Motor health state diagnosis system based on multi-sensor fusion

Abstract

The invention relates to the technical field of motor health state diagnosis, and particularly discloses a motor health state diagnosis system based on multi-sensor fusion, which comprises a multi-source sensing acquisition module, a high-precision time synchronization module, a dynamic feature extraction module, a self-adaptive fusion decision module and a health state output module, according to the invention, a high-precision time synchronization module is constructed, and a strategy of combining hardware timestamp injection with phase offset estimation based on cross-correlation and rotation modulation compensation is adopted, so that time alignment of three kinds of heterogeneous signals of vibration, current and temperature at a sub-microsecond level is realized; the problem of feature space distortion caused by sensor sampling frequency difference and a mechanical rotation modulation effect under a high-speed working condition is fundamentally solved, and the space-time consistency of original data is remarkably improved.

Description

Technical Field

[0001] This invention belongs to the field of motor health status diagnosis technology, specifically relating to a motor health status diagnosis system based on multi-sensor fusion. Background Technology

[0002] In the field of industrial automation and intelligent operation and maintenance, motors, as core power devices, directly affect the reliability and safety of production systems. With the popularization of predictive maintenance concepts, motor fault diagnosis technology based on sensor data has become a research hotspot. This technology collects multi-source signals such as vibration, current, and temperature to monitor and evaluate the motor's operating status in real time, aiming to identify typical faults such as bearing wear, rotor eccentricity, and winding short circuits at an early stage, thereby avoiding unplanned downtime and safety accidents.

[0003] Multi-sensor fusion-based motor health status diagnosis methods have attracted widespread attention due to their advantages in information complementarity and robustness. This method improves the accuracy and generalization ability of fault identification by simultaneously acquiring sensor signals from different physical dimensions and using data fusion algorithms to extract high-dimensional features. However, under actual high-speed operating conditions, the inconsistent sampling frequencies of different sensors due to differences in hardware characteristics can easily cause microsecond-level time phase shifts.

[0004] Existing technologies generally lack high-precision time alignment mechanisms when processing multi-source heterogeneous sensor data. Especially in high-speed motor rotation scenarios, even small phase deviations between vibration and current signals can be significantly amplified, causing distortion in the feature space and thus misleading the fusion model's judgment of bearing wear status. Furthermore, traditional fusion strategies often employ fixed weights or simple weighted averages, making it difficult to dynamically adapt to changes in operating conditions, further weakening the reliability of diagnostic results. Summary of the Invention

[0005] The purpose of this invention is to provide a motor health status diagnosis system based on multi-sensor fusion, so as to solve the technical problem in the prior art that the feature space is distorted due to the microsecond-level time phase shift of multi-source sensor signals under high-speed conditions, which in turn leads to misjudgment.

[0006] This invention provides a motor health status diagnosis system based on multi-sensor fusion, comprising: The system comprises a multi-source sensor acquisition module, a high-precision time synchronization module, a dynamic feature extraction module, an adaptive fusion decision module, and a health status output module. The multi-source sensor acquisition module synchronously acquires vibration signals, three-phase current signals, and temperature signals during motor operation. The high-precision time synchronization module performs sub-microsecond time alignment processing on the heterogeneous signals output by the multi-source sensor acquisition module. The dynamic feature extraction module extracts time-frequency domain features related to bearing wear, rotor eccentricity, and winding short circuits from the time-aligned signals. The adaptive fusion decision module dynamically adjusts the fusion weights of each feature channel according to the current motor operating conditions and generates a comprehensive health index. The health status output module maps the comprehensive health index to a preset health level and outputs the diagnostic results.

[0007] Furthermore, the multi-source sensing acquisition module includes a triaxial accelerometer mounted on the motor bearing housing, a high-bandwidth current transformer embedded in the power supply line, and a platinum resistance temperature sensor arranged at the end of the stator winding; the sampling frequency of the triaxial accelerometer is not less than 50 kHz, the sampling frequency of the current transformer is not less than 20 kHz, and the sampling frequency of the platinum resistance temperature sensor is not less than 10 Hz; all sensors are synchronously started at the initial sampling time through hardware trigger signals.

[0008] Furthermore, the high-precision time synchronization module includes a timestamp injection unit, a phase offset estimation unit, and a signal resampling unit. The timestamp injection unit injects a unified hardware timestamp into the data streams of each sensor at the beginning of each sampling period. The phase offset estimation unit calculates the maximum correlation lag point between the vibration signal and the current signal based on the cross-correlation function method, and corrects the equivalent time offset caused by the rotation modulation effect by combining the motor speed information. The signal resampling unit uses a cubic spline interpolation algorithm to perform non-uniform resampling of the current signal and the temperature signal based on the sampling clock of the vibration signal, so that the three types of signals are aligned at the sub-microsecond level on the time axis.

[0009] Furthermore, the dynamic feature extraction module includes a vibration feature submodule, a current feature submodule, and a temperature feature submodule. The vibration feature submodule performs empirical mode decomposition on the triaxial vibration signal to extract the energy entropy and kurtosis values ​​of the first three intrinsic mode functions. The current feature submodule performs Park transform on the three-phase current signal to obtain the dq-axis current components and calculates their harmonic distortion rate and negative sequence component amplitude. The temperature feature submodule performs first-order difference processing on the temperature signal to extract the temperature rise rate and steady-state fluctuation standard deviation. All features are updated in real time using a 100-millisecond sliding window.

[0010] Furthermore, the adaptive fusion decision module includes a working condition identification unit, a weight generation unit, and a comprehensive index calculation unit. The working condition identification unit constructs a three-dimensional working condition vector based on real-time speed, load current, and ambient temperature, and determines the current operating mode through a pre-trained working condition classifier. The operating modes include four categories: no-load low speed, rated load medium speed, overload high speed, and start-stop transient. The weight generation unit pre-sets a set of basic weight coefficients for each type of operating mode and dynamically adjusts these coefficients according to the signal-to-noise ratio of the current feature channel. The higher the signal-to-noise ratio, the greater the corresponding weight. The comprehensive index calculation unit performs a weighted summation of the normalized health scores of each feature channel to generate a comprehensive health index between 0 and 100, where 100 represents complete health and 0 represents severe fault.

[0011] Furthermore, the operating condition classifier adopts a support vector machine model, and its training samples are derived from operating condition segments annotated by experts in the historical operating database. The input features are a combination of speed change rate, effective value of load current and temperature gradient. The kernel function of the support vector machine model is a radial basis function, the penalty parameter C is set to 10, and the kernel function width γ is set to 0.1.

[0012] Furthermore, the health status output module divides the comprehensive health index into four levels: when the comprehensive health index is greater than or equal to 90, it is judged as a healthy state; when the comprehensive health index is less than 90 but greater than or equal to 75, it is judged as a slightly deteriorated state; when the comprehensive health index is less than 75 but greater than or equal to 60, it is judged as a moderate fault state; when the comprehensive health index is less than 60, it is judged as a serious fault state. Each level is associated with preset maintenance suggestion instructions, including continued monitoring, planned maintenance, and immediate shutdown.

[0013] Furthermore, the adaptive fusion decision module is also equipped with a feedback correction mechanism; when the health status output module outputs the same level for three consecutive diagnostic cycles and there is no trend of deterioration, the system automatically reduces the weight update frequency of the corresponding feature channel under this condition; when the comprehensive health index is detected to have decreased by more than 15 percentage points within 5 cycles, the system triggers the weight reset process to reinitialize the basic weight coefficient of the weight generation unit.

[0014] Furthermore, the phase offset estimation unit of the high-precision time synchronization module enables the rotational modulation compensation algorithm when the motor speed is higher than 3000 rpm. The algorithm calculates the electrical cycle based on the number of motor pole pairs and the current speed, and uses the first harmonic component caused by the unbalanced force in the vibration signal as the reference phase to perform phase alignment on the same frequency component in the current signal, thereby eliminating the equivalent time drift caused by mechanical-electrical coupling.

[0015] Compared with the prior art, the beneficial effects of the present invention are as follows: 1. This invention constructs a high-precision time synchronization module and adopts a strategy that combines hardware timestamp injection with phase offset estimation based on cross-correlation and rotation modulation compensation. This achieves sub-microsecond time alignment of three heterogeneous signals: vibration, current, and temperature. It fundamentally solves the problem of feature space distortion caused by sensor sampling frequency differences and mechanical rotation modulation effects under high-speed operating conditions, and significantly improves the spatiotemporal consistency of the original data. 2. The adaptive fusion decision module designed in this invention abandons the traditional fixed weight fusion method and instead dynamically adjusts the fusion weights based on the real-time operating condition identification results and the signal-to-noise ratio of each feature channel. This enables the system to automatically focus on the most discriminative feature dimension under different load, speed and environmental conditions, thereby enhancing the diagnostic model's adaptability to complex operating scenarios. 3. The feedback correction mechanism and comprehensive health index graded output strategy introduced in this invention not only realize the quantitative assessment of fault status, but also establish a closed-loop linkage between diagnostic results and maintenance decisions, effectively supporting the accurate execution of predictive maintenance. 4. While ensuring high diagnostic accuracy, this system significantly reduces the risk of false alarms and missed alarms caused by phase mismatch during long-term operation of high-speed motors, providing reliable technical support for the intelligent operation and maintenance of industrial motors. Attached Figure Description

[0016] Figure 1 This is a schematic diagram of the overall technical solution architecture of the motor health status diagnosis system based on multi-sensor fusion proposed in this invention; Figure 2 This is a schematic diagram of the core principle framework of the high-precision time synchronization module in this invention; Figure 3 This is a flowchart of the dynamic feature extraction module in this invention. Figure 4 This is a schematic diagram of the multi-level interaction relationship and data flow of the adaptive fusion decision module in this invention. Detailed Implementation

[0017] Example 1: The overall architecture of the motor health status diagnosis system based on multi-sensor fusion proposed in this invention is shown in the attached figure. Figure 1 To be continued Figure 4 As shown in the attached diagram, this system consists of five core functional units: a multi-source sensor acquisition module, a high-precision time synchronization module, a dynamic feature extraction module, an adaptive fusion decision-making module, and a health status output module. These units are tightly coupled with the control logic via a high-speed data bus, forming a closed-loop, adaptive health status diagnosis system. The following will refer to the attached diagram... Figure 1 To be continued Figure 4 The specific implementation methods, internal structure, data processing flow and interaction mechanism of each module are explained layer by layer.

[0018] First, the multi-source sensing acquisition module is used to acquire key physical quantities reflecting the mechanical, electrical, and thermal states of the motor in real time during operation. This module includes three heterogeneous sensing units: a triaxial accelerometer, a high-bandwidth current transformer, and a platinum resistance temperature sensor. The triaxial accelerometer is mounted on the surface of the motor bearing housing and is used to acquire vibration signals in the X, Y, and Z orthogonal directions. Its sampling frequency is set to 50 kHz to ensure complete capture of high-frequency impact components excited by early bearing wear or rotor eccentricity. The high-bandwidth current transformer is embedded in the three-phase main circuit of the motor power supply line to synchronously acquire three-phase current signals. Its sampling frequency is no less than 20 kHz, sufficient to cover the high-order harmonic components caused by winding short-circuit faults. The platinum resistance temperature sensor is located at the end of the stator winding near the slot opening to monitor the winding temperature rise. Its sampling frequency is 10 Hz, much lower than the previous two, but sufficient to reflect heat accumulation effects and abnormal heat dissipation. All sensors initiate initial sampling via the same hardware trigger signal, generated by the system's master clock and synchronously sent to the front-end conditioning circuits of each sensor at the start of each sampling period. This ensures that the three types of signals have an initial time alignment foundation at the source. The raw analog signals output by each sensor are processed by their respective independent anti-aliasing filters and analog-to-digital converters to form digital data streams, which are then sent to the high-precision time synchronization module via a parallel high-speed serial interface.

[0019] The high-precision time synchronization module eliminates microsecond-level time phase shifts introduced by differences in sensor physical characteristics, inconsistent signal transmission paths, and the modulation effect of high-speed motor rotation, ensuring strict time consistency for the three types of heterogeneous signals in the subsequent feature extraction stage. Please refer to the appendix. Figure 2 This module consists of three parts: a timestamp injection unit, a phase offset estimation unit, and a signal resampling unit. At the beginning of each sampling period, the timestamp injection unit generates a unified hardware timestamp with 64-bit nanosecond precision based on the master clock and embeds it into the header metadata area of ​​each sensor data frame as a reference anchor point for subsequent time alignment. The phase offset estimation unit receives the timestamped vibration signal and current signal, first performs bandpass filtering on both signals to retain the effective frequency band from 0.5 kHz to 10 kHz, and then uses the cross-correlation function method to calculate the maximum correlation hysteresis point between them. Let the vibration signal be... The current signal is The cross-correlation function is then defined as:

[0020] Through search Delay to reach the maximum value This allows for a preliminary estimation of the static phase shift between the two signals. However, under operating conditions where the motor speed exceeds 3000 rpm, the unbalanced force caused by mechanical rotation modulates the phase of the vibration signal, while the current signal on the electrical side is also affected by the same rotational frequency, resulting in an equivalent dynamic shift between the two signals caused by electromechanical coupling. Therefore, the phase shift estimation unit further introduces a rotation modulation compensation algorithm: based on the number of motor pole pairs... With real-time speed (Unit: revolutions per minute), calculate the electrical angular frequency. And extract the first harmonic component from the vibration signal (i.e. The corresponding frequency components are used as the reference phase; simultaneously, the same-frequency components in the d-axis or q-axis components are extracted from the Park transform results of the current signal, and the phase difference between them and the vibration reference phase is calculated. This is then converted into an equivalent time offset. The final total phase shift is The signal resampling unit uses the original sampling clock of the vibration signal as the time reference to perform non-uniform resampling on the current and temperature signals. Because the temperature signal has an extremely low sampling rate, linear interpolation is used for its resampling; while the current signal, due to its higher sampling rate and the need to maintain waveform integrity, is reconstructed using a cubic spline interpolation algorithm. After this processing, the three types of signals achieve sub-microsecond alignment on the time axis, with a typical alignment error of less than 0.8 microseconds, meeting the timing requirements for subsequent high-precision feature extraction.

[0021] The dynamic feature extraction module is responsible for extracting discriminative features strongly correlated with specific fault modes from time-aligned multi-source signals. Please refer to the appendix. Figure 3 The module is divided into three parallel processing channels: vibration feature submodule, current feature submodule and temperature feature submodule. Each submodule updates the feature vector in real time with a sliding window length of 100 milliseconds and a step size of 10 milliseconds.

[0022] The vibration feature submodule receives the triaxially aligned vibration signal, first performs zero-mean normalization and normalization, and then adaptively decomposes the signal into several intrinsic mode functions (IMFs) using the empirical mode decomposition method. For two typical mechanical faults—bearing wear and rotor eccentricity—the system extracts only the first three IMFs (IMF1, IMF2, and IMF3), as they contain the majority of the impact energy and modulation information. For each IMF, its energy entropy and kurtosis are calculated: the energy entropy reflects the degree of energy concentration in the time domain of that mode, and is calculated by first determining the energy distribution of each IMF. energy Then normalize to obtain the probability distribution Finally, calculate the Shannon entropy. The kurtosis value characterizes the impulse intensity of the signal and is defined as the ratio of the fourth central moment to the square of the variance. Finally, the vibration feature submodule outputs a 6-dimensional feature vector: the energy entropy (3-dimensional) and kurtosis value (3-dimensional) of IMF1 to IMF3.

[0023] The current feature submodule processes the three-phase aligned current signal. First, Clarke transformation is performed to obtain the components in the α-β stationary coordinate system, and then the components are determined based on the real-time rotor electrical angle. (Provided by encoder or observer) Perform Park transformation to obtain the DC component in the dq rotating coordinate system. and Winding short-circuit faults typically manifest as current waveform distortion and enhanced negative sequence components. Therefore, this submodule calculates two key indicators: one is the total harmonic distortion (THD) of the dq-axis current, achieved through... and Perform Fast Fourier Transform on each component to extract the fundamental amplitude. With each harmonic amplitude ( =2 to 50), calculate Secondly, the amplitude of the negative sequence current component is determined by constructing a symmetrical component method model to separate the negative sequence component from the three-phase current. The effective value is then taken. Finally, the current feature submodule outputs a 2D feature vector: the THD value and the negative sequence current amplitude.

[0024] The temperature feature submodule receives the aligned temperature sequence. First, a moving average filter is performed to suppress measurement noise, followed by the calculation of the first-order difference. This is used to characterize the temperature rise rate. Within a 100-millisecond window, two statistics are extracted: the maximum or mean of the temperature rise rate (reflecting changes in heat dissipation capacity) and the standard deviation of steady-state temperature fluctuation (reflecting thermal stability). When the motor is in the start-up, shutdown, or load change phase, the temperature rise rate increases significantly; while a local short circuit in the winding can lead to local overheating, causing increased temperature fluctuations. Therefore, the temperature feature submodule outputs a 2D feature vector: the temperature rise rate and the standard deviation of temperature fluctuation.

[0025] At this point, the dynamic feature extraction module has output a total of 10-dimensional original feature vectors, corresponding to health characteristics in the mechanical, electrical, and thermal dimensions, respectively.

[0026] The adaptive fusion decision module is the core intelligent unit of this system. Its goal is to dynamically adjust the fusion weights of each feature channel based on the current operating conditions to generate a comprehensive health index with physical meaning. Please refer to the appendix. Figure 4 The module consists of a three-level structure: a working condition identification unit, a weight generation unit, and a comprehensive index calculation unit, and is equipped with a feedback correction mechanism.

[0027] The operating condition identification unit receives real-time speed signals, effective values ​​of load current, and ambient temperature signals to construct a three-dimensional operating condition vector. ,in Rotational speed (unit: revolutions per minute). This represents the effective value of the three-phase current (unit: ampere). The ambient temperature gradient (unit: degrees Celsius per second) is used as input to a pre-trained operating condition classifier. This classifier employs a support vector machine (SVM) model, and its training samples are derived from operating condition segments annotated by domain experts in a historical operating database. These segments cover four typical operating modes: no-load low speed (speed < 1000 rpm, load < 10% of rated speed), rated load medium speed (speed 1000–3000 rpm, load 80%–100% of rated speed), overload high speed (speed > 3000 rpm, load > 110% of rated speed), and start-stop transient (speed change rate > 500 rpm per second). The SVM uses a radial basis function kernel with a penalty parameter C set to 10 and a kernel width γ set to 0.1. After cross-validation, the classification accuracy reaches 96.7%. The operating condition identification unit outputs the current operating mode label every 100 milliseconds.

[0028] The weight generation unit pre-sets a set of basic weight coefficient vectors for each of the four operating modes mentioned above. ( =1,2,3,4), each vector group contains 10 elements, corresponding to the initial weights of the 10-dimensional features. For example, in the no-load low-speed mode, the vibration feature weight is higher, while the current feature weight is lower; in the overload high-speed mode, the weights of the current harmonics and temperature rise features are increased. Based on this, the weight generation unit further introduces a dynamic signal-to-noise ratio correction mechanism: for each feature channel, its signal-to-noise ratio within the current window is calculated ( ), defined as the ratio of the feature mean to the standard deviation. If a feature channel If the value exceeds a preset threshold (e.g., 10 dB), its weight is increased proportionally; conversely, if it falls below, its weight is decreased. The specific correction formula is as follows:

[0029] in The sensitivity coefficient is set to 0.3. The threshold. The corrected weight vector. After normalization, the sum of all weights is ensured to be 1.

[0030] The comprehensive index calculation unit receives a 10-dimensional normalized health score vector. (Each feature is mapped to the 0–1 range using a minimum-maximum scaling method, where 1 represents the healthiest feature) and the weight vector. Perform a weighted summation operation:

[0031] Generate a comprehensive health index between 0 and 100. This index has a clear physical meaning: 100 represents a brand new, fault-free state, and 0 represents multiple serious faults stacked together.

[0032] The feedback correction mechanism monitors the historical diagnostic results of the health status output module. If the same health level is output for three consecutive diagnostic cycles (i.e., 300 milliseconds) and the comprehensive health index shows no downward trend (absolute value of change rate < 0.5% / cycle), the system determines that the current operating condition is stable and automatically extends the weight update frequency in this mode from 100 milliseconds to 500 milliseconds to reduce computational overhead. Conversely, if the comprehensive health index drops by more than 15 percentage points within five consecutive cycles, it is determined to be a sudden deterioration event, and the system immediately triggers the weight reset process: clearing the current weights and reloading the basic weight coefficients for the corresponding operating condition. It then initiates high-frequency weight updates (once every 20 milliseconds) until the health index stabilizes.

[0033] The health status output module will integrate health indices. The mapping is to four preset health levels, and associated with maintenance recommendation instructions. The specific rules are as follows: When When the value is ≥90, the system is considered healthy, and a "Continue Monitoring" command is output; when the value is ≤75, the system is considered healthy. If the value is less than 90, it is considered a minor degradation state, and a "planned maintenance" command is output, suggesting that the inspection be scheduled during the next downtime window; if the value is less than or equal to 60, it is considered a minor degradation state. When the value is less than 75, the system is classified as a moderate fault and a "strengthen monitoring and prepare for shutdown" command is output; when... When the value is less than 60, it is determined to be a serious fault state, and an "Immediate Stop" command is output. An emergency alarm signal is also sent to the host computer via an industrial communication protocol (such as Modbus TCP). All diagnostic results are recorded in local non-volatile memory and can be uploaded to the cloud operation and maintenance platform via Ethernet interface.

[0034] The entire system runs on an embedded real-time operating system. The main control chip adopts a multi-core ARM Cortex-A72 architecture, equipped with a hardware floating-point unit and DSP instruction set, ensuring that all signal processing and decision-making algorithms are completed within a 100-millisecond cycle. Upon system startup, a self-test program is executed to verify the communication status of each sensor, time synchronization accuracy, and model integrity. During long-term operation, the system supports online model updates: after accumulating sufficient new labeled data, the system can replace the condition classifier or adjust the basic weight coefficients through a secure firmware upgrade mechanism, enabling continuous evolution of diagnostic capabilities.

[0035] In summary, this embodiment eliminates phase mismatch of heterogeneous signals through high-precision time synchronization, focuses on fault-sensitive dimensions through dynamic feature extraction, optimizes the weight of operating condition perception through adaptive fusion decision-making, and constructs a closed-loop maintenance strategy through feedback correction and hierarchical output. Thus, it achieves highly robust and accurate diagnosis of motor health status in harsh industrial environments such as high speed, variable load, and complex electromagnetic interference.

[0036] Example 2: In certain special application scenarios, such as large wind turbines or ship propulsion motors, the operating environment is subject to strong electromagnetic interference, severe mechanical shock, and extreme temperature and humidity changes, which may cause temporary failure of some sensor signals or a sharp drop in the signal-to-noise ratio. To improve the system's fault tolerance under partial sensor failure conditions, this example enhances the adaptive fusion decision module based on Example 1 by introducing a feature channel redundancy evaluation and virtual feature reconstruction mechanism.

[0037] Specifically, the weight generation unit adds a channel reliability assessment subunit. This subunit continuously monitors the data integrity and statistical stability of each feature channel. For vibration feature channels, if any axis signal of the triaxial accelerometer is continuously lost for more than 50 milliseconds, or its kurtosis value abruptly drops to 0 (indicating signal saturation or disconnection), that axis is deemed unreliable; if two or more axes are unreliable, the entire vibration feature submodule is marked as failed. Similarly, if the current feature submodule detects a three-phase current imbalance exceeding 15%, or a non-physical jump in the dq-axis component after Park transformation (e.g., instantaneous amplitude exceeding three times the rated value), the current signal is deemed to be under strong interference, and its feature reliability is downgraded. If the temperature feature submodule does not receive valid data updates within 10 seconds, the sensor is considered offline.

[0038] When a feature channel is deemed unreliable, the system does not discard it directly but instead initiates a virtual feature reconstruction mechanism. This mechanism uses a Generative Adversarial Network (GAN) model trained on historical normal operating data, employing the current features of other reliable channels as conditional inputs to generate virtual feature values ​​for the failed channel. For example, when a vibration signal fails, the system uses five reliable features—current THD, negative sequence component, and temperature rise rate—as conditional vectors, inputting them into a pre-trained vibration feature generator, outputting estimates of IMF energy entropy and kurtosis. This generator undergoes adversarial training offline using multi-source data synchronously collected under normal operating conditions to ensure that the generated features are statistically consistent with the real features.

[0039] The reconstructed virtual features participate in subsequent weighted fusion, but their weight is forcibly limited to within 30% of the base weight to prevent model illusions from causing misjudgments. Simultaneously, the health status output module adds a "Some sensors have failed; results are based on reconstructed features" label to the diagnostic results and recommends that the corresponding sensors be repaired as soon as possible.

[0040] Furthermore, this embodiment expands the input of the operating condition identification unit to a six-dimensional vector, adding derived features such as the effective value of vibration, current imbalance, and the second derivative of the temperature change rate to improve the robustness of operating condition identification when the sensor partially fails. The support vector machine model is also retrained accordingly, incorporating synthetic data samples that include sensor failure scenarios.

[0041] Actual testing showed that even with the complete failure of a single sensor type, the diagnostic accuracy of this embodiment remained above 89%, a decrease of only about 5 percentage points compared to Embodiment 1, significantly outperforming the traditional fixed-weight fusion scheme (which saw a decrease of up to 22%). This enhancement mechanism effectively ensured the continuous availability of the system in harsh industrial environments.

[0042] Example 3: In response to the application requirements of ultra-high speed motors (speed > 10,000 rpm), this example deeply optimizes the high-precision time synchronization module, focusing on solving the problem of aggravated rotation modulation effect caused by extremely high speed.

[0043] In such motors, the electrical frequency can reach hundreds of hertz or even thousands of hertz. The phase relationship between the first harmonic component in the vibration signal and the same-frequency component in the current signal is highly dynamic, making it difficult for traditional phase alignment methods based on a single frequency point to meet sub-microsecond accuracy requirements. Therefore, this embodiment upgrades the phase offset estimation unit to a multi-band joint phase tracking architecture.

[0044] In its implementation, the system no longer relies solely on the first harmonic component, but simultaneously extracts the first, second, and third harmonic components from the vibration signal (corresponding to fault characteristic frequencies such as imbalance, misalignment, and looseness, respectively), and synchronously extracts the corresponding harmonic components from the dq-axis components of the current signal. For each frequency point... ( =1,2,3), calculate the phase difference between the vibration and current signals respectively. And construct the phase consistency cost function:

[0045] By minimizing Solving for the optimal time offset This method utilizes multi-frequency information redundancy to effectively suppress estimation bias caused by noise interference at a single frequency. Experiments show that, under operating conditions of 15,000 rpm, this method further reduces the time alignment error from 0.75 microseconds in Example 1 to 0.32 microseconds.

[0046] Meanwhile, the signal resampling unit is replaced with a fractional delay filter based on Lagrange interpolation. Its interpolation kernel can be dynamically adjusted according to the resampling point position, avoiding the phase nonlinear distortion introduced by cubic splines at extremely high sampling rates. This optimization improves the amplitude-frequency response flatness of the resampled current signal by 40% in the frequency band above 5 kHz.

[0047] With the above improvements, this embodiment achieves a detection rate of 92% for early bearing micro-pitting and 88% for winding inter-turn short circuits in the health diagnosis of ultra-high speed motors, which is significantly better than the existing commercial system (detection rates of 76% and 70%, respectively).

[0048] Through the progressive design of Examples 1 to 3, this invention comprehensively covers the health diagnosis needs of motors ranging from conventional industrial motors to special motors operating under extreme conditions. It has achieved technological breakthroughs in multiple dimensions, such as time synchronization accuracy, feature extraction effectiveness, adaptive fusion decision-making, and system fault tolerance, providing a solid technical foundation for predictive maintenance of motors.

Claims

1. A motor health condition diagnosis system based on multi-sensor fusion, characterized in that, include: Multi-source sensor acquisition module, high-precision time synchronization module, dynamic feature extraction module, adaptive fusion decision module, and health status output module; The multi-source sensor acquisition module is used to synchronously acquire vibration signals, three-phase current signals and temperature signals during motor operation; The high-precision time synchronization module is used to perform sub-microsecond time alignment processing on the heterogeneous signals output by the multi-source sensor acquisition module. The dynamic feature extraction module is used to extract time-frequency domain features related to bearing wear, rotor eccentricity and winding short circuit from the time-aligned signal, respectively. The adaptive fusion decision module is used to dynamically adjust the fusion weights of each feature channel according to the current motor operating conditions and generate a comprehensive health index. The health status output module is used to map the comprehensive health index to a preset health level and output the diagnostic results.

2. The multi-sensor fusion based motor health condition diagnostic system according to claim 1, wherein, The multi-source sensing acquisition module includes a triaxial accelerometer mounted on the motor bearing housing, a high-bandwidth current transformer embedded in the power supply line, and a platinum resistance temperature sensor arranged at the end of the stator winding; the sampling frequency of the triaxial accelerometer is not less than 50 kHz, the sampling frequency of the current transformer is not less than 20 kHz, and the sampling frequency of the platinum resistance temperature sensor is not less than 10 Hz; all sensors are synchronously started at the initial sampling time through hardware trigger signals.

3. The multi-sensor fusion based motor health condition diagnostic system according to claim 1, wherein, The high-precision time synchronization module includes a timestamp injection unit, a phase offset estimation unit, and a signal resampling unit; The timestamp injection unit injects a uniform hardware timestamp into the data stream of each sensor at the beginning of each sampling period; The phase offset estimation unit calculates the maximum correlation hysteresis point between the vibration signal and the current signal based on the cross-correlation function method, and corrects the equivalent time offset caused by the rotational modulation effect by combining the motor speed information. The signal resampling unit uses a cubic spline interpolation algorithm to perform non-uniform resampling of the current and temperature signals based on the sampling clock of the vibration signal, so that the three types of signals are aligned at the sub-microsecond level on the time axis.

4. The multi-sensor fusion based motor health condition diagnostic system according to claim 3, wherein, The phase offset estimation unit enables the rotational modulation compensation algorithm when the motor speed is higher than 3000 rpm; The algorithm calculates the electrical cycle based on the number of pole pairs of the motor and the current speed, and uses the first harmonic component in the vibration signal caused by the unbalanced force as the reference phase to align the phase of the same frequency component in the current signal, thereby eliminating the equivalent time drift caused by mechanical-electrical coupling.

5. The multi-sensor fusion based motor health condition diagnostic system according to claim 1, wherein, The dynamic feature extraction module includes a vibration feature submodule, a current feature submodule, and a temperature feature submodule; The vibration feature submodule performs empirical mode decomposition on the triaxial vibration signal and extracts the energy entropy and kurtosis values ​​of the first three intrinsic mode functions; the current feature submodule performs Park transform on the three-phase current signal, obtains the dq-axis current components, and calculates their harmonic distortion rate and negative sequence component amplitude. The temperature feature submodule performs first-order difference processing on the temperature signal to extract the temperature rise rate and the standard deviation of steady-state fluctuations; all features are updated in real time using a 100-millisecond sliding window.

6. The motor health status diagnosis system based on multi-sensor fusion according to claim 1, characterized in that, The adaptive fusion decision module includes a working condition identification unit, a weight generation unit, and a comprehensive index calculation unit. The operating condition identification unit constructs a three-dimensional operating condition vector based on real-time speed, load current and ambient temperature, and determines the current operating mode through a pre-trained operating condition classifier. The operating modes include four categories: no-load low speed, rated load medium speed, overload high speed and start-stop transient. The weight generation unit pre-sets a set of basic weight coefficients for each type of operating mode and dynamically corrects the coefficients according to the signal-to-noise ratio of the current feature channel. The higher the signal-to-noise ratio, the greater the corresponding weight. The comprehensive index calculation unit performs a weighted summation of the normalized health scores of each feature channel to generate a comprehensive health index between 0 and 100.

7. The multi-sensor fusion based motor health condition diagnostic system according to claim 6, wherein, The operating condition classifier adopts a support vector machine model, and its training samples are derived from operating condition segments annotated by experts in the historical operating database. The input features are a combination of speed change rate, effective value of load current and temperature gradient. The kernel function of the support vector machine model is a radial basis function.

8. The multi-sensor fusion based motor health condition diagnostic system according to claim 1, wherein, The health status output module divides the comprehensive health index into four levels: when the comprehensive health index is greater than or equal to 90, it is judged as a healthy state; when the comprehensive health index is less than 90 but greater than or equal to 75, it is judged as a slightly deteriorated state. When the overall health index is less than 75 but greater than or equal to 60, it is judged as a moderate fault state; when the overall health index is less than 60, it is judged as a severe fault state; each level is associated with preset maintenance suggestion instructions.

9. The motor health status diagnosis system based on multi-sensor fusion according to claim 6, characterized in that, The adaptive fusion decision module is also equipped with a feedback correction mechanism; when the health status output module outputs the same level for three consecutive diagnostic cycles and there is no trend of deterioration, the system automatically reduces the weight update frequency of the corresponding feature channel under this condition; when the comprehensive health index is detected to have decreased by more than 15 percentage points within 5 cycles, the system triggers the weight reset process to reinitialize the basic weight coefficient of the weight generation unit.

10. The multi-sensor fusion based motor health condition diagnostic system according to claim 5, wherein, The vibration feature submodule outputs features including the energy entropy and kurtosis values ​​corresponding to the first three intrinsic mode functions, totaling six-dimensional features; the current feature submodule outputs features including the total harmonic distortion rate of the dq-axis current component and the amplitude of the negative sequence current component, totaling two-dimensional features; the temperature feature submodule outputs features including the temperature rise rate and the standard deviation of steady-state temperature fluctuation, totaling two-dimensional features; the dynamic feature extraction module outputs a total of 10-dimensional original feature vectors.

Images