Electric bicycle motor health evaluation system based on multi-sensor fusion

Abstract

The application discloses a motor health evaluation system for electric bicycles based on multi-sensor fusion, and particularly relates to the technical field of electric bicycles, which comprises an intelligent sensor acquisition module for acquiring armature current, terminal voltage, winding temperature and magnetic field change; a data fusion processing module for generating a fusion signal sequence representing motor transient behavior through time domain-frequency domain collaborative fusion based on an adaptive weighted multi-source fusion algorithm; a motor health state modeling module for constructing a multi-physical field coupling health evaluation model of electromagnetic performance, thermal-electric coupling and vibration response, and generating a motor health measurement index through cross-consistency correction; and an adaptive diagnosis output module for generating a health grade and triggering a dynamic threshold recalibration mechanism in an abnormal state. The application realizes multi-dimensional and high-precision motor health evaluation, adapts to full-load working conditions and the whole life cycle of the motor, reduces the false alarm rate, provides accurate basis for maintenance, and solves the problems of single traditional evaluation method and poor adaptability.

Description

Technical Field

[0001] This invention relates to the field of electric bicycle technology, and more specifically, to an electric bicycle motor health assessment system based on multi-sensor fusion. Background Technology

[0002] As a convenient means of transportation, electric bicycles rely on their motors as core power components. The operating status of the motor directly affects the safety and reliability of the entire vehicle. With the increase in service life and the influence of complex working conditions (such as climbing, heavy loads, and frequent start-stop), motors are prone to failures such as winding aging, bearing wear, magnetic field distortion, and overheating. If these failures are not detected in time, they may lead to the expansion of the faults and even cause safety accidents.

[0003] Existing motor health assessment technologies suffer from the following main shortcomings: First, they rely heavily on data collected from a single sensor, reflecting only the state of a particular physical field of the motor (e.g., monitoring only temperature or current), making it difficult to comprehensively capture fault characteristics caused by the coupling of multiple physical fields, resulting in a one-sided assessment. Second, data processing methods are simplistic, often involving only single-dimensional analysis in the time or frequency domain, failing to fully extract transient and fault frequency information from the signal, and exhibiting weak anti-interference capabilities. Third, health assessment models are mostly based on a single physical mechanism or are purely data-driven, lacking theoretical interpretability and practical adaptability, and are easily affected by changes in operating conditions. Fourth, fixed judgment thresholds cannot adapt to the decline in the health baseline throughout the motor's life cycle (e.g., aging), leading to missed reports in later stages or false reports in earlier stages. Fifth, they do not consider the impact of load differences on fault characterization; potential faults are difficult to capture under low loads, while misjudgments are easily made due to signal fluctuations under high loads. These problems result in insufficient accuracy and adaptability of existing assessment systems, making it difficult to meet the health monitoring needs of electric bicycle motors under all operating conditions and throughout their entire life cycle. Summary of the Invention

[0004] To overcome the aforementioned deficiencies of the prior art, embodiments of the present invention provide an electric bicycle motor health assessment system based on multi-sensor fusion. The system addresses the problems of one-sided assessment, weak anti-interference ability, poor model adaptability, fixed thresholds, and insufficient load adaptability in the prior art as mentioned in the background section through the following solutions.

[0005] To achieve the above objectives, the present invention provides the following technical solution: a health assessment system for electric bicycle motors based on multi-sensor fusion, characterized in that it includes:

[0006] Intelligent sensor acquisition module: Acquires multi-source operating status data characterizing the motor's operating status;

[0007] Data fusion processing module: Based on an adaptive weighted multi-source fusion algorithm, the multi-source operating status data is subjected to time-domain and frequency-domain collaborative fusion processing to generate a fusion signal sequence for characterizing the transient behavior of the motor;

[0008] Motor health status modeling module: Based on the fused signal sequence, a multi-physics field coupled health assessment model is established, and motor health measurement indicators are generated by cross-consistency correction of the output results of the multi-physics field coupled health assessment model;

[0009] The multiphysics coupled health assessment model includes an electromagnetic performance sub-model, a thermo-electric coupling sub-model, and a vibration response sub-model.

[0010] Adaptive diagnostic output module: Generates motor health level based on the motor health metric, and triggers dynamic threshold recalibration mechanism when the health level reaches a preset abnormal threshold.

[0011] The technical effects and advantages of this invention are as follows:

[0012] 1. Multi-source data fusion improves the comprehensiveness and accuracy of assessment: By collecting multi-dimensional electromagnetic, thermal, and magnetic field signals and combining them with time-domain-frequency domain collaborative fusion algorithms, the transient characteristics and fault frequency information of motor operation are fully explored, avoiding the one-sidedness of single signal assessment; the adaptive weighting strategy dynamically allocates weights according to signal noise, stability, and correlation, improving the credibility of the fused signal and significantly reducing the assessment error caused by interference.

[0013] 2. Multi-physics coupling model for deep health characterization: Construct three sub-models: electromagnetic performance, thermo-electric coupling, and vibration response, to quantify the electromagnetic conversion, thermal balance, and mechanical structure state of the motor, respectively. Combined with a hybrid modeling strategy of "physical mechanism + data-driven", it ensures theoretical interpretability and improves prediction accuracy through neural network correction; cross-consistency correction eliminates the output bias of the sub-models and ensures the reliability of health measurement indicators.

[0014] 3. Adaptability to the entire life cycle and full load conditions: The dynamic threshold recalibration mechanism can automatically update the judgment threshold according to the decline in health baseline caused by motor aging and changes in operating conditions, avoiding false alarms and missed alarms caused by fixed thresholds; the load condition correction covers the entire load range of low, medium and high, enhancing the ability to detect potential faults under low load and the sensitivity to anomalies under high load, adapting to complex usage scenarios.

[0015] 4. Tiered early warning system clarifies maintenance needs: The health status is divided into 5 levels, and the dual judgment logic of "real-time value + short-term trend" is combined to avoid misjudgment due to momentary interference. Each level corresponds to clear fault risks and maintenance suggestions, which makes it easy for users to take targeted measures in a timely manner, reduce the risk of fault expansion, and extend the service life of the motor. Attached Figure Description

[0016] Figure 1 This is a schematic diagram of the overall system structure of the present invention;

[0017] Figure 2This is a schematic diagram of the fusion signal sequence acquisition process of the present invention;

[0018] Figure 3 This is a schematic diagram of the motor health indicator acquisition process of the present invention;

[0019] Figure 4 This is a flowchart illustrating the overall method steps of the present invention;

[0020] Figure 5 This is a schematic diagram illustrating the relationship between health indicators and load rate, and between health indicators and corrected health index, according to the present invention. Detailed Implementation

[0021] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0022] As attached Figures 1 to 5 The electric bicycle motor health assessment system based on multi-sensor fusion shown includes: an intelligent sensor acquisition module: which collects multi-source operating status data characterizing the motor's operating status;

[0023] It should be specifically noted that the multi-source operating status data includes armature current signal, terminal voltage signal, winding temperature signal, and magnetic field change signal;

[0024] The armature current signal (denoted by symbol I) a The unit is ampere (A): A closed-loop Hall current sensor, model ACS758LCB-050B, is selected; the measurement range is ±50A, and the accuracy class is ±0.5%FS (FS refers to full scale, i.e., 50A); it is installed in series in the phase line of the motor armature circuit, specifically on the connection cable between the motor controller and the motor stator winding. During installation, ensure that the center aperture of the sensor matches the cable, and that the cable and the sensor core are tightly fitted to reduce the influence of magnetic leakage; the acquired analog signal is filtered by a second-order Butterworth low-pass filter circuit (cutoff frequency 1kHz) to remove high-frequency noise, and then amplified to 0-3.3V by an instrumentation amplifier (AD8421);

[0025] The terminal voltage signal (denoted by the symbol U) t(Unit: Volt, abbreviated as V): A differential voltage sensor, model AVM500, is used; the measurement range is 0-100V, and the accuracy class is ±0.2%FS; it is connected to the motor power input terminal through a gold-plated test probe in parallel. The probe and the terminal are connected by a flexible crimp structure and equipped with an insulating protective sleeve to adapt to vibration environment and prevent short circuit; the acquired signal is stepped down to 0-3.3V by a resistor divider circuit (using a 0.1% accuracy metal film resistor), and after being protected against voltage spikes by a transient voltage suppressor diode (TVS) in parallel, it is isolated by a voltage follower (LM324);

[0026] The winding temperature signal (symbolized by) (Unit: Celsius, abbreviated as °C): A PT100 platinum resistance temperature sensor (accuracy class: A) is used; the measurement range is -50℃ to 200℃, and the accuracy is ±(0.15+0.002|t|)℃ (t is the measured temperature value); it is fixed to the end of the motor stator winding using a high-temperature resistant ceramic adhesive, and integrated with a miniature Bluetooth 5.0 wireless transmission module (transmission distance ≥10m, power consumption ≤10mA). The encapsulation shell is made of polyimide material with a temperature resistance ≥200℃; the motor is non-rotating. A wireless receiving module is installed at the end (such as the motor end cover). The received temperature data is converted from a resistance signal to a voltage signal (0.08075V for -50℃ and 0.177V for 200℃) by a constant current source circuit (REF200 chip provides 1mA constant current). The voltage signal is then amplified 18 times by a differential amplifier circuit (INA128) (the voltage range after amplification is 1.4535V~3.186V). The nonlinear error of PT100 is corrected by a temperature compensation circuit, and finally a standard voltage signal of 0-3.3V is output.

[0027] The magnetic field change signal (represented by the symbol B, with units of millitalas, abbreviated as B) The tunnel magnetoresistive (TMR) sensor, model TMR2300, was selected; the measurement range is ±200mT, and the sensitivity is [missing value]. The accuracy level is ±1%FS; three sensor units are arranged in an array, evenly distributed along the circumference of the motor (120° interval), installed inside the motor end cover, 5-8mm away from the motor rotor surface, with the sensitive axis direction consistent with the motor radial direction, and fixed by epoxy resin encapsulation; the weak bipolar signal (maximum output ±100mV, corresponding to ±200mT) is amplified 100 times by an instrumentation amplifier (amplified voltage range ±10V), first superimposed with a 10V DC voltage through a level-up circuit (converting the bipolar signal into a 0-20V unipolar signal, retaining the information of magnetic field reversal), then scaled to 0-3.3V through a precision resistor voltage divider circuit (voltage division ratio 3.3V / 20V) (adapted to ADC acquisition, mapping relationship: 0V→-200mT, 1.65V→0mT, 3.3V→+200mT), and finally filtered by a high-pass filter circuit (cutoff frequency 0.1Hz) to remove static magnetic field interference and retain the complete dynamic magnetic field change components;

[0028] Data fusion processing module: Based on an adaptive weighted multi-source fusion algorithm, the multi-source operating status data is subjected to time-domain and frequency-domain collaborative fusion processing to generate a fusion signal sequence for characterizing the transient behavior of the motor;

[0029] It should be specifically noted that the core function of the data fusion processing module is to receive multi-source operating status data (including armature current signal I) transmitted by the intelligent sensor acquisition module. a Terminal voltage signal U t Winding temperature signal T v The system uses the magnetic field change signal (B) to achieve deep fusion of multi-source signals through standardized preprocessing, dynamic weight calculation, and time-domain-frequency domain collaborative fusion strategy. This generates a fused signal sequence that can accurately characterize the transient operating behavior of the motor, providing high-quality and highly reliable feature inputs for subsequent motor health status modeling.

[0030] It should be specifically noted that the purpose of the data preprocessing is to eliminate outliers and noise interference in the original signal, achieve time axis alignment of multi-source data, and provide a clean, synchronized, high-quality data foundation for subsequent fusion calculations. The preprocessing process includes three core operations: "outlier removal → noise suppression → data alignment".

[0031] It should be further explained that the outlier removal adopts a dual judgment mechanism of "3σ criterion + gradient threshold": First, for a single signal, the mean is calculated within a 100ms sliding window (including 100 sampling points). with standard deviation It will exceed Values ​​within a certain range are marked as suspected outliers; then, the gradient of change between the suspected outlier and its adjacent sampling points is calculated. If the gradient exceeds a preset threshold... If the value is not found, it is considered an outlier and replaced using linear interpolation. The replacement formula is: In the formula: This is the value after outlier correction at time k. These are the normal sampled values ​​adjacent to time k, where i=1,2,3,4 correspond to the armature current signal, terminal voltage signal, winding temperature signal, and magnetic field change signal, respectively.

[0032] It should be further explained that the noise suppression employs differentiated filtering algorithms based on the noise characteristics of different signals: for armature current signals (I... a ) and terminal voltage signal (U t The high-frequency impulse noise it contains is processed using a 3-5 point adaptive median filter. When the signal is stable, a 3-point window is used, and when the signal fluctuates, the window is expanded to 5 points. The filtering rule is divided into two stages:

[0033] Phase 1 (3-point window): First, sort the sampled values ​​in the window from smallest to largest to obtain the minimum, median and maximum values; if the median is neither the minimum nor the maximum value, it means there is no impulse noise in the window, and the median is directly output; if the median is the minimum or the maximum value, it means there is suspected impulse noise, and the window is expanded to 5 points to enter Phase 2.

[0034] Second stage (5-point window): Similarly, sort the sampled values ​​within the 5-point window to obtain the minimum, median, and maximum values; if the median is still the minimum or maximum value at this time, it indicates that the noise interference is strong, and the sampled value at the current moment is output; if the median is no longer an extreme value, the median of the 5-point window is output;

[0035] For winding temperature signal The low-frequency drift noise it contains is processed using a first-order low-pass filter, with the following formula: In the formula: Let be the filtered value of the i-th type of signal at time k. For the original sampled value of the i-th type of signal at time k, This represents the filtered value of the i-th type of signal at time k-1.

[0036] For the magnetic field change signal (B), the static magnetic field residual noise it contains is processed using Kalman filtering, and the core formula is: In the formula: x s (k) represents the system state value at time k, w(k) represents the process noise (variance Q = 0.001), z(k) represents the observed value at time k, v(k) represents the observed noise (variance R = 0.01), and K(k) represents the filter gain at time k. Let P(k|k-1) be the state estimate at time k, P(k|k-1) be the prior covariance at time k, and P(k|k) be the posterior covariance at time k. Initial condition: the first observation is used as the state estimate at time 0. The posterior covariance at time 0 is P(0|0) = 0.1 (based on the initial noise level setting of the magnetic field signal).

[0037] It should be further noted that the data alignment is based on the smallest timestamp t0 (accurate to microseconds) of all input signals, and the step size is generated accordingly. Standard Timeline For signals missing sample values ​​on the standard time axis, corresponding values ​​are supplemented using linear interpolation to ensure that the four types of signals are fully aligned on the time axis with a time deviation of ≤0.1ms.

[0038] It should be specifically noted that the core of the adaptive weighted multi-source fusion algorithm is to dynamically calculate the fusion weight of each signal based on the real-time characteristics of the multi-source signals (noise distribution, sampling stability, signal correlation). The weight calculation period is 100ms, which is synchronized with the sliding window to ensure that the weight can adapt to signal changes in real time and improve the reliability of the fusion result.

[0039] It should be further explained that the implementation process of the adaptive weighted multi-source fusion algorithm includes two steps: "single-item weight calculation → weight normalization". The single-item weight evaluates signal reliability from three dimensions: weights based on noise distribution (W...). 1i Weights based on sampling stability (W) 2i ), weights based on signal correlation (W) 3i ).

[0040] It should be further explained that the weights (W) based on noise distribution 1i The variance of the signal within a 100ms sliding window ( Let the standard deviation of the window for the i-th type of signal characterize the noise distribution. The smaller the variance, the less noise the signal and the higher its reliability; the larger the corresponding weight, the calculation formula is: In the formula: i = 1, 2, 3, 4 correspond to the armature current signal, terminal voltage signal, winding temperature signal, and magnetic field change signal, respectively. Let be the window variance of the j-th type of signal. It is the sum of the reciprocals of the variances of the four types of signals;

[0041] The weights based on sampling stability (W) 2i The coefficient of variation of the signal (Reflecting the relative dispersion of signal sampling) Characterizing sampling stability, coefficient of variation ( Let be the standard deviation of the window for the i-th type of signal; (where is the window mean of the i-th type of signal). The smaller the coefficient of variation, the more stable the signal sampling and the higher the reliability, and the larger the corresponding weight. The calculation formula is: To avoid calculation errors caused by a mean of 0, when | When |<0.1, take | |=0.1 (corresponding to signal units);

[0042] The weights based on signal correlation (W) 3i Using Pearson correlation coefficient (Unitless, value range [-1, 1], the larger the absolute value, the stronger the linear correlation between the two types of signals) Characterizes the linear correlation between any two types of signals. The lower the correlation, the more independent information the signal contains and the higher its reliability, and the greater the corresponding weight. First, calculate the average correlation coefficient between the i-th type of signal and the other three types of signals. (Unitless, j=1,2,3,4 and j≠i), then calculate the weights using the following formula: In the formula: Let Pearson correlation coefficient be the sum of the i-th and j-th class signals, expressed as the covariance of the signals within the window. (The unit is the product of the units of the two corresponding signal types, such as I) a with U t The product of covariance (unit: A·V) and standard deviation The ratio (consistent with unit and covariance) is calculated, i.e. .

[0043] It should be further explained that the weight normalization is performed by weighting and summing the three types of individual weights, and then normalizing the result to obtain the final fused weight. (), unitless, value range [0,1]); where the weight coefficients of the weights based on noise distribution, sampling stability, and signal correlation are set to 0.4, 0.3, and 0.3 respectively (optimized based on motor health assessment experimental data, with noise distribution having the greatest impact on the fusion result), and the specific formula is: , In the formula: Let be the initial comprehensive weights for the i-th type of signal. The sum of the initial integrated weights for the four types of signals, and the final fusion weights satisfy the following: This ensures weight constraints for fusion computing.

[0044] It should be specifically noted that the time-domain and frequency-domain collaborative fusion processing adopts the strategy of "time-domain fusion to extract transient features → frequency-domain fusion to extract frequency features → time-domain and frequency-domain superposition enhancement" to fully explore the time-domain transient behavior and frequency-domain fault features of multi-source signals, and realize the comprehensive extraction and fusion of signal information.

[0045] It should be further explained that the time-domain fusion processing is based on the final fusion weight (Wi) to weight and superimpose the four types of preprocessed signals to generate a time-domain fused signal. Then, first-order differential processing is used to highlight the transient behavior of the motor (such as signal abrupt changes during startup and acceleration). The specific formula is as follows: , ,in The signal at time t is the time-domain fused signal (unitless, the standardized composite signal value); W1-W4: the final fusion weights (unitless) of the armature current signal, terminal voltage signal, winding temperature signal, and magnetic field change signal, respectively; I a,norm (t): Standardized armature current signal value at time t (unitless, range [0,1]), U t,norm (t), W4B norm (t) is also the standardized signal value; Transient characteristic signal at time t (unitless, reflecting the rate of change of the time-domain fused signal at time t-1, with a value range of [-1, 1]); Time-domain fused signal (unitless, value range [0,1]).

[0046] The frequency domain fusion process first performs a 1024-point Fast Fourier Transform (FFT) on each of the four standardized signals to convert the time-domain signal into a frequency-domain signal, with N FFT points. FFT =1024, frequency resolution Covering the main frequency range of motor operation ;

[0047] Then, amplitudes are extracted and fused according to the signal's physical characteristics across frequency ranges:

[0048] Electromagnetic frequency range: including the motor fundamental frequency f0 (unit: Hz, f0=n / 60, n is the motor speed, unit: r / min), 2 times the fundamental frequency 2f0, 3 times the fundamental frequency 3f0, 5 times the fundamental frequency 5f0, and 7 times the fundamental frequency 7f0 (the current fundamental frequency is the same as the motor fundamental frequency f0); only the frequency domain amplitude of the armature current, terminal voltage, and magnetic field change signals are extracted in this range, and a final fusion weighted sum is used, with the formula: A(f)=W1A Ia (f)+W2A Ut (f)+W4A B (f), (W1, W2, W4) are the final fusion weights of the armature current, terminal voltage, and magnetic field change signals, respectively.

[0049] Temperature-related low-frequency range (0-1Hz): This range is directly related to the thermal change characteristics of the motor. Only the frequency domain amplitude of the winding temperature signal is extracted, and the weight is the final fusion weight W3 of the winding temperature signal, i.e., A(f) = W3A.Tv (f); (Note: The final fusion weights of the four types of signals satisfy W1+W2+W3+W4=1) Where: A(f): the frequency domain fusion amplitude of the characteristic frequency point f (unit: Hz) (unitless, value range [0,1]); : The frequency domain amplitude of the standardized armature current signal at the characteristic frequency point f (unitless, value range [0,1]); , and A B (f) is also the frequency domain amplitude of the standardized signal at the characteristic frequency point f; W1-W4: the final fusion weights of the four types of signals (unitless).

[0050] Finally, the fused amplitudes of all characteristic frequency points are used to form a frequency domain fusion spectrum, which is then converted into a time-domain enhanced signal using an inverse fast Fourier transform (IFFT). (Unitless, value range [0,1]), realizes the mapping of frequency domain features to time domain.

[0051] It should be further explained that the collaborative fusion signal generation is the generation of time-domain transient feature signals. With frequency domain enhancement signal The signals are superimposed with a 1:1 weight to obtain the final fused signal sequence, which is then normalized (mapped to the [0,1] interval) for easy use by the subsequent health status modeling module. The specific formula is as follows: , In the formula: Moment-time coordinated fusion signal (unitless, value range [-0.5, 1.5]); Time-domain frequency-enhanced signal (unitless, value range [0,1]); Time-normalized fused signal (unitless, value range [0,1]); in the normalization formula, and For fixed values, It is -0.5. The value is set to 1.5 to ensure consistency of standardized results under different operating conditions.

[0052] It should be specifically noted that the data fusion processing module will standardize and fuse the signal sequence. The output is sent to the motor health status modeling module. The output data frame includes signal timestamps (accurate to microseconds) and fused signal values. and fusion weight information ( This facilitates the modeling module's ability to trace the fusion process and improves the interpretability of health assessments.

[0053] Motor health status modeling module: Based on the fused signal sequence, a multi-physics field coupled health assessment model is established, and motor health measurement indicators are generated by cross-consistency correction of the output results of the multi-physics field coupled health assessment model;

[0054] The multiphysics coupled health assessment model includes an electromagnetic performance sub-model, a thermo-electric coupling sub-model, and a vibration response sub-model.

[0055] It should be specifically noted that the motor health status modeling module is the core component of the system health assessment. Its core function is to transform the standardized fusion signal output by the data fusion processing module into characteristic parameters reflecting the state of the motor's electromagnetic, thermal, and mechanical physical fields. After quantifying the health level of each physical field through three sub-models, the output deviation between the sub-models is eliminated, and finally a unified and reliable health measurement index is generated, providing an accurate basis for subsequent diagnostic output. The module adopts a hybrid modeling strategy of "physical mechanism + data-driven", which not only ensures the theoretical interpretability of the model, but also improves the prediction accuracy through actual data correction. The model update is synchronized with the fusion signal output (updated once every 1 millisecond).

[0056] It should be further explained that the core of feature extraction is to filter out key information directly related to the three physical fields of the motor from the fused signal, forming the exclusive input for each sub-model; feature extraction uses a fixed window of 1 second (containing 1000 sampling points) to ensure that the features can reflect the short-term stable operating state of the motor; all extracted features are normalized and mapped to the 0-1 interval to eliminate the impact of magnitude differences on model calculation; the feature sets of different sub-models are constructed around the core characteristics of the corresponding physical fields, and the specific feature composition and extraction methods are as follows:

[0057] First, there is the electromagnetic performance sub-model input feature set. This feature set focuses on the electromagnetic energy conversion efficiency and magnetic field stability of the motor, and contains six features, all extracted from the current, voltage, and magnetic field components in the fused signal. Specifically, these include:

[0058] Equivalent effective value of current: Within a 1-second window, first acquire all standardized armature current signal values, square each value and sum them, divide the sum by the number of 1000 sampling points, and finally take the square root; the larger this value is, the heavier the current electromagnetic load on the motor, and the value range is 0-1;

[0059] Voltage-current phase coordination coefficient: used to reflect the power factor characteristics of the motor; in the calculation, first calculate the mean values ​​of the standardized terminal voltage and current within the window, then calculate the deviation of each signal value from the mean value, and obtain the correlation coefficient by summing the product of the deviations and taking the square root of the sum of the squares of the respective deviations. The closer the correlation coefficient is to 1, the better the phase coordination of voltage and current, and the higher the electromagnetic conversion efficiency. The original correlation coefficient (-1 to 1) is normalized and mapped to the 0-1 interval.

[0060] Magnetic field fluctuation kurtosis: Reflects the peak characteristics of the magnetic field signal; magnetic field distortion leads to increased kurtosis. The calculation first determines the mean of the standardized magnetic field signal within the window, then calculates the deviation of each signal value from the mean and raises it to the fourth power. The sum of these deviations is then divided by the number of sampling points in the window. Simultaneously, the sum of the squares of these deviations is divided by the number of sampling points, and the squared result is used as the denominator. The ratio of these two values ​​is the original kurtosis. Since the kurtosis of a normally distributed signal is approximately 3, a truncation strategy is adopted to avoid the normalized value exceeding the 0-1 interval: if the original kurtosis is greater than 3, the normalized value is 1.0; if the original kurtosis is less than 0 (outliers have been removed; this is for redundancy protection), the normalized value is 0; in other cases, the original kurtosis is divided by 3, ultimately yielding the characteristic value in the 0-1 interval.

[0061] Electromagnetic fundamental frequency amplitude ratio: Extract the motor fundamental frequency (calculated from the motor speed, the speed divided by 60 is the fundamental frequency) and the amplitude corresponding to twice the fundamental frequency from the frequency domain enhancement signal, and divide the fundamental frequency amplitude by the sum of the amplitudes of the fundamental frequency and twice the fundamental frequency; the closer this value is to 1, the more stable the electromagnetic characteristics are, and the value range is 0-1;

[0062] Total harmonic distortion of current: The amplitudes of the 5th harmonic, 7th harmonic, and fundamental frequency are extracted from the frequency domain characteristics of the current signal. During calculation, the amplitudes of the 5th and 7th harmonics are squared and summed, and the square root is taken to obtain the combined amplitude. This combined amplitude is then divided by the fundamental frequency amplitude. To avoid results exceeding the 0-1 range, if the calculated result is greater than 1.0, it is set to 1.0; if the fundamental frequency amplitude is 0 (extreme operating condition, after outlier removal), it is set to 0. The smaller this value, the closer the current waveform is to a sine wave, and the smaller the electromagnetic interference. The final value range is 0-1.

[0063] Transient current change rate: Within a 1-second window, find the maximum absolute value of the transient characteristic signal in the time domain (the first-order difference of the fused signal) and use it directly as the characteristic value; if the value is too large, it means that the current surge is severe when the motor starts or accelerates, which can easily cause winding damage. The value range is 0-1.

[0064] Next is the input feature set of the thermo-electric coupling sub-model. This feature set focuses on the balance between heat generation and dissipation in the motor, and consists of 5 features, with the core correlation being the thermal effect of winding temperature and current. Specifically, it includes:

[0065] Steady-state temperature value: The average value of all standardized winding temperature signals within a 1-second window is directly taken, which intuitively reflects the current thermal level of the motor; the larger the value, the higher the thermal risk, and the value range is 0-1;

[0066] Temperature change rate: The temperature change trend is evaluated by linear fitting; the slope of the fitted line is calculated with the sampling point number (1 to 1000) in the window as the horizontal axis and the standardized temperature value as the vertical axis; a positive slope indicates that the temperature is rising, and the larger the absolute value, the faster the change. The slope is normalized by dividing by 0.001 (maximum reasonable slope) to obtain the characteristic value in the 0-1 interval.

[0067] Current thermal effect coefficient: Calculated by summing the squares of all standardized armature current values ​​within a 1-second window and dividing by the number of sampling points; this value is proportional to the heat loss generated by the current, and the larger the value, the more severe the heat loss, with a range of 0-1.

[0068] Temperature-current correlation: The Pearson correlation coefficient is used to measure the positive correlation between temperature and current (under normal operating conditions, an increase in current should lead to an increase in temperature); the calculation method is consistent with the voltage-current phase coordination coefficient. The closer the correlation coefficient is to 1, the more normal the correlation is. If it is too low, it may indicate a failure in the heat dissipation system. After normalization, the value is taken as 0-1.

[0069] Low-frequency proportion in the temperature frequency domain: Calculate the sum of amplitudes in the 0-1Hz low-frequency band from the frequency domain characteristics of the temperature signal, and divide it by the sum of amplitudes in the 0-50Hz frequency band. Temperature changes are mainly low-frequency. The closer this value is to 1, the more stable the temperature changes are and the normal operation of the heat dissipation system. The value range is 0-1.

[0070] Finally, there is the vibration response sub-model input feature set. This feature set indirectly reflects the motor vibration state through magnetic field and current signals (magnetic field distortion and electromagnetic force imbalance can both cause vibration). It consists of 5 features, focusing on the stability of the mechanical structure, specifically including:

[0071] Vibration characteristic frequency amplitude: First calculate the characteristic frequency of motor bearing failure (speed multiplied by the number of bearing rolling elements and then divided by 120), extract the amplitude corresponding to this frequency from the frequency domain enhancement signal, and then divide it by the maximum amplitude in the 0-500Hz frequency band; the larger the value, the higher the risk of bearing failure, and the value range is 0-1.

[0072] Time-domain vibration kurtosis: First, calculate the first-order difference of the normalized magnetic field signal (the difference between two adjacent sampling points), and then use the same method to calculate the kurtosis of the magnetic field fluctuation; this value reflects the severity of the impact vibration, and after normalization, it takes a value of 0-1. The larger the value, the more severe the impact.

[0073] Current fluctuation energy: Calculated from the frequency domain characteristics of the current signal, the sum of squares of the amplitude in the 200-500Hz high-frequency band is divided by the sum of squares of the amplitude in the 0-500Hz full-frequency band. High-frequency fluctuations are directly related to electromagnetic vibrations; the larger the value, the more intense the electromagnetic vibrations. The value range is 0-1.

[0074] Vibration signal waveform factor: The effective value of the magnetic field differential signal (which is calculated in the same way as the equivalent effective value of the current) is divided by the average absolute value of the differential signal. The waveform factor of the sine wave is about 1.11. The further the value deviates from this value, the more severe the vibration waveform distortion. After normalization by dividing by 1.11, the value is taken as 0-1.

[0075] Transient vibration peak value: Within a 1-second window, find the maximum and minimum values ​​of the transient characteristic signal in the time domain. The difference between the two is the peak value minus the peak value. Divide the difference by 2 and normalize to obtain the characteristic value in the 0-1 range. The larger the value, the stronger the transient vibration impact. The value range is 0-1.

[0076] It should be noted that all three sub-models adopt a hybrid structure of "physical mechanism calculation + BP neural network correction": mechanism calculation ensures that the model conforms to the physical laws of motor operation, while neural network correction reduces the prediction error of the mechanism model through historical data; the output of each sub-model is a health index in the range of 0-1, where 1 represents the optimal state and 0 represents complete failure.

[0077] The first is the electromagnetic performance sub-model, which primarily assesses the health of the motor's electromagnetic energy conversion. The calculation is divided into two steps, specifically including:

[0078] Mechanism calculation: Based on electromagnetic loss theory, the electromagnetic loss coefficient is obtained by weighted summation of 6 input features; the weight allocation is based on the degree of influence of the feature on the electromagnetic performance: the current equivalent effective value has the highest weight (0.3) because it directly reflects the electromagnetic load; the voltage-current phase coordination coefficient is calculated in the form of (1-eigenvalue) (weight 0.25), and the better the phase coordination, the lower the loss; the magnetic field fluctuation kurtosis (0.2), the electromagnetic fundamental frequency amplitude ratio (1-eigenvalue, 0.15), and the current harmonic total distortion rate (0.1) decrease in that order; the larger the loss coefficient, the worse the electromagnetic health status, and the value is 0-1.

[0079] Neural Network Correction: A 3-layer BP neural network is constructed, with 6 feature nodes in the input layer, 10 nodes in the hidden layer, and 1 node in the output layer; the deviation between the calculated mechanism value and the measured electromagnetic health value in historical data is used as the training objective, enabling the network to learn the deviation pattern; the final electromagnetic health index (denoted as H) is... em H = 1 - (mechanism calculation value + correction coefficient of neural network output), and constraints are used to ensure that the result is within the range of 0-1; em Electromagnetic properties are good when ≥0.8, H em A value <0.3 indicates a serious abnormality.

[0080] Next is the thermo-electric coupling sub-model, which assesses the motor's thermal equilibrium state and the risk of thermal damage. Its calculation logic is consistent with the electromagnetic sub-model, and specifically includes:

[0081] Mechanism calculation: Based on the heat balance equation (heat generation = heat dissipation), the heat imbalance coefficient is obtained by weighting 5 characteristics; the steady-state temperature value has the highest weight (0.4) and directly determines the thermal risk benchmark; the temperature change rate (0.3) and the current thermal effect coefficient (0.2) reflect the trend of heat change and the source of heat generation, respectively; the temperature-current correlation degree participates in the calculation in the form of (1-eigenvalue) (weight 0.1), and abnormal correlation means that heat dissipation may fail; the larger the heat imbalance coefficient, the worse the thermal health status, and the value ranges from 0 to 1.

[0082] Neural network correction: A 3-layer backpropagation (BP) neural network with 5 nodes in the input layer, 8 nodes in the hidden layer, and 1 node in the output layer, with the training objective being the "deviation between the calculated mechanistic value and the measured thermal health value"; the final thermal-electrical health index (denoted as...) = 1 - (mechanism calculation value + correction coefficient), constrained in the interval 0-1; A value ≥0.8 indicates good thermal condition. There is a serious risk of overheating when the value is less than 0.25.

[0083] Finally, there is the vibration response sub-model, which assesses the health status of the motor's mechanical structure (bearings, rotor) through indirect feature evaluation. The calculation is divided into two steps, specifically including:

[0084] Mechanism calculation: Based on the vibration intensity standard, the vibration intensity coefficient is obtained by weighting five characteristics; the amplitude of the bearing fault characteristic frequency has the highest weight (0.35), which is directly related to the core mechanical components; vibration kurtosis (0.25) and current fluctuation energy (0.2) are next, reflecting the impact vibration and electromagnetic vibration levels, respectively; waveform factor (0.1) and transient vibration peak value (0.1) supplement the assessment of vibration stability and impact; the larger the intensity coefficient, the worse the mechanical health status, with a value of 0-1.

[0085] Neural network correction: A 3-layer backpropagation (BP) neural network with 5 nodes in the input layer, 8 nodes in the hidden layer, and 1 node in the output layer. The training objective is to determine the deviation between the calculated vibration health value and the measured vibration health value. The final vibration health index (denoted as...) = 1 - (mechanism calculation value + correction coefficient), constrained in the interval 0-1; A value ≥0.8 indicates a healthy mechanical structure. A value less than 0.2 indicates a serious risk of mechanical failure.

[0086] It should be specifically noted that the core purpose of the cross-consistency correction is to address potential deviations in the sub-model output (such as the output H of the electromagnetic sub-model). em Normal but oscillator model output (Anomalies) By judging the consistency and reliability among sub-models, the health index is weighted and corrected to ensure that the final result is consistent and reliable. The correction process consists of three steps: consistency judgment, confidence calculation, and weighted fusion. Here, Hem represents the health index of the electromagnetic performance sub-model. Represents the health index of the thermo-electric coupler model. This represents the health index of the vibration response sub-model.

[0087] It should be further explained that the consistency determination method specifically involves calculating the absolute difference between the health indices of any two sub-models. Compared with the preset reasonable deviation threshold (0.15, based on historical normal data statistics), there are three cases:

[0088] Completely consistent: all three differences are ≤0.15, indicating that the outputs of the sub-models are all reliable;

[0089] Local consistency: If two differences are ≤0.15 and one difference is >0.15, the sub-model with excessive deviation may have errors.

[0090] Completely inconsistent: All three differences are >0.15, requiring further judgment based on the quality of the fused signal.

[0091] The confidence level reflects the historical prediction accuracy of the sub-model and takes a value of 0-1, with a higher value indicating greater reliability.

[0092] The confidence level is calculated by selecting 1000 historical samples, summing the absolute differences between the predicted health value and the actual health value of the sub-model in each sample, dividing by the number of samples to obtain the historical average prediction error, and subtracting this error from 1 to get the confidence level; for example, if the historical average error of a sub-model is 0.05, then the confidence level is 0.95.

[0093] The weighted fusion employs different strategies based on the consistency determination result, specifically including:

[0094] Completely consistent: directly weighted by confidence level, the corrected health index = (electromagnetic confidence level × H) em +Thermo-Electrical Confidence Level × +Vibration confidence level × ) ÷ (the sum of the three confidence levels);

[0095] Local consistency: Identify sub-models with excessive bias, multiply their confidence by 0.5 to reduce their weight, and then calculate according to the above formula;

[0096] Completely inconsistent: First, calculate the overall quality coefficient of the fused signal. This coefficient is obtained by weighting three parts (each with a weight of 1 / 3): outlier ratio coefficient (1 minus the proportion of outliers in the signal), noise level coefficient (1 minus the ratio of the mean variance of the four signal windows to the maximum allowable variance), and sampling stability coefficient (1 minus the mean of the coefficients of variation of the four signals).

[0097] If the overall quality coefficient is ≥0.9 (good signal quality), then the sub-model index with the highest confidence (H) is used. em , or The core index is 0.7, supplemented by the average of two other indices (0.3).

[0098] If the overall quality coefficient is <0.9 (poor signal quality), further determine whether the noise level coefficient is <0.8 (noise-dominant error) or the sampling stability coefficient is <0.8 (stability-dominant error), and output the corresponding "noise exceeds the standard" or "sampling is unstable" prompts to trigger the sensor self-test accordingly.

[0099] It should be noted that the health measurement index is the final standardized index obtained by correcting the health index and combining it with the motor load conditions. The value ranges from 0 to 1, and the larger the value, the better the health status.

[0100] The load condition is directly characterized by the equivalent effective value of the current (the input characteristic of the electromagnetic performance sub-model), referred to as the load factor. The correction logic is as follows: enhance anomaly sensitivity under high load, and also improve sensitivity under low load (when faults are more concealed), achieving full load coverage. The specific correction method is as follows:

[0101] If the load rate is ≤30% (low load): H=[corrected health index×(1-0.2×load rate)+0.2×load rate×(corrected health index)²]×1.2 (sensitivity enhancement factor, based on low load fault experimental data fitting).

[0102] If 30% < load rate ≤ 70% (medium load): H = Corrected health index × (1 - 0.3 × load rate) + 0.3 × load rate × (corrected health index)²;

[0103] If the load factor is >70% (high load): H = Corrected health index × (1 - 0.4 × load factor) + 0.4 × load factor × (corrected health index)²;

[0104] This correction method covers the entire load range from low to high, ensuring that potential faults at low loads (such as slight aging of windings or insufficient bearing lubrication) can be accurately detected, and that severe anomalies at high loads (such as current surges or overheating) are not missed.

[0105] It should be further noted that some experimental data are shown in the table below:

[0106]

[0107] Adaptive diagnostic output module: Generates motor health level based on the motor health metric, and triggers dynamic threshold recalibration mechanism when the health level reaches a preset abnormal threshold.

[0108] It should be specifically noted that the motor health level is a concrete classification of the health measurement index H, used to intuitively reflect the health status of the motor. The level division is based on the principle of "progressive fault risk + clear maintenance needs", and is divided into 5 levels. The judgment logic of each level combines the dual dimensions of "real-time H value + short-term trend" to avoid misjudgment caused by a single instantaneous value.

[0109] It should be further explained that, based on motor failure statistics (covering 3 years of operating data from 1000 electric bicycle motors under different working conditions), a baseline correspondence between health levels and H values ​​is established, and core characteristics and risk descriptions for each level are defined:

[0110] Level 1 (Healthy): The criterion is H ≥ 0.8; the core characteristic is that the outputs of the electromagnetic, thermal, and vibrational sub-models are all within the healthy range (H). em ≥0.8 ≥0.8 (≥0.8), with no abnormal characteristics; the risk description is that the motor is in excellent operating condition with no risk of failure.

[0111] Level 2 (Sub-health): The criterion is 0.6 ≤ H < 0.8; the core feature is that the output of a single sub-model is close to the critical value (e.g., =0.75, or =0.78), the remaining sub-models are normal, and the short-term H value fluctuation range is ≤0.05; the risk description is that there is slight parameter drift, which needs to be strengthened for monitoring.

[0112] Level 3 (Minor Fault): The judgment condition is 0.4 ≤ H < 0.6; the core feature is that the output of a single sub-model is lower than the health threshold (e.g., H). em =0.55), or two or more sub-models are in the sub-healthy range ( =0.7、 =0.65), the H value shows a slow downward trend; the risk description is that there is a minor local fault that does not affect basic operation but requires planned maintenance.

[0113] Level 4 (Severe Fault): The judgment condition is 0.2 ≤ H < 0.4; the core feature is that the output of a single sub-model is in the fault range (e.g., =0.18), or the outputs of both sub-models are below the health threshold (Hem =0.45、 =0.3), H value decrease rate ≥0.02 / second; risk description: the fault has affected motor performance, continued operation may aggravate the damage, and immediate shutdown and inspection are required.

[0114] Level 5 (Fatal Failure): The judgment condition is H < 0.2; the core feature is that at least two sub-model outputs are in the failure range (e.g., =0.15、 =0.1), or the H value drops instantaneously to below 0.2; the risk description is that the core components of the motor have failed, posing a safety hazard, and requiring emergency shutdown for repair.

[0115] It should be further explained that, in order to avoid misjudgment caused by transient interference (such as a sudden drop in H value due to sudden load fluctuations), the module introduces the "verification of 3 consecutive sampling points" rule: the real-time received H value must meet a certain level judgment condition for 3 consecutive times (cumulative 3ms) before the final level can be output; if a single H value triggers a high level but subsequent sampling points return to normal, it is judged as "transient interference", the original level is output and the interference event is marked; for example: at a certain moment the H value suddenly drops to 0.35 (triggers level 4), but the H values ​​of the next two sampling points are 0.72 and 0.75 respectively (level 2), then the module judges it as transient interference, maintains level 2 output, and records "1 instance of level 4 interference triggered" in the system log.

[0116] It should be specifically noted that the dynamic threshold recalibration is the core adaptive capability of the module, which aims to solve problems such as "aging leading to a decline in the health baseline" and "changes in operating conditions causing judgment deviations" throughout the entire life cycle of the motor. When the health level reaches the preset abnormal threshold (level 4 and above), or when the duration of level 2 and 3 exceeds the set threshold (level 2 lasts for ≥24 hours, level 3 lasts for ≥1 hour), the calibration process is automatically triggered. The calibration object is the H value judgment threshold corresponding to the health level (such as the boundary threshold of level 1 and 2 of 0.8, the boundary threshold of level 2 and 3 of 0.6, etc.).

[0117] The module has two preset trigger conditions; recalibration will be initiated when either condition is met. These conditions include:

[0118] Emergency Trigger: If the health level reaches level 4 or above, or if there are 3 consecutive fluctuations from level 3 to 4 within 1 minute, immediately pause the level output and start a fast calibration (calibration time ≤ 50ms) to ensure accurate subsequent judgments.

[0119] Regular trigger: Level 2 lasts for ≥24 hours (the motor is in a sub-healthy state for a long time, which may be due to a decline in the health baseline), or Level 3 lasts for ≥1 hour (the fault is developing slowly and the threshold needs to be adjusted to avoid missed reports). In this case, regular calibration is started during the system's idle period (such as when the motor stops) (calibration time ≤100ms).

[0120] It should be further explained that dynamic threshold recalibration follows a four-step process: data screening, baseline calculation, threshold update, and verification. The specific implementation method is as follows:

[0121] Health data screening: Extract the "full-condition normal health data" of the motor from the system's historical database, and screen it in three operating condition intervals (extract no less than 300 data points for each interval, totaling ≥1000 data points): Low load interval (load rate ≤30%): ambient temperature 5-40℃, health level 1 and duration ≥30 minutes;

[0122] Medium load range (30% < load rate ≤ 70%): ambient temperature 5-40℃, health level 1 and duration ≥ 30 minutes;

[0123] High load range (load rate > 70%): ambient temperature 5-40℃, health level 1 and duration ≥ 15 minutes; the comprehensive health baseline is calculated based on the weight of each range according to "low load: medium load: high load = 2:5:3" to avoid baseline deviation caused by single operating condition data;

[0124] Health baseline calculation: Statistical analysis is performed on the screened H-value data to calculate its 95th percentile (denoted as H0). — This quantile represents the "stable health baseline" of the motor's current health status, avoiding the influence of extreme values; for example: if 95% of the selected 1000 H values ​​are ≤0.78, then... =0.78;

[0125] Threshold ladder update: Based on this, the judgment thresholds for each level are updated according to the principle of "equal-distance decreasing," and the update formula follows the principle of "the original threshold ratio remains unchanged." The original baseline threshold sequence is [0.8, 0.6, 0.4, 0.2] (the dividing point between levels 1 and 5). The updated threshold sequence is calculated as follows: New threshold = Original threshold × ( / 0.8) (0.8 is the initial health baseline); Example: If =0.78, then the new threshold sequence is 0.78 (0.8×0.78 / 0.8), 0.585 (0.6×0.78 / 0.8), 0.39 (0.4×0.78 / 0.8), 0.195 (0.2×0.78 / 0.8), that is, the boundary between level 1 and 2 becomes 0.78, the boundary between level 2 and 3 becomes 0.585, and so on;

[0126] Verification of effectiveness: Apply the new threshold sequence to nearly 100 historical health data (covering different levels). If the accuracy rate is ≥95% (compared with the results of manual annotation), the new threshold will take effect immediately; if the accuracy rate is <95%, the data will be re-screened (expanding the screening range to load rate ≤60%), and the calibration process will be repeated until the accuracy rate meets the standard.

[0127] Secondly: The accompanying drawings of the embodiments disclosed in this invention only involve the structures involved in the embodiments disclosed in this invention. Other structures can refer to the general design. In the absence of conflict, the same embodiment and different embodiments of this invention can be combined with each other.

[0128] In conclusion, the above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A health assessment system for electric bicycle motors based on multi-sensor fusion, characterized in that, include: Intelligent sensor acquisition module: Acquires multi-source operating status data characterizing the motor's operating status; Data fusion processing module: preprocesses multi-source operating status data, then evaluates signal purity based on the noise distribution characteristics of each signal, evaluates sampling stability based on the dispersion of the sampled data, evaluates information independence based on the correlation between signals, and then assigns initial weights to each signal according to the evaluation results. After weighted summation and normalization, the final fusion weights are obtained. The multi-source data is weighted and superimposed by the final fusion weights, and transient change features are extracted. Frequency intervals are divided according to the physical domain characteristics of the signals and weighted fusion is performed. Finally, a fusion signal sequence for characterizing the transient behavior of the motor is generated by the synergistic superposition of time domain and frequency domain features. Motor health status modeling module: Based on the fused signal sequence, a multi-physics coupled health assessment model is established, and motor health metrics are generated by cross-consistency correction of the output results of the multi-physics coupled health assessment model; the multi-physics coupled health assessment model includes an electromagnetic performance sub-model, a thermo-electric coupling sub-model, and a vibration response sub-model; The electromagnetic performance sub-model is used to quantify the electromagnetic-related features in the fused signal to assess the health level of the motor's energy conversion. This sub-model filters features related to current, voltage, and magnetic field from the fused signal, focusing on the motor's electromagnetic energy conversion efficiency and magnetic field stability. By quantifying the parameter deviations of these features, the health level of the motor's electromagnetic energy conversion is evaluated. The thermal-electric coupling sub-model is used to quantify the thermal-electric correlation features in the fused signal to assess the motor thermal balance health level. The sub-model filters features related to winding temperature and armature current thermal effects from the fused signal, focuses on the balance between motor heat generation and heat dissipation, and evaluates the motor thermal balance status and thermal damage risk by quantifying the parameter deviations of these features. The vibration response sub-model is used to quantify the health level of the motor mechanical structure by analyzing the magnetic field and mechanical correlation features in the fused signal. The sub-model selects features from the magnetic field changes and current fluctuation components of the fused signal, focuses on the stability of the mechanical structure, and evaluates the health status of the motor mechanical structure by quantifying the parameter deviations of these features. The cross-consistency correction is used to eliminate the deviations in the output results of the electromagnetic performance sub-model, the thermo-electric coupling sub-model, and the vibration response sub-model. Its specific implementation process includes three steps: First, consistency is determined by calculating the absolute difference between the health indices of any two sub-models and comparing it with a preset deviation range, classifying them into three cases: completely consistent, partially consistent, and completely inconsistent. Second, the confidence level of each sub-model is calculated by statistically analyzing the average error between the predicted health value and the actual health value of the sub-model based on historical sample data, and subtracting this error from 1 to obtain the confidence level. Finally, weighted fusion is performed. When completely consistent, the correction result is directly calculated by weighting according to the confidence level. When partially consistent, the confidence level of the sub-models with excessive deviations is reduced before weighting. When completely inconsistent, the weights are adjusted by combining the comprehensive quality coefficient of the fused signal. Adaptive diagnostic output module: Generates motor health level based on the motor health metric, and triggers dynamic threshold recalibration mechanism when the health level reaches a preset abnormal threshold.

2. The electric bicycle motor health assessment system based on multi-sensor fusion according to claim 1, characterized in that: The multi-source operating status data includes armature current signal, terminal voltage signal, winding temperature signal, and magnetic field change signal; The current and voltage signals are used to capture the electromagnetic domain operating characteristics of the motor, the temperature signal is used to capture the thermal domain energy balance characteristics of the motor, and the magnetic field change signal is used to capture the magnetic field domain distribution characteristics of the motor.

3. The electric bicycle motor health assessment system based on multi-sensor fusion according to claim 1, characterized in that: The motor health level is divided into 5 levels according to the failure risk and maintenance requirements, based on the health measurement index value and short-term trend, and verified by 3 consecutive sampling points.

Images