Dynamic fault diagnosis method and system for numerical control machine tool

Abstract

The application belongs to the technical field of production monitoring systems, and discloses a dynamic fault diagnosis method and system for a numerical control machine tool; the method comprises the following steps: generating a global time reference signal through a spindle encoder and a clock synchronization protocol; collecting vibration data of spindle bearings, current data of an electric cabinet and process parameters in each unit time, generating a preprocessed data sequence through transmission delay compensation and multi-rate up-conversion processing; inputting the vibration data and the current data into a preset mechanical-electrical transfer function model to calculate a time delay parameter; performing phase alignment on the preprocessed data sequence based on the time delay parameter to generate an aligned data sequence; inputting the aligned data sequence into a time series neural network to output a fusion feature vector; calculating a cross-correlation coefficient of the fusion feature vector, and generating a fault diagnosis result based on a preset cross-correlation threshold value; and the application solves the problem of fault feature extraction failure caused by data asynchrony in the prior art.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of production monitoring systems, and more particularly, to a dynamic fault diagnosis method and system for a numerical control machine tool. BACKGROUND

[0002] In the field of intelligent fault diagnosis of numerical control machine tools, multi-sensor fusion analysis is a core means to achieve early fault identification, which relies on high-precision time synchronization and causal feature correlation of multi-source signals such as vibration and current. However, in the prior art, machine tool multi-sensor data synchronization usually adopts software timestamp or hardware trigger mode. The former is affected by the operating system scheduling, and is prone to produce time deviation of microseconds to milliseconds in the process of data acquisition, transmission and processing, resulting in distortion of the phase relationship between vibration signals and current signals, and unable to accurately reflect the real-time coupling characteristics between mechanical vibration and electrical response. The latter relies on complex hardware wiring to build a synchronization topology, which not only increases the system complexity and cost, but also causes the synchronization accuracy to decline due to the differences in cable propagation delay and electromagnetic interference, and is difficult to adapt to the requirements of machine tool multi-working condition transformation and sensor expansion.

[0003] The inherent defects of the above synchronization technology result in that the subsequent fault diagnosis model based on time sequence neural network cannot effectively extract the time sequence dependent features of cross-modal signals. Due to the microsecond-level asynchronization of vibration data and current data, the vibration impact and current modulation signal caused by mechanical failure are misaligned on the time axis, making it difficult for the time sequence neural network to capture the causal correlation features based on the mechanical-electrical transfer function. The feature interaction of the fault sensitive frequency band is drowned in noise, eventually causing early fault missed detection or misjudgment. For example, in the early wear scenario of the main shaft bearing, the 1200Hz vibration impact caused by slight pitting and the 1150Hz current modulation signal cannot be correctly associated due to the time synchronization deviation. The traditional method can only rely on single signal threshold diagnosis, with high missed detection rate, and cannot meet the preventive maintenance needs of high-precision processing equipment.

[0004] Solving the problem of microsecond-level accurate synchronization of multi-source sensor data and cross-modal causal feature extraction has become a key bottleneck to improve the reliability of fault diagnosis of numerical control machine tools.

[0005] In view of this, the present application provides a dynamic fault diagnosis method and system for a numerical control machine tool to solve the above problems. SUMMARY

[0006] In order to overcome the above-mentioned defects of the prior art and achieve the above-mentioned purpose, the present application provides the following technical scheme: a dynamic fault diagnosis method for a numerical control machine tool, comprising:

[0007] A global time reference signal is generated by a spindle encoder and a clock synchronization protocol, and one zero pulse period of the global time reference signal is taken as a unit time;

[0008] Collect vibration data of main shaft bearing, current data of electric cabinet and process parameters of each unit time, generate pre-processing data sequence through transmission delay compensation and multi-rate up-sampling processing;

[0009] Input the vibration data and the current data into a preset mechanical-electrical transfer function model to calculate a time delay parameter; perform phase alignment on the pre-processing data sequence based on the time delay parameter to generate an aligned data sequence;

[0010] Input the aligned data sequence into a time sequence neural network to output a fusion feature vector;

[0011] Calculate a cross-correlation coefficient of the fusion feature vector, and generate a fault diagnosis result based on a preset cross-correlation threshold.

[0012] Further, the method for calculating the time delay parameter comprises:

[0013] Perform short-time Fourier transform on the vibration data and the current data respectively to generate a vibration frequency spectrum sequence and a current frequency spectrum sequence;

[0014] Based on a preset main shaft speed and bearing geometric parameters, determine a fault sensitive frequency band in the vibration frequency spectrum sequence and a modulation frequency band in the current frequency spectrum sequence, and uniformly mark the fault sensitive frequency band and the modulation frequency band as a feature frequency band;

[0015] Within the feature frequency band, solve parameters of a mechanical-electrical transfer function based on a parameter estimation method to generate a mechanical-electrical transfer function model;

[0016] Based on a phase response curve of the mechanical-electrical transfer function model, calculate a group time delay of the vibration signal to the current signal to generate the time delay parameter.

[0017] Further, the method for inputting the aligned data sequence into the time sequence neural network to output the fusion feature vector comprises:

[0018] Input the aligned data sequence into the time sequence neural network to generate a vibration feature sequence and a current feature sequence;

[0019] Based on the time delay parameter and a preset fluctuation range, construct a rectangular window function;

[0020] Based on the rectangular window function, perform attention weighting on the vibration feature sequence and the current feature sequence to generate a causal attention weight matrix;

[0021] Multiply the causal attention weight matrix with the current feature sequence to generate a weighted current feature sequence; the vibration feature sequence remains unchanged;

[0022] Perform global average pooling on the weighted current feature sequence and the vibration feature sequence to generate the fusion feature vector.

[0023] Further, the method for calculating the cross-correlation coefficient of the fusion feature vector, generating the fault diagnosis result based on the preset cross-correlation threshold value comprises:

[0024] Based on the vibration data and the current data in the normal state in the historical data, a vibration reference vector and a current reference vector are constructed;

[0025] The cosine similarity of the fusion feature vector and the vibration reference vector is calculated to generate a vibration similarity; the cosine similarity of the fusion feature vector and the current reference vector is calculated to generate a current similarity;

[0026] Based on the vibration similarity and the current similarity, the cross-correlation coefficient of the fusion feature vector is calculated;

[0027] The cross-correlation coefficient is compared with the preset cross-correlation threshold value; if the cross-correlation coefficient is less than the preset cross-correlation threshold value, a fault diagnosis result is generated; if the cross-correlation coefficient is greater than or equal to the preset cross-correlation threshold value, there is no fault.

[0028] Further, the method for performing phase alignment on the preprocessed data sequence based on the time delay parameter to generate an aligned data sequence comprises:

[0029] The current data is time-shifted according to the time delay parameter to generate time-delay-compensated current data;

[0030] The time-delay-compensated current data is interpolated and resampled to align the time stamp of the current data with the vibration data one by one to generate resampled current data;

[0031] Based on the zero-position pulse of the global time reference signal, the vibration data and the resampled current data are synchronously truncated to generate an aligned data sequence.

[0032] Further, the method for generating a global time reference signal through a main shaft encoder and a clock synchronization protocol comprises:

[0033] The A-phase, B-phase and Z-phase signals output by the main shaft encoder are distinguished and directionally processed to generate a main shaft position signal;

[0034] Based on the clock synchronization protocol, a local clock is synchronized with a network clock to generate a synchronous clock signal;

[0035] The main shaft position signal and the synchronous clock signal are time-stamped and associated to generate a global time reference signal.

[0036] Further, the method for generating a preprocessed data sequence comprises:

[0037] The vibration data of the main shaft bearing, the current data of the electric cabinet and the process parameters of each unit time are collected;

[0038] The vibration data and the current data are compensated for transmission delay based on a sensor installation position and a signal transmission path, and compensated data is generated;

[0039] The compensated data is interpolated and resampled to unify sampling rates of the vibration data and the current data, and a preprocessed data sequence is generated.

[0040] Further, the method for generating the mechanical-electrical transfer function model comprises:

[0041] The mechanical-electrical transfer function is assumed to be a first-order inertia model by using a least square method;

[0042] A target function is constructed by minimizing a mean square error between a current spectrum in a characteristic frequency band and a predicted value of the mechanical-electrical transfer function model, and a transfer function gain and a time constant reflecting a current mechanical-electrical coupling state are solved;

[0043] The mechanical-electrical transfer function model is constructed based on the first-order inertia model, the transfer function gain and the time constant.

[0044] Further, the method for calculating a group delay of the vibration signal to the current signal comprises:

[0045] The group delay of the vibration signal to the current signal is obtained by taking a negative derivative of a phase angle of a phase response curve of the mechanical-electrical transfer function model with respect to an angular frequency at a set angular frequency.

[0046] The dynamic fault diagnosis system of the numerical control machine tool comprises:

[0047] The hardware synchronization module generates a global time reference signal by a spindle encoder and a clock synchronization protocol, and takes one zero pulse period of the global time reference signal as a unit time;

[0048] The preprocessing module collects vibration data of the spindle bearing, current data of the electrical cabinet and process parameters in each unit time, and generates a preprocessed data sequence through transmission delay compensation and multi-rate up-sampling;

[0049] The dynamic modeling module inputs the vibration data and the current data into a preset mechanical-electrical transfer function model, calculates a time delay parameter, and performs phase alignment on the preprocessed data sequence based on the time delay parameter to generate an aligned data sequence;

[0050] The feature fusion module inputs the aligned data sequence into a time series neural network, and outputs a fusion feature vector;

[0051] The real-time diagnosis module calculates a cross-correlation coefficient of the fusion feature vector, and generates a fault diagnosis result based on a preset cross-correlation threshold.

[0052] Compared with the prior art, the technical effects and advantages of the numerical control machine tool dynamic fault diagnosis method and system of the present application are:

[0053] The present application generates a global time reference signal through the spindle encoder and clock synchronization protocol, realizes the deep integration of mechanical position signal and time reference, combines dynamic time delay modeling based on bearing geometric parameters and spindle speed, aligns the vibration data and current data at the microsecond level, and strengthens the causal correlation feature learning of cross-modal signals through the time sequence neural network with causal attention layer, solving the fault feature extraction failure problem caused by data asynchrony in the prior art, providing a breakthrough technical solution for realizing the precise diagnosis and prediction of early machine tool faults, and having important engineering significance for improving the reliability of processing equipment and reducing unplanned downtime losses. BRIEF DESCRIPTION OF DRAWINGS

[0054] Figure 1 The figure is a schematic diagram of the numerical control machine tool dynamic fault diagnosis system of the embodiment of the present application.

[0055] Figure 2 The figure is a flow chart of the numerical control machine tool dynamic fault diagnosis method of the embodiment of the present application.

[0056] Figure 3 The figure is a flow chart of the method for calculating the time delay parameter of the embodiment of the present application.

[0057] Figure 4 The figure is a flow chart of the method for outputting the fusion feature vector of the embodiment of the present application. DETAILED DESCRIPTION

[0058] The technical solutions in the embodiments of the present application will be described in detail, clearly and completely below with reference to the drawings in the embodiments of the present application. It should be particularly noted that the specific embodiments described below are only used to better illustrate and explain the technical solutions of the present application, and are intended to enable those skilled in the art to better understand and implement the present application, and should not be understood as limiting the scope of protection of the present application. Those skilled in the art can modify, adjust or equivalently replace the present application according to the content disclosed in the present application without departing from the spirit and essence of the present application, and these should be regarded as the protection scope of the present application.

[0059] Embodiment 1:

[0060] Please refer to Figure 1 As shown in the figure, the present embodiment discloses a numerical control machine tool dynamic fault diagnosis system, which comprises a hardware synchronization module, a preprocessing module, a dynamic modeling module, a feature fusion module and a real-time diagnosis module, each module is connected through wired and / or wireless connection to realize data transmission.

[0061] The hardware synchronization module generates a global time reference signal through the main shaft encoder and a clock synchronization protocol, and takes one zero pulse period of the global time reference signal as a unit time.

[0062] The method for generating a global time reference signal through the main shaft encoder and a clock synchronization protocol includes:

[0063] The A-phase, B-phase and Z-phase signals output by the main shaft encoder are distinguished and directionally processed to generate a main shaft position signal; based on the clock synchronization protocol, a local clock is synchronized with a network clock to generate a synchronized clock signal; and the main shaft position signal and the synchronized clock signal are time-stamped and associated to generate a global time reference signal.

[0064] Exemplarily, the embodiment provides a method for generating a global time reference signal through the main shaft encoder and a clock synchronization protocol, and specifically as follows:

[0065] The A-phase, B-phase and Z-phase signals output by the main shaft encoder are distinguished and directionally processed to generate a main shaft position signal. Specifically, a FPGA chip is used to realize signal processing, the rotating direction of the main shaft is determined through the 90°-phase-difference A-phase and B-phase signals, and the signals are 4-fold frequency distinguished and processed, so that the original encoder resolution is improved from 1024 pulses / revolution to 4096 pulses / revolution, to obtain higher-precision main shaft position information.

[0066] Based on the IEEE 1588 clock synchronization protocol, a local clock is synchronized with a network clock to generate a synchronized clock signal. In this step, the network switch is configured as a boundary clock, and through PTP protocol message interaction, the time synchronization of the local clock and the network clock is realized.

[0067] The main shaft position signal and the synchronized clock signal are time-stamped and associated to generate a global time reference signal. Specifically, a time-stamp counter is set in the FPGA, and whenever the Z-phase signal (main shaft zero pulse) is captured, the time stamp of the current synchronized clock is recorded, and the time stamp is bound with the main shaft position signal to form a time-stamp-position data pair. In this way, the mechanical position information is associated with the accurate time information to generate a unified global time reference signal.

[0068] The spindle encoder signal processing in the embodiment adopts a 4-fold frequency discrimination technology, which can improve the position resolution by 1 times compared with the traditional 2-fold frequency discrimination, and effectively reduces the synchronization error caused by insufficient encoder resolution. In addition, the application of IEEE 1588 clock synchronization protocol can improve the time synchronization accuracy to nanosecond level compared with the traditional NTP protocol (accuracy in millisecond level), which meets the requirement of microsecond level data alignment. In the prior art, the machine tool multi-sensor data synchronization usually adopts software timestamp or hardware trigger mode. The former has limited accuracy due to the influence of operating system scheduling, and cannot realize nanosecond level high-precision synchronization. The sensor data is easy to produce microsecond to millisecond level time deviation in the process of acquisition, transmission and processing due to thread preemption, interruption delay and other system level delay, which leads to the distortion of the phase relationship between vibration data and current data. For example, in the machine tool tool wear monitoring scene, the high-frequency vibration signal and the motor current sudden change signal at the moment of tool breakage may be incorrectly separated due to time deviation, which makes the subsequent phase alignment based on time delay parameter invalid, and further causes the time sequence neural network to fail to accurately capture the space-time correlation of fault features, and finally causes fault diagnosis false alarm or false alarm; the latter depends on complex hardware wiring and has poor expansibility. For example, in a multi-axis linkage machining center, if the spindle vibration sensor and the feed shaft current sensor adopt hardware trigger synchronization, the newly added tool temperature sensor needs to re-lay the trigger line and adjust the synchronization clock, which not only consumes time and effort, but also may introduce 50Hz power frequency interference due to the proximity of the trigger line to the motor cable, which reduces the synchronization accuracy by 1-2 orders of magnitude. The present application breaks through the inherent contradiction between accuracy and expansibility of the traditional method by using the zero position pulse of the global time reference signal to uniformly truncate data, the dynamic phase alignment mechanism based on time delay parameter, and the software defined synchronization strategy without hardware modification, which provides a solution for real-time synchronization of multi-source heterogeneous sensors, and significantly improves the timeliness and accuracy of fault diagnosis.

[0069] In the embodiment, the selection of the FPGA chip is based on its high-speed parallel processing capability, which can complete signal capture and timestamp generation within nanosecond level time, ensuring the correlation accuracy of the spindle position signal and the synchronization clock signal. The IEEE 1588 clock synchronization protocol parameter configuration is in E2E mode, and the master clock is selected as the GPS clock source. These settings are based on the requirement of high-precision time synchronization in the machine tool multi-sensor synchronization scene, which is more suitable for distributed sensor network compared with P2P mode, and the GPS clock source provides higher time reference reliability.

[0070] The preprocessing module collects the vibration data of the spindle bearing, the current data of the electric cabinet and the process parameters in each unit time, and generates a preprocessed data sequence after transmission delay compensation and multi-rate up-conversion processing.

[0071] The method for generating the preprocessed data sequence comprises:

[0072] Collect vibration data of main shaft bearing, current data of electric cabinet and process parameters of each unit time; based on the installation position of sensor and signal transmission path, transmission delay compensation is performed on the vibration data and current data to generate compensated data; interpolation resampling is performed on the compensated data to unify the sampling rate of vibration data and current data, and a preprocessed data sequence is generated.

[0073] Exemplarily, the embodiment provides a method for generating a preprocessed data sequence, which is specifically as follows:

[0074] Collect vibration data, current data and process parameters of each unit time, wherein the vibration data is collected by an ICP type accelerometer installed on the main shaft bearing seat, the sampling frequency is set to 50 kHz; the current data is collected by a Hall current sensor, the sampling frequency is set to 10 kHz; the process parameters are read from the machine tool PLC system and converted into digital signals, and the sampling period is set to 100 ms.

[0075] Based on the installation position of sensor and signal transmission path, transmission delay compensation is performed on the vibration data and current data to generate compensated data, which is specifically implemented by the following steps: the physical distance from the vibration sensor to the data acquisition card is calculated to be 1.2 meters, combined with the propagation speed of electrical signals in coaxial cables of about 2×10 8 meters / second, it is concluded that the transmission delay of vibration signals is 6 microseconds; the physical distance from the current sensor to the data acquisition card is calculated to be 0.8 meters, combined with the propagation speed of electrical signals in twisted pair wires of about 2.3×10 8 meters / second, it is concluded that the transmission delay of current signals is 3.5 microseconds; 6 microseconds of time advance is applied to the vibration data, and 3.5 microseconds of time advance is applied to the current data to realize transmission delay compensation.

[0076] Interpolation resampling is performed on the compensated data to unify the sampling rate of vibration data and current data, and a preprocessed data sequence is generated, which specifically adopts Lagrange interpolation method to improve the sampling rate of current data from 10 kHz to 50 kHz, so that the sampling points of vibration data and current data on the time axis correspond one by one.

[0077] The vibration data sampling frequency in this embodiment is set to 50 kHz, which is based on the fact that the fault characteristic frequency of the main shaft bearing is usually below 10 kHz. According to the Nyquist sampling theorem, the sampling frequency needs to be greater than twice the highest frequency of the signal. Therefore, 50 kHz is selected to fully retain the fault characteristic information. The current data sampling frequency is set to 10 kHz, because the fault characteristic frequency in the motor current signal is usually below 1 kHz, and the frequency response characteristics of the current sensor and the balance of the data volume are also considered. The process parameter sampling period is set to 100 ms, because the process parameters such as machine tool feed speed and cutting depth change slowly during normal machining, and a sampling period of 100 ms is sufficient to capture their change trend. The measurement of physical distance in transmission delay compensation uses a high-precision laser range finder. The accurate collection of these parameters provides a reliable foundation for microsecond-level transmission delay compensation.

[0078] In the prior art, multi-sensor data synchronization usually only considers the alignment of sampling time stamps, while ignoring the physical delay in the signal transmission process, resulting in an unavoidable deviation in time between different sensor data, which affects the accuracy of subsequent fault diagnosis. The present application improves the time alignment accuracy of multi-sensor data from milliseconds to microseconds by accurately calculating and compensating for transmission delay, significantly enhancing the reliability of fault feature extraction. In the interpolation and resampling process, the Lagrange interpolation method is chosen instead of the linear interpolation method, because the Lagrange interpolation method can better preserve the original characteristics of the signal, especially in the case of complex signal frequency components, it can effectively avoid the frequency distortion introduced in the interpolation process, which is particularly important for high-frequency vibration signal processing. The sampling frequency is unified to 50 kHz to meet the format requirements of the subsequent time series neural network for input data, ensuring that the vibration data and current data have the same resolution in the time dimension, thereby achieving more accurate feature fusion and fault diagnosis.

[0079] The dynamic modeling module inputs the vibration data and current data into a pre-set mechanical-electrical transfer function model to calculate the time delay parameter; and performs phase alignment on the pre-processed data sequence based on the time delay parameter to generate an aligned data sequence.

[0080] The method for calculating the time delay parameter comprises:

[0081] The vibration data and current data are respectively subjected to short-time Fourier transform to generate a vibration frequency spectrum sequence and a current frequency spectrum sequence; based on a pre-set main shaft speed and bearing geometric parameters, a fault sensitive frequency band in the vibration frequency spectrum sequence and a modulation frequency band in the current frequency spectrum sequence are determined, and the fault sensitive frequency band and the modulation frequency band are uniformly marked as a characteristic frequency band; within the characteristic frequency band, the parameters of the mechanical-electrical transfer function are solved based on a parameter estimation method to generate a mechanical-electrical transfer function model; and based on the phase response curve of the mechanical-electrical transfer function model, the group time delay of the vibration signal to the current signal is calculated to generate the time delay parameter.

[0082] Referring to Figure 3 The embodiment provides a method for calculating a time delay parameter, and the method comprises the following steps:

[0083] The vibration data and the current data are respectively subjected to short-time Fourier transform to generate a vibration spectrum sequence and a current spectrum sequence. In the implementation, a Hanning window is used as a window function, the window length is set to 1024 sampling points, the frame shift is set to 512 sampling points, the time domain signal is converted into a frequency domain signal through Fourier transform, and the vibration spectrum and the current spectrum corresponding to each time frame are obtained.

[0084] Based on a preset main shaft rotating speed and bearing geometric parameters, a fault sensitive frequency band in the vibration spectrum sequence and a modulation frequency band in the current spectrum sequence are determined, and the fault sensitive frequency band and the modulation frequency band are uniformly marked as a feature frequency band. The main shaft rotating speed is calculated through A / B phase signals of a main shaft encoder, and specifically, the time interval of two adjacent A phase pulses is measured, and the main shaft rotating speed is calculated through a formula "main shaft rotating speed = 60 / (pulse interval * encoder line number)", and the encoder line number is 2048 lines. The bearing geometric parameters include the number of rolling elements, the diameter of rolling elements, the pitch diameter and the contact angle, which are obtained from a design document of a machine tool main shaft component. For example, the number of rolling elements of a certain type of angular contact ball bearing is 8, the diameter of rolling elements is 6 mm, the pitch diameter is 35 mm, and the contact angle is 15°. According to a bearing fault characteristic frequency calculation formula, the inner ring fault characteristic frequency is calculated to be 1200 Hz, and the outer ring fault characteristic frequency is calculated to be 800 Hz, so that 800 Hz-1200 Hz in the vibration spectrum is set as the fault sensitive frequency band. The bearing fault characteristic frequency calculation formula is:

[0085] ;

[0086] In the formula, is the bearing fault characteristic frequency, which represents the vibration characteristic frequency generated when the inner ring of the bearing fails; GDT is the number of rolling elements of the bearing, which is determined by the bearing type, for example, the number of rolling elements of a certain angular contact ball bearing is 8; is the main shaft rotating frequency, that is, the number of rotations of the main shaft per second; d is the diameter of the rolling elements, which is provided by the design document; D is the pitch diameter of the bearing, that is, the diameter of the center circle of the rolling elements, which is provided by the design document; and α is the contact angle of the bearing, which represents the included angle between the normal line of the contact point of the rolling element and the raceway and the radial direction, which is provided by the design document.

[0087] The formula is based on the bearing dynamics principle and describes the frequency of the impact between the rolling elements and the inner ring raceway when the inner ring fails, and is one of the core characteristic frequencies for bearing fault diagnosis. The bearing fault characteristic frequency calculated through the formula can accurately locate the frequency component in the vibration signal related to the inner ring fault of the bearing, such as 800 Hz-1200 Hz, so as to avoid misjudgment of irrelevant frequencies such as gear meshing frequency or motor electromagnetic frequency as fault characteristics.

[0088] Exemplary, the embodiment provides a method for generating a mechanical-electrical transfer function model, specifically as follows:

[0089] According to the main shaft speed, the frequency band of 50Hz± of the current spectrum is set as the modulation frequency band, that is, 40Hz-60Hz. Within the characteristic frequency band, the parameters of the mechanical-electrical transfer function are solved based on the parameter estimation method to generate the mechanical-electrical transfer function model. Specifically, the least square method is used to set the transfer function as a first-order inertia link G(s)=K / (Ts+1), wherein K is the transfer function gain, reflecting the energy transfer efficiency of the mechanical vibration to the current response. The transfer function gain is stable in the healthy state, and decreases in the fault state. Ts is the time constant, representing the system inertia, which is related to the mass and damping of the mechanical components. The time constant is 500 microseconds in the healthy state, and Ts increases in the fault state. The process of the mechanical vibration of the main shaft system to the motor current is mainly affected by the bearing damping and the motor rotor inertia, so it can be approximated as a first-order inertia system, which can simplify the calculation and retain the core dynamic characteristics such as time delay and transfer function gain change, and balance between calculation efficiency and model accuracy.

[0090] The objective function is constructed in the frequency domain:

[0091]

[0092] In the formula, TZPD is the characteristic frequency band; is the spectrum value of the current signal at frequency f, which is obtained by short-time Fourier transform; is the spectrum value of the vibration signal at frequency f, which is obtained by short-time Fourier transform; is the imaginary unit, ; f is the frequency point in the characteristic frequency band, that is, the fault sensitive frequency band and the modulation frequency band determined by the bearing fault characteristic frequency calculation formula.

[0093] By minimizing the mean square error of the current spectrum and the predicted value of the mechanical-electrical transfer function model in the characteristic frequency band, the objective function is constructed to solve the transfer function gain and time constant that best reflect the current mechanical-electrical coupling state. The objective function is optimized only for the fault-related frequency, not for the full frequency band, reducing the interference of environmental noise such as cutting fluid impact and power grid harmonics on parameter solving.

[0094] The method for calculating the group time delay of the vibration signal to the current signal includes:

[0095] By taking the negative derivative of the phase angle of the phase response curve of the mechanical-electrical transfer function model and the angular frequency at the set angular frequency, the group time delay of the vibration signal to the current signal is obtained, specifically as follows:

[0096] Based on the phase response curve of the mechanical-electrical transfer function model, the group time delay of the vibration signal to the current signal is calculated :

[0097]

[0098] In the formula, is the phase angle of the transfer function at frequency 2 f, that is, wherein , is the angular frequency, is the arctangent function.

[0099] The group time delay in the characteristic frequency band is calculated, and the average value of the group time delay in the characteristic frequency band is taken as the final time delay parameter.

[0100] The window length and frame shift of the short-time Fourier transform in the embodiment are set based on the frequency characteristics and time resolution requirements of the vibration signal and the current signal. The Hanning window can effectively suppress spectral leakage, and the 1024-point window length corresponds to a time window of 20.48 ms at a sampling rate of 50 kHz, which meets the analysis accuracy of the bearing fault characteristic frequency. The frame shift of 512 points ensures that the adjacent time frames have 50% overlap, avoiding the loss of signal characteristics. The spindle speed is collected by using an encoder direct measurement method, which can reflect the actual speed of the spindle in real time compared with reading the set value from the PLC system, avoiding the speed deviation caused by the slipping of the transmission system. The bearing geometric parameters are directly obtained from the design document, ensuring the accuracy of the calculation of the fault characteristic frequency, and reducing the training time in the initial stage compared with the self-learning calibration method, which is more suitable for engineering applications.

[0101] In the prior art, the time delay calculation of multi-source data usually uses cross-correlation method or fixed empirical value. The former is prone to false peak values in a noisy environment, and the latter cannot adapt to working condition changes, resulting in large time delay calculation error. The present application determines the characteristic frequency band by combining the spindle speed and bearing geometric parameters, focuses on the frequency components related to the fault, reduces the interference of irrelevant frequency bands, and uses a parameter estimation method to construct a transfer function model, which can dynamically reflect the real-time characteristics of the mechanical-electrical system, and compared with the traditional method, the time delay calculation accuracy is improved from milliseconds to microseconds. In the parameter estimation process, a first-order inertia link is selected as the transfer function model, which is based on the fact that the main dynamic characteristics of the spindle system are the coupling of mechanical damping and electrical inertia, which can balance between calculation efficiency and model accuracy. The calculation of the group time delay uses the frequency domain differentiation method, which can more accurately reflect the phase lag characteristics of the system in the characteristic frequency band compared with the time domain cross-correlation method, providing reliable time delay parameters for subsequent phase alignment, which is crucial for the accurate correlation of microsecond asynchronous data, effectively solving the problem of fault feature misjudgment caused by inaccurate time delay calculation in the prior art.

[0102] The method for phase aligning the preprocessed data sequence based on the time delay parameter to generate an aligned data sequence comprises:

[0103] Time shifting the current data according to the time delay parameter to generate time delay compensated current data;

[0104] Interpolating and resampling the time delay compensated current data to align the time stamp of the current data with the vibration data one by one to generate resampled current data;

[0105] Synchronously truncating the vibration data and the resampled current data based on the zero position pulse of the global time reference signal to generate the aligned data sequence.

[0106] For example, the embodiment provides a method for phase aligning a preprocessed data sequence based on a time delay parameter to generate an aligned data sequence, which specifically comprises:

[0107] According to the calculated time delay parameter (for example, 350 μs), the current data is time shifted 350 μs to the left on the time axis to realize time delay compensation of the current data.

[0108] Interpolating and resampling the time delay compensated current data to align the time stamp of the current data with the vibration data one by one to generate resampled current data, which specifically adopts a cubic spline interpolation method to calculate the corresponding current value at each time stamp position of the vibration data, for example, at the vibration data time stamp positions t1, t2, t3, …, the corresponding current values I1, I2, I3, … are calculated through a cubic spline interpolation function, thereby generating resampled current data completely aligned with the vibration data time stamp.

[0109] Synchronously truncating the vibration data and the resampled current data based on the zero position pulse of the global time reference signal to generate the aligned data sequence, which is specifically realized through the following steps: detecting the Z signal in the global time reference signal, recording the time stamp t0 of the Z signal, and starting to intercept a data segment with a length of 10 seconds from the time t0 of the vibration data and the resampled current data to obtain an aligned data sequence with the same starting time and length.

[0110] The collection of the time delay parameter in the embodiment is obtained through bearing fault characteristic frequency calculation, transfer function parameter estimation and group time delay calculation, ensuring the accuracy and reliability of the time delay parameter. The time shift operation is realized by data shifting. In the FPGA chip, the current data is shifted by the number of sampling points corresponding to the time delay parameter through the setting of the FIFO buffer. For example, under a sampling rate of 50 kHz, 350 μs corresponds to 17.5 sampling points. The current data is delayed by 17.5 sampling points through the FIFO buffer, realizing accurate time shift. The selection of the cubic spline interpolation method is based on its ability to provide higher accuracy and smoothness than linear interpolation, especially in areas with dramatic signal changes, which can better preserve the original characteristics of the signal and avoid introducing false frequency components during the interpolation process.

[0111] In the prior art, multi-sensor data alignment usually adopts simple timestamp matching or linear interpolation, which cannot handle microsecond-level time delay differences and is prone to introduce errors during the interpolation process, resulting in inaccurate data alignment. The present application improves the data alignment accuracy from milliseconds to microseconds through accurate calculation of the time delay parameter and application of the cubic spline interpolation method, improving the accuracy of subsequent fault diagnosis. The synchronous truncation operation is based on the Z-phase signal of the global time reference signal, ensuring that each truncated data segment starts from the same position of the main shaft, eliminating data differences caused by different rotation angles of the main shaft, which is particularly important for fault diagnosis of rotating machinery, because bearing fault characteristics are usually closely related to the rotation angle of the main shaft.

[0112] In the present embodiment, the truncation length is set to 10 seconds, which is based on the fact that bearing fault characteristics usually need a long enough time window to be effectively extracted, and at the same time, considering the real-time requirement of data processing, a 10-second data segment can meet the needs of real-time diagnosis in industrial field while ensuring the effectiveness of fault feature extraction.

[0113] The feature fusion module inputs the aligned data sequence into the time sequence neural network and outputs a fusion feature vector.

[0114] The method of inputting the aligned data sequence into the time sequence neural network and outputting a fusion feature vector comprises:

[0115] The method of inputting the aligned data sequence into the time sequence neural network and outputting a fusion feature vector comprises:

[0116] A rectangular window function is constructed based on the time delay parameter and a preset fluctuation range.

[0117] The vibration feature sequence and the current feature sequence are subjected to attention weighting based on the rectangular window function, generating a causal attention weight matrix.

[0118] The causal attention weight matrix is multiplied by the vibration feature sequence and the current feature sequence respectively, generating a weighted feature sequence.

[0119] The global average pooling is performed on the weighted feature sequence to generate a fusion feature vector.

[0120] Referring to Figure 4 As shown in the figure, the embodiment provides a method for inputting an aligned data sequence into a time sequence neural network and outputting a fusion feature vector, which is specifically as follows:

[0121] The aligned data sequence is input into the time sequence neural network to generate a vibration feature sequence and a current feature sequence. Specifically, a bidirectional long short-term memory network (BiLSTM) is used as a time sequence feature extractor, the hidden layer dimension is set to 128, the aligned vibration data and current data are respectively input into two independent branches of the BiLSTM, the time sequence dependency is captured through forward and backward propagation, and a vibration feature sequence and a current feature sequence with a dimension of 128 are generated.

[0122] A rectangular window function is constructed based on the time delay parameter and the preset fluctuation range. Specifically, the following steps are implemented: the time delay parameter T is converted into the number of sampling points N, for example, the time delay parameter can be 350 μs, under a sampling rate of 50 kHz, N = 17.5 sampling points, the preset fluctuation range is set to ±10%, that is, the fluctuation range is ±1.75 sampling points, a rectangular window function with a width of 2N+1, that is, 36 sampling points is constructed, and the value of the window function at the center position corresponding to the time delay parameter T is 1, and the values of the remaining positions are 0.

[0123] Based on the rectangular window function, the vibration feature sequence and the current feature sequence are weighted by attention to generate a causal attention weight matrix. Specifically, the following steps are implemented: the cosine similarity between each time step in the vibration feature sequence and all time steps in the current feature sequence is calculated to obtain a similarity matrix, the rectangular window function is applied to the similarity matrix, the similarity values within the coverage range of the window function are retained, the remaining positions are set to 0, and the processed similarity matrix is row-normalized so that the sum of the elements in each row is 1 to generate a causal attention weight matrix.

[0124] The causal attention weight matrix is multiplied by the current feature sequence to generate a weighted current feature sequence; the vibration feature sequence remains unchanged.

[0125] The weighted current feature sequence and the weighted vibration feature sequence are subjected to global average pooling to generate a fusion feature vector. Specifically, the following steps are implemented: the vibration feature sequence and the weighted current feature sequence are respectively subjected to average pooling in the time dimension to obtain two feature vectors with a dimension of 128, and the two feature vectors are spliced into a fusion feature vector with a dimension of 256.

[0126] In this embodiment, the BiLSTM hidden layer dimension is set to 128, which is based on the experimental evaluation of the bearing fault feature representation ability. The 128-dimensional feature space can effectively capture the fault features in the vibration and current signals, while avoiding the overfitting problem caused by high dimension. The preset fluctuation range is set to ±10%, which takes into account the possible micro-delay fluctuations in the actual industrial environment, such as the subtle changes in sensor installation position or the changes in signal propagation speed caused by temperature changes, etc. The ±10% fluctuation range can provide certain fault tolerance ability while ensuring the causal relationship. The width of the rectangular window function is set to 2N+1, which ensures that there are N sampling points on both sides of the time delay parameter T, which can effectively capture the causal relationship between the vibration signal and the current signal.

[0127] In the prior art, multi-sensor data fusion usually uses simple feature splicing or averaging, ignoring the time delay and causal relationship between sensor data, resulting in fusion features that cannot accurately reflect fault features. The present application realizes causal attention weighting of vibration feature sequence and current feature sequence by constructing a rectangular window function based on time delay parameters, so that the fusion features can better capture the temporal dependence between sensor data, significantly improving the accuracy of fault diagnosis. In the generation process of the causal attention weight matrix, the application of the rectangular window function forces the model to only focus on the time steps that meet the causal relationship, avoiding the interference of non-causal relationships, which is particularly important for rotating machinery fault diagnosis, because there is a clear causal temporal relationship between vibration signals and current signals. The selection of global average pooling operation is based on its ability to effectively preserve statistical information of features while reducing feature dimension and improving the efficiency of subsequent fault classification.

[0128] In this embodiment, two 128-dimensional feature vectors are spliced into a 256-dimensional fusion feature vector, which not only preserves the feature information of vibration and current signals, but also realizes effective fusion of features, providing more rich and accurate feature representation for subsequent fault diagnosis.

[0129] The real-time diagnosis module calculates the cross-correlation coefficient of the fusion feature vector and generates the fault diagnosis result based on the preset cross-correlation threshold.

[0130] The method for calculating the cross-correlation coefficient of the fusion feature vector and generating the fault diagnosis result based on the preset cross-correlation threshold includes:

[0131] Based on the vibration data and current data under normal conditions in the historical data, vibration reference vectors and current reference vectors are constructed;

[0132] The cosine similarity between the fusion feature vector and the vibration reference vector is calculated to generate a vibration similarity, and the cosine similarity between the fusion feature vector and the current reference vector is calculated to generate a current similarity;

[0133] Calculate the cross-correlation coefficient of the fusion feature vector based on the vibration similarity and the current similarity.

[0134] Compare the cross-correlation coefficient with a preset cross-correlation threshold value. If the cross-correlation coefficient is less than the preset cross-correlation threshold value, generate a fault diagnosis result. If the cross-correlation coefficient is greater than or equal to the preset cross-correlation threshold value, there is no fault.

[0135] The embodiment provides a method for calculating the cross-correlation coefficient of the fusion feature vector, generating a fault diagnosis result based on a preset cross-correlation threshold value, and specifically as follows:

[0136] Based on the vibration data and the current data under normal conditions in the historical data, construct a vibration reference vector and a current reference vector. Specifically, the following steps are implemented: select the vibration data and the current data of the machine tool under normal operating conditions from the historical data, input these data into the time series neural network to generate corresponding vibration feature sequences and current feature sequences, and perform global average pooling on the vibration feature sequences and the current feature sequences respectively to obtain the vibration reference vector and the current reference vector.

[0137] Calculate the cosine similarity of the fusion feature vector and the vibration reference vector to generate a vibration similarity, specifically calculated by the formula "vibration similarity = (fusion feature vector x vibration reference vector) / (||fusion feature vector|| x ||vibration reference vector||)"; calculate the cosine similarity of the fusion feature vector and the current reference vector to generate a current similarity, specifically calculated by the formula "current similarity = (fusion feature vector x current reference vector) / (||fusion feature vector|| x ||current reference vector||)". Wherein, "||fusion feature vector||" is the vector module length of the fusion feature vector, The symbol represents the solution of the vector norm or module length, which is used to calculate the vector module length.

[0138] Based on the vibration similarity and the current similarity, calculate the cross-correlation coefficient of the fusion feature vector, specifically calculated by the formula "cross-correlation coefficient = (vibration similarity + current similarity) / 2".

[0139] Compare the cross-correlation coefficient with a preset cross-correlation threshold value. If the cross-correlation coefficient is less than the preset cross-correlation threshold value, generate a fault diagnosis result. If the cross-correlation coefficient is greater than or equal to the preset cross-correlation threshold value, there is no fault, specifically implemented by the following steps: compare the calculated cross-correlation coefficient with the preset cross-correlation threshold value (for example, 0.85). If the cross-correlation coefficient is less than 0.85, it is determined that the machine tool has a fault, and a fault diagnosis result is generated. If the cross-correlation coefficient is greater than or equal to 0.85, it is determined that the machine tool is running normally and there is no fault.

[0140] The fault diagnosis result includes bearing fault, main shaft asymmetry or imbalance, tool wear or damage, motor fault, load anomaly and coupling fault;

[0141] The detection logic of the bearing fault is: physically manifested as periodic impact of rolling elements and raceways; vibration waveform presents periodic impulse and vibration frequency contains bearing fault characteristic frequency; motor current is modulated by fault frequency, sideband appears; at the same time, at the impact moment, the current synchronously appears a transient peak, which is strictly aligned with the vibration pulse timestamp;

[0142] The detection logic of the main shaft asymmetry or imbalance is: physically manifested as periodic imbalance of the main shaft under stress; vibration waveform is a sine wave with a frequency of the main shaft rotation frequency; current amplitude periodically fluctuates with the rotation frequency, and the peaks and troughs of the vibration sine wave are in phase with the peaks and troughs of the current amplitude;

[0143] The detection logic of the tool wear or damage is: physically manifested as continuously increasing cutting force with wear, and decreasing cutting process stability; vibration waveform changes from a smooth sine wave to a burr, and the frequency contains cutting frequency harmonics; current amplitude continuously rises and appears cutting frequency modulation; at the peak of vibration harmonics, the peak of current modulation corresponds;

[0144] The detection logic of the motor fault is: physically manifested as uneven rotor mass distribution and periodic change of centrifugal force; vibration waveform is a sine wave with a frequency of the motor rotation frequency; current amplitude periodically fluctuates with the motor rotation frequency, and the frequencies of the vibration sine wave and the current fluctuation are both the motor rotation frequency, which is independent of the main shaft;

[0145] The detection logic of the load anomaly is: physically manifested as instantaneous mutation of cutting load; vibration waveform has a transient spike and no periodicity, and the frequency is a wide frequency distribution; current instantaneously jumps and has no modulation, only amplitude mutation; the timestamps of the vibration spike and the current jump are strictly synchronized;

[0146] The detection logic of the coupling fault is: physically manifested as periodic impact of coupling gap or misalignment every rotation; vibration waveform produces periodic impact pulses, and the frequency is the main shaft rotation frequency, which contains sidebands; current is modulated by the main shaft rotation frequency, and sidebands appear; the frequencies of the impact pulse and the current sideband are both the rotation frequency, and the impact moment corresponds to the extreme point of the current sideband.

[0147] The collection of historical data in the embodiment is based on the monitoring data of the machine tool in the normal running state. By collecting and analyzing the normal data of multiple cycles, the representativeness and reliability of the reference vector are ensured. In the construction process of the vibration reference vector and the current reference vector, the global average pooling operation is adopted, which can effectively retain the statistical information of the features, reduce the feature dimension, and improve the efficiency of the subsequent similarity calculation. The calculation of the cosine similarity is based on the inner product and the length of the vector, which can effectively measure the similarity of two vectors in the direction and avoid the influence of the vector length on the similarity calculation. The calculation formula of the cross-correlation coefficient adopts the average value of the vibration similarity and the current similarity, which can comprehensively consider the feature information of the vibration signal and the current signal, and improve the accuracy of fault diagnosis.

[0148] In the prior art, fault diagnosis is usually based on single sensor data or simple threshold judgment, which cannot fully utilize the complementary information of multi-sensor data, resulting in low accuracy and reliability of fault diagnosis. The present application realizes effective fusion of multi-sensor data by constructing vibration reference vector and current reference vector, and through the calculation of cosine similarity and cross-correlation coefficient, the difference between the current data and the normal data can be accurately measured, which significantly improves the accuracy and reliability of fault diagnosis. The preset cross-correlation threshold is set to 0.85, which is based on statistical analysis and verification of a large amount of experimental data. The threshold of 0.85 can effectively reduce the false alarm rate while ensuring the fault detection rate, balancing the accuracy and reliability of fault diagnosis. In actual application, the preset cross-correlation threshold can be adjusted according to the type, working condition and fault characteristics of the machine tool to adapt to different application scenarios and needs.

[0149] Embodiment 2:

[0150] Please refer to Figure 2 The present embodiment provides a dynamic fault diagnosis method for a numerical control machine tool, which comprises:

[0151] A global time reference signal is generated by the spindle encoder and the clock synchronization protocol, and one zero pulse period of the global time reference signal is taken as a unit time;

[0152] The vibration data of the spindle bearing, the current data of the electrical cabinet and the process parameters of each unit time are collected, and after transmission delay compensation and multi-rate up-conversion processing, a preprocessed data sequence is generated;

[0153] The vibration data and the current data are input into a preset mechanical-electrical transfer function model to calculate the time delay parameter; and the preprocessed data sequence is phase-aligned based on the time delay parameter to generate an aligned data sequence;

[0154] The aligned data sequence is input into a time series neural network to output a fusion feature vector;

[0155] The cross-correlation coefficient of the fusion feature vector is calculated, and a fault diagnosis result is generated based on a preset cross-correlation threshold.

[0156] The above merely describes the specific embodiments of the present application, but the protection scope of the present application is not limited thereto, any person skilled in the art can easily think of changes or replacements within the technical range disclosed by the present application, which should be covered within the protection scope of the present application. Therefore, the protection scope of the present application should be subject to the protection scope of the claims.

[0157] Finally: the above merely describes the preferred embodiments of the present application, and is not used to limit the present application, any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present application should be covered within the protection scope of the present application.

Claims

1. A dynamic fault diagnosis method for a numerical control machine tool, characterized by, The method comprises the following steps: generating a global time reference signal by a spindle encoder and a clock synchronization protocol, taking one zero pulse period of the global time reference signal as a unit time; collecting vibration data of a spindle bearing, current data of an electrical cabinet and process parameters in each unit time, generating a preprocessed data sequence through transmission delay compensation and multi-rate up-sampling processing; inputting the vibration data and the current data into a preset mechanical-electrical transfer function model to calculate a time delay parameter; performing phase alignment on the preprocessed data sequence based on the time delay parameter to generate an aligned data sequence; the method for calculating the time delay parameter comprises: performing short-time Fourier transform on the vibration data and the current data respectively to generate a vibration spectrum sequence and a current spectrum sequence; determining a fault-sensitive frequency band in the vibration spectrum sequence and a modulation frequency band in the current spectrum sequence based on a preset spindle speed and bearing geometric parameters, and uniformly marking the fault-sensitive frequency band and the modulation frequency band as a characteristic frequency band; solving parameters of the mechanical-electrical transfer function in the characteristic frequency band based on a parameter estimation method to generate a mechanical-electrical transfer function model; calculating a group time delay of the vibration signal to the current signal based on a phase response curve of the mechanical-electrical transfer function model to generate the time delay parameter; inputting the aligned data sequence into a time sequence neural network to output a fusion feature vector; calculating a cross-correlation coefficient of the fusion feature vector, and generating a fault diagnosis result based on a preset cross-correlation threshold.

2. The dynamic fault diagnosis method of a numerical control machine tool according to claim 1, characterized in that, The method for inputting the aligned data sequence into the time sequence neural network to output the fusion feature vector comprises: inputting the aligned data sequence into the time sequence neural network to generate a vibration feature sequence and a current feature sequence; constructing a rectangular window function based on the time delay parameter and a preset fluctuation range; performing attention weighting on the vibration feature sequence and the current feature sequence based on the rectangular window function to generate a causal attention weight matrix; multiplying the causal attention weight matrix with the current feature sequence to generate a weighted current feature sequence; the vibration feature sequence remains unchanged; performing global average pooling on the weighted current feature sequence and the vibration feature sequence to generate the fusion feature vector.

3. The dynamic fault diagnosis method of a numerical control machine tool according to claim 1, characterized in that, The method for calculating the cross-correlation coefficient of the fusion feature vector and generating the fault diagnosis result based on the preset cross-correlation threshold comprises: constructing a vibration reference vector and a current reference vector based on vibration data and current data in a normal state in historical data; calculating a cosine similarity of the fusion feature vector and the vibration reference vector to generate a vibration similarity; calculating a cosine similarity of the fusion feature vector and the current reference vector to generate a current similarity; calculating the cross-correlation coefficient of the fusion feature vector based on the vibration similarity and the current similarity; comparing the cross-correlation coefficient with the preset cross-correlation threshold, if the cross-correlation coefficient is less than the preset cross-correlation threshold, generating the fault diagnosis result; if the cross-correlation coefficient is greater than or equal to the preset cross-correlation threshold, there is no fault.

4. The dynamic fault diagnosis method of a numerical control machine tool according to claim 1, characterized in that, The method for performing phase alignment on the preprocessed data sequence based on the time delay parameter to generate the aligned data sequence comprises: time-shifting the current data according to the time delay parameter to generate time delay compensated current data; performing interpolation resampling on the time delay compensated current data to align the time stamp of the current data with the vibration data one by one to generate resampled current data; The zero-position pulse of the global time reference signal is used to synchronize and truncate the vibration data and the resampled current data to generate an aligned data sequence.

5. The dynamic fault diagnosis method of a numerical control machine tool according to claim 1, characterized in that, The method for generating a global time reference signal based on a spindle encoder and a clock synchronization protocol comprises: Discriminating and direction-determining the A-phase, B-phase and Z-phase signals output by the spindle encoder to generate a spindle position signal; Synchronizing a local clock with a network clock based on the clock synchronization protocol to generate a synchronized clock signal; Timestamping the spindle position signal and the synchronized clock signal to generate the global time reference signal.

6. The dynamic fault diagnosis method of a numerical control machine tool according to claim 1, characterized in that, The method for generating a preprocessed data sequence comprises: Collecting vibration data of the spindle bearing, current data of the electrical cabinet and process parameters in each unit time; Compensating for transmission delays in the vibration data and the current data based on the sensor installation positions and signal transmission paths to generate compensated data; Interpolating and resampling the compensated data to unify the sampling rates of the vibration data and the current data to generate a preprocessed data sequence.

7. The dynamic fault diagnosis method of a numerical control machine tool according to claim 1, characterized in that, The method for generating a mechanical-electrical transfer function model comprises: Assuming the mechanical-electrical transfer function as a first-order inertia model using the least squares method; Constructing an objective function by minimizing the mean square error between the current spectrum in the characteristic frequency band and the predicted value of the mechanical-electrical transfer function model to solve the transfer function gain and the time constant reflecting the current mechanical-electrical coupling state; Constructing the mechanical-electrical transfer function model based on the first-order inertia model, the transfer function gain and the time constant.

8. The dynamic fault diagnosis method of a CNC machine tool according to claim 1, characterized in that, The method for calculating the group delay of the vibration signal to the current signal comprises: Obtaining the group delay of the vibration signal to the current signal by taking the negative of the partial derivative of the phase angle of the phase response curve of the mechanical-electrical transfer function model with respect to the angular frequency at a set angular frequency.

9. A dynamic fault diagnosis system for a numerical control machine tool, for implementing the dynamic fault diagnosis method for a numerical control machine tool according to any one of claims 1 to 8, characterized in that, The method comprises: A hardware synchronization module generates a global time reference signal based on a spindle encoder and a clock synchronization protocol, and takes one zero-position pulse period of the global time reference signal as a unit time; A preprocessing module collects vibration data of the spindle bearing, current data of the electrical cabinet and process parameters in each unit time, compensates for transmission delays and performs multi-rate up-sampling to generate a preprocessed data sequence; A dynamic modeling module inputs the vibration data and the current data into a preset mechanical-electrical transfer function model to calculate time delay parameters; A phase alignment module aligns the preprocessed data sequence based on the time delay parameters to generate an aligned data sequence; The method for calculating the time delay parameters comprises: Performing short-time Fourier transform on the vibration data and the current data to generate vibration spectrum sequences and current spectrum sequences; Based on a preset spindle speed and bearing geometric parameters, determining a fault-sensitive frequency band in the vibration spectrum sequences and a modulation frequency band in the current spectrum sequences, and uniformly marking the fault-sensitive frequency band and the modulation frequency band as a characteristic frequency band; In the characteristic frequency band, estimating the parameters of the mechanical-electrical transfer function based on a parameter estimation method to generate a mechanical-electrical transfer function model; Based on the phase response curve of the mechanical-electrical transfer function model, calculating the group delay of the vibration signal to the current signal to generate the time delay parameters; A feature fusion module inputs the aligned data sequence into a time series neural network to output a fusion feature vector. The real-time diagnosis module calculates a cross-correlation coefficient of the fusion feature vector, and generates a fault diagnosis result based on a preset cross-correlation threshold.

Images