A water pump motor fault diagnosis method and system based on current harmonic characteristics

Abstract

The application relates to the technical field of electrical engineering and intelligent fault diagnosis, and discloses a water pump motor fault diagnosis method and system based on current harmonic characteristics, which comprises the following steps: collecting stator current, rotating speed and load signals; generating alpha beta current components through Clark transformation, constructing analytical signals, extracting instantaneous phases, and calculating instantaneous frequencies; combining the rotating speed to establish a mapping relationship, resampling the current signals to eliminate frequency dispersion; adopting multi-scale synchronous compression wavelet transformation to obtain a high-resolution harmonic energy distribution diagram; extracting the energy concentration degree, peak value offset and bandwidth expansion coefficient in a characteristic frequency band, inputting a pre-trained multilayer perception classifier, and outputting a fault type. Through high-precision time-frequency focusing and multi-dimensional feature fusion, the application significantly improves fault recognition accuracy and robustness, and can realize online diagnosis of various typical faults only by using current signals.

Description

Technical Field

[0001] This invention belongs to the field of electrical engineering and intelligent fault diagnosis technology, specifically relating to a method and system for diagnosing water pump motor faults based on current harmonic characteristics. Background Technology

[0002] With the rapid development of industrial automation and intelligent operation and maintenance technologies, motors, as core power equipment, are widely used in water pump systems, and their operating status directly affects water supply safety and energy efficiency. Motor faults (such as bearing damage, rotor bar breakage, stator eccentricity, etc.) often manifest as weak current harmonic anomalies in their early stages. Therefore, non-invasive fault diagnosis methods based on current signals have become a research hotspot. This method indirectly identifies mechanical and electrical faults by analyzing the harmonic characteristic frequencies and amplitude changes in the motor stator current, and has advantages such as convenient installation, low cost, and easy integration.

[0003] Fault diagnosis techniques based on current harmonic characteristics rely on spectral analysis of current signals to extract fault-sensitive components. Traditional methods commonly employ Fast Fourier Transform (FFT) for spectral estimation, identifying characteristic frequencies corresponding to specific faults (such as bearing fault frequencies, slip frequency sidebands, etc.) to determine the equipment's health status. However, FFT is limited by its inherent frequency resolution, especially under conditions of low-speed motor operation or frequent load fluctuations. In these conditions, fault characteristic frequencies highly overlap with the fundamental frequency and other harmonic components in the frequency domain, leading to spectral leakage and ambiguity, making it difficult to accurately separate weak fault information.

[0004] Insufficient spectral resolution of the FFT (Fourier Transform) obscures early, weak fault features, significantly reducing diagnostic sensitivity. Feature extraction is highly dependent on expert experience, requiring manual selection of harmonic amplitude, sideband spacing, energy ratio, and other parameters, which is not only inefficient but also susceptible to subjective judgment and has poor generalization ability. Although some studies have attempted to improve resolution by introducing window functions or zero-filling, these methods cannot fundamentally overcome the physical limitations of the Fourier transform. Furthermore, under complex operating conditions, harmonics generated by different fault modes exhibit cross-coupling effects, further exacerbating feature confusion. Summary of the Invention

[0005] This invention provides a method and system for fault diagnosis of water pump motors based on current harmonic characteristics. It aims to solve the technical problem that traditional fast Fourier transform (FFT) suffers from insufficient frequency resolution under low-speed and fluctuating load conditions, leading to aliasing of fault characteristic frequencies with the fundamental frequency and other harmonics, making accurate separation difficult. This method constructs a high-resolution time-frequency analysis model and combines it with motor operating state parameters to dynamically model and decouple harmonic components, achieving accurate identification of early mechanical and electrical faults in water pump motors.

[0006] This invention provides a method for fault diagnosis of a water pump motor based on current harmonic characteristics, comprising: acquiring three-phase stator current signals of the water pump motor, and simultaneously acquiring motor speed signals and load power signals; performing a Clarke transform on the three-phase stator current signals to generate α-axis and β-axis current components in a two-phase stationary coordinate system; constructing analytical signals based on the α-axis and β-axis current components and extracting their instantaneous amplitude and instantaneous phase; calculating an instantaneous frequency sequence using the instantaneous phase and establishing a speed-frequency mapping relationship in conjunction with the motor speed signal; and resampling the current signal according to the speed-frequency mapping relationship to eliminate speed ripples. The frequency dispersion effect caused by the motion is investigated; a multi-scale synchronous compressed wavelet transform is performed on the resampled current signal to obtain a harmonic energy distribution map with high time-frequency focusing capability; the energy concentration, peak offset, and bandwidth expansion coefficient in the characteristic frequency region corresponding to the typical fault mode of the water pump motor are extracted from the harmonic energy distribution map; the energy concentration, peak offset, and bandwidth expansion coefficient are used as fault feature vectors and input to a pre-trained multilayer perceptron classifier to output the fault type judgment result of the water pump motor, wherein the fault types include bearing outer ring damage, bearing inner ring damage, rotor bar breakage, stator winding inter-turn short circuit, and air gap eccentricity.

[0007] Furthermore, the acquisition of the three-phase stator current signal of the water pump motor is specifically achieved by a high-precision Hall current sensor installed in the motor power supply circuit, with a sampling frequency set to 20 kHz and a sampling duration of not less than 10 seconds; the motor speed signal is obtained by an encoder or a back EMF zero-crossing detection circuit, with a resolution of 60 pulses per revolution; the load power signal is measured in real time by a power analyzer, with an update cycle of 10 milliseconds.

[0008] Furthermore, the construction of the analytical signal and extraction of its instantaneous amplitude and instantaneous phase specifically includes: performing a Hilbert transform on the α-axis current component to generate its conjugate signal; combining the original α-axis current component with its conjugate signal to form a complex analytical signal; the instantaneous amplitude is the modulus of the analytical signal, and the instantaneous phase is its principal argument value.

[0009] Furthermore, the calculation of the instantaneous frequency sequence using the instantaneous phase specifically involves: performing a first-order difference operation on the instantaneous phase sequence, dividing it by the sampling time interval, and then performing low-pass filtering to suppress differential noise, thereby obtaining a smooth instantaneous frequency time sequence.

[0010] Furthermore, the establishment of the speed-frequency mapping relationship specifically includes: aligning the instantaneous frequency sequence with the synchronously acquired speed signal in time; calculating the theoretical value of the fundamental frequency corresponding to a unit speed, i.e., the inverse of the pole pair multiplied by 50 Hz; and establishing a linear proportionality coefficient between the measured instantaneous frequency and the theoretical fundamental frequency through least squares fitting, which is used for subsequent frequency normalization.

[0011] Furthermore, the resampling of the current signal to eliminate the frequency dispersion effect caused by speed fluctuation specifically includes: re-interpolating the time axis of the original current signal according to the linear scaling factor at constant electrical angle intervals; the interpolation method adopts cubic spline interpolation, and the length of the resampled signal remains consistent with the original signal, but its spectrum is no longer broadened by speed fluctuation.

[0012] Furthermore, the specific implementation steps of the multi-scale synchronous compressed wavelet transform include: selecting the Moray wavelet as the mother wavelet, setting the scale range to cover from the fundamental frequency to 15 times the fundamental frequency; calculating the complex conjugate gradient of the wavelet coefficients at each scale; compressing the energy of the wavelet coefficients along the frequency axis according to the gradient direction, so that the originally diffuse energy is concentrated on the real instantaneous frequency trajectory; and finally synthesizing all the compressed wavelet coefficients to form a high-resolution time-frequency representation matrix.

[0013] Furthermore, the extraction of energy concentration, peak offset, and bandwidth expansion coefficient within the characteristic frequency region specifically includes: defining a characteristic frequency band on the normalized frequency axis for each preset fault mode, for example, a bearing outer ring fault corresponds to 0.4 to 0.6 times the fundamental frequency, and a rotor bar breakage corresponds to 1.8 to 2.2 times the fundamental frequency; summing the frequency slices at all time points within each characteristic frequency band in the time-frequency representation matrix to obtain the total energy of that frequency band; calculating the absolute deviation between the frequency position corresponding to the maximum energy value within that frequency band and the theoretical fault frequency, as the peak offset; calculating the width covered by all frequency points within that frequency band whose energy exceeds 30% of the maximum value, as the bandwidth expansion coefficient; and defining the energy concentration as the ratio of the total energy of that frequency band to the sum of the total energy of the three adjacent non-fault frequency bands.

[0014] Furthermore, the pre-trained multilayer perceptron classifier includes an input layer, two hidden layers, and an output layer; the input layer has 15 neurons, corresponding to three feature parameters for each of the five fault types; the first hidden layer contains 30 neurons with a modified linear unit activation function; the second hidden layer contains 20 neurons with a hyperbolic tangent activation function; the output layer contains 5 neurons, corresponding to the five fault types respectively, and uses a soft maximum function to output the probability distribution; the classifier is trained under supervision using a historical fault sample dataset, which contains at least 500 sets of labeled current signals under different operating conditions and their corresponding real fault labels.

[0015] This invention also provides a water pump motor fault diagnosis system based on current harmonic characteristics, comprising: a current and state parameter acquisition module for acquiring the three-phase stator current signal of the water pump motor and simultaneously acquiring the motor speed signal and load power signal; a coordinate transformation and analytical signal generation module for performing Clarke transform on the stator three-phase current signal to generate α-axis and β-axis current components in a two-phase stationary coordinate system, and constructing an analytical signal based on the α-axis current component to extract instantaneous amplitude and instantaneous phase; an instantaneous frequency calculation and mapping module for calculating the instantaneous frequency sequence using the instantaneous phase and establishing a speed-frequency mapping relationship in conjunction with the motor speed signal; and a speed fluctuation compensation resampling module for... Based on the speed-frequency mapping relationship, the current signal is resampled to eliminate the frequency dispersion effect caused by speed fluctuations; a high-resolution time-frequency analysis module is used to perform multi-scale synchronous compressed wavelet transform on the resampled current signal to obtain a harmonic energy distribution map with high time-frequency focusing capability; a fault feature extraction module is used to extract the energy concentration, peak offset, and bandwidth expansion coefficient in the characteristic frequency region corresponding to the typical fault mode of the water pump motor from the harmonic energy distribution map; a fault type determination module is used to input the energy concentration, peak offset, and bandwidth expansion coefficient as fault feature vectors into a pre-trained multilayer perceptron classifier and output the fault type determination result of the water pump motor.

[0016] Furthermore, the current and status parameter acquisition module includes three Hall current sensors, an incremental photoelectric encoder, and a digital power meter; the Hall current sensors are installed on the three-phase power supply line of the motor, with a range of 0 to 100 amperes and a bandwidth of 0 to 25 kHz; the incremental photoelectric encoder is installed on the output shaft of the motor, with a resolution of 60 lines per revolution; the digital power meter is connected to the power supply circuit through a voltage transformer and a current transformer, with a sampling rate of 1000 Hz.

[0017] Furthermore, the coordinate transformation and analytical signal generation module, the instantaneous frequency calculation and mapping module, the speed fluctuation compensation resampling module, the high-resolution time-frequency analysis module, the fault feature extraction module, and the fault type determination module are all integrated into the embedded signal processing unit; the embedded signal processing unit adopts a digital signal processor chip with a main frequency of 800 MHz, is equipped with 512 megabytes of synchronous dynamic random access memory, runs a real-time operating system, and has a task scheduling cycle of 20 milliseconds.

[0018] Furthermore, when performing multi-scale synchronous compressed wavelet transform, the high-resolution time-frequency analysis module pre-configures the mother wavelet parameters, scale range, and compression threshold. The scale range is dynamically adjusted according to the number of motor pole pairs and rated speed to ensure coverage of all potential fault harmonic components. The compression threshold is set to 5% of the maximum amplitude of the wavelet coefficients, and coefficients below this threshold do not participate in frequency redistribution.

[0019] Furthermore, when defining the characteristic frequency band, the fault feature extraction module predetermines the theoretical characteristic frequency of various faults based on the number of pole pairs, rated speed, and bearing geometry in the motor nameplate parameters, using the fault frequency calculation formula, and then extends the tolerance band by ±10% as the actual analysis frequency band.

[0020] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0021] 1. This invention effectively eliminates the harmonic frequency dispersion phenomenon caused by motor speed fluctuations by introducing a resampling mechanism based on instantaneous frequency estimation, fundamentally improving the clarity of spectrum analysis.

[0022] 2. The multi-scale synchronous compressed wavelet transform is used to replace the traditional fast Fourier transform. While preserving time-domain information, it achieves sub-hertz-level frequency resolution, which enables the weak fault harmonic components adjacent to the fundamental wave to be clearly distinguished even under low-speed operating conditions.

[0023] 3. By constructing a multi-dimensional fault feature vector that integrates energy concentration, peak offset, and bandwidth expansion coefficient, the shortcomings of single-image criteria being susceptible to load interference are overcome, and the robustness and accuracy of fault diagnosis are significantly improved.

[0024] 4. The entire diagnostic process does not require the installation of additional vibration sensors. It can identify a variety of typical faults by relying solely on current signals, which reduces system costs and maintenance complexity. It is suitable for long-term online monitoring and early warning of water pump motors in industrial sites. Attached Figure Description

[0025] Figure 1 This is a schematic diagram of the overall technical solution architecture of the present invention;

[0026] Figure 2 This is a schematic diagram of the core principle framework of the present invention, which uses multi-scale synchronous compressed wavelet transform and speed fluctuation compensation to achieve high-resolution harmonic feature extraction.

[0027] Figure 3 This is a flowchart of the steps in this invention;

[0028] Figure 4 This is a schematic diagram of the multi-level interaction relationship and data flow between the various functional modules of the embedded diagnostic system and the external sensing unit in this invention. Detailed Implementation

[0029] Please refer to Figures 1 to 4This invention provides a method and system for fault diagnosis of water pump motors based on current harmonic characteristics. It aims to solve the technical problem that traditional fast Fourier transform (FFT) suffers from insufficient frequency resolution under low-speed and fluctuating load conditions, leading to aliasing of fault characteristic frequencies with the fundamental frequency and other harmonics, making accurate separation difficult. This method constructs a high-resolution time-frequency analysis model and combines it with motor operating state parameters to dynamically model and decouple harmonic components, achieving accurate identification of early mechanical and electrical faults in water pump motors.

[0030] The method includes the following steps:

[0031] S1, acquire the three-phase current signal of the water pump motor stator, and simultaneously acquire the motor speed signal and load power signal;

[0032] S2, perform Clarke transform on the stator three-phase current signal to generate α-axis and β-axis current components in a two-phase stationary coordinate system;

[0033] S3, Based on the α-axis and β-axis current components, construct an analytical signal and extract its instantaneous amplitude and instantaneous phase;

[0034] S4, calculate the instantaneous frequency sequence using the instantaneous phase, and establish a speed-frequency mapping relationship by combining it with the motor speed signal;

[0035] S5. Based on the speed-frequency mapping relationship, the current signal is resampled to eliminate the frequency dispersion effect caused by speed fluctuation;

[0036] S6 performs multi-scale synchronous compressed wavelet transform on the resampled current signal to obtain a harmonic energy distribution map with high time-frequency focusing capability;

[0037] S7. Extract the energy concentration, peak offset and bandwidth expansion coefficient of the characteristic frequency region corresponding to the typical fault mode of the water pump motor from the harmonic energy distribution map.

[0038] S8, the energy concentration, peak offset and bandwidth expansion coefficient are used as fault feature vectors and input to the pre-trained multilayer perceptron classifier, and the fault type judgment result of the water pump motor is output. The fault types include bearing outer ring damage, bearing inner ring damage, rotor bar breakage, stator winding turn short circuit and air gap eccentricity.

[0039] In step S1, the acquisition of the three-phase stator current signal of the water pump motor is specifically achieved through a high-precision Hall current sensor installed in the motor power supply circuit. The sampling frequency is set to 20 kHz, and the sampling duration is no less than 10 seconds. The Hall current sensor has a range of 0 to 100 amperes and a bandwidth covering 0 to 25 kHz, ensuring complete capture of the fundamental wave and its higher harmonic components. The synchronously acquired motor speed signal is obtained through an incremental photoelectric encoder or a back EMF zero-crossing detection circuit, with a resolution of 60 pulses per revolution and a timestamp accuracy better than 100 microseconds. The load power signal is measured in real time by a digital power meter with an update cycle of 10 milliseconds. This power meter is connected to the power supply circuit through a voltage transformer and a current transformer, with a sampling rate of 1000 Hz. The output includes parameters such as active power, reactive power, and power factor, among which only active power is used as the load power signal. All acquired signals are time-synchronized through the same high-precision clock source to ensure that each signal is strictly aligned on the time axis, with a maximum synchronization error not exceeding the current sampling period, i.e., 50 microseconds.

[0040] In step S2, the stator three-phase current signal is subjected to a Clarke transform to generate α-axis and β-axis current components in a two-phase stationary coordinate system. Let the original three-phase current signals be... , , ,satisfy The constraints. The Clark transform formula is defined as:

[0041]

[0042]

[0043] This transformation projects the three-phase currents onto a two-dimensional orthogonal coordinate system fixed to the stator, preserving all current information while reducing data dimensionality. The transformed... and All signals are real-valued continuous-time signals, synchronously sampled in the discrete domain at a frequency of 20 kHz, forming a discrete sequence of no less than 200,000 points. The purpose of this step is to simplify the subsequent analytical signal construction process, as only a single component needs to be processed to characterize the harmonic characteristics of the entire three-phase system.

[0044] In step S3, based on the α-axis and β-axis current components, an analytical signal is constructed and its instantaneous amplitude and instantaneous phase are extracted. In specific implementation, only... The components are used to construct an analytical signal. First, the components are analyzed... Perform a Hilbert transform to generate its conjugate signal. The Hilbert transform is implemented using a finite impulse response (FIR) filter of order 496, with a passband range of 0 to 9000 Hz and a stopband attenuation greater than 80 B, ensuring phase linearity and no significant group delay distortion. Subsequently, the original... Combined with its conjugate signal, it forms a complex analytic signal z(t):

[0045]

[0046] Where j is the imaginary unit. The instantaneous amplitude A(t) is defined as the modulus of the analytic signal:

[0047]

[0048] The instantaneous phase φ(t) is its principal argument value, and its range is limited to between negative π and positive π:

[0049]

[0050] The atan2 function ensures that the phase changes continuously across the entire quadrant, avoiding phase jumps caused by arctangent truncation. Both the instantaneous amplitude and instantaneous phase sequences are output at a frequency of 20 kHz for subsequent instantaneous frequency calculations.

[0051] In step S4, the instantaneous frequency sequence is calculated using the instantaneous phase, and a speed-frequency mapping relationship is established by combining it with the motor speed signal. Instantaneous frequency Obtained by performing a first-order difference operation on the instantaneous phase sequence:

[0052]

[0053] Where Ts is the current sampling period, equal to 50 microseconds. Since differential operations amplify high-frequency noise, the resulting instantaneous frequency sequence needs to be low-pass filtered. The low-pass filter used is a sixth-order Butterworth filter with a cutoff frequency set to 500 Hz, effectively suppressing noise introduced by the derivative while preserving the frequency variation trend caused by speed fluctuations. The filtered instantaneous frequency sequence is a smooth continuous-time function, reflecting the real-time changes in the motor's electrical frequency.

[0054] The synchronously acquired rotational speed signal r(t) is input in the form of 60 pulses per revolution. First, a continuous rotational speed curve is obtained by interpolation using pulse counting and timestamps. This curve is then converted into mechanical angular velocity. Then, based on the number of motor pole pairs Calculate the theoretical electrical frequency :

[0055]

[0056] This theoretical frequency represents the frequency position of the fundamental wave under ideal no-slip conditions. The measured instantaneous frequency... with f After time alignment, the least squares method is used to fit the linear relationship between the two:

[0057]

[0058] in This is the proportionality coefficient. This is the bias term. Under steady-state conditions, Approaching zero This reflects the actual slip rate. This mapping is used for subsequent frequency normalization and resampling compensation.

[0059] In step S5, the current signal is resampled according to the speed-frequency mapping relationship to eliminate the frequency dispersion effect caused by speed fluctuations. Specifically, the electrical angle at each moment is first calculated based on the fitted proportional coefficient k. :

[0060]

[0061] The integral is performed in the discrete domain using the trapezoidal rule, with the initial phase set to zero. Then, the original current signal... The time axis is spaced according to a constant electrical angle. Re-interpolate. Set as Divide by the length N of the resampled signal, where N is the original signal length to maintain data consistency. Cubic spline interpolation is used because its second derivative continuity effectively preserves the smoothness of the signal waveform and the integrity of the harmonic structure. The resampled signal... With uniform distribution in the electrical angular domain, its spectrum is no longer broadened by rotational speed fluctuations, and all harmonic components appear as sharp spectral lines in the frequency domain, laying the foundation for high-resolution time-frequency analysis.

[0062] In step S6, a multi-scale synchronous compressed wavelet transform is performed on the resampled current signal to obtain a harmonic energy distribution map with high time-frequency focusing capability. The mother wavelet selected is the Moray wavelet, which has excellent time-frequency localization performance and is suitable for harmonic transient feature extraction. Scale range to Dynamically set based on the motor's rated frequency f_n and the number of pole pairs p:

[0063]

[0064] Where f_s is the effective sampling frequency of the resampled signal, equal to , This represents the original sampling duration. Scale discretization uses logarithmic intervals, with a total of 128 scale points. At each scale... Calculate the wavelet transform coefficients. :

[0065]

[0066] in For the Moray wavelet function, This represents complex conjugate. The complex conjugate gradient ∇ is then calculated. The direction of this gradient indicates the local flow of energy in the time-frequency plane. The synchronous compression operation, based on this gradient direction, redistributes the wavelet coefficients along the frequency axis to the most probable true instantaneous frequency positions. The compression threshold is set to 5% of the maximum amplitude of the wavelet coefficients across all scales; coefficients below this threshold are considered noise and do not participate in the redistribution. Finally, all compressed wavelet coefficients are synthesized to form a time-frequency representation matrix. Its frequency resolution can reach the sub-hertz level, which is significantly better than the traditional fast Fourier transform.

[0067] In step S7, the energy concentration, peak offset, and bandwidth expansion coefficient within the characteristic frequency region corresponding to the typical fault modes of the water pump motor are extracted from the harmonic energy distribution map. First, based on the pole pair number p, rated speed n_r, and bearing geometric dimensions D, d, α in the motor nameplate parameters, the theoretical characteristic frequencies of various faults are pre-calculated using the standard fault frequency formula. For example, the failure frequency of the bearing outer ring is:

[0068]

[0069] in For the number of rolling elements, The rotational frequency is [value missing]. The characteristic frequency of the broken rotor bar is [value missing]. s is the slip ratio. The stator winding inter-turn short circuit exhibits sideband modulation with a center frequency of... Air gap eccentricity will trigger Harmonics, where m is an integer. For each type of fault, a characteristic frequency band on the normalized frequency axis is defined. For example, a bearing outer ring fault corresponds to 0.4 to 0.6 times the fundamental frequency, and a broken rotor bar corresponds to 1.8 to 2.2 times the fundamental frequency.

[0070] In the time-frequency representation matrix In the middle, for each characteristic frequency band [ , Summing the frequency slices at all time points within a given range yields the total energy of that frequency band. :

[0071]

[0072] The three adjacent non-faulty frequency bands are defined as and ,in It is 10% of the fundamental frequency. Calculate the sum of the total energy of these three frequency bands. Energy concentration Defined as:

[0073]

[0074] Peak offset The frequency corresponding to the maximum energy within this frequency band Compared with theoretical failure frequency Absolute deviation:

[0075]

[0076] The bandwidth expansion factor B is defined as the width covered by all frequency points within the band whose energy exceeds 30% of the maximum value.

[0077]

[0078] in and They are respectively satisfied The highest and lowest frequencies. These three parameters are extracted for each type of fault, forming a 15-dimensional fault feature vector.

[0079] In step S8, the energy concentration, peak offset, and bandwidth expansion coefficient are used as fault feature vectors and input to a pre-trained multilayer perceptron classifier, which outputs the fault type determination result of the water pump motor. The multilayer perceptron includes an input layer, two hidden layers, and an output layer. The input layer has 15 neurons, corresponding to three feature parameters for each of the five fault types. The first hidden layer contains 30 neurons, with a modified linear unit activation function, and its output... The second hidden layer contains 20 neurons, with the hyperbolic tangent activation function, and the output... The output layer contains 5 neurons, each corresponding to one of the five fault types, and uses a soft maximum function to output the probability distribution:

[0080]

[0081] in The classifier is trained under supervised supervision using a historical fault sample dataset. This dataset contains at least 500 sets of labeled current signals and their corresponding real fault labels under different operating conditions, covering normal conditions and five types of faults. Training employs a cross-entropy loss function and an adaptive moment estimation optimizer, with an initial learning rate of 0.001, a batch size of 32, and 500 training epochs. After training, the model is deployed on an embedded platform, with an inference latency of less than 10 milliseconds.

[0082] The system includes a current and state parameter acquisition module, a coordinate transformation and analytical signal generation module, an instantaneous frequency calculation and mapping module, a speed fluctuation compensation and resampling module, a high-resolution time-frequency analysis module, a fault feature extraction module, and a fault type determination module. The current and state parameter acquisition module includes three Hall current sensors, an incremental photoelectric encoder, and a digital power meter. The Hall current sensors are installed on the three-phase power supply line of the motor, with a range of 0 to 100 amperes and a bandwidth of 0 to 25 kHz. The incremental photoelectric encoder is installed on the motor output shaft, with a resolution of 60 lines per revolution. The digital power meter is connected to the power supply circuit through voltage and current transformers, with a sampling rate of 1000 Hz.

[0083] The coordinate transformation and analytical signal generation module, instantaneous frequency calculation and mapping module, speed fluctuation compensation and resampling module, high-resolution time-frequency analysis module, fault feature extraction module, and fault type determination module are all integrated into the embedded signal processing unit. This embedded signal processing unit uses an 800 MHz digital signal processor chip, is equipped with 512 MB of synchronous dynamic random access memory, runs a real-time operating system, and has a task scheduling cycle of 20 milliseconds. When performing multi-scale synchronous compressed wavelet transform, the high-resolution time-frequency analysis module pre-configures the mother wavelet parameters, scale range, and compression threshold. The scale range is dynamically adjusted according to the number of motor pole pairs and rated speed to ensure coverage of all potential fault harmonic components. The compression threshold is set to 5% of the maximum amplitude of the wavelet coefficients; coefficients below this threshold do not participate in frequency redistribution. When defining the characteristic frequency band, the fault feature extraction module pre-determines the theoretical characteristic frequencies of various faults based on the number of pole pairs, rated speed, and bearing geometry in the motor nameplate parameters, using the fault frequency calculation formula, and then extends this by a tolerance band of ±10% as the actual analysis frequency band.

[0084] The entire diagnostic process requires no additional vibration sensors; it relies solely on current signals to identify various typical faults, reducing system costs and maintenance complexity. It is suitable for long-term online monitoring and early warning of water pump motors in industrial settings. By introducing a resampling mechanism based on instantaneous frequency estimation, the harmonic frequency dispersion caused by motor speed fluctuations is effectively eliminated, fundamentally improving the clarity of spectrum analysis. The use of multi-scale synchronous compressed wavelet transform instead of traditional fast Fourier transform achieves sub-Hertz frequency resolution while preserving time-domain information, enabling clear differentiation of weak fault harmonic components adjacent to the fundamental frequency even at low speeds. Furthermore, by constructing a multi-dimensional fault feature vector that integrates energy concentration, peak offset, and bandwidth expansion coefficient, the vulnerability of single-frame criteria to load interference is overcome, significantly improving the robustness and accuracy of fault diagnosis.

Claims

1. A method for diagnosing water pump motor faults based on current harmonic characteristics, characterized in that, include: Collect the three-phase current signal of the water pump motor stator, and simultaneously acquire the motor speed signal and load power signal; The stator three-phase current signal is subjected to Clarke transform to generate α-axis and β-axis current components in a two-phase stationary coordinate system. Based on the α-axis and β-axis current components, an analytical signal is constructed and its instantaneous amplitude and instantaneous phase are extracted, specifically including: Perform a Hilbert transform on the α-axis current component to generate its conjugate signal; The original α-axis current component is combined with its conjugate signal to form an analytic signal in complex form; The instantaneous amplitude is the modulus of the analyzed signal, and the instantaneous phase is its principal argument value; The instantaneous frequency sequence is calculated using the instantaneous phase, and a speed-frequency mapping relationship is established by combining it with the motor speed signal, specifically as follows: The instantaneous phase sequence is subjected to a first-order difference operation, divided by the sampling time interval, and then low-pass filtered to suppress differential noise, resulting in a smooth instantaneous frequency time sequence. The instantaneous frequency sequence is time-aligned with the synchronously acquired rotational speed signal; Calculate the theoretical value of the fundamental frequency corresponding to a unit rotational speed, which is the reciprocal of the number of pole pairs multiplied by 50 Hz; A linear proportionality coefficient between the measured instantaneous frequency and the theoretical fundamental frequency is established by least squares fitting, which is used for subsequent frequency normalization. Based on the aforementioned speed-frequency mapping relationship, the current signal is resampled to eliminate the frequency dispersion effect caused by speed fluctuations, specifically including: Based on the linear scaling factor, the time axis of the original current signal is re-interpolated at constant electrical angle intervals; The interpolation method uses cubic spline interpolation, and the length of the resampled signal remains consistent with the original signal, but its spectrum is no longer broadened by the speed fluctuation. Perform multi-scale synchronous compressed wavelet transform on the resampled current signal to obtain a harmonic energy distribution map with high time-frequency focusing capability. The specific implementation steps include: The Morley wavelet was selected as the mother wavelet, and the scale range was set to cover from the fundamental frequency to 15 times the fundamental frequency. Calculate the complex conjugate gradient of the wavelet coefficients at each scale; The wavelet coefficients are compressed along the frequency axis according to the gradient direction, so that the originally diffuse energy is concentrated on the real instantaneous frequency trajectory. Finally, all compressed wavelet coefficients are synthesized to form a high-resolution time-frequency representation matrix; The energy concentration, peak offset, and bandwidth expansion coefficient within the characteristic frequency region corresponding to the typical fault modes of the water pump motor are extracted from the harmonic energy distribution map, specifically including: For each type of preset fault mode, define its characteristic frequency band on the normalized frequency axis; In the time-frequency representation matrix, the total energy of the frequency band is obtained by summing the frequency slices at all time points within each characteristic frequency band. Calculate the absolute deviation between the frequency position corresponding to the maximum energy value within the frequency band and the theoretical fault frequency, and use it as the peak offset. Calculate the width covered by all frequency points within the band whose energy exceeds 30% of the maximum value, and use this as the bandwidth expansion factor; Energy concentration is defined as the ratio of the total energy of the frequency band to the sum of the total energy of the three adjacent non-faulty frequency bands; The energy concentration, peak offset, and bandwidth expansion coefficient are used as fault feature vectors and input to a pre-trained multilayer perceptron classifier to output the fault type determination result of the water pump motor. The fault types include bearing outer ring damage, bearing inner ring damage, rotor bar breakage, stator winding inter-turn short circuit, and air gap eccentricity. A pre-trained multilayer perceptron classifier consists of an input layer, two hidden layers, and an output layer; The input layer has 15 neurons, corresponding to three feature parameters for each of the five types of faults; The first hidden layer contains 30 neurons, and the activation function is the modified linear unit; The second hidden layer contains 20 neurons, and the activation function is the hyperbolic tangent function; The output layer contains 5 neurons, each corresponding to one of the five fault types, and uses a soft maximum function to output the probability distribution. The classifier is trained under supervision using a historical fault sample dataset, which contains at least 500 sets of labeled current signals under different operating conditions and their corresponding real fault labels.

2. The water pump motor fault diagnosis method based on current harmonic characteristics according to claim 1, characterized in that, The acquisition of the three-phase stator current signal of the water pump motor is achieved by a high-precision Hall current sensor installed in the motor power supply circuit; the motor speed signal is obtained by an encoder or a back EMF zero-crossing detection circuit; and the load power signal is measured in real time by a power analyzer.

3. A water pump motor fault diagnosis system based on current harmonic characteristics, characterized in that, Using the method described in any one of claims 1 to 2 to perform fault diagnosis of a water pump motor includes: The current and status parameter acquisition module is used to acquire the three-phase current signal of the water pump motor stator, and simultaneously acquire the motor speed signal and load power signal; The coordinate transformation and analytical signal generation module is used to perform Clark transformation on the stator three-phase current signal to generate α-axis and β-axis current components in a two-phase stationary coordinate system, and to construct an analytical signal based on the α-axis current component to extract the instantaneous amplitude and instantaneous phase. The instantaneous frequency calculation and mapping module is used to calculate the instantaneous frequency sequence using the instantaneous phase and establish a speed-frequency mapping relationship in combination with the motor speed signal; The speed fluctuation compensation resampling module is used to resample the current signal according to the speed-frequency mapping relationship to eliminate the frequency dispersion effect caused by speed fluctuation; The high-resolution time-frequency analysis module is used to perform multi-scale synchronous compressed wavelet transform on the resampled current signal to obtain a harmonic energy distribution map with high time-frequency focusing capability. The fault feature extraction module is used to extract the energy concentration, peak offset and bandwidth expansion coefficient in the characteristic frequency region corresponding to the typical fault mode of the water pump motor from the harmonic energy distribution map. The fault type determination module is used to input the energy concentration, peak offset and bandwidth expansion coefficient as fault feature vectors into a pre-trained multilayer perceptron classifier and output the fault type determination result of the water pump motor.

Images