by cutting force.

[0030] Significantly improved signal-to-noise ratio: The wireless rotor vibration sensor directly collects the original vibration of the rotor, which can effectively filter out the noise in the transmission process. Experiments show that the signal-to-noise ratio of the weak fault signals it collects is 3-5 times that of traditional wired vibration sensors.

[0031] Outstanding sensitivity to weak faults: The wireless rotating vibration sensor integrates a high-sensitivity sensing module, while traditional wired vibration sensors, due to transmission loss, have a capture rate of less than 30% for such early weak fault signals.

[0032] Supports three-axis full-dimensional monitoring: The wireless rotation vibration sensor can simultaneously collect vibration acceleration signals in the X, Y, and Z directions, covering the radial and axial vibration characteristics during the rotation of the ball screw pair. Compared with traditional wired vibration sensors, it can more comprehensively reflect the fault state, such as the axial vibration of a loose nut and the radial impact of a cracked screw.

[0033] Strong adaptability to complex working conditions: Under dynamic working conditions such as variable speed and variable load, the signal collected by the infinitely rotating vibration sensor can reflect the dynamic vibration changes of the ball screw pair in real time. Combined with wavelet denoising and GADF image conversion in this application, it can adapt to the feature drift caused by working condition fluctuations and solve the problem that the fixed acquisition template of traditional wired vibration sensors cannot cope with changes in working conditions.

[0034] Multi-sensor parallel operation and low maintenance: The wireless rotation vibration sensor supports wireless parallel transmission and can be deployed on multiple sets of ball screw pairs or multiple rotating parts of the same device without considering cable layout compatibility; at the same time, the wireless design reduces failure points such as cable wear and loose interfaces, and the maintenance cost is reduced by more than 60% compared with traditional wired sensing systems.

[0035] Supports multimodal feature fusion: The high-quality temporal vibration signal collected by the infinite-tracking vibration sensor is converted into a two-dimensional image containing time-space features by GADF, and can be efficiently utilized by the CNN module and GRU module of the CNN-GRU-Attention model.

[0036] Comparative experiments were conducted to evaluate the performance of CNN-GRU-Attention, LeNet-5, AlexNet, and VGGNet models using a test set. The classification of various weak faults was analyzed using a confusion matrix, and the weak fault recognition rate was used as the core indicator to verify the model's effectiveness.

[0037] Experimental results are as follows Figure 8 As shown, LeNet-5, AlexNet, VGGNet and other comparative models have serious misclassification in weak fault identification, while the CNN-GRU-Attention model has almost no misclassified samples for weak faults, with an average diagnostic accuracy of 99.99% and a weak fault identification rate of 99.8%, which is significantly better than other comparative models.

[0038] Therefore, this invention adopts the aforementioned method for diagnosing weak faults in ball screw pairs based on wireless rotational vibration sensors. This method balances the effectiveness of weak fault feature extraction with the practicality of the model. By accurately acquiring weak fault signals through wireless rotational vibration sensors and combining GADF with the CNN-GRU-Attention model, it overcomes the shortcomings of traditional methods in nonlinear feature extraction and multimodal fusion of weak faults, achieving high-precision diagnosis of weak faults in ball screw pairs and providing reliable technical support for condition monitoring and intelligent maintenance of industrial equipment.

[0039] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and not to limit them. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can still be made to the technical solutions of the present invention, and these modifications or equivalent substitutions cannot cause the modified technical solutions to deviate from the spirit and scope of the technical solutions of the present invention.

Claims

1. A method for diagnosing minor faults in ball screw pairs based on wireless rotational vibration sensors, characterized in that, Includes the following steps: S1: Build a ball screw pair weak fault test platform and verify the simulation of various weak faults and normal operation status. S2: The one-dimensional vibration signal under different weak fault conditions is collected in real time by the wireless rotation vibration sensor in the S1 test platform, and the one-dimensional vibration signal is processed by wavelet noise reduction and the data scale is unified. S3: The processed one-dimensional vibration signal obtained in S2 is converted into a 64×64 pixel two-dimensional image by Gram angle difference domain image GADF encoding to form the initial dataset; S4: Expand the initial dataset generated in S3, and divide the expanded dataset into a training set and a test set according to a ratio of 70% and 30%, respectively. The proportion of weak fault samples in the test set shall not be less than 50%. S5: Construct a CNN-GRU-Attention weak fault classification model, including a CNN module, a GRU module, and an attention mechanism; the CNN module extracts image spatial features, the GRU module captures temporal dependencies, the attention mechanism optimizes the weight allocation of key features, performs structural optimization on weak fault features, and outputs the identification results of four fault categories; S6: Use the Adam optimizer to train the CNN-GRU-Attention weak fault classification model in S5 until the weak fault classification model converges on the training set obtained in S4. S7: Evaluate the performance of the CNN-GRU-Attention model using the test set in S4. Complete the classification and identification of four fault states using the confusion matrix and accuracy metrics. If the accuracy is not up to standard, return to S5 for retraining. If the accuracy is up to standard, complete the classification of fault signals.

2. The method for diagnosing minor faults in ball screw pairs based on a wireless rotational vibration sensor according to claim 1, characterized in that: In S2, the wireless rotational vibration sensor is installed at the end of the ball screw assembly via a sensor connecting rod to directly capture the one-dimensional vibration signal of the rotating component. The specific steps are as follows: S21: External testing equipment is used to test whether the wireless rotation vibration sensor's signal acquisition function is normal; if the signal acquisition function is abnormal, replace the wireless rotation vibration sensor until the signal acquisition function is normal, then proceed to the next step. S23: A connecting sleeve is set on the wireless rotation vibration sensor. The two ends of the connecting sleeve are connected to the sensor connecting rod and the 3D printed cylindrical shell, respectively. The 3D printed cylindrical shell houses the sensing unit, the processing unit and the battery. S24: The wireless rotational vibration sensor rotates synchronously with the ball screw pair to capture the one-dimensional vibration signal of the ball screw pair.

3. The method for diagnosing minor faults in ball screw pairs based on a wireless rotational vibration sensor according to claim 1, characterized in that: The specific process of expanding the initial dataset in S4 is as follows: S41: Define the core parameters of the initial dataset, use a weighted augmentation strategy to assign higher augmentation weights to weakly faulty samples and lower augmentation weights to normal samples, and calculate the proportion of weakly faulty samples. S42: Add ±5% Gaussian noise to the one-dimensional vibration signal to simulate the characteristics of weak fault signals under different noise environments; S43: Increase the diversity of the 64×64 pixel 2D image by random cropping to complete the initial dataset expansion.

4. The method for diagnosing minor faults in ball screw pairs based on a wireless rotational vibration sensor according to claim 3, characterized in that: The calculation logic for strengthening weak fault samples through a weighted augmentation strategy in S41 is as follows: S411: Set weighting coefficients: The augmentation weighting coefficients for weak fault samples are as follows: The augmentation weighting coefficient for normal state samples is ,satisfy Set the weight ratio: ,in k This is the weighting amplification factor. k =1.5~2.0; S412: Calculate the number of augmented samples per class: The target number of augmented samples for the normal state is... , ,in The number of samples in the normal state; the number of targets expanded in each category in the weak fault samples is... The calculation method is as follows: ,in, This indicates the initial number of minor ball bearing fault samples; the initial number of minor lead screw fault samples. Initial number of samples of minor nut faults Initial number of ball bearing minor fault samples same; S413: Calculate the proportion of weak fault samples after expansion.

5. The method for diagnosing minor faults in ball screw pairs based on a wireless rotational vibration sensor according to claim 4, characterized in that: The specific process for adding ±5% Gaussian noise to S42 is as follows: S421: Set the one-dimensional vibration signal as... , Given the number of signal sampling points, calculate the peak value of the one-dimensional vibration signal. : ; S422: Set the noise intensity range to within ±5%, and the standard deviation of the noise. The relationship with the peak value is: Based on the Gaussian distribution characteristics, 99.7% of the noise values ​​fall within... Within the range, the peak value of the one-dimensional vibration signal is ; The generated follows a mean of 0 and a standard deviation of . Gaussian noise sequence , ; S423: The one-dimensional vibration signal in S421 Compared with the Gaussian noise sequence obtained in S422 The augmented one-dimensional vibration signal is obtained by superposition. : ; The amplitude range is set at 10. -3 ~10 1 m / s 2 .

6. The method for diagnosing minor faults in ball screw pairs based on a wireless rotational vibration sensor according to claim 5, characterized in that: S43 performs the following operations: S431. Input a 64×64 pixel 2D image, and set the pixel matrix of the 2D image as follows. , For the row index of the original image, =1~64, For column indexes of the original image, =1~64, the range of pixel grayscale values ​​is set to [0, 255]; S432: Perform a standard crop, and the cropped image size is 70%~90% of the original size; S433: Calculate the clipping coordinates; S434: Pixel matrix of the cropped image Represented as: ； in, This is the row index of the cropped image. =1~56, For the column index of the cropped image, =1~56.

7. The method for diagnosing weak faults in ball screw pairs based on a wireless rotational vibration sensor according to claim 1, characterized in that: The CNN module in S5 contains two convolutional blocks, and the identification of four fault categories specifically includes the following steps: S51: Extracting spatial features: The first convolutional block uses a 5×5 convolutional kernel with 8 channels. Combined with the same padding method, after batch normalization and sigmoid activation, it is reduced in dimensionality by 2×2 max pooling. The second convolutional block uses a 5×5 convolutional kernel with 16 channels. It performs the same batch normalization, activation, and pooling dimensionality reduction operations as the first convolutional block. The extraction of weak fault spatial features is completed through multi-scale convolutional kernels. S52: Strengthen the weight allocation of key features of weak faults: After the extracted spatial features are flattened by the dropout layer, they are input into the GRU layer to capture the temporal dependence in the weak fault signal. The number of hidden layer neurons in the GRU layer is increased to 128. The weight allocation of key features of weak faults is strengthened by the dynamic weight allocation algorithm in the attention mechanism. S53: After being processed by 128 neurons in the fully connected layer and the sigmoid activation function, the softmax classification layer outputs the identification results of four weak fault categories.

8. The method for diagnosing minor faults in ball screw pairs based on a wireless rotational vibration sensor according to claim 7, characterized in that: In S6, the initial learning rate is set to 1e-3.

9. The method for diagnosing weak faults in ball screw pairs based on wireless rotational vibration sensors according to claim 7, characterized in that: The dropout rate in S52 is set to 0.5.

Images