Method and system for machine tool spindle fault diagnosis and prognosis

Abstract

The application relates to a predictive maintenance technology and discloses a machine tool spindle fault diagnosis and prediction method and system, which comprises the following steps: collecting spindle vibration data as a steady-state sample; decomposing the vibration data into high, medium and low frequency bands through wavelet packet transformation, adopting corresponding processing methods for different frequency bands, realizing phase consistency signal reconstruction, aligning the instruction domain, and calculating the spindle characteristic value; calculating the maximum value of the characteristic value of each instruction domain positive sample as a threshold value, comparing the instruction domain samples, and judging the spindle health state; using a sliding window to calculate the statistics of the vibration signal characteristic sequence and integral data; using the pretreatment data to train a time convolution network model; and using the trained model to perform fault prediction. The application reduces the characteristic value extraction error through multi-band physical constraint dynamic filtering, reduces the reasoning delay by adopting a TCN parallel computing architecture, and can realize high-precision diagnosis and real-time prediction only by using a single vibration sensor, and the device is simple and economical.

Description

Technical Field

[0001] This application relates to the field of mechanical equipment fault diagnosis, and in particular to predictive maintenance technology for machine tool spindles. Background Technology

[0002] As the core component of a machining center, the spindle's performance has a decisive impact on the overall machining accuracy and productivity of the machine tool. During machine tool operation, factors such as abnormal mechanical structure and failure of the cooling and lubrication system can lead to abnormal spindle conditions. Once the spindle experiences radial runout or axial movement, it will directly cause the workpiece's roundness and coaxiality to exceed tolerances, significantly deteriorating the quality of the machined surface, and in severe cases, even causing equipment damage and personnel injury. Therefore, real-time monitoring and fault prediction of machine tool spindle conditions are of significant practical importance.

[0003] Currently, spindle fault diagnosis technologies are mainly divided into two categories: one is diagnostic analysis technology based on spindle motor current signals; the other is to collect multiple signals such as machine tool current, speed, and temperature, and use artificial intelligence technology for fusion learning to achieve intelligent identification of the spindle status. However, these existing technologies all have significant shortcomings. Diagnostic analysis methods based on current signals have limited accuracy in acquiring signals, making it difficult to accurately predict the spindle status; while AI-based full-process diagnostic methods require large data volumes and high processor performance, resulting in high implementation costs and hindering widespread application in industrial settings.

[0004] Therefore, achieving rapid and low-cost spindle fault diagnosis and prediction while ensuring accurate spindle fault prediction has become an urgent technical problem to be solved. Specifically, the key technical challenges to be addressed include: how to improve the accuracy of spindle vibration signal processing to obtain more accurate characteristic values; how to achieve real-time fault prediction under limited computing resources; and how to simplify the system structure, reduce implementation costs, and make the technical solution more economical and practical. Summary of the Invention

[0005] The purpose of this application is to provide a method and system for diagnosing and predicting machine tool spindle faults, so as to solve the problems mentioned in the background art.

[0006] This application discloses a method for diagnosing and predicting machine tool spindle faults, including the following steps:

[0007] During machine tool operation, spindle vibration data collected by vibration sensors during the cutting process is used as a steady-state sample.

[0008] The vibration data is decomposed into three sub-bands—high frequency, mid frequency, and low frequency—using wavelet packet transform. Ensemble Empirical Mode Decomposition (EEMD) is used to separate the residual DC component of the low-frequency sub-band. Variational Mode Decomposition (VMD) is used to extract the impact component from the high-frequency sub-band signal, combined with adaptive threshold wavelet denoising. A dynamic parameter Butterworth filter is used for the mid-frequency sub-band signal, dynamically modulating the cutoff frequency based on the instantaneous frequency. A physically constrained Kalman filter is introduced for the low-frequency sub-band signal. The state equation is constructed using the integral relationship between acceleration, velocity, and displacement. Finally, the weight coefficients of each frequency band are dynamically adjusted based on a residual feedback mechanism to achieve phase-consistent signal reconstruction, align the command domain, and calculate the principal axis eigenvalues.

[0009] The maximum value of each feature value of the positive sample in each instruction domain is calculated as the corresponding threshold. The actual instruction domain data and the theoretical instruction domain data are compared to form the instruction domain samples. The health status of the machine tool spindle is judged based on the comparison results and the corresponding thresholds, and then displayed visually.

[0010] The mean and standard deviation of the vibration signal feature sequence obtained after processing and the velocity and displacement data obtained after integration are calculated using a sliding window, transient noise is removed, and normalization is performed to obtain preprocessed data.

[0011] The preprocessed data was used to train the Temporal Convolutional Network (TCN) model.

[0012] The trained Temporal Convolutional Network (TCN) model is used to predict faults in the machine tool spindle, and the prediction results are displayed on a graphical interface.

[0013] In a preferred embodiment, the characteristic values ​​include peak-to-peak acceleration, effective velocity, peak-to-peak displacement, and characteristic amplitudes of the bearing outer ring, inner ring, cage, and rollers.

[0014] In a preferred embodiment, the threshold values ​​for the characteristic amplitudes of the bearing outer ring, inner ring, cage, and roller are all -40dB.

[0015] In a preferred embodiment, the spindle vibration data is characterized by being collected as a steady-state sample by a vibration sensor during machine tool operation using an edge module.

[0016] In a preferred embodiment, the step of decomposing the vibration data into three sub-bands—high frequency, mid frequency, and low frequency—through wavelet packet transform employs an improved wavelet packet transform.

[0017] In a preferred embodiment, the Temporal Convolutional Network (TCN) model comprises the first four TCN layers as shared feature extractors, employing causal dilated convolution to capture common temporal features of multiple physical quantities, and three independent branches for acceleration, velocity, and displacement prediction, each branch containing two TCN residual blocks and a fully connected layer.

[0018] In a preferred embodiment, the step of using the trained Temporal Convolutional Network (TCN) model to predict the fault of the machine tool spindle and obtaining the prediction result is characterized in that the prediction result is the predicted value for the next three steps.

[0019] In a preferred embodiment, the vibration data is characterized by the following: before decomposing the vibration data into three sub-bands (high frequency, mid frequency, and low frequency), the vibration acceleration signal is converted into a frequency domain signal using a Fourier transform, and frequency domain integration is performed sequentially to obtain velocity and position signal data. The integration calculation formula is as follows:

[0020]

[0021] Where V(w) is the frequency domain representation of velocity, X(w) is the frequency domain representation of displacement, A(w) is the frequency domain representation of acceleration, w is the angular frequency, and j is the imaginary unit.

[0022] In a preferred embodiment, the cutoff frequency fc(t) of the Butterworth filter with dynamic parameters is calculated using the following formula:

[0023] f c (t)=α·max(f inst (t), 50Hz) Formula 3

[0024] Where α is the bandwidth expansion coefficient, and finst(t) is the instantaneous frequency extracted by the Hilbert transform.

[0025] In a preferred embodiment, the state equation of the physically constrained Kalman filter is:

[0026]

[0027] Where x k+1 Let x be the displacement at time k+1. k Let v be the displacement at time k. k Let a be the velocity at time k. k Let w be the acceleration at time k, Δt be the time step, and w be the acceleration at time k. k For process noise, z k These are the observed values.

[0028] In a preferred embodiment, the characteristic is that the calculation formula for the weighting coefficient Wi of each frequency band in the residual feedback mechanism is:

[0029]

[0030] Where εi 2 For frequency band residuals, σi 2 This represents the noise standard deviation.

[0031] In a preferred embodiment, the normalization process is characterized by the following formula:

[0032]

[0033] Where x norm σ represents the normalized data, x represents the data before normalization, u represents the global mean, and σ represents the standard deviation.

[0034] In a preferred embodiment, the characteristic is that: in step 300, the comparison results of each feature value and the corresponding threshold are combined according to the following rules to obtain the health status of the machine tool spindle: when none of the feature values ​​exceed the corresponding threshold, the spindle health status is normal; when any feature value exceeds the corresponding threshold, the spindle health status is abnormal.

[0035] In a preferred embodiment, the preprocessed data input to the temporal convolutional network model TCN in step 500 includes vibration signal feature sequences, velocity data, and displacement data.

[0036] In a preferred embodiment, the prediction result in step 600 is displayed on a graphical interface in the form of an instruction domain sample comparison chart.

[0037] In a preferred embodiment, the feature is that in step 100, the machining G-code is designed to make the spindle reach the rated speed of 4000 r / s, the CNC system runs the G-code, and the vibration sensor data is collected as a steady-state sample.

[0038] This application also discloses a system for diagnosing and predicting machine tool spindle faults, including:

[0039] The acquisition module is used to collect spindle vibration data during the cutting process as steady-state samples based on vibration sensors during machine tool operation;

[0040] The signal processing module decomposes the vibration data into three sub-bands—high frequency, mid frequency, and low frequency—using wavelet packet transform. It then employs ensemble empirical mode decomposition (EEMD) to separate the residual DC component of the low-frequency sub-band. For the high-frequency sub-band signal, variational mode decomposition (VMD) is used to extract the impact component, combined with adaptive threshold wavelet denoising. For the mid-frequency sub-band signal, a dynamic parameter Butterworth filter is applied, dynamically modulating the cutoff frequency based on the instantaneous frequency. For the low-frequency sub-band signal, a physical constraint Kalman filter is introduced, and a state equation is constructed using the integral relationship between acceleration, velocity, and displacement. Finally, based on a residual feedback mechanism, the weight coefficients of each frequency band are dynamically adjusted to achieve phase-consistent signal reconstruction, align the command domain, and calculate the principal axis eigenvalues.

[0041] The diagnostic module is used to calculate the maximum value of each feature value of the positive samples of each instruction domain as the corresponding threshold, compare the actual instruction domain data with the theoretical instruction domain data to make instruction domain samples, judge the health status of the machine tool spindle based on the comparison results and the corresponding threshold, and display the results visually.

[0042] The data preprocessing module is used to calculate the mean and standard deviation of the vibration signal feature sequence obtained after processing by the signal processing module and the velocity and displacement data obtained after integration using a sliding window, remove transient noise, and perform normalization processing to obtain preprocessed data.

[0043] The model training module is used to train the Temporal Convolutional Network (TCN) model using the preprocessed data.

[0044] The prediction module is used to predict the faults of the machine tool spindle using the trained Temporal Convolutional Network (TCN) model, obtain the prediction results, and display the prediction results on a graphical interface.

[0045] Compared with the prior art, the machine tool spindle fault diagnosis and prediction method and system provided in this application have the following advantages:

[0046] First, this application implements a command domain sample comparison function. The system performs multi-band physical constraint dynamic filtering on the spindle vibration data collected by the vibration sensor to obtain a feature sequence, and compares and analyzes the actual command domain data with the theoretical command domain data to achieve accurate spindle health status judgment. Each feature value (including peak-to-peak acceleration, effective velocity, peak-to-peak displacement, and bearing characteristic frequency amplitude, etc.) has its own independent threshold judgment standard, which improves the accuracy of diagnosis.

[0047] Secondly, this application significantly reduces the eigenvalue extraction error by using a multi-band physical constraint dynamic filtering method. Specifically, this method decomposes the vibration signal into three sub-bands: high frequency, mid frequency, and low frequency, and employs different processing methods accordingly: VMD is used to extract the impact component for the high-frequency signal, a dynamic parameter Butterworth filter is used for the mid-frequency signal, and a physical constraint Kalman filter is introduced for the low-frequency signal. By dynamically adjusting the weight coefficient Wi of each frequency band through a residual feedback mechanism, phase-consistent signal reconstruction is ultimately achieved, improving the spindle fault identification rate.

[0048] Furthermore, this application employs a parallel computing architecture of the Temporal Convolutional Network (TCN) model, significantly reducing inference latency. The first four TCN layers of this model serve as shared feature extractors, using causal dilated convolutions to capture common temporal features of multiple physical quantities. Three independent branches are designed for acceleration, velocity, and displacement prediction. Each branch contains two TCN residual blocks and a fully connected layer, enabling prediction of the health status of the main axis over the next three steps and supporting real-time monitoring.

[0049] Finally, this application only requires a single vibration sensor and an edge module (EM) to complete spindle condition monitoring, resulting in a simple and economical device structure. Real-time data acquisition via the edge module (EM), combined with efficient data processing algorithms and predictive models, ensures system reliability while reducing hardware costs.

[0050] The achievement of the above-mentioned technical effects enables this application to achieve rapid and low-cost spindle fault diagnosis and prediction while ensuring accurate prediction of spindle faults, thus solving the problems existing in the prior art.

[0051] The specification of this application contains numerous technical features distributed across various technical solutions. Listing all possible combinations of these technical features (i.e., technical solutions) would make the specification excessively lengthy. To avoid this problem, the various technical features disclosed in the above-described invention, the various technical features disclosed in the following embodiments and examples, and the various technical features disclosed in the accompanying drawings can be freely combined to form various new technical solutions (all of which are considered to have been described in this specification), unless such a combination of technical features is technically infeasible. For example, one example discloses feature A+B+C, and another example discloses feature A+B+D+E. Features C and D are equivalent technical means that serve the same function, and technically only one needs to be used; they cannot be used simultaneously. Feature E can technically be combined with feature C. Therefore, the solution A+B+C+D should not be considered as described because it is technically infeasible, while the solution A+B+C+E should be considered as described. Attached Figure Description

[0052] Figure 1 This is a flowchart illustrating a method for diagnosing and predicting machine tool spindle faults according to an embodiment of this application.

[0053] Figure 2 This is a schematic diagram illustrating the implementation principle of a machine tool spindle fault diagnosis and prediction method according to an embodiment of this application.

[0054] Figure 3 This is a schematic diagram illustrating the implementation principle of the machine tool spindle health status prediction method according to an embodiment of this application.

[0055] Figure 4 This is an overall flowchart of a method for diagnosing and predicting machine tool spindle faults according to an embodiment of this application.

[0056] Figure 5 This is a comparison diagram of machine tool spindle acceleration command domain samples in the method for diagnosing and predicting machine tool spindle faults according to an embodiment of this application.

[0057] Figure 6 This is a schematic diagram of the structure of a machine tool spindle fault diagnosis and prediction system according to an embodiment of this application. Detailed Implementation

[0058] In the following description, many technical details are presented to help the reader better understand this application. However, those skilled in the art will understand that the technical solutions claimed in this application can be implemented even without these technical details and various variations and modifications based on the following embodiments.

[0059] Explanation of some concepts:

[0060] Instruction domain: In a machine tool control system, this refers to the mathematical space comprised of the ideal machining trajectory or motion state defined by CNC commands. Unlike the time domain and frequency domain, the instruction domain represents the state that the machine tool should theoretically execute, and is used for comparison and analysis with actual measurement data.

[0061] Wavelet packet transform: a time-frequency analysis method, an extension of wavelet transform, capable of more refined signal decomposition and providing more flexible time-frequency resolution, particularly suitable for analyzing non-stationary signals. This application employs an improved wavelet packet transform to decompose the vibration signal into three sub-bands: high frequency, mid frequency, and low frequency.

[0062] Ensemble Empirical Mode Decomposition (EEMD): A signal decomposition method, an improved version of Empirical Mode Decomposition (EMD), which overcomes the mode mixing problem by repeatedly adding white noise to the original signal and averaging the results. In this application, it is used to separate the residual DC component in the low-frequency sub-band.

[0063] Variational Mode Decomposition (VMD): A non-recursive adaptive signal processing method capable of decomposing complex signals into a set of mode functions with specific sparse properties. In this application, it is used to extract impulse components from high-frequency sub-band signals.

[0064] Butterworth filter: A filter designed to produce the flattest possible passband frequency response. The dynamic parameter Butterworth filter in this application is used to process intermediate frequency sub-band signals by dynamically modulating the cutoff frequency according to the instantaneous frequency.

[0065] Kalman filtering: a recursive state estimation algorithm capable of estimating unknown variables from noisy measurements. The physical constraint Kalman filter in this application constructs state equations using the integral relationship between acceleration, velocity, and displacement, and is used to process low-frequency sub-band signals.

[0066] Temporal Convolutional Networks (TCNs): A deep learning architecture specifically designed for sequence modeling, combining causal convolution and dilated convolution to effectively capture long-term dependencies in time series data. In this application, it is used for spindle failure prediction.

[0067] Causal dilated convolution: a special convolution operation that expands the receptive field without increasing the number of parameters by introducing holes (skipping certain input values) into standard convolution, while maintaining causality (the current output depends only on past inputs). This technique is used in the TCN model of this application to capture the common temporal features of multiple physical quantities.

[0068] Residual feedback mechanism: A method for dynamically adjusting system parameters by calculating the difference (residual) between the actual output and the expected output and feeding it back into the system to adjust the parameters. In this application, it is used to dynamically adjust the weighting coefficients of each frequency band to achieve phase-consistent signal reconstruction.

[0069] Spindle characteristic values: key indicators reflecting the operating status of machine tool spindles. In this application, these include peak-to-peak acceleration, effective velocity, peak-to-peak displacement, as well as characteristic amplitudes of bearing outer ring, inner ring, cage, and rollers.

[0070] Edge module: A computing device deployed at the equipment site (edge) capable of processing sensor data in real time, reducing data transmission latency and bandwidth requirements. In this application, it is used to collect vibration data during the spindle cutting process.

[0071] Sliding window: A data processing technique that uses a fixed-size window to slide across time-series data, performing calculations on the data within the window to achieve data smoothing or feature extraction. In this application, it is used to calculate the mean and standard deviation of vibration data and remove transient noise.

[0072] TCN residual blocks are the basic building blocks in temporal convolutional networks, containing dilated convolutions, activation functions, and skip connections. They effectively alleviate the vanishing gradient problem and improve network training efficiency. In this application, each prediction branch contains two layers of TCN residual blocks.

[0073] The following is a brief summary of some of the innovative aspects of this application:

[0074] In summary, this application proposes a composite diagnostic prediction scheme based on multi-band physical constraint dynamic filtering and time series deep learning to address issues such as workpiece roundness and coaxiality deviations caused by radial runout or axial movement during actual machining of the machine tool spindle, a core component. This scheme organically integrates multiple signal processing methods, including improved wavelet packet transform, ensemble empirical mode decomposition (EEMD), variational mode decomposition (VMD), and physically constrained Kalman filtering, to achieve accurate decomposition and reconstruction of vibration signals across different frequency bands. In particular, this application solves the technical challenge of accurately extracting spindle feature values ​​using traditional frequency domain analysis methods by introducing a state equation based on the acceleration-velocity-displacement integral relationship for Kalman filtering in the low-frequency sub-bands and dynamically adjusting the weight coefficients of each frequency band using a residual feedback mechanism. Furthermore, this application creatively combines an instruction domain sample comparison mechanism with the hierarchical prediction architecture of a temporal convolutional network (TCN). By setting multiple independent prediction branches to model multiple physical quantities such as acceleration, velocity, and displacement, it ensures prediction accuracy while reducing computational resource requirements. This multi-layered and multi-dimensional technical feature works in synergy to overcome the limitations of existing technologies that rely on a single signal source or require a large amount of data support, thus achieving high-efficiency and high-precision spindle fault diagnosis and prediction.

[0075] To make the objectives, technical solutions, and advantages of this application clearer, the embodiments of this application will be described in further detail below with reference to the accompanying drawings.

[0076] The first embodiment of this application relates to a method for diagnosing and predicting machine tool spindle faults, the process of which is as follows: Figure 1 As shown, the method includes the following steps:

[0077] Step 100: During machine tool operation, spindle vibration data during the cutting process is collected based on vibration sensors as steady-state samples.

[0078] Step 200: The vibration data is decomposed into three sub-bands—high frequency, mid frequency, and low frequency—using wavelet packet transform. The residual DC component of the low-frequency sub-band is separated using ensemble empirical mode decomposition (EEMD). The impact component is extracted from the high-frequency sub-band signal using variational mode decomposition (VMD) and combined with adaptive threshold wavelet denoising. The mid-frequency sub-band signal is processed using a dynamic parameter Butterworth filter, and the cutoff frequency is dynamically modulated based on the instantaneous frequency. The low-frequency sub-band signal is processed using a physical constraint Kalman filter. The state equation is constructed using the integral relationship between acceleration, velocity, and displacement. Finally, the weight coefficients of each frequency band are dynamically adjusted based on the residual feedback mechanism to achieve phase-consistent signal reconstruction, align the command domain, and calculate the principal axis eigenvalues.

[0079] Step 300: Calculate the maximum value of each feature value of the positive samples in each instruction domain as the corresponding threshold. Compare the actual instruction domain data with the theoretical instruction domain data to determine the health status of the machine tool spindle based on the comparison results and the corresponding thresholds, and display the results visually.

[0080] Step 400: The mean and standard deviation of the vibration signal feature sequence obtained in step 200 and the velocity and displacement data obtained after integration are calculated using a sliding window, transient noise is removed, and normalization is performed to obtain preprocessed data.

[0081] Step 500: Train the Temporal Convolutional Network (TCN) model using the preprocessed data.

[0082] Step 600: Use the trained Temporal Convolutional Network (TCN) model to predict faults in the machine tool spindle, obtain the prediction results, and display the prediction results on a graphical interface.

[0083] Furthermore, the machine tool spindle fault diagnosis and prediction method proposed in this embodiment first collects spindle operating data through vibration sensors, then processes the vibration signal using a multi-band physical constraint dynamic filtering method, next determines the spindle health status through command domain sample comparison, and finally predicts the spindle health status using a temporal convolutional network model. This method achieves both high-precision diagnosis and real-time prediction, while also being simple and economical in its equipment, effectively solving the problem that existing spindle fault diagnosis methods cannot simultaneously achieve high efficiency, high precision, and economical equipment. Figure 2 and Figure 3 The following diagrams illustrate the implementation principles of real-time fault diagnosis and health status prediction for machine tool spindles as described in this embodiment. Figure 4 This is a flowchart of the method described in this embodiment.

[0084] Optionally, the characteristic values ​​include peak-to-peak acceleration, effective velocity, peak-to-peak displacement, and characteristic amplitudes of the bearing outer ring, inner ring, cage, and rollers.

[0085] Specifically, in this embodiment, the spindle characteristic values ​​are key indicators for measuring the spindle's operating status. The peak-to-peak acceleration reflects the magnitude range of spindle vibration intensity, the effective velocity value (also known as the root mean square value) characterizes the vibration energy level, and the peak-to-peak displacement directly reflects the maximum displacement of the spindle relative to the reference position. As a critical component of the spindle, the bearing's characteristic amplitude is an important indicator for judging its health status. This includes the characteristic amplitude of the outer ring (reflecting outer ring damage), the characteristic amplitude of the inner ring (reflecting inner ring damage), the characteristic amplitude of the cage (reflecting cage damage), and the characteristic amplitude of the rollers (reflecting roller damage). Comprehensive analysis of these characteristic values ​​can fully reflect the operating status of the spindle system, providing a reliable basis for fault diagnosis.

[0086] Optionally, the threshold values ​​for the characteristic amplitudes of the bearing outer ring, inner ring, cage, and roller are all -40dB.

[0087] Specifically, based on experimental statistics and engineering experience, this embodiment sets the threshold for the characteristic amplitude of each bearing component to -40 dB. This threshold is the logarithmic ratio of the amplitude relative to the reference signal in frequency domain analysis. When the characteristic amplitude of each bearing component exceeds -40 dB, it indicates that the corresponding component may be abnormal or damaged. -40 dB was chosen as the threshold because this value can effectively identify early bearing faults while avoiding excessive false alarms, and it has good applicability in actual working conditions. The characteristic amplitude of each bearing component is obtained by envelope demodulation and spectral analysis of the vibration signal, and its characteristic frequency is related to bearing geometric parameters, rotational speed, and other factors.

[0088] Optionally, in step 100, an edge module is used to collect the spindle vibration data.

[0089] Specifically, the edge module used in this embodiment is a data acquisition and preliminary processing unit installed locally on the machine tool. It can acquire and preprocess data from vibration sensors in real time without transmitting all data to a remote server. Using an edge module to acquire vibration data has the following advantages: First, it reduces network bandwidth pressure by uploading only the processed key feature data; second, it reduces data transmission latency and improves system response speed; finally, the edge module can continue to operate when the equipment is shut down or the network is interrupted, ensuring the continuity and reliability of data acquisition. In practical applications, edge modules typically integrate high-speed A / D converters, signal conditioning circuits, and embedded processors, with sampling frequencies generally reaching 50kHz or higher, meeting the requirements for high-frequency vibration signal acquisition.

[0090] Optionally, step 200 uses an improved wavelet packet transform.

[0091] Specifically, the improved wavelet packet transform in this embodiment is an optimization based on the traditional wavelet packet transform, mainly reflected in two aspects: First, the selection of basis functions is optimized. Based on the characteristics of the spindle vibration signal, basis functions with good time-frequency localization characteristics, such as Daubechies wavelets or Symlet wavelets, are selected. Second, an adaptive decomposition depth strategy is adopted. The number of decomposition layers in different frequency bands is dynamically adjusted according to the signal spectrum characteristics, performing more detailed decomposition of the high-frequency region and appropriately merging the low-frequency region, thereby improving the ability to identify different motion modes of the spindle. This improved wavelet packet transform can more accurately decompose vibration data into three sub-bands: high frequency (above 1kHz), mid frequency (20Hz-1kHz), and low frequency (below 20Hz). Each sub-band corresponds to different physical phenomena in the spindle system, such as high frequency corresponding to impact response, mid frequency corresponding to bearing characteristic frequency, and low frequency corresponding to structural resonance and rotation frequency.

[0092] Optionally, the Temporal Convolutional Network (TCN) model includes the first four TCN layers as shared feature extractors, employing causal dilated convolutions to capture common temporal features of multiple physical quantities, and three independent branches for acceleration, velocity, and displacement prediction. Each branch contains two TCN residual blocks and a fully connected layer.

[0093] Specifically, the Temporal Convolutional Network (TCN) model in this embodiment adopts a specific architectural design, such as... Figure 3 As shown, the first four layers of the TCN use causal dilated convolutions as a shared feature extractor. This convolutional method effectively captures temporal dependencies in long-sequence data by expanding the receptive field without increasing the number of parameters. Causal dilated convolutions ensure that the model only uses current and past information during prediction, avoiding information leakage. After the shared feature extractor, three independent prediction branches are formed, corresponding to acceleration, velocity, and displacement, respectively. Each branch contains two TCN residual blocks and a final fully connected layer. The advantages of this design are: firstly, the shared feature extractor can learn the common features and interrelationships between multiple physical quantities; secondly, the independent branches can specifically model the characteristics of different physical quantities. The TCN residual blocks introduce skip connections, which alleviates the gradient vanishing problem during network training and improves the model's convergence and generalization ability.

[0094] Optionally, the prediction result in step 600 is the prediction value for the next 3 steps.

[0095] Specifically, the TCN model in this embodiment predicts values ​​for the next three time steps. In practical applications, this typically corresponds to the spindle state over the next 30ms to 300ms, depending on the sampling frequency. The selection of three prediction steps is based on the following considerations: First, too few steps may not provide sufficient reaction time for preventative maintenance; second, too many steps increase prediction uncertainty and reduce accuracy; finally, three-step prediction achieves a good balance between computational cost and prediction effectiveness. The prediction results include acceleration, velocity, and displacement data for the next three steps, displayed through a graphical interface, allowing operators to intuitively understand the spindle's state changes and promptly identify potential problems. The system typically combines the current health status with the predicted values ​​to comprehensively assess the spindle's future failure risk, providing support for maintenance decisions.

[0096] Optionally, in step 200, the vibration acceleration signal is converted into a frequency domain signal through Fourier transform, and the velocity and position signal data are obtained by frequency domain integration. The integration calculation formula is as follows:

[0097]

[0098] Where V(w) is the frequency domain representation of velocity, X(w) is the frequency domain representation of displacement, A(w) is the frequency domain representation of acceleration, w is the angular frequency, and j is the imaginary unit.

[0099] Specifically, in this embodiment, the time-domain vibration acceleration signal is converted to the frequency domain using Fourier transform, and then the integral relationship is used for calculation in the frequency domain. This method avoids the cumulative error problem in time-domain integration. Equation 1 represents the integral relationship from acceleration to velocity, that is, in the frequency domain, the velocity signal V(w) is equal to the acceleration signal A(w) divided by jw (where j is the imaginary unit and w is the angular frequency). Equation 2 represents the integral relationship from velocity to displacement, or directly the second-order integral relationship from acceleration to displacement, that is, the displacement signal X(w) is equal to the acceleration signal A(w) divided by (jw)². It should be noted that when w = 0 (i.e., the DC component), the above formulas are not applicable and require special handling. Usually, the DC component is set to zero because the actual vibration signal should not contain static displacement. After frequency domain integration, the signal is converted back to the time domain using inverse Fourier transform to obtain the time-domain representation of velocity and displacement. This frequency domain integration method is particularly suitable for processing periodic or quasi-periodic vibration signals and can effectively suppress the low-frequency drift problem during the integration process.

[0100] Optionally, the cutoff frequency fc(t) of the Butterworth filter with dynamic parameters is calculated using the following formula:

[0101] f c (t)=α·max(f inst (t), 50Hz) Formula 3

[0102] Where α is the bandwidth expansion coefficient, and finst(t) is the instantaneous frequency extracted by the Hilbert transform.

[0103] Specifically, the dynamic parameter Butterworth filter in this embodiment is an innovative filtering method. Its core feature is that the cutoff frequency fc(t) is not fixed but dynamically adjusted over time. As shown in Equation 3, the cutoff frequency fc(t) is calculated based on the instantaneous frequency finst(t) extracted by the Hilbert transform, with a minimum value of 50Hz set as a lower limit. The bandwidth expansion coefficient α is typically taken between 1.2 and 1.5 to appropriately expand the passband width, ensuring that the effective components of the signal are not filtered out. This method of dynamically adjusting the cutoff frequency has significant advantages: on the one hand, it can adaptively track the frequency changes of the signal, making it particularly suitable for processing non-stationary signals; on the other hand, it effectively avoids the problems of insufficient or excessive filtering that may occur with traditional fixed parameter filters when the signal frequency changes. In spindle vibration signal processing, the main frequency components of the signal may change due to variations in spindle speed and cutting conditions. Using a dynamic parameter Butterworth filter can more accurately retain useful signal components and improve the accuracy of feature extraction.

[0104] Optionally, the state equation for the physical constraint Kalman filter is:

[0105]

[0106] Where x k+1 Let x be the displacement at time k+1. k Let v be the displacement at time k. k Let a be the velocity at time k. k Let w be the acceleration at time k, Δt be the time step, and w be the acceleration at time k. k For process noise, z k These are the observed values.

[0107] Specifically, the physical constraint Kalman filter in this embodiment is an optimization method that incorporates physical motion laws into state estimation, primarily used to handle integral drift problems in low-frequency signals. The state equation in Equation 4 consists of two parts: the first equation describes the state transition process, i.e., the displacement x at time k+1 according to Newton's laws of motion. k+1 Displacement x at time k k Speed ​​v k acceleration a k The equation is derived from kinematic relationships, where Δt is the sampling time interval, 0.5ak(Δt)² represents the displacement change caused by acceleration, and w... kThe first equation represents process noise, indicating modeling error and external disturbances; the second equation describes the observation process, i.e., the observed value z. k It is acceleration a k and velocity v k The combination of these factors, along with traditional Kalman filtering, offers the advantage of physically constrained Kalman filtering: it explicitly incorporates the physical integral relationship between acceleration, velocity, and displacement into the filtering process, ensuring that the estimation results conform to both sensor observations and physical motion laws, effectively suppressing integral drift. This method is particularly suitable for processing low-frequency vibration signals, as these signals are more prone to drift during integration. In spindle health monitoring, accurate displacement information is crucial for assessing spindle radial runout and axial axial movement.

[0108] Optionally, the formula for calculating the weighting coefficient Wi of each frequency band in the residual feedback mechanism is as follows:

[0109]

[0110] Where ε i 2 For frequency band residuals, σ i 2 This represents the noise standard deviation.

[0111] Specifically, the residual feedback mechanism in this embodiment is an adaptive signal reconstruction method used to optimize the synthesis effect of the multi-band filtered signal. As shown in Equation 5, the frequency band weight coefficient Wi is based on the frequency band residual ε. i 2 and noise standard deviation σ i 2 The frequency band residual ε is obtained through dynamic calculation. i 2 This indicates the degree of difference between the filtered signal and the original signal in the corresponding frequency band, expressed as the noise standard deviation σ. i 2 This represents the intensity estimate of the noise in that frequency band. The formula uses the square root of the sum of squares as the denominator, conforming to the minimum mean square error criterion. The core idea of ​​this weighting calculation method is: when the residual and noise levels of a frequency band are both low, the signal quality of that band is high, and it should be assigned a larger weight; conversely, when the residual or noise levels are high, the signal quality of that band is poor, and it should be assigned a smaller weight. Through this dynamic weight adjustment, the filtering results of each frequency band can be adaptively integrated to achieve phase-consistent signal reconstruction, ensuring that the signal maintains good phase characteristics after different filtering processes, avoiding signal distortion and feature loss. In spindle health monitoring, accurately preserving the phase characteristics of the signal is particularly important for fault diagnosis, because many fault modes exhibit characteristic vibration patterns in specific phases.

[0112] Optionally, the normalization process is calculated using the following formula:

[0113]

[0114] Where x norm σ represents the normalized data, x represents the data before normalization, u represents the global mean, and σ represents the standard deviation.

[0115] Specifically, the normalization process in this embodiment is a standard data preprocessing method used to eliminate the influence of differences in the dimensions and scales of different features. As shown in Equation 6, the normalization process first calculates the mean u and standard deviation σ of the original data x, then subtracts the mean from the original data and divides by the standard deviation to obtain the normalized data x. norm This Z-score normalization method gives the processed data a mean of 0 and a standard deviation of 1, making it suitable for subsequent deep learning model training. Normalization in this method has the following important implications: First, different characteristics of the spindle vibration signal (such as acceleration, velocity, and displacement) differ significantly in units and numerical ranges; normalization can eliminate these differences from negatively impacting model training. Second, normalization can accelerate the convergence process of the deep learning model, improving training efficiency. Finally, normalization improves the model's generalization ability to different working conditions and equipment, enhancing the method's practicality. In practical applications, the normalization parameters (mean and standard deviation) are usually calculated based on the training dataset and reused in subsequent prediction processes to ensure consistency.

[0116] Optionally, in step 300, the comparison results of each feature value with the corresponding threshold are combined according to the following rules to obtain the health status of the machine tool spindle: when none of the feature values ​​exceed the corresponding threshold, the spindle health status is normal; when any feature value exceeds the corresponding threshold, the spindle health status is abnormal.

[0117] Specifically, this embodiment employs a comprehensive judgment rule to assess the health status of the machine tool spindle. This rule is based on the "weakest link" principle, meaning that the reliability of a system depends on its weakest link. For the spindle system, any critical characteristic value (such as peak-to-peak acceleration, effective velocity, peak-to-peak displacement, or bearing characteristic amplitude) exceeding the normal range may indicate a potential fault. Therefore, this method uses "AND logic" for comprehensive judgment: the spindle is considered healthy only when all characteristic values ​​are within their respective threshold ranges; if any characteristic value exceeds its threshold, the spindle is judged to be abnormal. This judgment method has high sensitivity, enabling early detection of potential problems and facilitating preventative maintenance. In practical applications, the system further analyzes the specific characteristics exceeding the thresholds to provide a basis for fault type identification. For example, if the bearing outer ring characteristic amplitude exceeds the threshold, it may indicate bearing outer ring damage; if the peak-to-peak displacement exceeds the threshold, it may indicate spindle runout problems, etc. Furthermore, the system records the degree to which each characteristic value exceeds the threshold as a reference for assessing the severity of the fault.

[0118] Optionally, the preprocessed data input to the Temporal Convolutional Network (TCN) model in step 500 includes vibration signal feature sequences, velocity data, and displacement data.

[0119] Specifically, in this embodiment, the preprocessed data input to the TCN model constitutes a multi-dimensional feature space, comprehensively reflecting the dynamic characteristics of the spindle system. The vibration signal feature sequence mainly refers to the acceleration time series obtained after multi-band physical constraint dynamic filtering, which contains high-frequency dynamic information of the spindle vibration; velocity data is obtained through integration, reflecting the energy distribution and mid-frequency dynamic characteristics of the spindle system; displacement data is obtained through quadratic integration, containing low-frequency information about spindle position changes. These three types of data each carry information from different frequency bands, complementing each other and jointly constituting a comprehensive description of the spindle state. Inputting these different types of data into the TCN model simultaneously helps the model learn the changing patterns of the spindle state from multiple perspectives and capture the interrelationships and intrinsic connections between different physical quantities. In model implementation, these three types of data are typically organized into a three-dimensional tensor form [batch size, time step, feature dimension], where the feature dimension contains multiple features of each physical quantity. The normalization operation during preprocessing ensures the balance between different physical quantities, preventing any one type of data from dominating the model training due to its larger value.

[0120] Optionally, the prediction results in step 600 are displayed on the graphical interface as a comparison chart of instruction domain samples.

[0121] Specifically, in this embodiment, the prediction results are displayed in the form of an instruction domain sample comparison chart, such as... Figure 5As shown, the command domain sample comparison chart is a visualization method that compares actual measurement data with theoretical commands or historical normal data on the same coordinate system. In the graphical interface, the horizontal axis usually represents the command number or time series, and the vertical axis represents the corresponding characteristic value, such as acceleration, velocity, or displacement. The chart typically contains two curves: one representing the actual measured or predicted data, and the other representing the data from the theoretical command or historical normal sample. The difference between the two curves intuitively reflects the degree of deviation between the current spindle state and the normal state. In addition, the interface also displays a warning threshold line. When the actual data exceeds the threshold range, the relevant area will be marked in red or other conspicuous colors to remind the operator. This visualization method has the following advantages: First, it is intuitive and clear, making it easy for operators to quickly judge the spindle state; second, the comparative display method helps to identify abnormal patterns and trends; finally, the time-series display format facilitates the observation of problem development trends, providing a time window for preventive maintenance.

[0122] Optionally, in step 100, the machining G-code is designed to make the spindle reach the rated speed of 4000 r / s, the CNC system runs the G-code, and the vibration sensor data is collected as a steady-state sample.

[0123] Specifically, in this embodiment, the spindle operation is controlled by a specially designed G-code to ensure that the spindle is in a stable working state during data acquisition. The rated speed of 4000 r / s (equivalent to 240,000 rpm) was chosen as the test speed because the dynamic characteristics of the spindle are relatively stable at this speed, and it can fully stimulate potential fault characteristics. G-code is the programming language of CNC machine tools, which can precisely control the spindle's start-up, acceleration, constant speed operation, and stopping processes. In actual operation, typical G-code sequences include: G90 (absolute coordinate mode), S4000 (set spindle speed to 4000 r / s), M03 (spindle forward rotation), etc. After the CNC system executes these commands, the spindle will reach the rated speed according to the predefined acceleration curve and maintain stable operation. During this period, the vibration sensor continuously collects data to form a steady-state sample. This standardized testing method ensures the consistency and comparability of the collected data, establishing a reliable foundation for subsequent analysis. It should be noted that the rated spindle speed may vary depending on the actual situation of different machine tools and should be adjusted appropriately according to the parameters of the specific equipment.

[0124] Through the detailed description above of the machine tool spindle fault diagnosis and prediction method, this embodiment achieves high-precision diagnosis and real-time prediction of spindle health status, offering the following advantages compared to existing technologies: First, the multi-band physical constraint dynamic filtering method significantly improves the accuracy of feature extraction; second, the instruction domain sample comparison technology provides an intuitive and reliable basis for health status assessment; third, the parallel computing architecture of the temporal convolutional network effectively reduces inference latency and supports real-time prediction; finally, the entire system requires only a single vibration sensor, making the device simple, economical, and easy to promote and apply in industrial environments. These advantages collectively address the shortcomings of existing spindle fault diagnosis technologies in terms of efficiency, accuracy, and economy, providing a new technical solution for intelligent maintenance in the manufacturing industry.

[0125] To make the technical solution of this application clearer, the preferred embodiments of this application will now be described in detail with reference to the accompanying drawings. However, it should be understood that the described embodiments are merely exemplary and not restrictive.

[0126] Reference Figure 2 This paper illustrates the implementation principle of real-time fault diagnosis for machine tool spindles in this embodiment. As shown in the figure, the machine tool spindle is first driven to run at a preset speed via G-code, and vibration sensors collect vibration data during the spindle cutting process. The collected vibration data undergoes multi-band physical constraint dynamic filtering, and feature values ​​are calculated after alignment with the command domain. Subsequently, the processed feature values ​​are compared with preset thresholds for command domain samples. The spindle health status is determined based on the comparison results, and the diagnostic results are displayed on the graphical interface.

[0127] Reference Figure 3 The figure illustrates the implementation principle of machine tool spindle health status prediction in this embodiment. As shown, the system first performs multi-band physical constraint dynamic filtering on steady-state data samples, then removes transient noise and performs normalization. The processed data is used to train the Temporal Convolutional Network (TCN) model. After training, the model can predict the future health status of the spindle based on the real-time input feature values ​​and display the prediction results on a graphical interface.

[0128] Reference Figure 4 This document illustrates the complete flowchart of the machine tool spindle fault diagnosis and prediction method in this embodiment. The flowchart begins with the execution of G-code and proceeds sequentially through steady-state sample extraction, multi-band physical constraint dynamic filtering, instruction domain alignment, instruction domain sample comparison, health status display, sample data upload, transient noise removal, normalization processing, TCN model training, and finally displays the predicted health status.

[0129] Reference Figure 5This diagram illustrates an example of a comparison chart of machine tool spindle acceleration command domain samples from an embodiment of this application. The horizontal axis represents the command number, and the vertical axis represents the acceleration value. The diagram contains two curves: the blue curve represents the actual collected acceleration data, and the red curve represents the theoretical command domain acceleration data. By comparing the differences between the two curves, the health status of the spindle can be intuitively determined. When the deviation between the actual data and the theoretical data exceeds a preset threshold, it indicates that the spindle may have an abnormality.

[0130] As illustrated and explained above, the machine tool spindle fault diagnosis and prediction method provided in this application can achieve high-precision fault diagnosis and real-time prediction, providing reliable technical support for preventive maintenance of equipment.

[0131] To better understand the technical solution of this application, a specific example is provided below. The details listed in this example are mainly for ease of understanding and are not intended to limit the scope of protection of this application.

[0132] This example provides a spindle fault diagnosis and prediction method based on command domain vibration data analysis and neural network prediction, in order to solve the problem that existing spindle fault diagnosis methods cannot simultaneously achieve high efficiency, high accuracy, and economical equipment.

[0133] The method in this example is based on the real-time acquisition of vibration data in the command domain of the vibration sensor during the machining process by the edge module for machine tool spindle fault diagnosis and prediction. The method designs a multi-band physical constraint dynamic filtering method based on the vibration characteristics of the spindle, performs integral filtering on the acquired data, obtains feature values ​​after aligning the command domain, compares command domain samples to determine the health status of the spindle, and then trains the TCN model for prediction after data preprocessing.

[0134] Specifically, the steps include the following:

[0135] Step S1: Data Acquisition

[0136] During machine tool operation, vibration data from the spindle cutting process is collected using an edge module as a steady-state sample. In specific implementation, the machining G-code is designed to make the spindle reach the rated speed of 4000 r / s. The CNC system runs the G-code, and the vibration sensor data is collected in real time by the edge module as a steady-state sample.

[0137] Step S2: Fault Diagnosis

[0138] Acceleration data from vibration sensors reflecting the spindle motion characteristics are extracted from steady-state samples. These data are then integrated and filtered to obtain velocity and position data. The data are then aligned to the command domain to obtain eigenvalues. The maximum eigenvalue of the positive sample in the command domain is calculated as a threshold, and the spindle's health status is obtained through comparison. This data is then displayed through a graphical interface. Finally, the sample data and its health status are uploaded to the sample database.

[0139] S21: Signal Conversion and Integration

[0140] The vibration acceleration signal is converted into a frequency domain signal using Fourier transform, and the velocity and position signal data are obtained by frequency domain integration. The calculation formula is as follows:

[0141]

[0142] Where V(w) is the frequency domain representation of velocity, X(w) is the frequency domain representation of displacement, A(w) is the frequency domain representation of acceleration, w is the angular frequency, and j is the imaginary unit.

[0143] S22: Multi-band Physically Constrained Dynamic Filtering

[0144] Multi-band physical constraint dynamic filtering of acceleration signal data, velocity signal data, and position signal data consists of the following three steps:

[0145] Wavelet packet-EEMD joint frequency domain decomposition:

[0146] The original signal is decomposed into three sub-bands—high frequency (1kHz), mid frequency (20Hz-1kHz), and low frequency (<20Hz)—using an improved wavelet packet transform.

[0147] For the low-frequency band, ensemble empirical mode decomposition (EEMD) is further used to separate the residual DC component from the effective low-frequency vibration modes.

[0148] Frequency band adaptive filtering:

[0149] For high-frequency signals, variational mode decomposition (VMD) is used to extract impulse components, which are then eliminated by adaptive threshold wavelet denoising.

[0150] For designing a Butterworth filter with dynamic parameters for mid-frequency signals, the cutoff frequency is modulated based on the instantaneous frequency, as shown in the following formula:

[0151] f c (t)=α·max(f inst (t), 50Hz) Formula 3

[0152] Where α is the bandwidth expansion coefficient, and finst(t) is the instantaneous frequency extracted by the Hilbert transform.

[0153] Physically constrained Kalman filtering is introduced for low-frequency signals. The state equation is constructed using the integral relationship between acceleration, velocity, and displacement, as shown in the following formula. The integral drift is then modified in real time through state estimation:

[0154]

[0155] Where x k+1 Let x be the displacement at time k+1. k Let v be the displacement at time k. k Let a be the velocity at time k. k Let w be the acceleration at time k, Δt be the time step, and w be the acceleration at time k. k For process noise, z k These are the observed values.

[0156] Multi-scale signal reconstruction:

[0157] The filtering results of each frequency band are superimposed in a time-frequency synchronization manner to maintain phase consistency.

[0158] Reconstruction errors are detected through a residual feedback mechanism, and the weighting coefficients of each frequency band are dynamically adjusted.

[0159]

[0160] Where εi 2 For frequency band residuals, σi 2 This represents the noise standard deviation.

[0161] S23: Eigenvalue Calculation and Health Status Assessment

[0162] The eigenvalues ​​of the filtered signal are calculated, and the maximum eigenvalue of the positive sample in the instruction domain is used as a threshold. The results are compared, and if the value is within the threshold, the spindle is considered healthy. If the value exceeds the threshold, the spindle is considered abnormal. Finally, the results are displayed in the graphical interface, and the sample data and health status are uploaded to the sample database.

[0163] The threshold values ​​for each feature are as follows:

[0164] Eigenvalues threshold unit Peak-to-peak acceleration 0.2 g RMS speed 1.3 mm / s peak displacement 21 um Bearing outer ring characteristic amplitude -40 dB Characteristic amplitude of bearing inner ring -40 dB Bearing cage characteristic amplitude -40 dB characteristic amplitude of bearing rollers -40 dB

[0165] Step S3: Fault Prediction

[0166] The data after integral filtering is cleaned and then used as input to train the TCN model, which can predict the health status of the spindle and display it through a graphical interface.

[0167] S31: Data Preprocessing

[0168] For the filtered acceleration data, the mean and standard deviation are calculated using a sliding window (10ms) to remove transient noise exceeding ±3δ. Then, the data is divided into input sequences within a fixed time window and normalized according to the following formula:

[0169]

[0170] Where xnorm is the normalized data, x is the data before normalization, u is the global mean, and σ is the standard deviation.

[0171] S32: TCN Model Training and Prediction

[0172] The basic TCN model architecture is adopted. The first four TCN layers serve as shared feature extractors. Causal dilated convolutions are used to capture the common temporal features of multiple physical quantities. After the shared layers, three independent branches (acceleration / velocity / displacement prediction branches) are constructed. Each branch contains two TCN residual blocks + a fully connected layer, outputting the predicted values ​​for the next three steps, which are displayed on a graphical interface.

[0173] System Implementation

[0174] According to another aspect of this example, a machine tool spindle fault diagnosis and prediction system is provided, including a spindle health status evaluation module and a processor. When the spindle health status evaluation module is invoked by the processor, it implements the above-mentioned machine tool spindle fault diagnosis and prediction method.

[0175] Compared with existing technologies, the machine tool spindle fault diagnosis and prediction method and system provided in this example achieves at least the following beneficial effects:

[0176] This example implements instruction domain sample comparison and uploads relevant data and health status to the sample database, accumulating industrial data for fault diagnosis.

[0177] This example uses multi-band physical constraint dynamic filtering to reduce the feature value extraction error and improve the spindle fault identification rate.

[0178] This example effectively reduces inference latency and supports real-time prediction by using the TCN parallel computing architecture.

[0179] This example only requires a single vibration sensor, making the device simple and economical.

[0180] The second embodiment of this application relates to a system for diagnosing and predicting machine tool spindle faults, the structure of which is as follows: Figure 6 As shown, the system for diagnosing and predicting machine tool spindle faults includes:

[0181] The acquisition module is used to collect spindle vibration data during the cutting process as steady-state samples based on vibration sensors during machine tool operation;

[0182] The signal processing module decomposes the vibration data into three sub-bands—high frequency, mid frequency, and low frequency—using wavelet packet transform. It then employs ensemble empirical mode decomposition (EEMD) to separate the residual DC component of the low-frequency sub-band. For the high-frequency sub-band signal, variational mode decomposition (VMD) is used to extract the impact component, combined with adaptive threshold wavelet denoising. For the mid-frequency sub-band signal, a dynamic parameter Butterworth filter is applied, dynamically modulating the cutoff frequency based on the instantaneous frequency. For the low-frequency sub-band signal, a physical constraint Kalman filter is introduced, and a state equation is constructed using the integral relationship between acceleration, velocity, and displacement. Finally, based on a residual feedback mechanism, the weight coefficients of each frequency band are dynamically adjusted to achieve phase-consistent signal reconstruction, align the command domain, and calculate the principal axis eigenvalues.

[0183] The diagnostic module is used to calculate the maximum value of each feature value of the positive samples of each instruction domain as the corresponding threshold, compare the actual instruction domain data with the theoretical instruction domain data to make instruction domain samples, judge the health status of the machine tool spindle based on the comparison results and the corresponding threshold, and display the results visually.

[0184] The data preprocessing module is used to calculate the mean and standard deviation of the vibration signal feature sequence obtained after processing by the signal processing module and the velocity and displacement data obtained after integration using a sliding window, remove transient noise, and perform normalization processing to obtain preprocessed data.

[0185] The model training module is used to train the Temporal Convolutional Network (TCN) model using the preprocessed data.

[0186] The prediction module is used to predict the faults of the machine tool spindle using the trained Temporal Convolutional Network (TCN) model, obtain the prediction results, and display the prediction results on a graphical interface.

[0187] The first embodiment is a method embodiment corresponding to this embodiment. The technical details in the first embodiment can be applied to this embodiment, and the technical details in this embodiment can also be applied to the first embodiment.

[0188] The above embodiments have the following technical effects:

[0189] First, the above embodiments implement the command domain sample comparison function. The system performs multi-band physical constraint dynamic filtering on the spindle vibration data collected by the vibration sensor to obtain a feature sequence, and compares and analyzes the actual command domain data with the theoretical command domain data to achieve accurate spindle health status judgment. Each feature value (including peak-to-peak acceleration, effective velocity, peak-to-peak displacement, and bearing characteristic frequency amplitude, etc.) has its own independent threshold judgment standard, which improves the accuracy of diagnosis.

[0190] Secondly, the above embodiments significantly reduce the feature extraction error by using a multi-band physical constraint dynamic filtering method. Specifically, this method decomposes the vibration signal into three sub-bands: high frequency, mid frequency, and low frequency, and employs different processing methods accordingly: VMD is used to extract the impact component for the high-frequency signal, a dynamic parameter Butterworth filter is used for the mid-frequency signal, and a physical constraint Kalman filter is introduced for the low-frequency signal. By dynamically adjusting the weight coefficient Wi of each frequency band through a residual feedback mechanism, phase-consistent signal reconstruction is ultimately achieved, improving the spindle fault identification rate.

[0191] Furthermore, the above embodiments employ a parallel computing architecture of the Temporal Convolutional Network (TCN) model, significantly reducing inference latency. The first four TCN layers of this model serve as shared feature extractors, using causal dilated convolutions to capture common temporal features of multiple physical quantities. Three independent branches are designed for acceleration, velocity, and displacement prediction. Each branch contains two TCN residual blocks and a fully connected layer, enabling prediction of the health status of the main axis over the next three steps and supporting real-time monitoring.

[0192] Finally, the above embodiments only require a single vibration sensor and an edge module (EM) to complete spindle condition monitoring, resulting in a simple and economical device structure. Real-time data acquisition via the edge module (EM), combined with efficient data processing algorithms and predictive models, ensures system reliability while reducing hardware costs.

[0193] The aforementioned technical effects enable the above embodiments to achieve rapid and low-cost spindle fault diagnosis and prediction while ensuring accurate prediction of spindle faults, thus solving the problems existing in the prior art.

[0194] It should be noted that in this patent application, relational terms such as "first" and "second" are used merely to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitations, an element defined by the phrase "comprising one" does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes said element. In this patent application, if it refers to performing an action according to an element, it means performing the action at least according to that element, including two cases: performing the action only according to that element, and performing the action according to that element and other elements. Expressions such as "multiple," "repeatedly," and "various" include two, two times, two kinds, and more than two, more than two times, and more than two kinds.

[0195] All documents mentioned in this application are considered to be incorporated in their entirety into the disclosure of this application so that they can serve as a basis for modifications if necessary. Furthermore, it should be understood that after reading the foregoing disclosure of this application, those skilled in the art can make various alterations or modifications to this application, and these equivalent forms also fall within the scope of protection claimed in this application.

Claims

1. A method for diagnosing and predicting machine tool spindle faults, characterized in that, Includes the following steps: During machine tool operation, spindle vibration data collected by vibration sensors during the cutting process is used as a steady-state sample. The vibration data is decomposed into three sub-bands—high frequency, mid frequency, and low frequency—using wavelet packet transform. Ensemble Empirical Mode Decomposition (EEMD) is used to separate the residual DC component of the low-frequency sub-band. Variational Mode Decomposition (VMD) is used to extract the impact component from the high-frequency sub-band signal, combined with adaptive threshold wavelet denoising. A dynamic parameter Butterworth filter is used for the mid-frequency sub-band signal, dynamically modulating the cutoff frequency based on the instantaneous frequency. A physically constrained Kalman filter is introduced for the low-frequency sub-band signal. The state equation is constructed using the integral relationship between acceleration, velocity, and displacement. Finally, the weight coefficients of each frequency band are dynamically adjusted based on a residual feedback mechanism to achieve phase-consistent signal reconstruction, align the command domain, and calculate the principal axis eigenvalues. The maximum value of each feature value of the positive sample in each instruction domain is calculated as the corresponding threshold. The actual instruction domain data and the theoretical instruction domain data are compared to form the instruction domain samples. The health status of the machine tool spindle is judged based on the comparison results and the corresponding thresholds, and then displayed visually. The mean and standard deviation of the vibration signal feature sequence obtained after processing and the velocity and displacement data obtained after integration are calculated using a sliding window, transient noise is removed, and normalization is performed to obtain preprocessed data. The preprocessed data was used to train the Temporal Convolutional Network (TCN) model. The trained Temporal Convolutional Network (TCN) model is used to predict faults in the machine tool spindle, and the prediction results are displayed on a graphical interface.

2. The method for diagnosing and predicting machine tool spindle faults according to claim 1, characterized in that: The characteristic values ​​include peak-to-peak acceleration, effective velocity, peak-to-peak displacement, and characteristic amplitudes of the bearing outer ring, inner ring, cage, and rollers.

3. The method for diagnosing and predicting machine tool spindle faults according to claim 2, characterized in that: The threshold values ​​for the characteristic amplitudes of the bearing outer ring, inner ring, cage, and roller are all -40dB.

4. The method for diagnosing and predicting machine tool spindle faults according to claim 1, characterized in that: In the step of collecting spindle vibration data during the cutting process based on vibration sensors as steady-state samples during machine tool operation, the spindle vibration data is collected using an edge module.

5. The method for diagnosing and predicting machine tool spindle faults according to claim 1, characterized in that: In the step of decomposing the vibration data into three sub-bands of high frequency, medium frequency, and low frequency through wavelet packet transform, an improved wavelet packet transform is used.

6. The method for diagnosing and predicting machine tool spindle faults according to claim 1, characterized in that: The Temporal Convolutional Network (TCN) model includes the first four TCN layers as shared feature extractors, employing causal dilated convolutions to capture common temporal features of multiple physical quantities, and three independent branches for acceleration, velocity, and displacement prediction. Each branch contains two TCN residual blocks and a fully connected layer.

7. The method for diagnosing and predicting machine tool spindle faults according to claim 1, characterized in that: In the step of using the trained Temporal Convolutional Network (TCN) model to predict faults in the machine tool spindle and obtaining the prediction results, the prediction results are the predicted values ​​for the next three steps.

8. The method for diagnosing and predicting machine tool spindle faults according to claim 1, characterized in that: Before decomposing the vibration data into three sub-bands—high frequency, mid frequency, and low frequency—the vibration acceleration signal is converted into a frequency domain signal using a Fourier transform. Frequency domain integration is then performed sequentially to obtain the velocity and position signal data. The integration formula is as follows: Where V(w) is the frequency domain representation of velocity, X(w) is the frequency domain representation of displacement, A(w) is the frequency domain representation of acceleration, w is the angular frequency, and j is the imaginary unit.

9. The method for diagnosing and predicting machine tool spindle faults according to claim 1, characterized in that: The formula for calculating the cutoff frequency fc(t) of the Butterworth filter with dynamic parameters is as follows: f c (t)=α·max(f inst (t), 50Hz) Formula 3 Where α is the bandwidth expansion coefficient, and finst(t) is the instantaneous frequency extracted by the Hilbert transform.

10. The method for diagnosing and predicting machine tool spindle faults according to claim 1, characterized in that: The state equation for the physically constrained Kalman filter is: Where x k+1 Let x be the displacement at time k+1. k Let v be the displacement at time k. k Let a be the velocity at time k. k Let w be the acceleration at time k, Δt be the time step, and w be the acceleration at time k. k For process noise, z k These are the observed values.

11. The method for diagnosing and predicting machine tool spindle faults according to claim 1, characterized in that: The formula for calculating the weighting coefficient Wi of each frequency band in the residual feedback mechanism is as follows: Where ε i 2 For frequency band residuals, σ i 2 This represents the noise standard deviation.

12. The method for diagnosing and predicting machine tool spindle faults according to claim 1, characterized in that: The calculation formula for the normalization process is as follows: Where x norm denoted as σ, where is the normalized data, x is the data before normalization, u is the global mean, and σ is the standard deviation.

13. The method for diagnosing and predicting machine tool spindle faults according to claim 1, characterized in that: In step 300, the comparison results of each feature value with the corresponding threshold are combined according to the following rules to obtain the health status of the machine tool spindle: when none of the feature values ​​exceed the corresponding threshold, the spindle health status is normal; when any feature value exceeds the corresponding threshold, the spindle health status is abnormal.

14. The method for diagnosing and predicting machine tool spindle faults according to claim 1, characterized in that: The preprocessed data input into the Temporal Convolutional Network (TCN) model in step 500 includes vibration signal feature sequences, velocity data, and displacement data.

15. The method for diagnosing and predicting machine tool spindle faults according to claim 1, characterized in that: The prediction results in step 600 are displayed on the graphical interface as a comparison chart of instruction domain samples.

16. The method for diagnosing and predicting machine tool spindle faults according to claim 1, characterized in that: In step 100, the design and machining G-code enables the spindle to reach the rated speed of 4000 r / s. The CNC system runs the G-code and collects vibration sensor data as steady-state samples.

17. A system for diagnosing and predicting machine tool spindle faults, characterized in that, include: The acquisition module is used to collect spindle vibration data during the cutting process as a steady-state sample based on vibration sensors during machine tool operation; The signal processing module decomposes the vibration data into three sub-bands—high frequency, mid frequency, and low frequency—using wavelet packet transform. It then employs ensemble empirical mode decomposition (EEMD) to separate the residual DC component of the low-frequency sub-band. For the high-frequency sub-band signal, variational mode decomposition (VMD) is used to extract the impact component, combined with adaptive threshold wavelet denoising. For the mid-frequency sub-band signal, a dynamic parameter Butterworth filter is applied, dynamically modulating the cutoff frequency based on the instantaneous frequency. For the low-frequency sub-band signal, a physical constraint Kalman filter is introduced, and a state equation is constructed using the integral relationship between acceleration, velocity, and displacement. Finally, based on a residual feedback mechanism, the weight coefficients of each frequency band are dynamically adjusted to achieve phase-consistent signal reconstruction, align the command domain, and calculate the principal axis eigenvalues. The diagnostic module is used to calculate the maximum value of each feature value of the positive samples of each instruction domain as the corresponding threshold, compare the actual instruction domain data with the theoretical instruction domain data to make instruction domain samples, judge the health status of the machine tool spindle based on the comparison results and the corresponding threshold, and display the results visually. The data preprocessing module is used to calculate the mean and standard deviation of the vibration signal feature sequence obtained after processing by the signal processing module and the velocity and displacement data obtained after integration using a sliding window, remove transient noise, and perform normalization processing to obtain preprocessed data. The model training module is used to train the Temporal Convolutional Network (TCN) model using the preprocessed data. The prediction module is used to predict the faults of the machine tool spindle using the trained Temporal Convolutional Network (TCN) model, obtain the prediction results, and display the prediction results on a graphical interface.

Images