Method for failure prediction and health management of a spinning frame spindle bearing

Abstract

The application discloses a fault prediction and health management method for a spinning frame spindle bearing and belongs to the technical field of intelligent operation and maintenance and fault diagnosis of textile equipment. Through a two-stage strategy of source domain migration pre-training + CGAN data enhancement, a reliable fault prediction model can still be constructed under the condition that the number of target spinning frame fault samples is less than 10 or even zero, so that a textile enterprise can be deployed without a long fault sample accumulation period. Compared with a fixed threshold method, a rotating speed partition dynamic threshold + temperature correction mechanism reduces the false alarm rate and eliminates a large number of invalid alarms during the working condition transition period such as the speed-up section and the doffing section, thereby improving the trust degree of maintenance personnel to the system.

Description

Technical Field

[0001] This invention relates to the field of intelligent operation and maintenance and fault diagnosis technology for textile equipment, specifically to a method for fault prediction and health management of spindle bearings in spinning frames. Background Technology

[0002] The spinning frame is the core production equipment of textile enterprises, used to further draw and twist roving into fine yarn of a specified fineness. A single spinning frame is typically equipped with 200 to 1500 spindles, and large spinning workshops can have dozens of spinning frames operating simultaneously, with a total of over 10,000 spindles in the entire workshop. The spindle bearings (usually deep groove ball bearings or angular contact ball bearings) operate at ultra-high speeds of 12,000 to 20,000 rpm for extended periods, enduring tens of thousands of cyclic impact loads per minute.

[0003] According to industry statistics, the average annual failure rate of spindle bearings in spinning machines in textile enterprises reaches 1.5% to 3%. On a 1008-spindle spinning machine, an estimated 15 to 30 bearing failures occur annually. The consequences of bearing failure include: increased yarn breakage rate (30% to 120%), decreased yarn evenness, and increased equipment vibration due to spindle eccentricity, which can lead to bearing seizure and even fiber combustion. The current main maintenance methods used by enterprises and their drawbacks are shown in the table below:

[0004] Table 1. Current Main Maintenance Methods and Their Drawbacks

[0005]

[0006] Existing data-driven fault diagnosis methods face three major bottlenecks in textile spinning machine scenarios:

[0007] (1) Small sample dilemma: Spindles of spinning machines rarely fail during normal operation, and historical data on early minor faults (such as initial pitting) are almost zero. Traditional CNN / LSTM models require a large number of labeled fault samples (more than 500 per class), which is not feasible in the spindle bearing scenario. At the same time, due to assembly errors, lubrication differences, and yarn tension differences, vibration characteristics vary at the spindle level between different spindle positions, resulting in a weak transfer foundation.

[0008] (2) Difficulty in adapting to different operating conditions: Within a production cycle (approximately 70-90 minutes), a spinning machine needs to go through multiple operating condition stages, including starting up and accelerating (0→18000rpm), spinning at a constant speed (18000rpm), decelerating when full yarn (18000→16000rpm), stopping the machine after doffing, and restarting. The vibration energy and frequency characteristics differ significantly in each stage. A fixed threshold will inevitably generate a large number of false alarms in the acceleration stage, while the sensitivity is insufficient in the low-speed stage, making it impossible to balance both.

[0009] (3) Lack of interpretability: Existing deep learning diagnostic systems typically only output binary conclusions of normal / abnormal, failing to inform maintenance personnel which bearing spindle should be checked, whether the problem lies with the inner or outer ring or rolling elements. This necessitates that maintenance personnel still rely on experience to make judgments, significantly diminishing the practical value of automated diagnostics.

[0010] Therefore, a method and system for fault prediction and health management of spindle bearings in spinning machines are proposed. Summary of the Invention

[0011] The technical problem to be solved by this invention is: how to achieve early identification of various faults (inner ring, outer ring, rolling elements, cage) of newly installed spindle bearings through cross-machine transfer learning and generative data augmentation; how to construct a set of health benchmark thresholds that are dynamically adjusted according to the real-time speed of the spinning frame and the bearing temperature, so as to control the false alarm rate to below 5% and the missed alarm rate to below 3% within the complete production cycle of the spinning frame; and to provide a method for fault prediction and health management of spinning frame spindle bearings.

[0012] The present invention solves the above-mentioned technical problems through the following technical solution, and the present invention includes the following steps:

[0013] S1: Signal Preprocessing

[0014] The original vibration signals collected from each spindle position of the fine sand machine are sequentially processed by detrending term, anti-aliasing filtering, resampling and segmentation to obtain vibration segments, and the average spindle speed, average bearing seat temperature and segment start timestamp metadata corresponding to each vibration segment are recorded simultaneously.

[0015] S2: Transfer Learning Modeling

[0016] A feature extraction model based on 1D-CNN-LSTM-Attention was constructed. The source domain of the model was pre-trained using the source domain bearing fault dataset, and then the target domain was fine-tuned based on the healthy vibration samples of the target spinning machine.

[0017] S3: Dynamic Benchmark Building

[0018] For each spindle position, a Gaussian mixture model at the spindle position level is established, and a graded alarm threshold is generated by combining the real-time average spindle speed and the average temperature of the bearing housing.

[0019] S4: Anomaly Detection Alarm

[0020] The vibration segment to be detected is input into the feature extraction model to obtain the feature vector. The health score is calculated based on the Gaussian mixture model. The health score is compared with the corresponding graded alarm threshold to determine the abnormal state and trigger the corresponding level alarm.

[0021] S5: Fault Location Explanation

[0022] For vibration segments identified as abnormal, interpretable faulty components are located, and structured diagnostic information is output.

[0023] Furthermore, in step S1, the specific processing procedure is as follows:

[0024] S11: Detrending Items

[0025] The trend term of the original vibration signal is removed by using a least-squares fitting polynomial to obtain the residual signal;

[0026] S12: Anti-aliasing filter

[0027] Use an elliptic filter to perform anti-aliasing filtering on the residual signal;

[0028] S13: Resampling

[0029] A multiphase filtering resampling method is used to sample the filtered signal to a fixed sampling rate, with a downsampling ratio of 1 / 5.

[0030] S14: Fragmentation and Metadata Extraction

[0031] Vibration segments are obtained by segmenting the data into segments with a 10-second window and a 50% overlap rate. The average spindle speed, average bearing housing temperature, and segment start timestamp metadata corresponding to each vibration segment are recorded simultaneously.

[0032] Furthermore, in step S2, the feature extraction model based on 1D-CNN-LSTM-Attention includes an input layer, a first convolutional layer, a second convolutional layer, a third convolutional layer, an SENet attention layer, a two-layer LSTM layer, a multi-head temporal attention layer, a fully connected layer, and a replaceable output layer connected in sequence. The replaceable output layer includes a 4-class Softmax classification head enabled during the source domain pre-training stage and a Deep SVDD hypersphere mapping layer enabled during the target domain fine-tuning stage. The two types of output layers are not enabled simultaneously.

[0033] Furthermore, in step S2, the specific processing procedure is as follows:

[0034] S21: Source Domain Pre-training Phase

[0035] The output layer of the feature extraction model uses a 4-class Softmax classification head and is trained on the source domain bearing dataset with the cross-entropy loss function to achieve four-class recognition of normal state, inner race fault, outer race fault, and rolling element fault.

[0036] S22: Target Domain Fine-tuning Stage

[0037] The parameters of the first two convolutional layers of the feature extraction model are frozen, and the output layer is replaced with a Deep SVDD hypersphere mapping layer. Using healthy vibration segments of the target spinning machine as fine-tuning data, the Deep SVDD loss function is employed for fine-tuning. The Deep SVDD loss function is as follows:

[0038] ;

[0039] in, The characteristic center of a healthy vibration segment is defined as N, where N is the number of healthy vibration segments in the batch. For the i-th input sample, It is a feature mapping function;

[0040] S23: Fault Sample Augmentation and Classification Model Adaptation

[0041] When the number of fault samples of the target spinning machine is less than a preset threshold, a conditional generative adversarial network (cGAN) is used to synthesize fault samples to supplement the training set and train a fault type classifier adapted to the target domain for subsequent fault component localization.

[0042] Furthermore, in step S23, when the number of fault samples of the target spinning machine is less than a preset threshold, a conditional generative adversarial network (cGAN) is used to synthesize fault samples to supplement the training set. The conditional generative adversarial network (cGAN) adopts the WGAN-GP loss function, the gradient penalty coefficient λ=10, the generator inputs 128-dimensional noise and 4-dimensional conditional labels, and outputs a synthesized vibration signal of the same length as the real vibration segment, which is used to train a fault type classifier adapted to the target domain.

[0043] Furthermore, in step S3, the specific process of establishing the ingot-level Gaussian mixture model is as follows:

[0044] S31: Input multiple healthy vibration segments from a single spindle into the fine-tuned feature extraction model to obtain a set of 128-dimensional feature vectors. ,in, For the number of healthy vibration segments and ≥200;

[0045] S32: Estimating Gaussian mixture model parameters using the EM algorithm:

[0046] ;

[0047] in, Let be the mixing coefficient of the k-th component. The mean of the k-th component. Let K be the diagonal covariance moment, and K be the total number of Gaussian components.

[0048] S33: Select the optimal number of components within the range K=2,...,6 using the Bayesian Information Criterion (BIC). Minimum, as follows:

[0049] ;

[0050] in, Let K be the total number of parameters of the K Gaussian components. The maximum likelihood value;

[0051] S34: Define the health score of the vibration segment as follows:

[0052] ;

[0053] in, The feature vector of the vibration segment to be detected. It is the minimum Mahalanobis distance from the eigenvector to all Gaussian components.

[0054] Furthermore, in step S3, the process of generating the graded alarm threshold is as follows:

[0055] S35: Divide the spindle rotation speed into multiple equally spaced rotation speed intervals, where the i-th interval is represented as... ,in, For the first The lower limit boundary of each speed range. For the first The upper limit boundary of a speed range; within the speed range The first percentile of the health score for each health segment is used as the warning threshold. The 0.5 percentile is used as the severe alarm threshold. ;

[0056] S36: The real-time average spindle speed is within the speed range. When the threshold is near the boundary range, linear interpolation is used to calculate the continuous threshold to avoid threshold jumps at the interval boundaries.

[0057] ;

[0058] in, This is the dynamic threshold based on the current real-time average spindle speed. This represents the current real-time average spindle speed. Warning threshold Or critical alarm threshold ;

[0059] S37: When the average temperature T of the bearing housing exceeds 60 degrees Celsius, both the warning threshold and the severe alarm threshold are simultaneously corrected for temperature. The corrected threshold is:

[0060] ;

[0061] in, , is the correction factor.

[0062] Furthermore, in step S4, the abnormal state and alarm triggering logic are specifically as follows:

[0063] When the health score of a single vibration segment is lower than the warning threshold under the current average spindle speed condition, it is marked as a suspected abnormality;

[0064] When the average health score of three consecutive vibration segments is lower than the warning threshold, a warning-level alarm is triggered.

[0065] When the average health score of three consecutive vibration segments is lower than the critical alarm threshold, a critical alarm is triggered and the historical data of the corresponding spindle position is locked.

[0066] Furthermore, in step S5, the specific process for locating the interpretable faulty component is as follows:

[0067] S51: Perform Hilbert transform on the vibration segment determined to be in an abnormal state to extract the envelope spectrum, divide the envelope spectrum into multiple frequency bands and calculate the energy of each frequency band;

[0068] S52: The KernelSHAP method is used to calculate the marginal contribution of each frequency band to the health score deviation, and the top 3 frequency bands with the highest contribution are selected.

[0069] S53: Calculate the outer ring passing frequency in real time based on the frequency corresponding to the current spindle speed and the bearing structure parameters. Inner ring frequency With the rotational frequency of the rolling element :

[0070] ;

[0071] ;

[0072] ;

[0073] Where Z is the number of rolling elements. The frequency corresponding to the rotational speed. The diameter of the rolling element, The diameter of the pitch circle. Contact angle;

[0074] S54: Compare the center frequencies of the top 3 frequency bands with the highest contribution with BPFO, BPFI, BSF and their 2nd to 5th harmonics one by one. If the deviation is within ±5%, it is determined to be the corresponding fault type, and the confidence level is output.

[0075] Furthermore, in step S5, the structured diagnostic information includes spindle position, health score, dynamic threshold, alarm level, fault prediction probability, maximum contributing frequency band, and suggested maintenance actions.

[0076] This invention also provides a fault prediction and health management system for spindle bearings of spinning frames, applied to the above-mentioned method, including:

[0077] The signal preprocessing module is used to process the raw vibration signals collected from each spindle position of the fine sand machine by performing detrending term, anti-aliasing filtering, resampling and segmentation processing in sequence to obtain vibration segments, and simultaneously record the average spindle speed, average bearing seat temperature and segment start timestamp metadata corresponding to each vibration segment.

[0078] The transfer learning modeling module is used to build a feature extraction model based on 1D-CNN-LSTM-Attention. The model is pre-trained in the source domain using the source domain bearing fault dataset, and then fine-tuned in the target domain based on the healthy vibration samples of the target spinning machine.

[0079] The dynamic benchmark construction module is used to build a Gaussian mixture model at the spindle level for each spindle position, and dynamically generate graded alarm thresholds by combining the real-time average spindle speed and the average temperature of the bearing housing.

[0080] The anomaly detection and alarm module is used to input the vibration segment to be detected into the feature extraction model to obtain the feature vector, calculate the health score based on the Gaussian mixture model, compare the health score with the corresponding graded alarm threshold, determine the abnormal state and trigger the corresponding level of alarm.

[0081] The fault location and interpretation module is used to locate interpretable faulty components in vibration segments determined to be in an abnormal state and output structured diagnostic information.

[0082] The present invention has the following advantages over the prior art:

[0083] (1) Solving the small sample dilemma: Through a two-stage strategy of source domain transfer pre-training + cGAN data augmentation, a reliable fault prediction model can still be built even when the number of fault samples of the target spinning machine is less than 10 or even zero, so that textile enterprises can deploy it without a long period of fault sample accumulation.

[0084] (2) Significantly reduced false alarm rate: Compared with the fixed threshold method, the speed zone dynamic threshold + temperature correction mechanism reduces the false alarm rate from 23.4% to 6.5%, eliminating a large number of invalid alarms during the working condition transition period such as the speed-up section and the doffing section, and improving the maintenance personnel's trust in the system.

[0085] (3) Early fault warning: The system can issue an early warning 8 to 12 hours in advance when early pitting occurs in the outer ring / inner ring / rolling element of the bearing (vibration energy changes by only 2 to 5%), which is 8 times earlier than the traditional vibration energy threshold method, providing sufficient time for maintenance window arrangement.

[0086] (4) Provides interpretable component location: SHAP frequency band contribution analysis accurately locates faulty bearing components (inner ring / outer ring / rolling element) with an accuracy of 84.6%, filling the gap in interpretability of existing black-box deep learning methods, enabling maintenance personnel to perform accurate disassembly and inspection directly based on system recommendations, reducing unnecessary disassembly procedures by about 60%.

[0087] (5) General scalability: Although this invention is designed for spinning machines, the transfer learning framework, dynamic GMM threshold method and SHAP frequency band interpretation strategy are all general and can be extended to similar textile equipment such as spinning machines, twisting machines and looms, and even extended to bearing health management scenarios of other high-speed rotating equipment (paper machine rollers, compressors, etc.). Attached Figure Description

[0088] Figure 1 This is a deployment architecture diagram of the fault prediction and health management system for spindle bearings of a spinning frame in an embodiment of the present invention;

[0089] Figure 2 This is a schematic diagram of the signal preprocessing process in an embodiment of the present invention;

[0090] Figure 3 This is a schematic diagram of the 1D-CNN-LSTM-Attention model structure in an embodiment of the present invention;

[0091] Figure 4 This is a flowchart illustrating the method for fault prediction and health management of spindle bearings in a spinning machine according to an embodiment of the present invention. Detailed Implementation

[0092] The embodiments of the present invention are described in detail below. These embodiments are implemented based on the technical solution of the present invention, and provide detailed implementation methods and specific operation processes. However, the scope of protection of the present invention is not limited to the following embodiments.

[0093] This embodiment provides a technical solution: a fault prediction and health management system for spindle bearings of spinning machines, the details of which are as follows:

[0094] 1. System Overall Architecture

[0095] The system consists of six functional modules, with data flowing as follows: spindle data acquisition terminal → edge computing gateway → cloud / server inference → electronic dashboard and mobile terminal, forming a closed loop of perception-analysis-decision-feedback. The overall system architecture is as follows: Figure 1 As shown.

[0096] 2. Signal preprocessing module

[0097] Signal preprocessing process as follows Figure 2 As shown, it includes four serial processing steps, all of which are executed in real time at the edge gateway with a latency of no more than 200ms.

[0098] 2.1 Step 1: Remove trend items (remove drift)

[0099] A second-order polynomial fitting method is employed. For the original signal of length N... , =0,...,N−1, least squares fitting polynomial ,in , , The coefficients of the polynomial to be fitted are the residual signal after drift removal. Defined as the difference between the original signal and the fitted trend term: The coefficients are solved by minimizing the sum of squared residuals, and the objective function is: In the edge gateway of this invention, instead of directly solving the above matrix, a lightweight numerical library, such as numpy.polyfit in embedded Python, is called, and the matrix is ​​passed in... and Specifying an order of 2 directly returns the coefficient vector. The calculation takes approximately 2-5 ms per iteration on a 10-second segment (25600 points), meeting the real-time processing requirements at the edge. The second-order polynomial is sufficient to remove the low-frequency DC offset caused by temperature drift and gravity components, while not losing the effective signal in the mid-frequency band (100-5000Hz) where bearing fault characteristics are located.

[0100] Parameter selection criteria: Compared to the first-order (only removing DC), the second-order can simultaneously handle linear trends and low-frequency ramps introduced by the accelerometer mounting angle; compared to higher-order (third-order and above), it avoids overfitting and erroneously removing effective low-frequency signals. Experimental comparisons show that the signal-to-noise ratio is improved by 3.2 dB after second-order processing (compared to the first-order).

[0101] 2.2 Step 2: Anti-aliasing elliptic filtering

[0102] An elliptic filter is used to filter the signal processed in step 1. Parameters: order 8, maximum passband ripple 0.1dB, minimum stopband attenuation 60dB. The cutoff frequency is set to 0.8 times the resampling Nyquist frequency. For example, with an original sampling rate of 12.8kHz and a target resampling rate of 2.56kHz: Nyquist frequency = 1280Hz, cutoff frequency = 1024Hz. The elliptic filter has the narrowest transition band for its order, making it suitable for implementation on resource-constrained edge hardware (ARM Cortex-M7) using fixed-point arithmetic.

[0103] 2.3 Step 3: Multiphase resampling

[0104] Polyphase resampling is used to resample the signal processed in step 2, with a downsampling ratio L / M = 2560 / 12800 = 1 / 5. Polyphase decomposition breaks down the long-term convolution into multiple short sub-filters that operate in parallel, avoiding the computational redundancy introduced by direct resampling and reducing CPU usage by approximately 80%. This is equivalent to `scipy.signal.resample_poly(x,1,5)`. The resampled output has a fixed sampling rate of 2560Hz, with 2560 samples per second.

[0105] 2.4 Step 4: Fragment Segmentation and Metadata Extraction

[0106] The vibration segments were divided into segments with a 10-second window (25,600 points) and 50% overlap (12,800-point step size), generating approximately two valid segments per unit time. Each vibration segment synchronously recorded: (a) the average spindle speed (rpm) read from the spinning machine PLC within the corresponding time period; (b) the average bearing housing temperature (degrees Celsius) measured by the temperature sensor; and (c) the segment start timestamp (UTC second-level accuracy). These three metadata fields are necessary inputs for subsequent dynamic threshold lookup and temperature correction.

[0107] 3. Small Sample Transfer Learning Module

[0108] 3.1 1D-CNN-LSTM-Attention Model Structure

[0109] The model has 11 layers, and the complete network structure is as follows: Figure 3 As shown in Table 2, the output dimensions of each layer are calculated as follows:

[0110] Table 2 Output Dimensions for Each Layer

[0111]

[0112] During the source domain pre-training phase, the output layer uses a 4-class Softmax classification head (normal / inner ring fault / outer ring fault / rolling body fault), and the loss function is cross-entropy loss. In the target domain adaptation phase, the Softmax head is replaced with a Deep SVDD hypersphere mapping layer, outputting a 128-dimensional feature vector, and the loss function is Deep SVDD loss. ,in, The health sample feature centers are defined (initialized from the forward propagation mean of the health samples), and N is the number of health samples in the batch. For the i-th input sample, This is the feature mapping function. The two stages share the parameters of the first 9 layers (Conv1 to the temporal attention layer). The source domain head and the target domain head are two mutually exclusive output layers and do not exist in the network simultaneously.

[0113] 3.2 Source Domain Pre-training Configuration

[0114] Source datasets: Case Western Reserve University (CWRU) bearing dataset (4 classes, 4000 segments per class) + self-collected historical fault data of the same model of spinning machine from a textile factory (3 classes, 200-800 segments per class). After merging the datasets, the training / validation / test sets were divided in a 6:2:2 ratio.

[0115] Training hyperparameters: Optimizer Adam (beta1=0.9, beta2=0.999), initial learning rate 0.001, cosine annealing to a minimum of 1e. -6 Batch size 64, maximum number of rounds 100, early stop patience = 10, L2 weight decay 1e -4 .

[0116] Source domain test set performance: Overall classification accuracy is 97.2%, confusion matrix shows that the precision of each category is >95%, and the recall rate for outer ring faults (the most common fault type) is 98.7%.

[0117] 3.3 Target Domain Adaptive Fine-tuning

[0118] Frozen layers: Conv1, Conv2 (all parameters of the first two convolutional layers are frozen and do not participate in gradient calculation, preserving the ability to extract low-level features from the source domain).

[0119] Trainable layers: Conv3, SENet attention, LSTM layer, temporal attention layer, fully connected layer (approximately 620K parameters in total), with the output header replaced by a Deep SVDD hypersphere mapping layer.

[0120] Fine-tuning data: Healthy samples from all spindle positions of the target spinning machine, at least 200 segments / spindle position, to ensure the quality of GMM (Gaussian Mixture Model) modeling. Learning rate 1×10⁻⁶. -4(This is 1 / 100th of the pre-trained value to prevent significant deviation from the source domain feature space). Fine-tune the training for 20 epochs, stopping if the loss on the validation set does not decrease in each epoch.

[0121] Fault Sample Augmentation and Classification Model Adaptation: When the number of fault samples is insufficient to meet the preset threshold N (N=10 recommended), the cGAN generator is activated to supplement synthetic samples. cGAN architecture: Generator G (5 layers of transposed convolutions, input noise 128-dimensional + conditional label 4-dimensional, output 25600-dimensional vibration signal), discriminator D (5 layers of 1D convolutions + Spectral Normalization), loss function WGAN-GP (gradient penalty coefficient λ=10). After achieving the required quality (FID<50), 1000 synthetic samples / fault classes are generated to train the fault type classifier, with the learning rate further reduced to 1×10⁻⁶. -6 .

[0122] 4. GMM Health Benchmark and Dynamic Threshold Module

[0123] 4.1 Establishment of GMM at the spindle level

[0124] Core idea: Due to mechanical assembly errors (coaxiality deviation ±0.02mm), differences in lubricant filling amount (±15%), and differences in yarn tension (±0.3N), the fundamental frequency and harmonic distribution of vibration vary among different spindle positions. Sharing the global distribution would cause normal individual differences to be misjudged as faults. Therefore, a Gaussian mixture model is established separately for each spindle position.

[0125] Modeling process: (1) Using the fine-tuned feature extractor (Conv1 to the temporal attention layer, up to the layer before the fully connected layer), extract the N of the spindle. h ≥200 healthy vibration segments are mapped to a set of 128-dimensional feature vectors. (2) Estimating GMM parameters using the EM algorithm: ,in The mixing coefficient, The mean of the k-th component. The diagonal covariance matrix is ​​used to reduce the number of parameters and avoid overfitting. (3) The number of components (total number of Gaussian components) K is automatically selected by the Bayesian Information Criterion (BIC) within the range of K=2,...,6, minimizing the number of components. ,in, The total number of parameters for the K components. This represents the maximum likelihood value. The optimal K=3 for the spindle position of a typical spinning machine is determined by actual measurements.

[0126] Health score definition: For newly acquired vibration segments Feature extraction Calculate the minimum Mahalanobis distance from this feature to all Gaussian components: Health score The larger the S value (the smaller the absolute value of the negative value), the closer x is to the healthy distribution; the smaller the S value (the larger the absolute value), the further away from the healthy state.

[0127] 4.2 Generation of Dynamic Thresholds for Rotation Speed ​​Zones

[0128] Partitioning scheme: The spindle speed of 0~20000rpm is divided into 10 equally spaced intervals (2000rpm per interval). The low speed range (0~2000rpm) is merged with 2000~4000rpm due to the scarcity of data, so 9 effective intervals are actually used.

[0129] Threshold calculation method: During the healthy operation accumulation period (it is recommended to accumulate at least 30 days of data), the set of health scores of healthy segments collected from each speed range. Calculate the warning threshold Severe alarm threshold ,in, , For the i-th interval The corresponding threshold, in the i-th interval middle, For the first The lower limit boundary of each speed range. For the first The upper limit boundary of each speed range. The α percentile, which is the score of the lowest α% of samples, serves as the lower bound. Below this value, it means that the segment has deviated from the lowest 1% (0.5%) of the healthy distribution under that speed condition, triggering the corresponding level of alarm.

[0130] Interpolation processing: When the average spindle rotation speed R is in the i-th interval of and Near the boundary (±200 rpm), use linear interpolation: This ensures the continuity of the threshold change with rotational speed and avoids threshold jumps at interval boundaries.

[0131] Temperature correction: When the average bearing housing temperature T > 60 degrees Celsius, the threshold is dynamically lowered: correction factor. Corrected threshold (The increase in vibration energy caused by temperature rise has the same effect on both thresholds in the same direction and with the same amplitude, so both thresholds need to be corrected.) The basis for this correction is that: rising temperature leads to a decrease in lubricating oil viscosity and an increase in frictional vibration energy. The vibration signal under healthy conditions will deviate from the normal temperature healthy baseline. Therefore, lowering the threshold can avoid false alarms caused by this factor.

[0132] Alarm triggering logic: (1) Health score of a single vibration segment (2) The average health score of three consecutive vibration segments was recorded as [suspected abnormality]; Triggering a [warning] push to the maintenance worker App, retaining the maintenance personnel's decision-making authority; (3) the average health score of three consecutive vibration segments. If a critical alarm is triggered, the corresponding spindle position will be marked with a flashing red indicator on the workshop's large screen, and the historical data window for that spindle position will be locked for subsequent analysis.

[0133] 5. Explainable Fault Location Module

[0134] 5.1 Visualization of Time Attention

[0135] The multi-head temporal attention layer following the LSTM (4 heads, 128-dimensional key / query vector) outputs the time step weight distribution. t=1,...,12 (corresponding to the 12 time steps of the Conv3 output, each step covers approximately 2500 points of the original signal, i.e., about 1 second). The time step with the highest weight corresponds to the period when the fault impact is most concentrated, which can be displayed on the original signal waveform as a heatmap at the front end.

[0136] 5.2 SHAP Envelope Spectrum Band Contribution Analysis

[0137] Envelope spectrum extraction: A Hilbert transform is performed on the vibration segments identified as abnormal in the alarm triggering logic to extract the envelope of the analytical signal. An FFT is then performed on the envelope to obtain the envelope spectrum, with a frequency range of 0-500Hz and a frequency resolution of 0.1Hz (1 / 10 second segment). The envelope spectrum is divided into 20 frequency bands (25Hz each), and the energy of each band is calculated as interpretive features. .

[0138] KernelSHAP calculation: Using KernelExplainer from the SHAP library, the background dataset consists of the envelope spectral features of the spindle's health samples (100 samples), and the samples to be explained consist of the envelope spectral features of the current alarm segment. The marginal contribution of each frequency band to the health score deviation is calculated. ,satisfy ,in The expected score for the health benchmark.

[0139] Fault frequency matching rule: Based on spindle model (bearing model 6000 series, contact angle) The value is 0, the number of rolling elements Z=9, the inner diameter d=10mm, the outer diameter D=26mm, and the pitch circle diameter. =18mm, rolling element diameter =5mm) and current frequency (Hz), Dynamic calculation: Frequency of the outer ring Inner ring frequency Rolling element rotation frequency The center frequencies of the top three frequency bands with the largest SHAP values ​​(calculated using the KernelSHAP method) are compared one by one with BPFO, BPFI, BSF and their 2nd to 5th harmonics. If the deviation is within ±5%, the corresponding fault type is determined, and the confidence level (the proportion of the absolute value of the SHAP value to the total value) is output.

[0140] Example of diagnostic output JSON: The system ultimately outputs the spindle position, health score, dynamic thresholds (specifically the currently effective "critical alarm threshold" and "early warning threshold" after speed interpolation and temperature correction) in structured JSON, alarm level, fault prediction probability (e.g., outer ring fault probability 0.87, inner ring fault probability 0.09), maximum contributing frequency band (e.g., 125~150Hz, SHAP value 0.42, corresponding to an outer ring passing frequency of 135Hz), and suggested maintenance actions (e.g., checking the outer ring of spindle bearing No. 427), for display on the maintenance worker's App and large screen dashboard.

[0141] like Figure 4 As shown, this embodiment also provides a method for fault prediction and health management of spindle bearings in spinning machines. Based on the above system implementation, it includes the following steps:

[0142] Step 1: Signal Preprocessing

[0143] The original vibration signals collected from each spindle position of the fine sand machine are sequentially processed by detrending term, anti-aliasing filtering, resampling and segmentation to obtain vibration segments, and the average spindle speed, average bearing seat temperature and segment start timestamp metadata corresponding to each vibration segment are recorded simultaneously.

[0144] Step 2: Transfer Learning Modeling

[0145] A feature extraction model based on 1D-CNN-LSTM-Attention was constructed. The model was pre-trained using the source domain bearing fault dataset, and then adaptively fine-tuned based on the healthy vibration samples of the target spinning machine.

[0146] Step 3: Dynamic Benchmark Construction

[0147] For each spindle position, a Gaussian mixture model at the spindle position level is established, and a graded alarm threshold is dynamically generated by combining the real-time average spindle speed and the average temperature of the bearing housing.

[0148] Step 4: Anomaly Detection Alarm

[0149] The vibration segment to be detected is input into the feature extraction model to obtain the feature vector. The health score is calculated based on the Gaussian mixture model. The health score is compared with the corresponding graded alarm threshold to determine the abnormal state and trigger the corresponding level alarm.

[0150] Step 5: Explanation of Fault Location

[0151] For vibration segments identified as abnormal, interpretable faulty components are located, and structured diagnostic information is output.

[0152] It should be noted that the specific processing procedures in each step are detailed in the specific processing procedures of each module in the above system, and will not be repeated here.

[0153] In summary, the fault prediction and health management method and system for spindle bearings of spinning frames described in the above embodiments, under the condition that the number of fault samples of the target spinning frame is less than 10 (or even zero), achieves early identification of various faults of newly installed spindle bearings through cross-machine transfer learning and generative data augmentation, so that the model recognition rate is not less than 85%; a set of health benchmark thresholds that are dynamically adjusted with the real-time speed of the spinning frame and the bearing temperature are constructed, and the false alarm rate is controlled below 5% and the false alarm rate is controlled below 3% within the complete production cycle of the spinning frame; without sacrificing the model accuracy, interpretable fault location outputs at the spindle position level and component level are provided, indicating which bearing component has a problem with an accuracy of not less than 80%, and giving a quantitative confidence level.

[0154] Although embodiments of the present invention have been shown and described above, it is understood that the above embodiments are exemplary and should not be construed as limiting the present invention. Those skilled in the art can make changes, modifications, substitutions and variations to the above embodiments within the scope of the present invention.

Claims

1. A method for fault prediction and health management of spindle bearings in spinning frames, characterized in that, Includes the following steps: S1: Signal Preprocessing The original vibration signals collected from each spindle position of the fine sand machine are sequentially processed by detrending term, anti-aliasing filtering, resampling and segmentation to obtain vibration segments, and the average spindle speed, average bearing seat temperature and segment start timestamp metadata corresponding to each vibration segment are recorded simultaneously. S2: Transfer Learning Modeling A feature extraction model based on 1D-CNN-LSTM-Attention was constructed. The source domain of the model was pre-trained using the source domain bearing fault dataset, and then the target domain was fine-tuned based on the healthy vibration samples of the target spinning machine. S3: Dynamic Benchmark Building For each spindle position, a Gaussian mixture model at the spindle position level is established, and a graded alarm threshold is generated by combining the real-time average spindle speed and the average temperature of the bearing housing. S4: Anomaly Detection Alarm The vibration segment to be detected is input into the feature extraction model to obtain the feature vector. The health score is calculated based on the Gaussian mixture model. The health score is compared with the corresponding graded alarm threshold to determine the abnormal state and trigger the corresponding level alarm. S5: Fault Location Explanation For vibration segments identified as abnormal, interpretable faulty components are located, and structured diagnostic information is output.

2. The method for fault prediction and health management of spindle bearings in spinning frames according to claim 1, characterized in that, In step S1, the specific processing procedure is as follows: S11: Detrending Items The trend term of the original vibration signal is removed by using a least-squares fitting polynomial to obtain the residual signal; S12: Anti-aliasing filter Use an elliptic filter to perform anti-aliasing filtering on the residual signal; S13: Resampling A multiphase filtering resampling method is used to sample the filtered signal to a fixed sampling rate, with a downsampling ratio of 1 / 5. S14: Fragmentation and Metadata Extraction Vibration segments are obtained by segmenting the data into segments with a 10-second window and a 50% overlap rate. The average spindle speed, average bearing housing temperature, and segment start timestamp metadata corresponding to each vibration segment are recorded simultaneously.

3. The method for fault prediction and health management of spindle bearings in spinning frames according to claim 1, characterized in that, In step S2, the feature extraction model based on 1D-CNN-LSTM-Attention includes an input layer, a first convolutional layer, a second convolutional layer, a third convolutional layer, an SENet attention layer, a two-layer LSTM layer, a multi-head temporal attention layer, a fully connected layer, and a replaceable output layer, which are connected in sequence. The replaceable output layer includes a 4-class Softmax classification head enabled during the source domain pre-training stage and a Deep SVDD hypersphere mapping layer enabled during the target domain fine-tuning stage. The two types of output layers are not enabled at the same time.

4. The method for fault prediction and health management of spindle bearings in a spinning frame according to claim 3, characterized in that, In step S2, the specific processing procedure is as follows: S21: Source Domain Pre-training Phase The output layer of the feature extraction model uses a 4-class Softmax classification head and is trained on the source domain bearing dataset with the cross-entropy loss function to achieve four-class recognition of normal state, inner race fault, outer race fault, and rolling element fault. S22: Target Domain Fine-tuning Stage The parameters of the first two convolutional layers of the feature extraction model are frozen, and the output layer is replaced with a Deep SVDD hypersphere mapping layer. Using healthy vibration segments of the target spinning machine as fine-tuning data, the Deep SVDD loss function is employed for fine-tuning. The Deep SVDD loss function is as follows: ； in, The characteristic center of a healthy vibration segment is defined as N, where N is the number of healthy vibration segments in the batch. For the i-th input sample, It is a feature mapping function; S23: Fault Sample Augmentation and Classification Model Adaptation When the number of fault samples of the target spinning machine is less than a preset threshold, a conditional generative adversarial network (cGAN) is used to synthesize fault samples to supplement the training set and train a fault type classifier adapted to the target domain for subsequent fault component localization.

5. The method for fault prediction and health management of spindle bearings in a spinning frame according to claim 4, characterized in that, In step S23, when the number of fault samples of the target spinning machine is less than a preset threshold, a conditional generative adversarial network (cGAN) is used to synthesize fault samples to supplement the training set. The conditional generative adversarial network (cGAN) adopts the WGAN-GP loss function, the gradient penalty coefficient λ=10, the generator inputs 128-dimensional noise and 4-dimensional conditional labels, and outputs a synthesized vibration signal of the same length as the real vibration segment, which is used to train a fault type classifier adapted to the target domain.

6. The method for fault prediction and health management of spindle bearings in a spinning frame according to claim 5, characterized in that, In step S3, the specific process of establishing the ingot-level Gaussian mixture model is as follows: S31: Input multiple healthy vibration segments from a single spindle into the fine-tuned feature extraction model to obtain a set of 128-dimensional feature vectors. ,in, For the number of healthy vibration segments and ≥200; S32: Estimating Gaussian mixture model parameters using the EM algorithm: ； in, Let be the mixing coefficient of the k-th component. The mean of the k-th component. Let K be the diagonal covariance moment, and K be the total number of Gaussian components. S33: Select the optimal number of components within the range K=2,...,6 using the Bayesian Information Criterion (BIC). Minimum, as follows: ； in, Let K be the total number of parameters of the K Gaussian components. The maximum likelihood value; S34: Define the health score of the vibration segment as follows: ； in, The feature vector of the vibration segment to be detected. It is the minimum Mahalanobis distance from the eigenvector to all Gaussian components.

7. The method for fault prediction and health management of spindle bearings in a spinning frame according to claim 6, characterized in that, In step S3, the process of generating the graded alarm threshold is as follows: S35: Divide the spindle rotation speed into multiple equally spaced rotation speed intervals, where the i-th interval is denoted as... ,in, For the first The lower limit boundary of each speed range. For the first The upper limit boundary of a speed range; within the speed range The first percentile of the health score for each health segment is used as the warning threshold. The 0.5 percentile is used as the severe alarm threshold. ; S36: The real-time average spindle speed is within the speed range. When the threshold is near the boundary range, linear interpolation is used to calculate the continuous threshold to avoid threshold jumps at the interval boundaries. ； in, This is the dynamic threshold based on the current real-time average spindle speed. This represents the current real-time average spindle speed. Warning threshold Or critical alarm threshold ; S37: When the average temperature T of the bearing housing exceeds 60 degrees Celsius, both the warning threshold and the severe alarm threshold are simultaneously corrected for temperature. The corrected threshold is: ； in, , is the correction factor.

8. The method for fault prediction and health management of spindle bearings in a spinning frame according to claim 7, characterized in that, In step S4, the abnormal state and alarm triggering logic are as follows: When the health score of a single vibration segment is lower than the warning threshold under the current average spindle speed condition, it is marked as a suspected abnormality; When the average health score of three consecutive vibration segments falls below the warning threshold, a warning-level alarm is triggered. When the average health score of three consecutive vibration segments is lower than the critical alarm threshold, a critical alarm is triggered and the historical data of the corresponding spindle position is locked.

9. The method for fault prediction and health management of spindle bearings in a spinning frame according to claim 8, characterized in that, In step S5, the specific process for locating the interpretable faulty component is as follows: S51: Perform Hilbert transform on the vibration segment determined to be in an abnormal state to extract the envelope spectrum, divide the envelope spectrum into multiple frequency bands and calculate the energy of each frequency band; S52: The KernelSHAP method is used to calculate the marginal contribution of each frequency band to the health score deviation, and the top 3 frequency bands with the highest contribution are selected. S53: Calculate the outer ring passing frequency in real time based on the frequency corresponding to the current spindle speed and the bearing structure parameters. Inner ring frequency With the rotational frequency of the rolling element : ； ； ； Where Z represents the number of rolling elements. The frequency corresponding to the rotational speed. The diameter of the rolling element, The diameter of the pitch circle. Contact angle; S54: Compare the center frequencies of the top 3 frequency bands with the highest contribution with BPFO, BPFI, BSF and their 2nd to 5th harmonics one by one. If the deviation is within ±5%, it is determined to be the corresponding fault type, and the confidence level is output.

10. A fault prediction and health management system for spindle bearings of spinning frames, characterized in that, The method applied to any one of claims 1 to 9 includes: The signal preprocessing module is used to process the raw vibration signals collected from each spindle position of the fine sand machine by performing detrending term, anti-aliasing filtering, resampling and segmentation processing in sequence to obtain vibration segments, and simultaneously record the average spindle speed, average bearing seat temperature and segment start timestamp metadata corresponding to each vibration segment. The transfer learning modeling module is used to build a feature extraction model based on 1D-CNN-LSTM-Attention. The model is pre-trained in the source domain using the source domain bearing fault dataset, and then fine-tuned in the target domain based on the healthy vibration samples of the target spinning machine. The dynamic benchmark construction module is used to build a Gaussian mixture model at the spindle level for each spindle position, and dynamically generate graded alarm thresholds by combining the real-time average spindle speed and the average temperature of the bearing housing. The anomaly detection and alarm module is used to input the vibration segment to be detected into the feature extraction model to obtain the feature vector, calculate the health score based on the Gaussian mixture model, compare the health score with the corresponding graded alarm threshold, determine the abnormal state and trigger the corresponding level of alarm. The fault location and interpretation module is used to locate interpretable faulty components in vibration segments determined to be in an abnormal state and output structured diagnostic information.

Images