Motor fault diagnosis algorithm based on multi-sensor fusion

Abstract

The invention relates to the technical field of motor fault diagnosis, in particular to a motor fault diagnosis algorithm based on multi-sensor fusion, and the algorithm comprises the steps: injecting a step excitation signal into a motor, synchronously collecting the original response waveforms of vibration and current sensors, and calculating the inherent response delay. Establishing a mapping relation library of delay values and current sensor filtering parameters, calling the delay values in real time according to the filtering parameters, performing reverse time offset compensation on a current harmonic signal time sequence, performing time alignment on the two types of data, finally performing cross-domain coupling analysis on the aligned data, extracting vibration pulse peak frequency and current harmonic fluctuation quantity, and determining the vibration pulse peak frequency and the current harmonic fluctuation quantity. Early faults are judged by combining the bearing outer ring fault characteristic frequency band and the load rate dynamic threshold value, graded alarm is generated by tracking characteristics, the problem of fault false judgment and missed judgment caused by sensor data space-time dislocation is solved, and the early fault diagnosis accuracy of the motor is improved.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of motor fault diagnosis, in particular to a motor fault diagnosis algorithm based on multi-sensor fusion. BACKGROUND

[0002] Motor fault diagnosis is an important technology. Under the background of the continuous improvement of industrial automation, this technology is the key support to ensure the stable operation of the motor and reduce the operation and maintenance cost. It can capture the characteristic signals of mechanical faults of the motor through vibration sensors and perceive the abnormalities of the electrical system through current sensors, avoiding the limitations of single sensor monitoring. At the same time, it can detect early faults in advance, prevent faults from expanding to cause the motor to stop, reduce production interruption losses, and adapt to the operation and maintenance needs of various industrial motors such as high-voltage motors and servo motors. The existing motor fault diagnosis technology based on multi-sensor fusion faces the core problem of time and space asynchronization of vibration and current sensor data in actual application, which leads to deviation in fault feature coupling analysis, misjudgment or missed judgment of early faults. Traditional technology directly collects data of two types of sensors for fusion without considering the inherent response difference of the sensors. The current sensor needs to be filtered to process high-frequency noise, which will cause a response delay, and the delay time is different under different filtering parameters. The vibration sensor responds faster, and the fault feature signals collected by the two sensors have a time difference. At the same time, there is a lack of calibration mechanism for response delay, and it is impossible to dynamically compensate the time sequence deviation according to the actual filtering parameters of the current sensor. This defect causes the multi-source data with time and space misalignment to be unable to be accurately associated, the corresponding relationship between the vibration pulse peak value and the current harmonic amplitude is destroyed, and normal current fluctuations may be misjudged as fault features or the weak coupling signals caused by early faults are missed. Secondly, misjudgment will cause unnecessary shutdown for maintenance, increasing the operation and maintenance cost, and missed judgment will make early faults develop into serious faults, which not only increases the difficulty of maintenance, but also may cause the motor to burn out and production interruption safety accidents. In the long run, it is difficult to meet the needs of high precision and early detection of motor fault diagnosis in industrial scenarios, and it is also difficult to fully play the technical advantages of multi-sensor fusion, resulting in waste of monitoring resources. In order to solve this technical problem, we provide a motor fault diagnosis algorithm based on multi-sensor fusion. SUMMARY

[0003] The purpose of the present application is to provide a motor fault diagnosis algorithm based on multi-sensor fusion to solve the problems raised in the background technology.

[0004] To achieve the above purpose, one of the purposes of the present application is to provide a motor fault diagnosis algorithm based on multi-sensor fusion, comprising the following steps: S1, by injecting a step excitation signal to the motor system, synchronously collecting the original response waveforms of the vibration sensor and the current sensor, calculating the inherent response delay based on the time difference between the rising edge timestamp of the vibration sensor and the filtered stable value timestamp of the current sensor, and establishing a mapping relationship library of the inherent response delay value and the filtering parameter of the current sensor; S2, in the real-time monitoring stage, the corresponding inherent response delay value is retrieved from the mapping relationship library according to the actual filtering parameter of the current sensor, and the time sequence of the harmonic signal collected by the current sensor is compensated by reverse time offset, so that the compensated current harmonic feature timestamp is aligned with the vibration sensor pulse peak timestamp; S3, the time-aligned vibration pulse peak frequency and the current harmonic amplitude are analyzed by cross-domain coupling, when the vibration pulse peak frequency is in the bearing outer ring fault characteristic frequency band and the fluctuation of the current harmonic amplitude in the synchronous time window exceeds the threshold value, it is determined that the early bearing fault occurs, wherein the length of the time window is set according to the vibration pulse peak frequency and is centered on the vibration pulse peak point.

[0005] Compared with the prior art, the beneficial effects of the present application are: The present application synchronously collects the original response waveforms of the vibration and current sensors by injecting a step excitation signal to the motor, calculates the inherent response delay and establishes a mapping relationship library of the inherent response delay and the filtering parameter of the current sensor, retrieves the delay value in real time to compensate the time sequence of the current harmonic signal by reverse time offset to realize the accurate time alignment of the two types of data, then extracts the vibration pulse peak frequency by fast spectrum analysis on the aligned data, demodulates the amplitude envelope of the current harmonic to calculate the fluctuation, combines the bearing outer ring fault characteristic frequency band and the dynamic threshold of the load rate to make cross-domain coupling determination, sets the length of the time window according to the vibration pulse period and adjusts the window length by integer multiple according to the signal-to-noise ratio, continuously tracks the frequency stability and current harmonic fluctuation in the fault characteristic period to generate a graded alarm, achieves the technical effects of accurate time alignment of sensor data, accurate identification of early bearing fault and improvement of the robustness of feature extraction and the reliability of alarm, and effectively solves the problem of deviation in fault feature coupling analysis and misjudgment of early fault caused by the asynchronization of vibration and current sensor data in time and space. BRIEF DESCRIPTION OF DRAWINGS

[0006] Figure 1 The overall workflow of the present application is shown in the figure. DETAILED DESCRIPTION

[0007] The technical solutions in the embodiments of the present application will be described clearly and completely below with reference to the drawings in the embodiments of the present application. Obviously, the described embodiments are only part of the embodiments of the present application, not all. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative labor are within the scope of protection of the present application.

[0008] Referring to Figure 1 As shown in the embodiment, the motor fault diagnosis algorithm based on multi-sensor fusion is provided, including the following steps: S1, by injecting a step excitation signal to the motor system, synchronously collecting the original response waveforms of the vibration sensor and the current sensor, calculating the inherent response delay based on the time difference between the rising edge timestamp of the vibration sensor and the filtered stable value timestamp of the current sensor, and establishing a mapping relationship library of the inherent response delay value and the current sensor filtering parameter; S2, in the real-time monitoring stage, the corresponding inherent response delay value is retrieved from the mapping relationship library according to the actual filtering parameter of the current sensor, and the time sequence of the harmonic signal collected by the current sensor is compensated by reverse time offset, so that the compensated current harmonic feature timestamp is aligned with the vibration sensor pulse peak timestamp; S3, the time-aligned vibration pulse peak frequency and current harmonic amplitude are analyzed by cross-domain coupling, when the vibration pulse peak frequency is in the bearing outer ring fault characteristic frequency band and the fluctuation amount of the current harmonic amplitude in the synchronous time window exceeds the threshold value, it is determined that the early bearing fault occurs, wherein the time window length is set according to the vibration pulse peak frequency and is centered on the vibration pulse peak point.

[0009] In S1, the step excitation signal is injected to the motor system, and the specific steps are as follows: In the no-load running state of the motor, a voltage step signal with a preset time length is injected to the three-phase stator winding through the motor driver, the amplitude of the step signal is set to a certain proportion of the rated voltage and the duration is greater than the maximum response delay of the current sensor; In the synchronous acquisition process, a data acquisition card with high-precision clock synchronization function is used to record the original voltage waveform output by the vibration sensor and the original current waveform output by the current sensor at a sampling rate not less than the highest response frequency of the vibration sensor.

[0010] The operation of calculating the inherent response delay based on the time difference between the rising edge timestamp of the vibration sensor and the filtered stable value timestamp of the current sensor includes the following specific steps: The starting time point of the rising edge of the step response is extracted from the vibration sensor waveform as the reference timestamp, which is determined by detecting the position where the waveform first exceeds the environmental noise threshold and continuously rises; At the same time, the arrival timestamp of the filtered stable value is extracted from the current sensor waveform, which is defined as the starting time when the current waveform enters the set error band range and maintains for more than the minimum stable time, and the difference between the two timestamps is taken as the inherent response delay value under the test condition.

[0011] The specific steps of establishing the mapping relationship library of the inherent response delay value and the current sensor filtering parameter are as follows: The low-pass filter cutoff frequency of the current sensor is dynamically adjusted through a programmable resistance-capacitance network, step excitation signals are repeatedly injected into the motor system under multiple discrete filter parameter combinations, a mapping relationship is established between the inherent response delay value obtained from each group of tests and the corresponding filter parameters, and a delay amount lookup table indexed by the filter time constant is formed, which covers all typical configurations within the working parameter range of the current sensor.

[0012] The operation of reverse time offset compensation for the timing of the harmonic signal collected by the current sensor in S2 is as follows: According to the real-time acquired current sensor filter parameter configuration, the corresponding inherent response delay value is retrieved from the mapping relationship library, and a ring-shaped data buffer is created to store the original sampling points of the current harmonic signal. The overall translation of the timing data is realized by adjusting the buffer read pointer position, and the translation amount is the sampling point number corresponding to the retrieved inherent response delay value.

[0013] The verification mechanism for aligning the compensated current harmonic feature timestamp with the vibration sensor pulse peak timestamp is as follows: After completing the reverse time offset compensation, the periodic pulse peak point time sequence caused by bearing failure in the vibration signal is extracted, and the energy distribution symmetry of the compensated current harmonic signal within a set time range before and after each pulse peak point is calculated. When the deviation between the energy peak position and the vibration pulse peak timestamp is less than the synchronization error tolerance, it is determined that the alignment is successful, otherwise the translation amount is dynamically corrected according to the deviation direction.

[0014] The specific steps of cross-domain coupling analysis are as follows: At the time-aligned vibration pulse peak point, the fundamental component of the vibration pulse is extracted as the vibration pulse peak frequency through fast spectrum analysis, and the compensated current harmonic signal is amplitude envelope demodulated within a time window centered on the peak point, and the peak-to-peak value of the current harmonic amplitude within the time window is calculated as the fluctuation amount feature.

[0015] The conditions for determining early bearing failure are refined as follows: When the vibration pulse peak frequency is detected to fall within the bearing outer ring failure characteristic frequency band, the current harmonic amplitude fluctuation amount analysis is triggered, and the fluctuation amount threshold is dynamically adjusted according to the current load rate of the motor. Specifically, by querying the pre-established load rate-current harmonic baseline amplitude reference table, the baseline amplitude under the corresponding load rate in the table is multiplied by a set coefficient to obtain the dynamic threshold.

[0016] It also includes a fault determination verification mechanism: When the fault condition is first detected, the frequency of the vibration pulse peak appearing in the subsequent set period is continuously tracked, and if the frequency stability index exceeds the set threshold and the current harmonic fluctuation continues to exceed the dynamic threshold, a hierarchical alarm signal is generated, wherein the frequency stability index is obtained by calculating the statistical variance of the frequency change rate between consecutive pulses.

[0017] The time window length is specifically set as: According to the fault pulse period corresponding to the current vibration pulse peak frequency, an integer multiple of the period is taken as the reference length of the time window, and the center point of the time window is strictly aligned with the pulse peak time stamp detected by the vibration sensor, that is, the window starting point is offset by one-half of the window length from the forward peak time stamp, and the window ending point is offset by one-half of the window length from the backward peak time stamp. The integer multiple is automatically adjusted according to the signal-to-noise ratio, and the integer multiple coefficient is increased when the noise level is high to improve the robustness of feature extraction.

[0018] It needs to be further explained that in the motor fault diagnosis algorithm based on multi-sensor fusion, the synchronization of sensor response is the premise to ensure the accuracy of fault feature extraction. If there is a response delay deviation between the current sensor and the vibration sensor, it will cause the time and space dislocation of multi-source data, and misjudge the fault feature (such as mistaking the time difference of the signal caused by delay as the rotor imbalance fault of the motor). Therefore, the inherent response delay of the sensor needs to be calibrated through step signal injection and synchronous acquisition, and the specific implementation is as follows: First, the test is carried out in the motor no-load running state, which means that the motor is disconnected from the load device and only relies on its own stator winding to drive the rotor to rotate. In this state, the motor runs smoothly and there is no interference from load fluctuations. During the test, the motor driver injects a pre-set voltage step signal into the three-phase stator winding. The voltage step signal refers to a voltage value that jumps from the current stable value (such as 0V or low voltage) to the target amplitude in a very short time (usually <1μs), maintaining a constant electric signal. Its core function is to apply a sudden excitation to the motor system to trigger the response changes of vibration and current, facilitating the capture of the dynamic characteristics of the sensor. The amplitude of the step signal needs to be set to 30%-50% of the rated voltage of the motor (such as 150V for a rated voltage of 380V). An amplitude that is too high may cause the motor winding to overheat or the rotor to impact too much, while an amplitude that is too low may not trigger a significant response signal. The duration needs to be greater than the maximum response delay of the current sensor (usually 10-20ms according to the sensor technical manual, so the duration is set to 25ms), ensuring that it can cover the entire process from receiving the excitation to outputting a stable signal. During the synchronous acquisition process, a data acquisition card with high-precision clock synchronization function needs to be used. This acquisition card needs to be connected to both the vibration sensor and the current sensor. The vibration sensor, such as a piezoelectric acceleration sensor, is installed radially on the motor end cover to capture the motor vibration response triggered by the step excitation. The current sensor, such as a Hall current sensor, is connected in series with the A-phase stator winding to record the change in winding current. The sampling rate of the acquisition card needs to be not less than the highest response frequency of the vibration sensor, which is 5kHz, so the sampling rate is set to 10kHz to meet the Nyquist sampling theorem and avoid signal aliasing. The original voltage waveforms output by the two types of sensors are recorded simultaneously. The vibration sensor outputs a 0-5V voltage signal, and the current sensor outputs a 0-10V voltage signal, corresponding to a 0-50m / s² vibration acceleration and a 0-50A current, respectively. The start time of acquisition is aligned with the injection time of the voltage step signal through the synchronous trigger signal of the data acquisition card and the motor driver (trigger delay <10μs), ensuring the consistency of vibration and current data in the time dimension. After acquiring the original waveforms, the response timestamps of the two types of sensors need to be determined through feature extraction to calculate the inherent response delay. First, the starting time point of the rising edge of the step response is extracted from the vibration sensor waveform as the reference timestamp. This process needs to exclude environmental noise interference, with the environmental noise threshold defined as twice the average noise voltage output by the vibration sensor when the motor is not injected with a step signal (such as 0.1V,25V), the starting point (100th sampling point) is the reference timestamp (corresponding to the actual time = sampling point number x sampling period, such as 100 x 0.1 ms = 10 ms), which can avoid misjudging single noise spikes as response start. At the same time, the arrival timestamp of the filtered stable value in the current sensor waveform is extracted. First, the current original waveform is low-pass filtered (cutoff frequency 1 kHz, eliminating high-frequency noise), and then the filtered stable value (such as the filtered current waveform finally stabilizes at 15V corresponding to 15A current) is determined. The error band range is defined as ±3% of the filtered stable value (i.e. 15A ± 0.45A, corresponding to 15V ± 0.45V), which is used to determine whether the current reaches a stable state. The minimum stable time is defined as the shortest time for the current waveform to maintain within the error band (set to 5 sampling periods, i.e. 5 x 0.1 ms = 0.5 ms), which avoids misjudging short fluctuations as stable. The arrival timestamp is the starting time when the current waveform first enters the error band range (such as the 180th sampling point voltage 14.6V, which enters the 14.55-15.45V interval), and the continuous 5 sampling points are maintained within the interval (the 180th sampling point, corresponding to the actual time 18 ms). The difference between the two timestamps (18 ms-10 ms = 8 ms) is taken as the inherent response delay value under the test conditions, which reflects the response lag of the current sensor relative to the vibration sensor, and provides a calibration basis for subsequent multi-source data space-time alignment. To comprehensively cover the inherent response delay of the current sensor under different working parameters, the low-pass filter cutoff frequency needs to be dynamically adjusted and a delay mapping relationship needs to be established: the cutoff frequency is adjusted through a programmable resistance-capacitance network, which is a circuit module composed of a digital potentiometer and a programmable capacitance array, integrated in the signal conditioning circuit of the current sensor. The resistance value (R) and capacitance value (C) of the network are controlled by the MCU sending I2C instructions, and the low-pass filter cutoff frequency is determined by the RC time constant (f = 1 / (2πRC)). For example, to adjust the cutoff frequency from 1 kHz to 2 kHz, the MCU controls the digital potentiometer to reduce the resistance from 15.9 kΩ to 7.96 kΩ (the capacitance remains 10 nF), and the cutoff frequency is dynamically switched by changing the RC value, without the need to manually replace the resistance and capacitance, improving test efficiency. Repeat the step excitation signal injection test under multiple discrete filter parameter combinations, which means that the resistance and capacitance parameters are divided into multiple RC configurations according to a fixed step, each combination corresponds to a different cutoff frequency, covering all working frequency ranges (100 Hz-10 kHz) marked in the current sensor manual. The test process is: For each RC combination, the parameter is loaded through the programmable network first, and then the voltage step signal injection (amplitude 150V, duration 25ms) and synchronous acquisition process are repeated, the vibration reference timestamp and current stable timestamp are extracted, the corresponding inherent response delay value is calculated, the inherent response delay value obtained by each group of tests is mapped with the corresponding filter parameter, the filter time constant is taken as the index, the corresponding cutoff frequency, resistance value, capacitance value, inherent response delay value are sorted as structured data, and finally a delay lookup table indexed by filter time constant is formed, which needs to cover all typical configurations in the working parameter range of the current sensor. In the subsequent fault diagnosis process, only the filter time constant calculated from the current sensor RC parameters needs to be read, and the corresponding inherent response delay can be quickly retrieved from the lookup table for time alignment of multi-source sensor data, ensuring the accuracy of fusion diagnosis.

[0019] The specific implementation of real-time delay retrieval and ring buffer time sequence alignment is as follows after the delay lookup table indexed by the filter time constant is established and the dynamic adjustment of the current sensor filter parameters is realized through the programmable resistance-capacitance network. The motor fault diagnosis system needs to use these pre-calibration results in real time to realize the time sequence alignment of vibration and current data. If the corresponding delay value cannot be matched according to the real-time filter parameter, the pre-calibration will lose its significance, and the traditional linear buffer is easy to affect the continuity due to data overflow, so real-time delay retrieval and ring buffer time sequence alignment are needed to realize accurate time sequence compensation. The specific implementation is as follows: First, according to the real-time acquisition of current sensor filter parameter configuration retrieval corresponding inherent response delay value, current sensor through the parameter feedback interface, real-time transmission to the system main unit under the current working state of resistance (R) and capacitance (C) parameters, main unit receives the parameter, immediately calculate the filter time constant τ=RxC (such as 10kΩx10nF=100μs), and τ as index, query the delay amount lookup table (table τ=100μs corresponding inherent response delay value 8ms); According to the sampling rate of data acquisition card, the delay value is converted into the corresponding sampling point number (translation amount=delay value÷sampling period=8ms÷0.1 ms = 80 sample points), which is the core basis for subsequent time series data translation, ensures the quantization accuracy of delay compensation, and synchronously creates a ring-shaped data buffer to store the original current harmonic signal sample points. The ring-shaped data buffer is a memory data area using a circular storage structure, and the circular writing and reading of data is achieved through the head-to-tail address mapping, without the need for frequent memory cleaning, avoiding data overflow and saving storage resources. The buffer size is set to twice the number of sample points corresponding to the maximum inherent response delay (for example, a maximum delay of 15 ms corresponds to 150 sample points, and the buffer size is set to 300 sample points), which not only ensures that the current harmonic data of sufficient duration can be completely stored, but also avoids memory waste. When creating, the system divides a continuous address space in memory, and associates the start address with the end address through logical mapping. When writing data, it is stored in order from the start address, and when the write address exceeds the end address, it automatically jumps to the start address to cyclically overwrite old data. Only the old data that has been read is overwritten, and the unread data is retained. After the current harmonic signal is collected in real time by the data acquisition card, it is written into the ring-shaped buffer in the order of sampling. Each sample point is stored as 16-bit binary data, matching the output accuracy of the sensor and ensuring that the details of the original signal are not lost. The overall translation of time series data is achieved by adjusting the position of the buffer read pointer. The ring-shaped buffer has a separate write pointer and read pointer. The write pointer increases in real time as data is written. Every time a sample point is written, the write pointer is incremented by 1, and if it exceeds the buffer length, it is taken modulo. The write pointer always points to the address where the next data is written. The initial state of the read pointer is synchronized with the write pointer, that is, the latest data is read in real time. When the translation amount is retrieved, the system shifts the read pointer backward by the address corresponding to the translation amount. At this time, the data read by the read pointer is the current harmonic signal translated forward by the translation amount of sample points. It is equivalent to advancing the current data in the time dimension to align with the reference timestamp of the vibration sensor, achieving reverse time offset compensation of the time series data of the two, laying a time synchronization foundation for the subsequent correlation analysis of fault features. Firstly, the time series of periodic impulse peak points caused by bearing fault in vibration signal is extracted. The periodic vibration impulse of motor end cover is caused by motor bearing fault. The extraction process needs to filter the original vibration signal with band-pass filter first. The filter range is set to 1-3 kHz, which is the typical frequency interval of bearing fault impulse. Low frequency mechanical noise and high frequency electromagnetic interference can be filtered. The threshold value of impulse peak is set to 3 times the average amplitude of the filtered vibration signal, which is used to distinguish fault impulse from ordinary vibration. The periodic impulse peak point is defined as the first point in the signal that exceeds the threshold value, and the amplitude of the subsequent 3 sampling points first increases and then decreases. The sampling point number corresponding to the peak point is recorded. The actual time stamp is calculated combined with the sampling period. All peak points are arranged in chronological order to form a periodic impulse peak time series. Then the energy distribution symmetry of compensated current harmonic signal in a set time range before and after each impulse peak point is calculated. The time range is set to ± 5 sampling points according to the duration of the bearing fault impulse, which corresponds to 0.5 ms, ensuring that the current response before and after the impulse is covered. The compensated current harmonic signal of the first 5 sampling points (t-0.5 ms to t) and the last 5 sampling points (t to t+0.5 ms) is taken with the peak point timestamp as the center. The energy distribution symmetry is realized by calculating the ratio of the energy sum of current harmonic signals in the two time ranges. The current harmonic energy is calculated by summing the square of each sampling point voltage value. The energy is proportional to the square of the voltage. For example, the energy sum of the first 5 sampling points E1=0.8²+0.9²+1.0²+1.1²+1.2², and the energy sum of the last 5 sampling points E2=1.2²+1.1²+1.0²+0.9²+0.8 2, symmetry coefficient = E1 / E2 (the coefficient approaches 1 when ideal symmetry, the closer the coefficient to 1, the more symmetrical the current harmonic energy distribution before and after the pulse peak, indirectly reflecting the higher time alignment, the synchronization error tolerance is set to 1 sampling period, corresponding to the time deviation of 1 sampling point, the tolerance is set based on the sampling accuracy of the sensor, to ensure the accuracy of the fault feature correlation, when the energy peak position (the timestamp corresponding to the maximum current harmonic energy) and the vibration pulse peak timestamp deviation is less than the synchronization error tolerance, it is determined that the alignment is successful, and the subsequent fault feature extraction link can be entered; if the deviation is greater than the tolerance, the shift amount is dynamically corrected according to the deviation direction, if the current energy peak timestamp lags behind the vibration pulse peak (the deviation is positive), it means that the current shift amount is insufficient, and the shift amount needs to be increased, if the current energy peak is ahead (the deviation is negative), the shift amount is reduced, the current data in the ring buffer is read again after correction, the energy symmetry calculation and alignment determination are repeated until the deviation meets the tolerance requirement, to ensure the accurate correlation of the vibration fault pulse and the current harmonic response. After the precise time alignment of the vibration and current data, the feature parameters directly related to the motor fault need to be extracted from the two types of data. The periodic pulse fundamental frequency of the vibration signal can reflect the fault type, and the amplitude fluctuation of the current harmonic can reflect the influence of the fault on the electrical system. Therefore, frequency spectrum analysis and envelope demodulation are needed to extract the features. The specific implementation is as follows: First, the fundamental frequency component of the vibration pulse is extracted as the vibration pulse peak frequency by fast spectrum analysis at the time-aligned vibration pulse peak point. The fast spectrum analysis takes the vibration pulse peak point as the center and takes the previous and next 20 sampling points, corresponding to 2 ms. This time window length can cover the complete pulse waveform and avoid introducing too much irrelevant signal vibration signal as the analysis object. The analysis process first performs windowing processing on the vibration signal in the time window, uses the Hanning window to reduce spectrum leakage, and smoothes the signal at both ends through the weighting coefficient to avoid high-frequency interference caused by sudden changes. Then, the time-domain signal is converted into a frequency-domain spectrum through fast Fourier transform (FFT). The fundamental frequency component refers to the frequency component with the largest amplitude in the spectrum, excluding the direct current component and high-frequency noise. For example, the amplitude corresponding to the 100 Hz frequency in the spectrum is the largest, and this frequency is close to the characteristic frequency of the motor bearing inner ring fault, which is calculated by the motor speed, such as the inner ring fault frequency of about 102 Hz corresponding to the speed of 1500 r / min. Therefore, 100 Hz is determined as the vibration pulse peak frequency. This frequency is directly related to the fault type and provides a core basis for subsequent fault diagnosis. At the same time, in the time window centered on the peak point, the 2 ms time window is consistent with the spectrum analysis, ensuring the uniformity of the time dimension. The amplitude envelope demodulation is performed on the compensated current harmonic signal. The amplitude envelope demodulation is used to extract the amplitude variation trend of the current harmonic signal, eliminate high-frequency carrier interference, and retain the amplitude fluctuation information related to the fault. The specific process is realized by using Hilbert transform. First, the current harmonic signal is subjected to Hilbert transform to obtain the corresponding analytic signal. Then, the analytic signal is subjected to modulus operation to obtain the amplitude envelope curve of the current harmonic signal. The envelope curve smoothly reflects the slow change of the amplitude and filters out the rapid fluctuations of the high-frequency harmonics. If Hilbert transform is not used, envelope demodulation can also be realized by sliding window peak extraction. A 3-sample-point sliding window is used to take the maximum value in each window as the envelope point, and all envelope points are connected in turn to form the amplitude envelope curve. Both methods can effectively extract the amplitude variation characteristics, and the former has higher precision and the latter has smaller calculation amount. The system can be selected according to the algorithm power. Finally, the peak-to-peak value of the current harmonic amplitude in the time window is calculated as the fluctuation quantity characteristic. The peak-to-peak value refers to the difference between the maximum and minimum values of the amplitude envelope curve in the time window. For example, the maximum value of the envelope curve is 1.5 V and the minimum value is 0.8 V. The peak-to-peak value is 1.5 V-0.8 V=0.7 V. This characteristic reflects the degree of current harmonic amplitude fluctuation caused by motor faults. When the motor is running normally, the current harmonic amplitude is stable and the peak-to-peak value is small (e.g., within 0.2 V). When bearing faults or winding faults occur, mechanical vibration caused by faults will cause changes in the electromagnetic induction of the stator winding, resulting in an increase in the current harmonic amplitude fluctuation and a significant increase in the peak-to-peak value (e.g., more than 0.5 V). Therefore, the peak-to-peak value can be used as a quantitative indicator of fault severity, and it can be used in conjunction with the vibration pulse peak frequency to form a complete fault feature set, providing data support for motor fault type identification and fault level evaluation.

[0020] The specific implementation of the current harmonic fluctuation amount analysis triggered by the bearing outer ring fault is as follows: after the basic feature extraction of the vibration pulse peak frequency and the current harmonic fluctuation amount is completed, directional verification needs to be carried out for specific fault types. It has been determined that the vibration pulse peak frequency can be used as the preliminary basis for judging the fault type. When the frequency falls into the characteristic frequency band of the bearing outer ring fault, only the vibration characteristics cannot completely rule out false signals. For example, the low-frequency vibration caused by the loosening of the motor base is easy to be confused with the outer ring fault frequency. Therefore, the depth analysis of the current harmonic amplitude fluctuation amount needs to be combined with the dynamic threshold to realize accurate judgment. The specific implementation is as follows: When the vibration pulse peak frequency is detected to fall into the bearing outer ring fault characteristic frequency band, first of all, the definition logic of the band is clear. The bearing outer ring fault characteristic frequency band is a special frequency range calculated based on the motor speed and bearing structure parameters. For example, the rated speed of the motor is 1500 r / min, which corresponds to 25 revolutions per second, i.e. 25 Hz. The characteristic frequency of the bearing outer ring fault is calculated as 0.3 x motor speed frequency (25 Hz), which is 7.5 Hz in theory. Considering that the motor speed may fluctuate by ±5% in actual operation, the frequency band width is set to 7-8 Hz. If the extracted vibration pulse peak frequency (such as 7.2 Hz) falls within this interval, the current harmonic amplitude fluctuation amount analysis is immediately triggered. The first step in the analysis process is to obtain the current load rate of the motor. The load rate is calculated based on the real-time current feedback data of the motor driver, i.e. the ratio of the actual operating current to the rated current. For example, if the actual operating current is 12 A and the rated current is 20 A, the load rate is 60%. This data is updated every 100 ms to ensure that it reflects the real-time operating state of the motor. The second step is to query the pre-established load rate-current harmonic baseline amplitude table. This table is the baseline data collected by the motor under normal conditions through step load testing (from 20% load rate to 100%, with each load level stable for 30 minutes). For example, the baseline amplitude corresponding to a load rate of 20% is 0.5 V, 40% is 0.6 V, 60% is 0.7 V, 80% is 0.9 V, and 100% is 1.1 V. The data in the table is recalibrated every quarter to avoid baseline drift caused by motor aging. The third step is to multiply the baseline amplitude corresponding to the load rate in the table by a set coefficient. The coefficient is set to 1.4, which is determined by historical fault case statistics. This value can filter normal operating current fluctuations and capture small fluctuations caused by early faults, generating a dynamic threshold (for example, the dynamic threshold corresponding to a load rate of 60% = 0.7 V x 1.4 = 0.98 V). Finally, the current harmonic amplitude fluctuation amount (such as 1.2 V) is compared with the dynamic threshold. If the fluctuation amount exceeds the threshold, it means that the current harmonic fluctuation is not caused by normal load changes, but is related to the mechanical vibration caused by the bearing outer ring fault. It is preliminarily determined that there is a risk of bearing outer ring fault. The specific implementation of the fault characteristic tracking and graded alarm signal generation is as follows: Firstly, the number of set periods for continuous tracking is determined, which is determined according to the period of the vibration pulse, for example, the vibration pulse frequency is 7.5 Hz (corresponding to a period of about 0.133 s), and 5 periods (total duration of about 0.665 s) are set for tracking, which can cover enough fault feature points, and will not delay early fault disposal due to too long tracking time. During the tracking process, the peak frequency of the vibration pulse in each period is extracted in real time, and the frequency stability index is calculated. The frequency stability index is obtained by the statistical variance of the frequency change rate between consecutive pulses. The calculation steps are as follows: first, calculate the frequency change rate of adjacent two pulses, such as the first pulse frequency 7.2 Hz, the second pulse frequency 7.3 Hz, the change rate = (7.3-7.2) / 7.2 ≈ 0.0139; the frequency of the second and the third is 7.25 Hz, the change rate = (7.25-7.3) / 7.3 ≈ -0.0068, and the change rate of 4 groups in 5 periods is calculated in turn); secondly, calculate the statistical variance of the 4 groups of change rates (the smaller the variance, the more stable the frequency change, and the more stable the fault feature), if the variance calculation result is 0.0004, the set frequency stability threshold is 0.001, at this time the stability index (0.0004) is lower than the threshold, it is judged that the frequency is stable; at the same time, the current harmonic fluctuation is continuously monitored in each period, if the fluctuation in 5 periods is more than the dynamic threshold (such as 0.98V, the actual fluctuation is 1.2V, 1.15V, 1.25V, 1.18V, 1.22V respectively), it is indicated that the fault feature exists continuously, and the transient interference is excluded. According to the frequency stability and the continuous situation of the fluctuation exceeding the threshold, a hierarchical alarm signal is generated: if the frequency stability index meets the requirements and the fluctuation exceeds 10%-30% (such as 0.98V-1.27V) of the threshold, a first-level alarm (mild fault, prompting the operation and maintenance personnel to strengthen the monitoring frequency) is generated; more than 30%-50% (1.27V-1.47V) generates a second-level alarm (moderate fault, suggesting to shut down within 24 hours for inspection); more than 50% (>1.47V) generates a third-level alarm (severe fault, triggering an emergency shutdown process), the alarm signal is uploaded to the motor operation and maintenance platform at the same time, including fault type (bearing outer ring fault), current frequency (7.2 Hz), fluctuation (1.2V) and continuous time (0.665s), which provides complete fault information for operation and maintenance. The specific implementation of the dynamic time window length setting directly determines the accuracy and robustness of feature extraction in the process of vibration pulse peak frequency extraction, current harmonic fluctuation analysis and fault feature tracking. If the window length is too short, the complete waveform of the fault pulse may be missed, and if the window length is too long, irrelevant noise may be introduced. Therefore, the time window needs to be dynamically adjusted according to the fault pulse period and signal-to-noise ratio. The specific implementation is as follows: The time window length setting is first based on the fault pulse period corresponding to the current vibration pulse peak frequency. The fault pulse period is the inverse of the vibration pulse peak frequency. For example, if the extracted peak frequency is 7.5 Hz, the period = 1 / 7.5 ≈ 0.133 s, and an integer multiple of the period is taken as the time window reference length. The initial integer multiple is set to 1 (reference window length 0.133 s), which ensures complete coverage of the waveform of a single fault pulse (from the rising edge to the falling edge). The center point of the time window is strictly aligned with the pulse peak timestamp detected by the vibration sensor, i.e., the window start point is offset by half the window length from the peak timestamp, and the window end point is offset by half the window length from the peak timestamp. For example, if the peak timestamp is 10.0 s and the reference window length is 0.133 s, the start point = 10.0 s - 0.133 s / 2 ≈ 9.933 s, and the end point = 10.0 s + 0.133 s / 2 ≈ 10.067 s. This symmetrical setting can balance the waveform data before and after the pulse peak, avoiding feature loss due to window position offset (such as only collecting the waveform after the peak, losing the rising edge fault information). The integer multiple of the window length is automatically adjusted according to the signal-to-noise ratio of the vibration signal. The signal-to-noise ratio is calculated by the ratio of the fault pulse amplitude to the environmental noise amplitude in the vibration signal (e.g., fault pulse amplitude 1.8 V, environmental noise amplitude 0.3 V, signal-to-noise ratio = 6). This ratio is calculated every 50 ms to reflect the current signal quality. When the noise level increases (e.g., signal-to-noise ratio decreases to 3, possibly due to motor load fluctuations or environmental vibration enhancement), the system automatically increases the integer multiple coefficient from 1 to 2 (window length becomes 0.266 s), which includes more fault pulse data (e.g., 2 consecutive pulses) to offset noise interference and improve the robustness of feature extraction. If the noise level decreases (signal-to-noise ratio > 8, signal quality is good), the integer multiple coefficient is reduced from 1 to 0.5 (window length 0.066 s), which reduces the introduction of irrelevant data (e.g., only the core waveform near the pulse peak is retained) and improves the efficiency of feature extraction. For example, when the motor operating environment has short-term strong noise (signal-to-noise ratio 2.5), the window length is automatically adjusted to 0.266 s, and the vibration signal from 9.866 s to 10.132 s is collected. Through the superposition analysis of two pulses, the peak frequency of 7.2 Hz can still be accurately extracted, avoiding frequency misjudgment caused by noise and ensuring stable acquisition of fault features under different signal qualities.

[0021] The application synchronously collects vibration and current sensor original response waveforms by injecting a step excitation signal into the motor, calculates inherent response delay, establishes a mapping relationship library of delay values and current sensor filtering parameters, and real-time retrieves delay values according to filtering parameters to perform reverse time offset compensation on current harmonic signal timing, so that the two types of data are time-aligned, and finally, cross-domain coupling analysis is performed on the aligned data to extract vibration pulse peak frequency and current harmonic fluctuation, combined with bearing outer ring fault characteristic frequency band and load rate dynamic threshold to determine early fault, track features to generate graded alarms, solve the problem of fault misjudgment caused by time and space dislocation of sensor data, and improve the accuracy of early motor fault diagnosis.

[0022] The basic principles, main features and advantages of the present application are shown and described above. Those skilled in the art should understand that the present application is not limited by the above examples, and the above examples and descriptions in the specification are only preferred examples of the present application and are not intended to limit the present application. Without departing from the spirit and scope of the present application, various changes and improvements can be made to the present application, and these changes and improvements all fall within the scope of the claimed present application. The scope of protection of the present application is defined by the appended claims and their equivalents.

Claims

1. A motor fault diagnosis algorithm based on multi-sensor fusion, characterized in that: The method comprises the following steps: S1, by injecting a step excitation signal to the motor system, synchronously collecting the original response waveforms of the vibration sensor and the current sensor, calculating the inherent response delay based on the time difference between the rising edge timestamp of the vibration sensor and the filtered stable value timestamp of the current sensor, and establishing a mapping relationship library of the inherent response delay value and the filtering parameter of the current sensor; S2, in the real-time monitoring stage, the corresponding inherent response delay value is retrieved from the mapping relationship library according to the actual filtering parameter of the current sensor, and the harmonic signal time sequence collected by the current sensor is compensated by reverse time offset, so that the compensated current harmonic feature timestamp is aligned with the vibration sensor pulse peak timestamp; S3, the time-aligned vibration pulse peak frequency and the current harmonic amplitude are coupled and analyzed across domains, and when the vibration pulse peak frequency is in the bearing outer ring fault characteristic frequency band and the fluctuation amount of the current harmonic amplitude in the synchronous time window exceeds the threshold value, the early bearing fault is determined, wherein the time window length is set according to the vibration pulse peak frequency and is centered on the vibration pulse peak point.

2. The multi-sensor fusion based motor fault diagnostic algorithm according to claim 1, characterized in that: In S1, the step excitation signal is injected to the motor system, and the specific steps are as follows: In the no-load running state of the motor, a voltage step signal with a preset time length is injected to the three-phase stator winding through the motor driver, the amplitude of the step signal is set as a certain proportion of the rated voltage, and the duration is greater than the maximum response delay of the current sensor; In the synchronous collection process, a data acquisition card with high-precision clock synchronization function is used to record the original voltage waveform output by the vibration sensor and the original current waveform output by the current sensor at a sampling rate not lower than the highest response frequency of the vibration sensor.

3. The multi-sensor fusion based motor fault diagnostic algorithm according to claim 1, characterized in that: The operation of calculating the inherent response delay based on the time difference between the rising edge timestamp of the vibration sensor and the filtered stable value timestamp of the current sensor comprises the following steps: The starting time point of the rising edge of the step response is extracted from the vibration sensor waveform as the reference timestamp, which is determined by detecting the position where the waveform first exceeds the environmental noise threshold and continuously rises; At the same time, the arrival timestamp of the filtered stable value is extracted from the current sensor waveform, which is defined as the starting time when the current waveform enters the set error band range and maintains for a minimum stable time, and the difference between the two timestamps is taken as the inherent response delay value under the test condition.

4. The multi-sensor fusion based motor fault diagnostic algorithm according to claim 3, characterized in that: The specific steps of establishing the mapping relationship library of the inherent response delay value and the filtering parameter of the current sensor are as follows: The low-pass filter cutoff frequency of the current sensor is dynamically adjusted through a programmable resistance-capacitance network, the step excitation signal is repeatedly injected to the motor system under multiple discrete filtering parameter combinations, the inherent response delay value obtained from each test is mapped with the corresponding filtering parameter, a delay amount lookup table indexed by the filtering time constant is formed, and the lookup table covers all typical configurations in the working parameter range of the current sensor.

5. The multi-sensor fusion based motor fault diagnostic algorithm according to claim 1, characterized in that: The operation of compensating the harmonic signal time sequence collected by the current sensor by reverse time offset in S2 comprises the following steps: According to the real-time acquired current sensor filtering parameter configuration, a corresponding inherent response delay value is retrieved from a mapping relationship library, and a ring data buffer is created to store current harmonic signal original sampling points, and the overall time sequence data is shifted by adjusting the buffer read pointer position, and the shift amount is the sampling point number corresponding to the retrieved inherent response delay value.

6. The multi-sensor fusion based motor fault diagnostic algorithm according to claim 5, characterized in that: The verification mechanism for aligning the compensated current harmonic feature timestamp with the vibration sensor pulse peak timestamp is as follows: After completing the reverse time offset compensation, the periodic pulse peak point time sequence caused by bearing failure in the vibration signal is extracted, the energy distribution symmetry of the compensated current harmonic signal in a set time range before and after each pulse peak point is calculated, and when the deviation of the energy peak position from the vibration pulse peak timestamp is less than the synchronization error tolerance, it is determined that the alignment is successful, otherwise the shift amount is dynamically corrected according to the deviation direction.

7. The multi-sensor fusion based motor fault diagnostic algorithm according to claim 1, characterized in that: The specific steps of cross-domain coupling analysis are as follows: At the vibration pulse peak point of time alignment, the fundamental frequency component of the vibration pulse is extracted as the vibration pulse peak frequency through fast spectrum analysis, and at the same time, the amplitude envelope demodulation of the compensated current harmonic signal is carried out in a time window centered on the peak point, and the peak-to-peak value of the current harmonic amplitude in the time window is calculated as the fluctuation amount feature.

8. The multi-sensor fusion based motor fault diagnostic algorithm according to claim 7, characterized in that: The condition for determining early bearing failure is refined as follows: When it is detected that the vibration pulse peak frequency falls into the bearing outer ring failure characteristic frequency band, the current harmonic amplitude fluctuation amount analysis is triggered, and the fluctuation amount threshold is dynamically adjusted according to the current load rate of the motor, specifically by querying the pre-established load rate-current harmonic baseline amplitude comparison table, multiplying the baseline amplitude under the corresponding load rate in the table by a set coefficient to obtain the dynamic threshold.

9. The multi-sensor fusion based motor fault diagnostic algorithm according to claim 8, characterized in that: It also includes a fault determination verification mechanism: When the fault condition is first detected, the vibration pulse peak frequency appearing in the subsequent set period is continuously tracked, and if the frequency stability index exceeds the set threshold and the current harmonic fluctuation amount continuously exceeds the dynamic threshold, a graded alarm signal is generated, wherein the frequency stability index is obtained by calculating the statistical variance of the frequency change rate between consecutive pulses.

10. The multi-sensor fusion based motor fault diagnostic algorithm according to claim 7, characterized in that: The specific setting of the time window length is as follows: According to the fault pulse period corresponding to the current vibration pulse peak frequency, an integer multiple of the period is taken as the reference length of the time window, and the center point of the time window is strictly aligned with the pulse peak timestamp detected by the vibration sensor, that is, the window starting point is offset by one-half of the window length from the peak timestamp, and the window ending point is offset by one-half of the window length from the peak timestamp, and the integer multiple of the window length is automatically adjusted according to the signal-to-noise ratio, and when the noise level increases, the integer multiple coefficient is increased to improve the feature extraction robustness.

Images