Spindle bending common characteristic diagnosis system based on stacking integration method

Abstract

The invention belongs to the field of fault diagnosis, and particularly relates to a main shaft bending common characteristic diagnosis system based on a stacking integration method, and the system comprises the following modules: a sensor module which is used for collecting a vibration signal of a main shaft in real time; the data acquisition and processing module is used for converting an analog signal into a digital signal and preprocessing the signal acquired by the sensor; the feature extraction module is used for extracting features from the preprocessed vibration signals; the machine learning diagnosis module analyzes and classifies the extracted features by using stack integration based on a plurality of trained machine learning models, and judges the health state of the main shaft; and the output module is used for outputting the diagnosis result and sending out an alarm signal when abnormality is detected. Real-time monitoring and automatic diagnosis of the main shaft bending fault are achieved, compared with a traditional method depending on expert experience and manual detection, the fault detection efficiency and accuracy can be greatly improved, and misjudgment and missed judgment risks caused by human factors are reduced.

Description

Technical Field

[0001] The invention belongs to the field of fault diagnosis, and in particular relates to a spindle bending common feature diagnosis system based on a stacking integration method. Background Art

[0002] The spindle is a vital component in mechanical equipment, and its health status directly affects the working accuracy and life of the equipment. Spindle bending is one of the common fault types, and traditional fault diagnosis methods mainly rely on expert experience and regular manual inspection. However, with the development of automation and intelligence in the manufacturing industry, how to achieve real-time and automatic diagnosis of spindle bending faults has become an urgent problem to be solved.

[0003] In recent years, with the advancement of machine learning technology, it has become possible to realize intelligent diagnosis of mechanical faults through automated data analysis. The machine learning-based diagnosis method can effectively process a large amount of sensor data, extract the features related to the fault, and thus accurately identify the bending fault of the spindle.

[0004] In the prior art, there is a Chinese invention patent with the publication number CN116701984B and the name "Automatic Fault Diagnosis System for CNC Machine Tools Based on Machine Learning". This patent has the following deficiencies: 1. It mainly relies on independent models for fault diagnosis, and its robustness and accuracy are insufficient. 2. At the same time, this patent mainly focuses on current fault detection and diagnosis, and lacks the function of predicting future fault trends. This system can only provide information on the current health status of the equipment and cannot help users make preventive maintenance decisions in advance. 3. There is no adaptive learning unit and output module.

[0005] Another example is the Chinese invention patent application with publication number CN110674842A, entitled "Patent for a Method for Predicting Faults in the Main Shaft Bearing of a Wind Turbine Unit". This patent has the following shortcomings: 1. The data preprocessing is just simple noise reduction, normalization and other operations, and no in-depth physical mining of the data is performed. 2. Although the ensemble learning algorithm XGBoost is also used in the machine learning module, the weak classifiers integrated by XGBoost are all decision trees, which is essentially an ensemble learning algorithm that continuously optimizes decision trees. 3. There is no adaptive learning unit and output module. Summary of the invention

[0006] In view of the above-mentioned problems existing in the prior art, the present invention proposes a common feature diagnosis system for spindle bending based on the stacking integration method. By real-time monitoring of the vibration signal of the spindle and combining advanced signal processing technology with machine learning algorithms, it can automatically identify the common features of spindle bending and accurately judge the health status of the spindle.

[0007] In order to achieve the above effects, the technical solution adopted by the present invention is as follows:

[0008] A spindle bending common feature diagnosis system based on a stacking integration method includes the following modules:

[0009] Module M 1 .Sensor module: used to collect the vibration signal of the spindle in real time;

[0010] The sensors used in the sensor module include vibration sensors, speed sensors, displacement sensors, and laser sensors. The sensors are installed around the bearing seat of the spindle or at the front end of the spindle to capture the dynamic changes of the spindle during operation.

[0011] Module M 2 Data acquisition and processing module: converts analog signals into digital signals and preprocesses the signals collected by the sensor, including noise removal, filtering, and normalization.

[0012] Module M 3 .Feature extraction module: extract features from the preprocessed vibration signal, including time domain features, frequency domain features and time-frequency domain features, and quantize the extracted features.

[0013] Module M 4 .Machine learning diagnosis module: Based on multiple trained machine learning models, stacking is used to analyze and classify the extracted features to determine the health status of the spindle. The basic machine learning models of stacking include support vector machine (SVM), K nearest neighbor, neural network, decision tree, random forest, linear regression model and naive Bayes algorithm;

[0014] Module M 5 Output module: used to output diagnostic results and send out alarm signals when abnormalities are detected. Output is achieved through a display, alarm or remote communication module, and a log is used to record diagnostic results.

[0015] Furthermore, the sensor module specifically includes the following submodules:

[0016] Module M 11 .Sensor: installed around the bearing seat of the spindle or the front section of the spindle, used to capture the vibration signal of the spindle and sense the dynamic changes of the spindle during operation;

[0017] Module M 12 .Sensor mounting base: Provides a stable mounting base for the piezoelectric accelerometer, ensuring that the sensor can accurately capture vibration signals;

[0018] Module M 13 .Signal amplifier: used to amplify the weak signal output by the sensor for subsequent processing;

[0019] Module M 14 .Cable: used to connect the sensor and data acquisition module to ensure complete signal transmission.

[0020] Furthermore, the data acquisition and processing module specifically includes the following submodules:

[0021] Module M 21 .Data collector: used to receive analog signals from vibration sensors in real time and digitize them;

[0022] Module M 22 .Noise removal unit: Use filtering technology to remove high-frequency noise and environmental interference in vibration signals, leaving signals within the effective frequency band;

[0023] Module M 23 .Normalization processor: Perform amplitude normalization on the signal to ensure signal consistency and comparability.

[0024] Module M 24 .Data buffer: temporarily stores processed data for use by subsequent feature extraction modules.

[0025] Furthermore, the feature extraction module specifically includes the following submodules:

[0026] Module M 31 .Time domain feature extractor: extracts time domain features from the preprocessed vibration signal. The time domain features include mean, variance, peak, skewness and kurtosis;

[0027] Module M 32 .Frequency domain feature extractor: extracts the spectral features of the signal through Fourier transform. The spectral features include the main frequency, harmonic components and frequency band energy;

[0028] Module M 33 .Time-frequency domain feature extractor: It uses methods such as wavelet transform or short-time Fourier transform to extract time-frequency domain features. The time-frequency domain features include wavelet coefficients and instantaneous frequency;

[0029] Module M 34 .Feature vector generator: Combines the extracted features to form a feature vector for use by the machine learning diagnosis module.

[0030] Furthermore, the machine learning diagnosis module specifically includes the following submodules:

[0031] Module M 41 .Feature preprocessing unit: normalize, reduce dimension, select features, smooth, transform logarithmically, transform Box-Cox, generate polynomial features, oversample and / or undersample the extracted features;

[0032] Module M 42 Machine learning model library: Contains training models of support vector machine (SVM), random forest, neural network, decision tree, random forest, linear regression model and naive Bayes algorithm for diagnosis;

[0033] Module M 43 .Classifier: Use stacking integration and include multiple basic models. Classify the feature vectors according to the results of multi-model integration to determine the health status of the spindle;

[0034] Module M 44 .Adaptive learning unit: dynamically updates the model based on actual operating data to improve diagnostic accuracy;

[0035] Module M 45 .Fault predictor: predicts the possible fault type of the spindle and its development trend based on the classification results.

[0036] Furthermore, the output module specifically includes the following submodules:

[0037] Module M 51 .Diagnostic result display: The diagnostic results and the health status of the spindle are displayed in real time on the display.

[0038] Module M 52 .Alarm unit: emits audible and visual alarm signals to alert operators when abnormalities are detected.

[0039] Module M 53 .Remote communication module: Send the diagnostic results to the remote monitoring center or the maintenance personnel's equipment through the network.

[0040] Module M 54 .Logger: Records the diagnostic process and results for subsequent analysis and reference.

[0041] The advantages of this application are:

[0042] 1. The present invention realizes real-time monitoring and automatic diagnosis of spindle bending faults. Compared with the traditional method that relies on expert experience and manual detection, it can greatly improve the efficiency and accuracy of fault detection and reduce the risk of misjudgment and missed judgment caused by human factors.

[0043] 2. Through multi-level feature extraction (including time domain features, frequency domain features and time-frequency domain features), the present invention can fully capture the important information in the spindle vibration signal and ensure the full extraction of fault features. Compared with existing methods, the system can more accurately reflect the health status of the spindle and improve the reliability of the diagnosis results.

[0044] 3. The stacking method is used to combine multiple basic models (such as SVM, KNN, neural network) for classification. Compared with the single model method, the present invention significantly improves the accuracy and robustness of diagnosis through the integration of multiple models, and can better cope with different types of spindle bending failures.

[0045] 4. The present invention introduces an adaptive learning unit that can dynamically update the machine learning model based on real-time data, ensuring that the model is always in the best state and adapts to the ever-changing operating environment and data characteristics. Compared with the diagnostic system with a fixed model, the present invention can continuously optimize the diagnostic ability and reduce the misdiagnosis rate.

[0046] 5. In addition to being able to diagnose the current spindle fault status in real time, the present invention also has the ability to predict faults, and can predict the possible future spindle fault types and their development trends based on the current classification results. Compared with the prior art, the prediction function of the present invention provides strong support for preventive maintenance, which helps to extend the life of the equipment and reduce unexpected downtime.

[0047] 6. The present invention adopts modular design, each module is independent and has good compatibility, which is convenient for integration or expansion of application with the existing spindle monitoring system. Compared with the traditional fault detection system with a single function, the present invention has stronger flexibility and scalability and can adapt to the needs of different application scenarios.

[0048] 7. Compared with a single machine learning model, the model of this application has improved robustness; since the stacking integration method can combine different types of basic models (for different scenarios), the system can easily adapt to different types of equipment and working environments.

[0049] 8. A single complex model (such as a deep neural network) is prone to overfitting on limited training data, but this risk can be reduced by stacking and integrating multiple models. The integrated model combines the prediction results of multiple models to reduce the bias and overfitting of a single model on specific data.

[0050] 9. By analyzing the contribution of each basic model to the final result, it can help technicians better understand the logic behind the model decision, thereby optimizing the overall process of fault diagnosis and improving the model's interpretability.

[0051] 10. Compared with the traditional approach of directly transmitting vibration signals or extracting a small number of features, the physical features extracted by this application are more comprehensive and complex, and the physical state reflected is more realistic. And this more complex feature can only be well utilized in conjunction with the subsequent stacked integration model. In module M4, the stacked integration technology is used to simultaneously extract the prediction capabilities of multiple machine learning models. At the same time, a large number of features are fully extracted in the previous feature extraction and feature preprocessing modules, and a large number of differentiated preprocessing are performed. These allow the stacked integration technology to better play its ability to handle complex problems and also play the respective advantages of each basic module. At the same time, the module also has the ability to update parameters according to current data, that is, adaptive ability, and can continuously update parameters. At the same time, the fault predictor can also predict possible future faults or calculate the bearing life after detecting the current fault. BRIEF DESCRIPTION OF THE DRAWINGS

[0052] Figure 1 FIG. 1 is a schematic diagram of a system structure according to an embodiment of the present invention.

[0053] Figure 2 Workflow diagram of the present invention. DETAILED DESCRIPTION

[0054] In order to make the purpose, technical solution and advantages of the embodiments of the present application clearer, the technical solution in the embodiments of the present application will be clearly and completely described below in conjunction with the drawings in the embodiments of the present application. Obviously, the described embodiments are part of the embodiments of the present application, rather than all the embodiments. The components of the embodiments of the present application described and shown in the drawings here can be arranged and designed in various different configurations.

[0055] Therefore, the following detailed description of the embodiments of the present application provided in the accompanying drawings is not intended to limit the scope of the present application for which protection is sought, but merely represents selected embodiments of the present application. Based on the embodiments in the present application, all other embodiments obtained by ordinary technicians in the field without creative work are within the scope of protection of the present application.

[0056] It should be noted that similar reference numerals and letters denote similar items in the following drawings, and therefore, once an item is defined in one drawing, further definition and explanation thereof is not required in subsequent drawings.

[0057] In the description of the present application, it should be noted that the terms "upper", "vertical", "inner", "outer", etc. indicate the orientation or position relationship based on the orientation or position relationship shown in the drawings, or the orientation or position relationship in which the product of the application is usually placed when in use, or the orientation or position relationship commonly understood by those skilled in the art. They are only for the convenience of describing the present application and simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, be constructed and operated in a specific orientation, and therefore cannot be understood as a limitation on the present application. In addition, the terms "first", "second", etc. are only used to distinguish the description, and cannot be understood as indicating or implying relative importance.

[0058] In the description of this application, it should also be noted that, unless otherwise clearly specified and limited, the terms "set", "install", and "connect" should be understood in a broad sense, for example, it can be a fixed connection, a detachable connection, or an integral connection; it can be a mechanical connection or an electrical connection; it can be a direct connection, or it can be an indirect connection through an intermediate medium, or it can be the internal communication of two components. For ordinary technicians in this field, the specific meanings of the above terms in this application can be understood according to specific circumstances.

[0059] Example 1

[0060] A spindle bending common feature diagnosis system based on a stacking integration method includes the following modules:

[0061] Module M 1 .Sensor module: used to collect the vibration signal of the spindle in real time;

[0062] The sensor module can use any sensor that can test vibration signals, including vibration sensors, such as piezoelectric accelerometers; speed sensors, such as electromagnetic speed sensors; displacement sensors, such as eddy current sensors; laser sensors, etc. The sensor is installed around the bearing seat of the spindle or at the front end of the spindle to capture the dynamic changes of the spindle during operation; the installation position is usually near the bearing seat. If the possible fault is mainly caused by cutting force, spindle imbalance or tool failure, the key position is the front end of the spindle.

[0063] Module M 2 .Data acquisition and processing module: converts analog signals into digital signals and preprocesses the signals collected by the sensor. The preprocessing includes noise removal, filtering, normalization and other processing to ensure the quality and availability of the signal.

[0064] Module M 3.Feature extraction module: extract features from the preprocessed vibration signal, including time domain features (such as mean, variance, peak), frequency domain features (such as spectral components, harmonics), time-frequency domain features (such as wavelet transform coefficients), etc., and quantize the extracted features.

[0065] Module M 4 .Machine learning diagnosis module: Based on multiple trained machine learning models, stacking is used to analyze and classify the extracted features to determine the health status of the spindle. The basic machine learning models of stacking include support vector machine (SVM), K nearest neighbor, neural network (multilayer perceptron MLP, extreme learning machine, convolutional neural network CNN, recurrent neural network RNN, long short-term memory network LSTM, Transformer), decision tree, random forest, linear regression model and naive Bayes algorithm;

[0066] Module M 5 Output module: used to output diagnostic results and send out alarm signals when abnormalities are detected. Output is achieved through a display, alarm or remote communication module, and a log is used to record diagnostic results.

[0067] Example 2

[0068] As a preferred embodiment of the present invention, the present invention includes a common feature diagnosis system for spindle bending based on machine learning, including the following modules:

[0069] Module M 1 .Vibration sensor module: used to collect the vibration signal of the spindle in real time. The sensor can be a piezoelectric accelerometer installed at a key position of the spindle to capture the dynamic changes of the spindle during operation.

[0070] Module M 2 .Data acquisition and processing module: converts analog signals into digital signals and pre-processes the signals collected by the vibration sensor, including noise removal, filtering, normalization and other processing to ensure the quality and availability of the signal. Specifically, it includes the following steps:

[0071] Module M 21 .Data collector: Receives analog signals from vibration sensors in real time. The collected vibration signals are time domain signals, which are expressed as voltage or acceleration that changes with time. The form is usually expressed as s(t), where t is time.

[0072] After the data is collected, it is digitized through a high-precision analog-to-digital converter (ADC) to ensure the accuracy and integrity of the data collection:

[0073] s[n]=s(n·T s )

[0074] s(t) is a continuous time signal; s[n] is a discrete time signal, representing the value at the nth sampling point; T s is the sampling period.

[0075] Module M 22 .Noise removal unit: removes high-frequency noise and environmental interference from the signal through filtering technology to ensure data purity. This unit uses an adaptive filter to adjust the filtering parameters according to the real-time signal, optimize the noise suppression effect, and retain the signal within the effective frequency band related to the health status of the spindle.

[0076] Module M 23 .Normalization processor: Perform amplitude normalization processing on the signal to ensure the consistency and comparability of the signal and eliminate the impact of signal amplitude differences under different working conditions on subsequent feature extraction.

[0077] Module M 24 .Data buffer: Temporarily stores pre-processed data to facilitate batch processing and analysis by subsequent feature extraction modules.

[0078] Module M 3 .Feature extraction module: extract relevant features from the preprocessed vibration signal, including time domain features (such as mean, variance, peak), frequency domain features (such as spectral components, harmonics), time-frequency domain features (such as wavelet transform coefficients), etc. and quantize the extracted features.

[0079] Module M 4 .Machine learning diagnosis module: Based on multiple trained machine learning models, stacking is used to analyze and classify the extracted features to determine the health status of the spindle. The basic models of stacking include support vector machine (SVM), K nearest neighbor or neural network algorithms.

[0080] Module M 5 Output module: used to output diagnostic results and send out alarm signals when abnormalities are detected. The output can be achieved through a display, alarm or remote communication module. At the same time, a log is used to record the diagnostic results.

[0081] Example 3

[0082] As a preferred embodiment of the present invention, the present invention includes a common feature diagnosis system for spindle bending based on machine learning, including the following modules:

[0083] Module M 1 .Vibration sensor module: used to collect the vibration signal of the spindle in real time. The sensor can be a piezoelectric accelerometer installed at a key position of the spindle to capture the dynamic changes of the spindle during operation.

[0084] Module M 2 .Data acquisition and processing module: converts analog signals into digital signals and pre-processes the signals collected by the vibration sensor, including noise removal, filtering, normalization and other processing to ensure the quality and availability of the signal.

[0085] Module M 3 .Feature extraction module: extract relevant features from the preprocessed vibration signal, including time domain features (such as mean, variance, peak), frequency domain features (such as spectral components, harmonics), time-frequency domain features (such as wavelet transform coefficients), etc. and quantize the extracted features. The extracted features include:

[0086] Time domain features: Time domain features are key indicators directly extracted from time series, mainly including mean, variance, peak value, skewness, etc. The mean represents the average level of the signal within a window, the variance measures the volatility of the signal, and the peak value is used to detect shocks or drastic changes in the signal. Skewness is a measure of the symmetry of the signal, reflecting whether the signal tends to be higher or lower values, while kurtosis evaluates the degree of sharp peaks in the signal, indicating the kurtosis characteristics of the signal. Among them:

[0087]

[0088] Peak Value=max(s[i])for i=1, 2,…,N

[0089]

[0090] Where N is the number of samples; s[i] is the signal value of the i-th sample point; max(s[i]) means finding the maximum value among all sample points; σ is the standard deviation of the signal, defined as

[0091] Frequency domain features: Frequency domain features are extracted by converting the time domain signal to the frequency domain through fast Fourier transform (FFT), focusing on the characteristics of frequency components. The main frequency is the frequency component with the largest amplitude in the spectrum, which is usually related to the rotation speed or fault frequency and can reveal the operating status of the mechanical equipment. Harmonic components are multiple frequency components related to the main frequency and are often used to identify faults such as imbalance or poor alignment. Band energy represents the vibration intensity within a specific frequency range and can effectively indicate the energy concentration of a specific fault frequency band.

[0092] Time-frequency domain features: Time-frequency domain features are extracted through wavelet transform or short-time Fourier transform (STFT), which can reflect the local characteristics of the signal in time and frequency. Wavelet coefficients are obtained through wavelet transform, which can reveal the characteristics of the signal at different scales, and are particularly suitable for capturing the characteristics of transient events or non-stationary signals. The instantaneous frequency describes the change of signal frequency over time, which is very useful for detecting dynamically changing fault characteristics. Through the analysis of time-frequency domain features, the characteristics of complex signals can be more fully understood, especially when dealing with nonlinear and non-stationary signals.

[0093] Module M 4 .Machine learning diagnosis module: Based on multiple trained machine learning models, stacking is used to analyze and classify the extracted features to determine the health status of the spindle. The basic models of stacking include support vector machine (SVM), K nearest neighbor or neural network algorithms.

[0094] Module M 5 Output module: used to output diagnostic results and send out alarm signals when abnormalities are detected. The output can be achieved through a display, alarm or remote communication module. At the same time, a log is used to record the diagnostic results.

[0095] Example 4

[0096] As a preferred embodiment of the present invention, the present invention includes a common feature diagnosis system for spindle bending based on machine learning, including the following modules:

[0097] Module M 1 .Vibration sensor module: used to collect the vibration signal of the spindle in real time. The sensor can be a piezoelectric accelerometer installed at a key position of the spindle to capture the dynamic changes of the spindle during operation.

[0098] Module M 2 .Data acquisition and processing module: converts analog signals into digital signals and pre-processes the signals collected by the vibration sensor, including noise removal, filtering, normalization and other processing to ensure the quality and availability of the signal.

[0099] Module M 3 .Feature extraction module: extract relevant features from the preprocessed vibration signal, including time domain features (such as mean, variance, peak), frequency domain features (such as spectral components, harmonics), time-frequency domain features (such as wavelet transform coefficients), etc. and quantize the extracted features.

[0100] Module M 4.Machine learning diagnosis module: Based on multiple trained machine learning models, stacking is used to analyze and classify the extracted features to determine the health status of the spindle. The basic models of stacking include support vector machine (SVM), K nearest neighbor or neural network algorithms. Module M 4 It includes the following sub-modules:

[0101] Module M 41 .Feature preprocessing unit: The feature preprocessing unit is used to further process the feature data obtained from the feature extraction module to adapt to different machine learning algorithms. The feature data will first be normalized again to make the numerical range of different features consistent, so as to avoid the uneven impact of some features on the model due to the large numerical range. Then, in the case of high feature dimensions, the feature preprocessing unit will perform dimensionality reduction through techniques such as principal component analysis (PCA) or linear discriminant analysis (LDA) to retain the feature subset that best represents the health status of the spindle. In addition, the system will also screen out the most important features through feature selection methods such as chi-square test and information gain to improve the efficiency and accuracy of the model.

[0102] Module M 42 .Machine learning model library: Contains different types of basic training models such as support vector machine (SVM), K-nearest neighbors (KNN), neural network, etc. for diagnosis. The working principle of the model is as follows:

[0103] The support vector machine separates samples of different categories in the input space by finding an optimal classification hyperplane W·x+b=0, where w is the weight vector, x is the input feature vector, and b is the bias term. The goal of SVM is to maximize the classification interval, and the optimization objective function is At the same time, the constraints y i (w·x i +b)≥1 for all samples i where y i is the class label of the sample (+1 or -1). In nonlinear classification problems, SVM uses the kernel function K(x i ,x j ) maps the input features to a high-dimensional space so that the samples are linearly separable in this space.

[0104] The K-Nearest Neighbors (KNN) classification algorithm calculates the test sample x test For each sample x in the training set i The Euclidean distance between Select the K samples with the closest distance {x (1) ,x (2) ,...,x (K)} as neighbors, and make voting decisions based on the categories of these neighbors y(1), y(2), …, y(K), and finally classify the test sample as y^ test Set to the category that appears most frequently among these neighbors, i.e. y^ test =mode(y(1),y(2),…,y(K)).

[0105] Neural network: The neural network calculates the activation value a of each layer of neurons through forward propagation (l) =σ(w (l) ·a (l-1) +b (l) ), where a (l) is the activation value of layer l, w (l) and b (l) are the weight and bias of the layer, and the σ activation function (such as ReLU, Sigmoid, etc.). Through the back propagation algorithm, the neural network calculates the partial derivative of the loss function L(y,y^) with respect to each parameter, and updates the weights and biases layer by layer. Where η is the learning rate. Ultimately, through iterative optimization, the network can learn the complex nonlinear mapping relationship between input features and output targets, and achieve high-precision prediction of fault detection.

[0106] Module M 43 .Classifier: The classifier is a key component in the machine learning diagnosis module. It uses a stacked model as a whole to classify the input feature vectors according to the selected basic model and meta-model to determine the health status of the spindle. Specific workflow:

[0107] First, multiple basic models (such as support vector machines, K nearest neighbors, neural networks, etc.) are trained independently on the original training data. Each basic model classifies the input data according to its own learning method and outputs the corresponding prediction results. The probability values ​​output by each basic model are collected to form a new feature set, called "meta-features" or "secondary features" (Meta-Features). In this model, the basic model outputs the probability that a sample is normal, slightly curved, or severely curved;

[0108] Assume that there are N basic models, and each basic model predicts M samples in the training set, then each sample will generate N prediction results (meta-features). These meta-features will form a new feature matrix as the input for training the meta-model. The meta-model can choose models such as logistic regression, linear regression, and random forest to combine the prediction results of the basic models. Take the logistic regression model as an example: the logistic regression meta-model learns how to best combine the prediction results of the basic models by fitting the relationship between the meta-features and the target variable. It assigns a weight to the meta-feature output by each basic model, and combines these weights to make the final prediction result.

[0109] Module M 44 .The adaptive learning unit is used to dynamically adjust and update the machine learning model to adapt to the real-time changing operating environment and data characteristics. The unit has the function of online learning, can accept new training data in real time, and adjust the parameters and structure of the model according to these data, so that the model can adapt to the changing spindle operating status. Online learning means that the model can gradually learn the incoming data without retraining the entire model from scratch. Online learning can use algorithms such as perceptron, stochastic gradient descent (SGD), online SVM, etc. to achieve real-time update of model parameters.

[0110] In addition, the adaptive learning unit can regularly update the model based on the latest data and fault conditions collected during the diagnosis process to ensure the continuous improvement of diagnostic accuracy. Through the data feedback mechanism, the system will compare and analyze the diagnosis results with the actual fault conditions. If the model is found to have misjudged or missed, the adaptive learning unit will optimize the model based on the feedback information to further improve the accuracy and stability of the diagnosis.

[0111] Module M 45 .Fault predictor: Based on the output of the classifier and historical data, the fault type and development trend of the spindle are predicted to provide a basis for preventive maintenance. Through trend analysis of historical diagnostic data and current health status, the fault predictor can identify the changing trend of the spindle state, such as whether a slight bend may develop into a severe bend. Based on the trend analysis results, the fault predictor can warn of possible failures in advance and remind operators to take preventive maintenance measures to avoid serious failures. The unit also combines multiple features and historical data for a comprehensive multi-factor analysis, such as considering factors such as operating time, load conditions, and ambient temperature to improve the accuracy and reliability of the prediction.

[0112] Module M 5 Output module: used to output diagnostic results and send out alarm signals when abnormalities are detected. The output can be achieved through a display, alarm or remote communication module. Module M 5 It includes the following sub-modules:

[0113] Module M 51 .Diagnostic result display: The diagnostic results and the health status of the spindle are displayed in real time on the display.

[0114] Module M 52 .Alarm unit: emits audible and visual alarm signals to alert operators when abnormalities are detected.

[0115] Module M 53.Remote communication module: Send the diagnostic results to the remote monitoring center or the maintenance personnel's equipment through the network.

[0116] Module M 54 .Logger: Records the diagnostic process and results for subsequent analysis and reference.

[0117] Example 5

[0118] A spindle bending common feature diagnosis system based on a stacking integration method includes the following modules:

[0119] Module M 1 .Sensor module: used to collect the vibration signal of the spindle in real time;

[0120] The sensor module can use any sensor that can test vibration signals, including vibration sensors, such as piezoelectric accelerometers; speed sensors, such as electromagnetic speed sensors; displacement sensors, such as eddy current sensors; laser sensors, etc. The sensor is installed around the bearing seat of the spindle or at the front end of the spindle to capture the dynamic changes of the spindle during operation; the installation position is usually near the bearing seat. If the possible fault is mainly caused by cutting force, spindle imbalance or tool failure, the key position is the front end of the spindle.

[0121] Module M 2 .Data acquisition and processing module: converts analog signals into digital signals and preprocesses the signals collected by the sensor. The preprocessing includes noise removal, filtering, normalization and other processing to ensure the quality and availability of the signal.

[0122] Module M 3 .Feature extraction module: extract features from the preprocessed vibration signal, including time domain features (such as mean, variance, peak), frequency domain features (such as spectral components, harmonics), time-frequency domain features (such as wavelet transform coefficients), etc., and quantize the extracted features.

[0123] Module M 4 .Machine learning diagnosis module: Based on multiple trained machine learning models, stacking is used to analyze and classify the extracted features to determine the health status of the spindle. The basic machine learning models of stacking include support vector machine (SVM), K nearest neighbor, neural network (multilayer perceptron MLP, extreme learning machine, convolutional neural network CNN, recurrent neural network RNN, long short-term memory network LSTM, Transformer), decision tree, random forest, linear regression model and naive Bayes algorithm;

[0124] Module M 5Output module: used to output diagnostic results and send out alarm signals when abnormalities are detected. Output is achieved through a display, alarm or remote communication module, and a log is used to record diagnostic results.

[0125] Furthermore, the sensor module specifically includes the following submodules:

[0126] Module M 11 .Sensor: installed around the bearing seat of the spindle or the front section of the spindle, used to capture the vibration signal of the spindle and sense the dynamic changes of the spindle during operation;

[0127] Module M 12 .Sensor mounting base: Provides a stable mounting base for the piezoelectric accelerometer, ensuring that the sensor can accurately capture vibration signals;

[0128] Module M 13 .Signal amplifier: used to amplify the weak signal output by the sensor for subsequent processing;

[0129] Module M 14 .Cable: used to connect the sensor and data acquisition module to ensure complete signal transmission.

[0130] Furthermore, the data acquisition and processing module specifically includes the following submodules:

[0131] Module M 21 .Data collector: used to receive analog signals from vibration sensors in real time and digitize them;

[0132] Module M 22 .Noise removal unit: Use filtering technology to remove high-frequency noise and environmental interference in vibration signals, leaving signals within the effective frequency band;

[0133] Module M 23 .Normalization processor: Perform amplitude normalization on the signal to ensure signal consistency and comparability.

[0134] Module M 24 .Data buffer: temporarily stores processed data for use by subsequent feature extraction modules.

[0135] Furthermore, the feature extraction module specifically includes the following submodules:

[0136] Module M 31 .Time domain feature extractor: extracts time domain features from the preprocessed vibration signal. The time domain features include mean, variance, peak, skewness and kurtosis;

[0137] Module M 32.Frequency domain feature extractor: extracts the spectral features of the signal through Fourier transform. The spectral features include the main frequency, harmonic components and frequency band energy;

[0138] Module M 33 .Time-frequency domain feature extractor: It uses methods such as wavelet transform or short-time Fourier transform to extract time-frequency domain features. The time-frequency domain features include wavelet coefficients and instantaneous frequency;

[0139] Module M 34 .Feature vector generator: Combines the extracted features to form a feature vector for use by the machine learning diagnosis module.

[0140] Furthermore, the machine learning diagnosis module specifically includes the following submodules:

[0141] Module M 41 .Feature preprocessing unit: normalize, reduce dimension, select features, smooth, logarithm transform, Box-Cox transform, generate polynomial features, oversample or / and undersample the extracted features to adapt to different machine learning algorithms. Since tree models such as decision trees and random forests are sensitive to noise and irrelevant features, feature selection, such as recursive feature elimination (RFE) is required; for time series models such as LSTM and RNN, smoothing, such as exponential weighted moving average (EWMA) is required; for linear regression, logistic regression, support vector machine (SVM), logarithm transform or Box-Cox transform may be required; generate polynomial features: enhance the adaptability of the model to complex data by generating square, cube or feature interaction terms of features; oversampling: such as SMOTE, balance the data by synthesizing new minority class samples; undersampling: reduce majority class samples to make the data set more balanced.

[0142] Module M 42 Machine learning model library: Contains training models of support vector machine (SVM), random forest, neural network, decision tree, random forest, linear regression model and naive Bayes algorithm for diagnosis;

[0143] Module M 43 .Classifier: Use stacking integration and include multiple basic models. Classify the feature vectors according to the results of multi-model integration to determine the health status of the spindle;

[0144] Module M 44 .Adaptive learning unit: dynamically updates the model based on actual operating data to improve diagnostic accuracy;

[0145] Module M 45 .Fault predictor: predicts the possible fault type of the spindle and its development trend based on the classification results.

[0146] Furthermore, the output module specifically includes the following submodules:

[0147] Module M 51 .Diagnostic result display: The diagnostic results and the health status of the spindle are displayed in real time on the display.

[0148] Module M 52 .Alarm unit: emits audible and visual alarm signals to alert operators when abnormalities are detected.

[0149] Module M 53 .Remote communication module: Send the diagnostic results to the remote monitoring center or the maintenance personnel's equipment through the network.

[0150] Module M 54 .Logger: Records the diagnostic process and results for subsequent analysis and reference.

[0151] Example 6

[0152] 1. Piezoelectric accelerometer collects vibration signals: The acceleration signal a(t) in the spindle vibration is converted into a voltage signal V(t). This signal is the electrical representation of the vibration and directly reflects the dynamic changes of the spindle.

[0153] 2. The signal amplifier amplifies the weak signal output by the sensor for subsequent processing:

[0154] V amp (t) = G·V(t)

[0155] G: Amplifier gain (Gain). This is a proportional factor that indicates the amplifier's amplification of the input signal.

[0156] 3. The data collector receives the analog signal from the vibration sensor in real time. The collected vibration signal is a time domain signal, which is expressed as a voltage or acceleration that changes with time. It is usually expressed as s(t), where t is time.

[0157] s(t)=V amp (t)

[0158] After the data is collected, it is digitized through a high-precision analog-to-digital converter (ADC) to ensure the accuracy and integrity of the data collection:

[0159] s[n]=s(n·T s )

[0160] s(t) is a continuous time signal; s[n] is a discrete time signal, representing the value at the nth sampling point; T s is the sampling period.

[0161] 4. In the noise removal unit, high-frequency noise and environmental interference in the signal are removed through filtering technology to ensure the purity of the data. The unit uses an adaptive filter to adjust the filtering parameters according to the real-time signal, optimize the noise suppression effect, and retain the signal in the effective frequency band related to the health status of the spindle.

[0162] Convolve the signal s[n] using the adaptive filter h[n]:

[0163]

[0164] L: The length of the filter h[n], which indicates the range of the filter action, that is, how many filter coefficients are used for the convolution operation.

[0165] h[n]: The impulse response of the filter, i.e. the coefficients of the filter, which describe how the filter affects the input signal. h[k] is the weight of the filter at the kth sample position.

[0166] n: time index, indicating the current moment. At moment n, the output signal sf[n] depends on the convolution of the input signal s[n] and the filter h[n].

[0167] k: An intermediate variable in the convolution operation, used to traverse all the coefficients of the filter h[n]. It represents the relative displacement of the signal and the filter at different time points.

[0168] According to the real-time signal, the filter coefficient h[n] is adjusted using an algorithm (such as the LMS algorithm):

[0169] h[n+1]=h[n]+μ·e[n]·s[n]

[0170] Where μ is the step size parameter, e[n] is the error signal, h[n] is the weight value of the filter at time step n. e[n] is the error at time step n, that is, the difference between the expected output and the actual output, and s[n] is the input signal at time step n.

[0171] 5. Perform amplitude normalization on the signal in the normalization processor to ensure the consistency and comparability of the signal and eliminate the impact of signal amplitude differences under different working conditions on subsequent feature extraction.

[0172]

[0173] The preprocessed data is then temporarily stored to facilitate batch processing and analysis in the subsequent feature extraction module.

[0174] 6. In the feature extraction module, relevant features are extracted from the preprocessed vibration signal, including time domain features (such as mean, variance, peak), frequency domain features (such as spectral components, harmonics), time-frequency domain features (such as wavelet transform coefficients), etc. The extracted features are quantized. The extracted features specifically include:

[0175] Time domain features: Time domain features are key indicators directly extracted from time series, mainly including mean, variance, peak value, skewness, etc. The mean represents the average level of the signal within a window, the variance measures the volatility of the signal, and the peak value is used to detect shocks or drastic changes in the signal. Skewness is a measure of the symmetry of the signal, reflecting whether the signal tends to be higher or lower values, while kurtosis evaluates the degree of sharp peaks in the signal, indicating the kurtosis characteristics of the signal. Among them:

[0176]

[0177] Where N is the number of samples; s[i] is the signal value of the i-th sample point; max(s[i]) means finding the maximum value among all sample points; o is the standard deviation of the signal, defined as

[0178] Frequency domain features: Frequency domain features are obtained by converting the time domain signal to the frequency domain through fast Fourier transform (FFT):

[0179]

[0180] Mainly focus on the characteristics of frequency components: The main frequency is the frequency component with the largest amplitude in the spectrum, which is usually related to the rotation speed or fault frequency and can reveal the operating status of mechanical equipment:

[0181]

[0182] The corresponding frequency is

[0183] Harmonic components are multiple frequency components related to the main frequency, and are often used to identify faults such as imbalance or poor alignment. The harmonic frequency in the analysis spectrum is: f n =n·f bas n: represents the order or multiple of the harmonic. n=1 represents the fundamental frequency, n=2 represents the second harmonic, and so on. The harmonic frequency is an integer multiple of the fundamental frequency.

[0184] Frequency band energy indicates the vibration intensity within a specific frequency range, which can effectively indicate the energy concentration of a specific fault frequency band: (Calculate energy within a specific frequency range [k1, k2])

[0185] Time-frequency domain features: Time-frequency domain features are extracted through wavelet transform or short-time Fourier transform (STFT), which can reflect the local characteristics of the signal in time and frequency.

[0186] Wavelet coefficients are obtained through wavelet transform, which can reveal the characteristics of signals at different scales. They are particularly suitable for capturing the characteristics of transient events or non-stationary signals. Discrete wavelet transform (DWT) is used to decompose the signal:

[0187]

[0188] where c 1 and d 1 are the approximation coefficient and detail coefficient, ψ 1 (n) and φ 1 (n) is the wavelet function.

[0189] The instantaneous frequency describes the change of signal frequency over time, which is very useful for detecting dynamically changing fault characteristics. Through the analysis of time-frequency domain characteristics, the characteristics of complex signals can be more fully understood, especially when dealing with nonlinear and non-stationary signals.

[0190] 6. The feature preprocessing unit further processes the feature data obtained from the feature extraction module to adapt to different machine learning algorithms. For example: (1) Since tree models such as decision trees and random forests are more sensitive to noise and irrelevant features, feature selection is required, such as recursive feature elimination (RFE) (2) For time series models such as LSTM and RNN, smoothing processing is required, such as exponentially weighted moving average (EWMA):

[0191] EWMA t =αx t +(1-α)EWMA t-1

[0192] α is the smoothing coefficient

[0193] (3) For linear regression, logistic regression, support vector machine (SVM), etc., logarithmic transformation or Box-Cox transformation may be required:

[0194] Logarithmic transformation: x′=log(x∈), where ∈ is to prevent the problem of logarithm zero.

[0195] Box-Cox transformation:

[0196] λ: Parameter in Box-Cox transformation, controlling the type of transformation. When λ=0, natural logarithmic transformation is used, which is suitable for cases where the data has positive skewness. When λ≠0, power transformation is used, and the transformation form is x λ, and the results are scaled.

[0197] (4) It is also possible to generate polynomial features: by generating squares, cubes, or feature interactions of features, the model's adaptability to complex data is enhanced. (5) Oversampling: such as SMOTE, by synthesizing new minority class samples to balance the data. Undersampling: reducing majority class samples to make the data set more balanced.

[0198] 7. The data enters the fault detector. First, multiple basic models classify the input data and output the corresponding prediction results. The probability values ​​output by each basic model are collected to form a new feature set, called "meta-features" or "secondary features" (Meta-Features). In this model, the basic model outputs the probability that a sample is normal, slightly bent, or severely bent;

[0199] The meta-model then integrates the classification results of the basic models: Assuming there are N basic models, each basic model predicts M samples in the training set, then each sample will generate N prediction results (meta-features). These meta-features will form a new feature matrix Z:

[0200]

[0201] where z ij is the predicted probability of the ith sample under the jth basic model, and M is the number of basic models.

[0202] As the input of the training meta-model. The meta-model can choose models such as logistic regression, linear regression, random forest, etc. to combine the prediction results of the basic models. Take the logistic regression model as an example: the logistic regression meta-model learns how to best combine the prediction results of the basic models by fitting the relationship between the meta-features and the target variable. It assigns a weight to the meta-feature output by each basic model, and combines the final prediction results through these weights:

[0203] The model form is:

[0204] β 0 : represents the intercept term (bias), which is the constant term in the model. It controls the baseline probability when all feature values ​​are 0.

[0205] β j : represents feature z j The coefficients or weights of each β j Weighing the feature z j The magnitude of the effect on the outcome y.

[0206] The minimization loss function is:

[0207] y i : represents the actual label of the i-th sample (0 or 1)

[0208] P(y i ∣Z i ): indicates that the logistic regression model predicts that sample i belongs to the positive class (y i =1).

[0209] The machine learning model library may include different types of basic training models such as support vector machine (SVM), K-nearest neighbors (KNN), neural network, etc. for diagnosis. The working principle of the model is as follows:

[0210] The support vector machine separates samples of different categories in the input space by finding an optimal classification hyperplane W·x+b=0, where w is the weight vector, x is the input feature vector, and b is the bias term. The goal of SVM is to maximize the classification interval, and the optimization objective function is At the same time, the constraints y i (w·x i +b)≥1 for all samples i where y i is the class label of the sample (+1 or -1). In nonlinear classification problems, SVM uses the kernel function K(x i ,x j ) maps the input features to a high-dimensional space so that the samples are linearly separable in this space.

[0211] The K-Nearest Neighbors (KNN) classification algorithm calculates the test sample x test For each sample x in the training set i The Euclidean distance between Select the K samples with the closest distance {x (1) ,x (2) ,...,x (K)} as neighbors, and make voting decisions based on the categories of these neighbors y(1), y(2), …, y(K), and finally classify the test sample as y^ test Set to the category that appears most frequently among these neighbors, i.e. y^ test =mode(y(1),y(2),…,y(K)).

[0212] Neural network: The neural network calculates the activation value a of each layer of neurons through forward propagation (l) =σ(w (l) ·a (l-1) +b (l) ), where a (l) is the activation value of layer l, w (l)and b (l) are the weight and bias of the layer, and the σ activation function (such as ReLU, Sigmoid, etc.). Through the back propagation algorithm, the neural network calculates the partial derivative of the loss function L(y,y^) with respect to each parameter, and updates the weights and biases layer by layer. Where η is the learning rate. Ultimately, through iterative optimization, the network can learn the complex nonlinear mapping relationship between input features and output targets, and achieve high-precision prediction of fault detection.

[0213] 8. While diagnosing faults, the adaptive learning unit is used to dynamically adjust and update the machine learning model to adapt to the real-time changing operating environment and data characteristics. The unit has the function of online learning, can accept new training data in real time, and adjust the parameters and structure of the model according to these data, so that the model can adapt to the changing spindle operating status. Online learning means that the model can gradually learn the incoming data without retraining the entire model from scratch. Online learning can use algorithms such as perceptron, stochastic gradient descent (SGD), online SVM, etc. to achieve real-time update of model parameters.

[0214] In addition, the adaptive learning unit can regularly update the model based on the latest data and fault conditions collected during the diagnosis process to ensure the continuous improvement of diagnostic accuracy. Through the data feedback mechanism, the system will compare and analyze the diagnosis results with the actual fault conditions. If the model is found to have misjudged or missed, the adaptive learning unit will optimize the model based on the feedback information to further improve the accuracy and stability of the diagnosis.

[0215] 9. After the fault diagnosis is completed, based on the output results of the classifier and historical data, the possible types of spindle faults and their development trends are predicted to provide a basis for preventive maintenance. Through trend analysis of historical diagnostic data and current health status, the fault predictor can identify the changing trend of the spindle state, such as whether a slight bend may develop into a severe bend. Based on the trend analysis results, the fault predictor can warn of possible faults in advance and remind operators to take preventive maintenance measures to avoid serious faults. The unit also combines multiple features and historical data to perform a comprehensive multi-factor analysis, such as considering factors such as operating time, load conditions, and ambient temperature to improve the accuracy and reliability of the prediction.

[0216] 10. Output the diagnostic results and send out an alarm signal when an abnormality is detected. The output can be achieved through a display, an alarm or a remote communication module.

[0217] In summary, after reading the present invention document, ordinary technicians in this field can make various other corresponding transformation schemes based on the technical scheme and technical concept of the present invention without creative mental labor, which all fall within the scope of protection of the present invention.

Claims

1. A spindle bending common feature diagnosis system based on stacking integration method, characterized in that: Includes the following modules: Sensor module: used to collect the vibration signal of the spindle in real time; The sensors used in the sensor module include vibration sensors, speed sensors, displacement sensors, and laser sensors. The sensors are installed around the bearing seat of the spindle or at the front end of the spindle to capture the dynamic changes of the spindle during operation. Data acquisition and processing module: converts analog signals into digital signals and preprocesses the signals collected by the sensor, including noise removal, filtering, and normalization. Feature extraction module: extracts features from the preprocessed vibration signal, including time domain features, frequency domain features and time-frequency domain features, and quantizes the extracted features; Machine learning diagnosis module: Based on multiple trained machine learning models, stacking integration is used to analyze and classify the extracted features to determine the health status of the spindle. The basic machine learning models of stacking integration include support vector machine, K nearest neighbor, neural network, decision tree, random forest, linear regression model and naive Bayes algorithm; Output module: used to output diagnostic results and send out alarm signals when abnormalities are detected. The output is realized through a display, alarm or remote communication module, and a log is used to record the diagnostic results.

2. According to claim 1, a spindle bending common feature diagnosis system based on stacking integration method is characterized in that: The sensor module specifically Includes the following submodules: Sensor: installed around the spindle bearing seat or the front section of the spindle to capture the vibration signal of the spindle and sense the dynamic changes of the spindle during operation; Sensor mounting base: provides a stable mounting base for the piezoelectric accelerometer, ensuring that the sensor can accurately capture vibration signals; Signal amplifier: used to amplify the weak signal output by the sensor for subsequent processing; Cable: used to connect the sensor and data acquisition module to ensure complete signal transmission.

3. The spindle bending common feature diagnosis system based on stacking integration method according to claim 1 is characterized in that: The data acquisition and processing module specifically includes the following submodules: Data collector: used to receive analog signals from vibration sensors in real time and digitize them; Noise removal unit: Use filtering technology to remove high-frequency noise and environmental interference in vibration signals, leaving signals within the effective frequency band; Normalization processor: normalize the signal amplitude to ensure the consistency and comparability of the signal; Data buffer: temporarily stores processed data for use by subsequent feature extraction modules.

4. According to claim 1, a spindle bending common feature diagnosis system based on stacking integration method is characterized in that: The feature extraction module specifically Includes the following submodules: Time domain feature extractor: extracts time domain features from the preprocessed vibration signal. The time domain features include mean, variance, peak, skewness and kurtosis. Frequency domain feature extractor: extracts the spectral features of the signal through Fourier transform. The spectral features include main frequency, harmonic components and frequency band energy. Time-frequency domain feature extractor: uses methods such as wavelet transform or short-time Fourier transform to extract time-frequency domain features, which include wavelet coefficients and instantaneous frequency; Feature vector generator: Combines the extracted features to form a feature vector for use by the machine learning diagnosis module.

5. The spindle bending common feature diagnosis system based on stacking integration method according to claim 1 is characterized in that: The machine learning diagnosis module specifically includes the following submodules: Feature preprocessing unit: normalizes, reduces dimension, selects features, smoothes, transforms logarithmically, transforms Box-Cox, generates polynomial features, and performs oversampling and / or undersampling operations on the extracted features; Machine learning model library: Contains training models of support vector machines, random forests, neural networks, decision trees, random forests, linear regression models, and naive Bayes algorithms for diagnostic use; Classifier: Use stacking integration and include multiple basic models. Classify the feature vectors according to the results of multi-model integration to determine the health status of the spindle; Adaptive learning unit: dynamically updates the model based on actual operating data to improve diagnostic accuracy; Fault predictor: predicts the possible fault type of the spindle and its development trend based on the classification results.

6. The spindle bending common feature diagnosis system based on stacking integration method according to claim 1 is characterized in that: The output module specifically includes the following submodules: Diagnostic result display: The diagnostic result and spindle health status are displayed in real time on the display; Alarm unit: emits sound and light alarm signals to alert operators when abnormalities are detected; Remote communication module: Sends diagnostic results to the remote monitoring center or maintenance personnel's equipment through the network; Logger: Records the diagnostic process and results for subsequent analysis and reference.

7. A spindle bending common feature diagnosis system based on a stacking integration method according to any one of claims 1 to 6, characterized in that: The specific process of the system is as follows: 1). Piezoelectric accelerometer collects vibration signals: converts the acceleration signal a(t) in the spindle vibration into a voltage signal V(t); 2). The signal amplifier amplifies the weak signal output by the sensor: V amp (t)=G·V(t) G is the gain of the amplifier, which is a proportional factor and represents the amplifier's amplification factor for the input signal; 3). The data collector receives the analog signal from the vibration sensor in real time. The collected vibration signal is a time domain signal, which is expressed as a voltage or acceleration that changes with time. It is expressed as s(t), where t is time: s(t)=V amp (t) After the acquisition is completed, it is digitized through an analog-to-digital converter: s [n]=s(n·T s ) s(t) is a continuous time signal; s[n] is a discrete time signal, representing the value at the nth sampling point; T s is the sampling period; 4). In the noise removal unit, high-frequency noise and environmental interference in the signal are removed by filtering technology, and adaptive filters are used to adjust the filtering parameters according to the real-time signal to retain the signal in the effective frequency band related to the health status of the spindle; Convolve the signal s[n] using the adaptive filter h[n]: L: the length of the filter h[n], which indicates the range of the filter; h[n]: the impulse response of the filter; h[k] is the weight of the filter at the kth sample position; n: the time index, indicating the current moment; at time n, the output signal sf[n] depends on the convolution of the input signal s[n] and the filter h[n]; k: an intermediate variable in the convolution operation, used to traverse all the coefficients of the filter h[n]; it represents the relative displacement of the signal and the filter at different time points; According to the real-time signal, adjust the filter coefficient h[n]: h[n+1]=h[n]+μ·[n]·s[n] Where μ is the step size parameter, e[n] is the error signal, h[n] is the weight value of the filter at time step n; e[n] is the error at time step n, and s[n] is the input signal at time step n; 5). Perform amplitude normalization on the signal in the normalization processor. The preprocessed data is then temporarily stored to facilitate batch processing and analysis by the subsequent feature extraction module; 6). In the feature extraction module, relevant features are extracted from the preprocessed vibration signal, including time domain features, frequency domain features, and time-frequency domain features, and the extracted features are quantized; The extracted features include: Time domain features: Time domain features are key indicators directly extracted from time series, mainly including mean, variance, peak, and skewness; mean represents the average level of the signal within a window, variance measures the volatility of the signal, and peak is used to detect shocks or drastic changes in the signal; skewness is a measure of signal symmetry, reflecting whether the signal tends to be higher or lower values, while kurtosis evaluates the degree of sharp peaks in the signal, indicating the kurtosis characteristics of the signal; 7) The feature preprocessing unit processes the feature data obtained from the feature extraction module to adapt to different machine learning algorithms; 8) When data enters the fault detector, multiple basic models first classify the input data and output corresponding prediction results; the probability values ​​output by each basic model are collected to form a new feature set, called "meta-features" or "secondary features"; The meta-model then integrates the classification results of the basic models: assuming there are N basic models, each of which predicts M samples in the training set, then each sample will generate N prediction results, namely meta-features; these meta-features will form a new feature matrix Z: where z ij is the predicted probability of the i-th sample under the j-th basic model, and M is the number of basic models; As input to train the meta-model, the meta-model selects logistic regression, linear regression or random forest to combine the prediction results of the base models; 9). While diagnosing faults, the adaptive learning unit is used to dynamically adjust and update the machine learning model to adapt to the real-time changing operating environment and data characteristics. The unit has the function of online learning, can accept new training data in real time, and adjust the parameters and structure of the model according to these data, so that the model can adapt to the changing spindle operating status; online learning means that the model can gradually learn the incoming data without retraining the entire model from scratch. Online learning can use the perceptron algorithm, stochastic gradient descent, and online SVM algorithm to achieve real-time update of model parameters; The adaptive learning unit can regularly update the model based on the latest data and fault conditions collected during the diagnosis process to ensure continuous improvement in diagnostic accuracy. Through the data feedback mechanism, the system will compare and analyze the diagnosis results with the actual fault conditions. If the model is found to have misjudged or missed a fault, the adaptive learning unit will optimize the model based on the feedback information. 10). After the fault diagnosis is completed, based on the output results of the classifier and historical data, the possible fault types and development trends of the spindle are predicted, providing a basis for preventive maintenance; through trend analysis of historical diagnostic data and current health status, the fault predictor can identify the change trend of the spindle status; based on the trend analysis results, the fault predictor can warn of possible faults in advance and remind operators to take preventive maintenance measures; 11). Output diagnostic results and send out alarm signals when abnormalities are detected.

8. The spindle bending common feature diagnosis system based on stacking integration method according to claim 7 is characterized in that: In 6), the mean, variance, peak, and skewness in the time domain features are specifically: Where N is the number of samples; s[i] is the signal value of the i-th sample point; max(s[i]) means finding the maximum value among all sample points; σ is the standard deviation of the signal, defined as The frequency domain features are as follows: The frequency domain feature is to convert the time domain signal to the frequency domain through fast Fourier transform to obtain the spectrum S[k]: Mainly focus on the characteristics of frequency components: the main frequency is the frequency component with the largest amplitude in the spectrum, which is related to the rotation speed or fault frequency and can reveal the operating status of the mechanical equipment: The corresponding frequencies are: The harmonic component is the frequency component related to the main frequency. The harmonic frequency in the analysis spectrum is: f n =n·f base n: represents the order or multiple of the harmonic; n=1 represents the fundamental frequency, n=2 represents the second harmonic, and so on; the harmonic frequency is an integer multiple of the fundamental frequency; Frequency band energy indicates the vibration intensity within a specific frequency range, which can effectively indicate the energy concentration of a specific fault frequency band: Calculate the energy within a specific frequency range [k1, k2]; The time-frequency domain features are as follows: The time-frequency domain features are extracted through wavelet transform or short-time Fourier transform, which can reflect the local characteristics of the signal in time and frequency; Wavelet coefficients are obtained through wavelet transform, which can reveal the characteristics of signals at different scales. They are suitable for capturing the characteristics of transient events or non-stationary signals. Discrete wavelet transform is used to decompose the signal: where c l and d l are the approximation coefficient and detail coefficient, ψ l (n) and φ l (n) is the wavelet function; The instantaneous frequency describes the change of signal frequency over time. By analyzing the time-frequency domain characteristics, we can have a more comprehensive understanding of the characteristics of complex signals.

9. The spindle bending common feature diagnosis system based on stacking integration method according to claim 7 is characterized in that: In 7), the following methods are used to process various machine learning algorithms: a). Since decision trees and random forest models are sensitive to noise and irrelevant features, feature selection is required; b). For LSTM and RNN time series models, smoothing and exponentially weighted moving average are required: EWMA t =αx t +(2-a)EWMA t-1 α is the smoothing coefficient c) Perform logarithmic transformation or Box-Cox transformation on linear regression, logistic regression, and support vector machine: Logarithmic transformation: x′=log(x+∈), ∈ is to prevent the problem of zero logarithm; Box-Cox transformation: λ: Parameter in Box-Cox transformation, controlling the type of transformation; when λ=0, natural logarithmic transformation is used, which is suitable for cases where the data has positive skewness; when λ≠0, power transformation is used, and the transformation form is x λ , and the results were scaled; d) Perform polynomial feature generation: Enhance the model's adaptability to complex data by generating square, cube or feature interaction terms of features; e). Oversampling: Use SMOTE to balance the data by synthesizing new minority class samples; Undersampling: Reduce majority class samples to make the data set more balanced.

10. The spindle bending common feature diagnosis system based on stacking integration method according to claim 7, characterized in that: 8) Metamodel selection logistic regression: The logistic regression metamodel learns how to best combine the predictions of the base models by fitting the relationship between the meta-features and the target variable; a weight is assigned to the meta-feature output by each base model, and the final prediction result is combined through these weights: The model form is: β0: represents the intercept term (bias), which is the constant term in the model, which controls the baseline probability when all feature values ​​are 0; β j : represents feature z j The coefficients or weights of each β j Weighing the feature z j The magnitude of the effect on outcome y; The minimization loss function is: y i : represents the actual label (0 or 1) of the i-th sample; P(y i ∣Z i ): indicates that the logistic regression model predicts that sample i belongs to the positive class (y i =1) probability; The machine learning model library includes: support vector machine, K nearest neighbor, neural network basic training model for diagnosis use. The model works as follows: The support vector machine separates samples of different categories in the input space by finding an optimal classification hyperplane W·x+b=0, where w is the weight vector, x is the input feature vector, and b is the bias term; the goal of the support vector machine is to maximize the classification interval, and the optimization objective function is At the same time, the constraints y i (w·x i +b)≥1, for all samples i where y i is the category label of the sample; in nonlinear classification problems, the support vector machine uses the kernel function K(x i ,x j ) Map the input features to a high-dimensional space so that the samples are linearly separable in the space; The K nearest neighbor classification algorithm calculates the test sample x test For each sample x in the training set i The Euclidean distance between Select the K samples with the closest distance {x (1) ,x (2) ,...,x (K) } as neighbors, and make voting decisions based on the categories of these neighbors y(1), y(2), …, y(K), and finally classify the test sample as y^ test Set to the category that appears most frequently among these neighbors, i.e. y^ test =mode(y(1),y(2),…,y(K)); Neural network: The neural network calculates the activation value a of each layer of neurons through forward propagation (l) =σ(w (l) ·a (1-1) +b (1) ), where a (1) is the activation value of layer l, w (1) and b (l) are the weight and bias of the layer respectively, and the σ activation function is ReLU or Sigmoid; through the back propagation algorithm, the neural network calculates the partial derivative of the loss function L(y,y^) with respect to each parameter, and updates the weight and bias layer by layer Where η is the learning rate, and ultimately through iterative optimization, the network can learn the complex nonlinear mapping relationship between input features and output targets.

Images