Intelligent monitoring alarm method and system of motor controller

Abstract

The invention relates to an intelligent monitoring alarm method and system for a motor controller, and the method comprises the steps: obtaining an internal circuit signal and a body vibration signal of the motor controller, carrying out the combined noise analysis, and generating a dynamic noise baseline; performing component fault identification based on the dynamic noise baseline and the internal circuit signal to obtain a fault precursor current feature; performing component impact response separation based on the dynamic noise baseline and the body vibration signal to obtain hidden fault vibration characteristics; performing component association mapping on the fault precursor current feature and the hidden fault vibration feature to obtain an internal fault component identifier; and performing alarm analysis based on the fault component identifier, the fault precursor current feature and the hidden fault vibration feature to obtain a graded alarm decision. According to the invention, early fault features can be captured more sensitively, and false alarms and missing alarms are reduced.

Description

Technical Field

[0001] The present invention relates to the technical field of motor control, and in particular to an intelligent monitoring and alarm method and system for a motor controller. Background Art

[0002] As a core component of industrial automation and electric equipment, the reliability of motor controllers directly impacts production efficiency and equipment safety. Traditional methods typically analyze current or vibration signals individually, failing to effectively leverage the collaborative characteristics of multi-source signals, resulting in insufficient sensitivity to potential faults. Existing technologies lack dynamic modeling of the coupling relationship between circuit noise and mechanical vibration, making fault feature extraction susceptible to environmental interference, leading to false positives or missed reports. This fragmented analysis model makes it difficult to identify the associated features of complex faults, especially in the early stages of hidden faults such as aging of high-frequency switching devices and microcracks in mechanical structures. Existing systems are often unable to capture diagnostic precursor information in a timely manner. Summary of the Invention

[0003] The main purpose of the present invention is to provide an intelligent monitoring and alarm method and system for a motor controller, which can more keenly capture early fault characteristics and reduce false alarms and missed alarms.

[0004] To achieve the above object, the present invention provides an intelligent monitoring and alarm method for a motor controller, comprising: Obtain the internal circuit signal and body vibration signal of the motor controller, perform joint noise analysis, and generate a dynamic noise baseline; performing component fault identification based on the dynamic noise baseline and the internal circuit signal to obtain a fault precursor current characteristic; Separating component impact responses based on the dynamic noise baseline and the body vibration signal to obtain a hidden fault vibration signature; Performing component association mapping on the fault precursor current characteristics and the hidden fault vibration characteristics to obtain an internal fault component identifier; An alarm analysis is performed based on the fault component identifier, the fault precursor current characteristics and the hidden fault vibration characteristics to obtain a graded alarm decision.

[0005] Furthermore, the acquisition of the internal circuit signal and the body vibration signal of the motor controller and the joint noise reduction to generate a dynamic noise baseline includes: collecting current from a DC bus circuit of the motor controller to obtain the internal circuit signal; Performing triaxial vibration acquisition on the power device installation area of ​​the motor controller housing to obtain the body vibration signal; performing bandpass filtering on the internal circuit signal to obtain a current waveform in a target frequency band; Cross-correlation analysis and blind source separation processing are performed on the target frequency band current waveform and the body vibration signal to obtain the dynamic noise baseline.

[0006] Furthermore, the performing component fault identification based on the dynamic noise baseline and the internal circuit signal to obtain a fault precursor current feature includes: Performing multi-scale time-frequency decomposition on the internal circuit signal to obtain a decomposition signal set; Performing frequency band noise suppression on the decomposed signal set according to the dynamic noise baseline to obtain a target frequency band signal; Performing waveform segment interception processing on the target frequency band signal to obtain candidate fault waveform segments; Performing multi-dimensional feature matching on the candidate fault waveform segment based on a preset fault feature library to obtain a multi-dimensional feature vector; Perform similarity matching on the multidimensional feature vector according to the fault template of the preset fault feature library, and output a probability distribution of the fault type; The fault type probability distribution is screened using a confidence threshold to obtain the fault precursor current characteristics.

[0007] Furthermore, the component impact response separation based on the dynamic noise baseline and the body vibration signal to obtain the hidden fault vibration characteristics includes: Suppressing environmental coupling noise on the body vibration signal according to the dynamic noise baseline to obtain a denoised vibration signal; Performing multi-scale resonance frequency band extraction on the denoised vibration signal to obtain fault-related frequency band signals including bearing wear, shaft eccentricity and gear meshing abnormality; Performing envelope spectrum analysis on the fault-related frequency band signal to obtain impact periodicity characteristics; Performing physical characteristic frequency band analysis on the impact periodicity characteristics based on preset mechanical component structural parameters to obtain fault impact energy distribution; The hidden fault vibration characteristics are obtained by performing a probabilistic judgment based on the impact periodic characteristics and the fault impact energy distribution.

[0008] Furthermore, the performing component association mapping on the fault precursor current feature and the latent fault vibration feature to obtain an internal fault component identifier includes: Performing physical characteristic frequency band analysis on the fault precursor current characteristics and the latent fault vibration characteristics to obtain a current vibration joint characteristic; Performing topological constraint matching on the current vibration joint features according to a preset component fault mapping relationship library to obtain a set of candidate fault components; Performing fault characteristic frequency band distribution verification on the candidate fault component set to obtain a valid fault component set; Based on the component failure mapping relationship library, primary and secondary conflicts are resolved on the valid fault component set to obtain a primary fault component identification set; Fault mode reverse mapping is performed on the main fault component identifier set to obtain the internal fault component identifier.

[0009] Furthermore, the alarm analysis based on the fault component identifier, the fault precursor current characteristics and the hidden fault vibration characteristics is performed to obtain a graded alarm decision, including: Performing fault type analysis based on the preset component type information and the fault component identifier to obtain electrical fault type information and mechanical fault type information; Performing multi-modal feature fusion on the fault precursor current feature and the hidden fault vibration feature to obtain a comprehensive fault feature vector; Performing a hazard diffusion assessment on the electrical fault type information and the mechanical fault type information according to the comprehensive fault feature vector to obtain a fault hazard level score; The fault hazard level score is subjected to decision mapping based on a preset alarm strategy library to obtain the graded alarm decision.

[0010] Furthermore, the performing of a hazard diffusion assessment on the electrical fault type information and the mechanical fault type information according to the comprehensive fault feature vector to obtain a fault hazard level score includes: Extracting fault parameters from the comprehensive fault feature vector to obtain a global fault intensity index; Performing graph structure modeling on the gas fault type information and the mechanical fault type information based on a preset internal component dependency relationship to obtain a fault propagation path network; Performing path propagation simulation on the global fault intensity index based on the fault propagation path network to obtain a fault diffusion evaluation parameter; A fault evaluation is performed on the fault diffusion evaluation parameters according to a preset historical fault case library to obtain the fault hazard level score.

[0011] The present invention further provides an intelligent monitoring and alarm system for a motor controller, which is applied to any of the above-mentioned intelligent monitoring and alarm methods for a motor controller, comprising: An acquisition module is used to acquire internal circuit signals and body vibration signals of the motor controller, and perform joint noise analysis to generate a dynamic noise baseline; an analysis module, configured to identify component faults based on the dynamic noise baseline and the internal circuit signal to obtain a pre-fault current characteristic; a correlation module, configured to separate component impact responses based on the dynamic noise baseline and the body vibration signal to obtain a hidden fault vibration signature; a processing module configured to perform component association mapping on the precursory current characteristics and the latent fault vibration characteristics to obtain an internal fault component identifier; A control module is used to perform alarm analysis based on the fault component identifier, the fault precursor current characteristics and the hidden fault vibration characteristics to obtain a hierarchical alarm decision.

[0012] The present invention provides an intelligent monitoring and alarm method and system for a motor controller, which has the following beneficial effects: Joint noise analysis of internal circuit signals and body vibration signals using a dynamic noise baseline overcomes the limitations of traditional methods that rely solely on a single signal, enabling more sensitive detection of early fault signatures and reducing false alarms and missed detections. Component fault identification and impulse response separation based on the dynamic noise baseline effectively distinguishes normal operating noise from fault signatures, reducing the impact of environmental interference and improving the detection rate of hidden faults (such as aging of high-frequency switching devices and mechanical microcracks). Correlation analysis of current and vibration signatures using a component fault mapping library accurately identifies faulty components, avoiding misjudgments caused by traditional signal fragmentation analysis and providing a reliable basis for predictive maintenance. Intelligent alarm analysis combines faulty component identification, current signatures, and vibration signatures to achieve graded warnings, ranging from minor anomalies to severe faults. This enables more targeted maintenance responses, reduces unnecessary downtime and maintenance, and improves equipment operational efficiency. BRIEF DESCRIPTION OF THE DRAWINGS

[0013] Figure 1 This is a flow chart of an intelligent monitoring and alarm method for a motor controller provided by the present invention; Figure 2 This is a structural diagram of an intelligent monitoring and alarm system for a motor controller provided by the present invention.

[0014] The purpose, features and advantages of the present invention will be further described with reference to the accompanying drawings and in conjunction with the embodiments. DETAILED DESCRIPTION

[0015] In order to make the purpose, technical solutions and advantages of the present invention more clearly understood, the present invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present invention and are not intended to limit the present invention.

[0016] The present invention will be further described below with reference to the accompanying drawings and specific embodiments.

[0017] Reference Figure 1As shown, the present invention provides an intelligent monitoring and alarm method for a motor controller, comprising: Step S1: Acquire the internal circuit signal and the body vibration signal of the motor controller, perform joint noise analysis, and generate a dynamic noise baseline; Step S2: Component fault identification is performed based on the dynamic noise baseline and the internal circuit signal to obtain a fault precursor current characteristic; Step S3: Separate the component impact response based on the dynamic noise baseline and the body vibration signal to obtain the hidden fault vibration characteristics; Step S4: performing component association mapping on the fault precursor current characteristics and the latent fault vibration characteristics to obtain the internal fault component identification; Step S5: Perform alarm analysis based on the fault component identification, the fault precursor current characteristics, and the latent fault vibration characteristics to obtain a graded alarm decision.

[0018] Based on the above steps, the detailed process is as follows: Step S1: The internal circuit signal acquisition of the motor controller requires the configuration of high-precision current sensors and voltage probes. Measurement points are arranged at key nodes of the power device to ensure that the switching transient and ripple characteristics can be captured. The body vibration signal is acquired using a three-axis accelerometer, and the installation location is selected at the structural resonance sensitive point of the motor controller, such as the heat sink fixing part and the power module mounting surface. Sampling synchronization needs to be considered during the signal acquisition process to ensure that the electrical signal and the mechanical vibration signal have a unified time base. The signal preprocessing stage includes the necessary hardware filtering and gain adjustment to eliminate the inherent noise and range mismatch introduced by the sensor.

[0019] The core of joint noise analysis lies in establishing a correlation model between electrical and mechanical signals. By comparing multiple sets of historical data under normal operating conditions, the characteristic frequency bands and amplitude ranges of circuit noise and mechanical vibration are extracted. A dynamic noise baseline is constructed using statistical learning methods to analyze the distribution patterns of signal characteristics under different operating conditions, forming a reference benchmark that includes mean, variance, and extreme values. The baseline update mechanism needs to account for the impact of ambient temperature changes and load fluctuations, using a sliding time window to regularly adjust the reference value. The validity of the noise baseline is verified by injecting known fault signals to ensure that normal fluctuations can be distinguished from true fault characteristics.

[0020] Step S2: Fault feature extraction of the current signal focuses on the operating status of the switching devices and power modules. The real-time current waveform is compared with the dynamic noise baseline to identify abnormal current ripple, rising edge distortion, and other features. Aging of power devices manifests as an increase in on-resistance, which is reflected in the current waveform as a change in slope during the on-phase. Degradation of the DC bus capacitor can lead to an increase in ripple current amplitude, requiring comprehensive analysis in conjunction with the voltage signal. Abnormalities in the drive circuit can cause changes in dead time, manifesting as a change in the symmetry of the current waveform.

[0021] Fault precursor characteristics are quantified using feature engineering methods, extracting key indicators from the time and frequency domains. Time-domain features include statistical parameters such as crest factor and crest factor, while frequency-domain features focus on the energy contribution of specific harmonic components. The feature selection process considers the sensitivity of different fault modes to ensure that the extracted features have clear physical meaning and discriminative power. Fault injection experiments are used during the feature verification phase to establish a correlation between feature parameters and fault severity. Feature data is standardized to eliminate the influence of different dimensions, providing a unified data foundation for subsequent correlation analysis.

[0022] Step S3: Impact response analysis of the vibration signal requires distinguishing between structural resonance and fault impact. The real-time vibration spectrum is compared with the dynamic noise baseline to identify abnormal frequency components and energy distribution. The characteristic frequency of bearing defects is calculated based on speed information and bearing geometry parameters, and is reflected in the spectrum as the presence of characteristic frequencies and their harmonics. The vibration characteristics of mechanical looseness are characterized by increased energy across a wide frequency band, requiring analysis of the periodicity of impact events in conjunction with time-domain waveforms. An imbalance fault in a cooling fan is reflected in the spectrum as an increase in the amplitude of the speed frequency, which may be accompanied by higher-order harmonics.

[0023] Hidden fault features are extracted using signal processing methods to separate fault-related components from the vibration waveform. Envelope analysis is used to extract fault features modulated on the carrier frequency, making it particularly suitable for detecting early, weak faults. Time-frequency analysis reveals the transient characteristics of the vibration signal and captures non-stationary impact events. Feature enhancement processing highlights fault-related frequency bands and suppresses background noise interference. Feature validity is verified by comparing known fault samples to ensure that the extracted features reflect the extent of the fault. Dimensionality reduction processing of the feature data retains the most discriminative parameters, providing input for subsequent multi-source information fusion.

[0024] Motor controller fault diagnosis requires combining electrical and mechanical signatures to accurately identify the specific faulty component. The fault mapping library is constructed based on historical failure cases, simulation analysis, and experimental data, covering typical failure modes of key components such as power devices (such as IGBTs and MOSFETs), drive circuits, DC bus capacitors, cooling fans, and mechanical connectors. Each fault mode corresponds to a specific combination of current and vibration signatures, for example: IGBT aging: The rising edge of the current waveform slows down, the conduction loss increases, and the high-frequency vibration component increases.

[0025] Capacitor degradation: The ripple current amplitude increases and the low-frequency vibration energy increases.

[0026] Mechanical looseness: Broadband shocks appear in the vibration signal, but the current characteristics change slightly.

[0027] The correlation mapping process uses a multi-feature fusion strategy to match the extracted pre-fault current signatures and latent fault vibration signatures with standard patterns in a mapping library. The matching algorithm considers feature weights; for example, current signatures are more sensitive to power device failures, while vibration signatures are more effective for mechanical failures. The matching results generate a list of possible faulty components and calculate confidence levels to eliminate false positives. Finally, by combining fault development patterns (such as the evolution trend of characteristic parameters), the most likely faulty component and its severity are determined, and a clear fault identification is output.

[0028] The core of alarm decision-making is to assess the severity and development trend of the fault in order to provide reasonable maintenance recommendations. Hierarchical alarms are usually divided into three levels: Warning Level (Low Risk): The fault characteristics are minor and may be in the early stages of degradation, such as a slight current ripple exceeding the limit or a slight abnormality in the vibration spectrum. In this case, data recording and enhanced monitoring are recommended; immediate shutdown is not necessary.

[0029] Warning (Medium Risk): Significant fault characteristics are present, such as a significant increase in IGBT conduction losses or a sustained increase in bearing vibration amplitude, potentially impacting system stability. Scheduled maintenance is recommended to prevent further deterioration of the fault.

[0030] Emergency Alarm (High Risk): Severe fault characteristics, such as severe current waveform distortion or severe mechanical shock, may cause equipment damage or a safety incident. Immediately shut down the machine for maintenance to prevent further damage.

[0031] Alarm decisions are made based on a comprehensive consideration of the criticality of the faulty component (e.g., a power device failure is more urgent than a cooling fan failure), the rate of characteristic parameter deterioration (e.g., whether vibration amplitude is accelerating), and historical maintenance records (e.g., whether similar components have been recently replaced). The final output includes an alarm message containing the fault location, severity, and recommended actions. This information can be pushed to maintenance personnel via the platform, enabling intelligent equipment health management.

[0032] The present invention provides an intelligent monitoring and alarm method for a motor controller. It uses a dynamic noise baseline to perform joint noise analysis on internal circuit signals and body vibration signals, overcoming the limitation of traditional methods that rely only on a single signal. It can more keenly capture early fault characteristics and reduce false alarms and missed alarms. Component fault identification and impulse response separation are performed based on the dynamic noise baseline, effectively distinguishing normal operating noise from fault characteristics, reducing the impact of environmental interference, and improving the detection rate of hidden faults (such as aging of high-frequency switching devices, mechanical microcracks, etc.). Utilizing a component fault mapping relationship library to perform correlation analysis on current characteristics and vibration characteristics, it is possible to accurately identify faulty components, avoid misjudgments caused by signal segmentation analysis in traditional methods, and provide a reliable basis for predictive maintenance. Intelligent alarm analysis is performed in combination with fault component identification, current characteristics, and vibration characteristics to achieve graded warnings from minor anomalies to serious faults, making operation and maintenance responses more targeted, reducing unnecessary downtime and maintenance, and improving equipment operation efficiency.

[0033] In one embodiment, obtaining an internal circuit signal and a body vibration signal of a motor controller and performing joint noise reduction to generate a dynamic noise baseline includes: The DC bus circuit current is acquired using an isolated high-frequency Rogowski coil current sensor. The sensor's toroidal core is directly attached to the negative copper busbar of the DC busbar to prevent interference from the positive high-voltage potential. An RF anti-aliasing filter is connected in series with the sensor's secondary output, with a cutoff frequency set to 0.4 times the sampling frequency (fixed at 100kHz) to eliminate electromagnetic radiation noise above 40kHz. The filtered analog signal is input into a 24-bit delta-sigma analog-to-digital converter module for digitization using synchronous sampling mode. The converted raw current data stream is transferred to the processing unit via direct memory access. The circuit analysis process is based on the physical topology of the power circuit: the geometric dimensions and insulation parameters of the DC busbar are extracted, and the distributed inductance and interlayer capacitance are calculated. The switching characteristic curves of the insulated gate bipolar transistor module are analyzed from the data sheet to calibrate the nonlinear change model of the junction capacitance during the turn-on and turn-off processes. The equivalent series resistance and parasitic inductance of the busbar capacitor are measured. A sixth-order equivalent circuit network is constructed, incorporating the distributed conductor parameters, device junction capacitance, and busbar capacitance parasitic parameters. The network is fed with raw current data for forward simulation, generating a theoretical current waveform. The residuals of the measured and theoretical waveforms are compared, and the distributed inductance, junction capacitance, and equivalent series resistance / equivalent series inductance parameters are iteratively optimized using a gradient descent algorithm until the residuals are minimized. The optimized model is then used to reversely compensate the measured current, isolating the resonant spikes caused by parasitic line oscillations, reconstructing the true current trajectory generated by the power device switching, and outputting an internal circuit signal. This signal fully preserves the switching transient current rise and fall characteristics, the freewheeling diode reverse recovery current pulse, and the bus capacitor charge and discharge ripple.

[0034] A piezoelectric accelerometer is used for three-axis vibration acquisition. The sensor is fixed to the surface of the controller housing directly above the heat sink of the insulated gate bipolar transistor module via rigid ceramic base bolts. The sensor's X-axis is perpendicular to the housing plane (corresponding to the direction of thermal expansion of the power device chip), the Y-axis is parallel to the direction of the DC bus (corresponding to the direction of electromagnetic force), and the Z-axis extends along the direction of the heat sink fins (corresponding to the vibration axis caused by the coolant flow). The original vibration signal is amplified by the constant current source power supply circuit of the integrated circuit piezoelectric sensor and digitized at a 50kHz sampling rate using a 24-bit synchronous acquisition card. Multi-level interference suppression is required to extract the body vibration signal: Ambient vibration suppression: A reference sensor of the same model is installed in a non-stressed area of ​​the controller to collect ambient vibration signals. A normalized least mean square adaptive filter is used, using the ambient vibration signal as the reference signal and the original vibration signal as the input. The filter coefficients are iteratively updated to minimize the mean square error between the ambient vibration signal and the ambient vibration components in the original vibration signal. The filter then outputs an estimated ambient vibration value. Vector subtraction is performed between the original vibration signal and the estimated ambient vibration value to eliminate the ambient vibration floor.

[0035] Fan disturbance elimination: A Hall effect speed sensor is installed at the cooling fan bearing to capture the fan speed pulse train in real time. The instantaneous speed is calculated based on the pulse interval, generating a set of sine and cosine template signals for the fundamental frequency and its 2nd to 5th harmonics. A multi-channel adaptive notch filter is constructed, using the template signal set as a reference to perform coherent filtering on the vibration signal. This filter dynamically tracks and eliminates fan disturbance spectrum lines within a speed fluctuation range of ±10%.

[0036] Electromagnetic coupling noise isolation: During current acquisition, the leading edge of the electromagnetic pulse generated by the switching action of the power device is detected. Using the leading edge as the time reference, vibration signal segments within a ±5 microsecond time window are extracted. 1000 trigger stacking and averaging are performed to enhance the signal-to-noise ratio of the switching stress wave and suppress random electromagnetic interference. The resulting output body vibration signal includes high-frequency mechanical waves generated by thermal stress deformation of the insulated gate bipolar transistor chip, broadband vibration caused by solder layer fatigue, and forced vibration induced by the interaction between the busbar current and the permanent magnet's magnetic field.

[0037] The target frequency band is dynamically set according to the switching characteristics of the power device: the lower limit frequency is 0.7 times the switching frequency (covering the main lobe of the switching transient energy), and the upper limit frequency is 4 times the switching frequency (including the highest order mode of parasitic oscillation). The bandpass filter adopts a dual-channel parallel processing architecture: Channel A (Phase-Sensitive Filtering): Design a linear-phase finite-length unit impulse response filter. Use a Blackman-Harris window function, with the order determined by the transition band requirements. The passband extends from the lower frequency limit to the upper frequency limit, and the stopband attenuation is greater than 80 dB. Zero-phase-delay convolution filtering is performed on internal circuit signals to preserve the time alignment of the switching current waveforms.

[0038] Channel B (Transient Feature Enhancement): A wavelet packet decomposition tree is constructed, using the Daubechies 10 wavelet basis for a 6-layer decomposition. Nodes containing the switching frequency are selected for reconstruction to highlight the switching transient edge features. The reconstructed signal is then combined with the output of Channel A in a time-domain weighted fusion, with the weight coefficient dynamically adjusted based on the local kurtosis value of the signal. Channel B is given a higher weight at current breakpoints to enhance transient details, while Channel A is used during stable periods to ensure waveform smoothness. The final current waveform in the target frequency band includes the current spike at the turn-on moment of the insulated gate bipolar transistor; the tail current oscillation during the turn-off process; and the resonance caused by the reverse recovery of the freewheeling diode.

[0039] This waveform has strong coupling characteristics with the mechanical vibration of the power device in the time-frequency domain, providing a physical correlation basis for subsequent joint analysis.

[0040] The current waveform in the target frequency band is used as reference signal A, and the body vibration signal is used as signal B. The cross-correlation function RAB(τ) between A and B is calculated, and the time delay τ is swept (within a range of ±1 switching cycle). The global maximum of the cross-correlation function is located. The corresponding τ value is the transmission delay of the current signal relative to the vibration signal (reflecting the delay in electro-mechanical energy conversion). Signal B is time-shifted and calibrated based on the τ value to ensure strict synchronization between the current and vibration signals, generating a synchronized signal group {A, B}.

[0041] Blind source separation processing: The synchronization signal group {A, B} is regarded as a two-dimensional observation signal, assuming that it is a linear mixture of three source signals: Source S1: the actual vibration response of the power device switching current (target signal); Source S2: pseudo vibration induced by electromagnetic noise inside the controller (disturbance 1); Source S3: High-frequency residual vibration transmitted by external shock (disturbance 2).

[0042] Independent component analysis (ICA) is used to solve the unmixing matrix W: {A, B} is preprocessed by centering and whitening. A fixed-point iterative algorithm is used to maximize the non-Gaussianity of the signal and estimate the separation matrix W. W is applied to the observed signal to extract independent components S1, S2, and S3. The component with the largest kurtosis value is selected as the noise-dominant component S1 (the switching current vibration response has an impulsive characteristic and is most non-Gaussian).

[0043] Dynamic noise baseline generation: An adaptive sliding window is applied to the dominant noise component S1. The window width W is positively correlated with the motor controller's real-time load factor (W increases by 5% for every 10% increase in load factor). The root mean square value of S1 is calculated within the window and used as the noise intensity baseline for the current window. The window is slid with a 50% overlap, and the baseline values ​​of adjacent windows are linearly interpolated to generate a continuous and smooth dynamic noise baseline. This baseline represents the natural vibration noise level of the power device in a healthy state and dynamically adjusts with load changes.

[0044] This embodiment uses high-frequency current sampling and circuit parasitic parameter modeling to accurately separate the true current trajectory of the power device switching action, eliminate the resonant distortion caused by the line distributed parameters, and thus obtain a pure circuit signal reflecting the switching characteristics of the device. Using three-axis vibration synchronous acquisition and environmental interference compensation, combined with electromagnetic pulse trigger superposition averaging technology, it effectively removes environmental vibration, fan disturbance and random electromagnetic noise, and extracts the main mechanical vibration characteristics induced by thermal stress and switch stress of the power device. Based on the dynamic design of the switching frequency, a dual-channel bandpass filter architecture is integrated with linear phase filtering and wavelet packet transient enhancement to fully retain key frequency band components such as switching spikes, tail oscillations and diode resonance in the current waveform, and strengthen the physical coupling correlation of the current-vibration signal. Through cross-correlation delay calibration and blind source separation, the mixed noise components of the current and vibration signals are decoupled, and a dynamic noise baseline that is adaptively adjusted with the load rate is generated, providing essential characteristic parameters for the health status of the power device.

[0045] In one embodiment, component fault identification is performed based on a dynamic noise baseline and internal circuit signals to obtain a pre-fault current signature, including: Multi-scale time-frequency decomposition uses a layered analysis strategy to process current signals. The dual-tree complex wavelet transform (DWT) is selected as the core tool, due to its phase-preserving and directional selectivity. Eight layers of decomposition are set to cover the entire frequency spectrum: the first layer resolves high-frequency noise components above 20 kHz, the second to third layers focus on the 5-10 kHz switching transient band, the fourth to sixth layers capture the 1-5 kHz resonant characteristics, and the seventh and eighth layers retain the low-frequency components of the fundamental modulation. Each decomposition layer generates a complex coefficient matrix, with row vectors corresponding to scale indices and column vectors corresponding to time indices. A short-time Fourier transform is performed simultaneously to increase frequency resolution. The Blackman-Harris window is used as the window function, and the window length is dynamically adjusted based on the period of the current fundamental wave. The dual-tree complex wavelet coefficient matrix and the time-spectrum spectrum together constitute the decomposed signal set, forming a three-dimensional feature space structure: the scale dimension represents signal resolution, the frequency dimension quantifies energy distribution, and the time dimension records the feature evolution process.

[0046] Frequency band noise suppression constructs an adaptive filter bank based on the statistical characteristics of a dynamic noise baseline. The noise baseline is analyzed to obtain the energy distribution model for each frequency band, and the signal-to-noise ratio (SNR) of each frequency point in the decomposed signal set is calculated. Nonlinear compression is performed on low SNR frequency bands: soft thresholding is applied to the wavelet domain coefficients to eliminate background noise, and Wiener filtering is used in the time-frequency spectrum to suppress random interference. The target frequency band signal is selected based on the motor controller failure mechanism: power device failure corresponds to the 2-20kHz switching harmonic frequency band, DC bus capacitor degradation manifests as a 100-500Hz low-frequency resonance, and filter capacitor breakdown produces a 5-10kHz characteristic oscillation. The output signal is reconstructed through inverse transformation to obtain the target frequency band signal. The reconstruction process strictly maintains phase continuity to ensure the time positioning accuracy of transient features.

[0047] The waveform segment capture adopts a collaborative mechanism of envelope detection and dynamic threshold. The Hilbert transform extracts the target signal envelope trajectory, and the dynamic detection threshold is calculated based on sliding window statistics. When the signal envelope exceeds the threshold, the event start point is marked, and continuous monitoring is carried out until the envelope falls back below the threshold to determine the end point. Additional zero-crossing point density verification eliminates false triggering: the ratio of the number of zero-crossing points to the duration in the event window is calculated, and events below the set threshold are considered invalid. The valid event capture window extends forward by a quarter of the fundamental wave cycle and backward by half a fundamental wave cycle. All captured segments perform phase synchronization processing: cyclic shift is performed based on the zero-crossing point of the first-order derivative of the signal to eliminate the influence of randomness of the switching cycle. The output candidate fault waveform segment contains complete metadata: peak current, rising edge slope, oscillation duration and other characteristic parameters.

[0048] Multi-dimensional feature matching constructs a five-dimensional fault feature space system. The time domain extracts the waveform rise time parameter, defined as the duration from 10% to 90% of the current amplitude. The energy domain calculates the normalized pulse energy integral to eliminate the effects of amplitude fluctuations. The frequency domain identifies the main resonant frequency and locates the peak of the spectral energy. The waveform domain calculates the kurtosis coefficient to quantify the sharpness of the pulse. The correlation domain analyzes the correlation coefficient between the candidate waveform and the noise baseline to assess feature independence.

[0049] A preset fault signature library stores typical fault template data: an open-circuit fault in an insulated gate bipolar transistor corresponds to a nanosecond rising edge and kilohertz oscillation, a breakdown in a freewheeling diode manifests as a microsecond delayed turn-on, and capacitor aging causes low-frequency resonance. The matching process calculates the weighted Euclidean distance between the candidate waveform and each template, generating a multidimensional feature vector to represent the degree of match.

[0050] Similarity matching utilizes a radial basis probability transformation model. The template matching function is defined as an exponential decay function, with the Euclidean distance between the input feature vector and the ideal vector determining the decay rate. Probability normalization ensures that the sum of all fault type probabilities is 1. Historical diagnostic accuracy is incorporated as a template weighting factor: templates with verified accuracy receive increased weighting, while templates with high misdiagnosis rates receive reduced weighting. The dynamic weighting factor update mechanism is continuously optimized based on field operation data.

[0051] The elements of the output probability distribution vector correspond to the fault type index in the preset fault signature library. The probability values ​​quantify the confidence level of the match between the current signal and each type of fault. The shape of the probability distribution reflects the clarity of the fault signature: a unimodal distribution indicates a typical fault, while a multimodal distribution suggests the possibility of a complex fault.

[0052] Confidence threshold screening implements a three-level decision-making mechanism. First-level probability threshold decision: When the probability of a single fault type exceeds 0.8, the diagnosis result is directly output. Second-level gradient decision: For fault types with probability values ​​between 0.5 and 0.8, the probability gradient difference between them and adjacent faults is calculated. When the gradient difference exceeds 0.2, the dominant fault type is output. Third-level temporal persistence verification: For suspected faults that do not reach the threshold, three consecutive detection cycles are monitored. If the probability of the same fault type continues to increase, the output is confirmed.

[0053] The final fault precursor current characteristics are packaged into structured data: the fault type coding adopts the International Electrotechnical Commission standard classification, the confidence level is divided into three levels: high / medium / low, and the characteristic waveform fingerprint is generated by the wavelet coefficient hash algorithm to generate a unique identifier. This characteristic data provides the input basis for subsequent component-level positioning This embodiment constructs a three-dimensional feature space through multi-scale time-frequency decomposition to achieve synchronous analysis of the transient and steady-state characteristics of the current signal, thereby fully preserving the correlation between high-frequency switching events and low-frequency modulation characteristics. Based on the adaptive frequency band suppression of the dynamic noise baseline, sensitive frequency bands are selected for noise filtering according to different fault mechanisms, effectively improving the characteristic signal-to-noise ratio under variable working conditions. A waveform interception mechanism that cooperates with envelope detection and dynamic threshold is adopted, combined with phase synchronization processing to ensure the time positioning accuracy and morphological integrity of the fault characteristic waveform. A five-dimensional feature space matching system is established to quantitatively evaluate the similarity between the waveform and the typical fault template, providing a measurable feature basis for probabilistic diagnosis.

[0054] In one embodiment, component shock response separation is performed based on the dynamic noise baseline and the body vibration signal to obtain the hidden fault vibration characteristics, including: A dynamic noise baseline serves as a reference input to the adaptive noise cancellation system. The system establishes a secondary channel transfer function model, simulating the physical path characteristics of circuit switching noise transmitted through the controller's mechanical structure to the vibration sensor. The transfer function model consists of a third-order infinite-length unit impulse response filter structure, whose coefficients are calibrated using white noise excitation vibration table experiments. The dynamic noise baseline is used as the reference signal input to the filter, while the body vibration signal serves as the primary input. A normalized least mean square algorithm is used to iteratively update the filter coefficients to minimize the mean square error between the filter output signal and the circuit noise component in the vibration signal.

[0055] Each iteration calculates the cross-correlation vector between the reference signal autocorrelation matrix and the main input signal, and the update step size is set to 0.01 to ensure convergence stability. The filter output signal is subtracted from the body vibration signal to eliminate the mechanical resonance interference component synchronized with the circuit switch. The residual signal is input into the Butterworth band-stop filter group, and the stopband center frequency is dynamically set according to the real-time speed of the cooling fan: the fundamental frequency is the speed divided by 60, the second harmonic is twice the fundamental frequency, and the stopband width is set to ±7% of the speed fluctuation range. The filter order is fixed at 6th order, and the stopband attenuation is greater than 45 decibels. The final output denoised vibration signal only retains the pure impact response components of mechanical components such as bearings, gears, and shafting, completely stripping away circuit switching noise and rotating mechanical disturbances.

[0056] A three-level wavelet packet decomposition tree is constructed to denoise the vibration signal. The first level uses the Daubechies 12 wavelet basis function for an 8-level full-band decomposition, dividing the 0-10 kHz band into 256 equally wide subbands. The second level predefines sensitive frequency bands based on the physical characteristics of the mechanical fault: the characteristic frequency band for bearing outer race damage is 1.2-3.5 kHz (covering harmonics of the ball passing frequency), the characteristic frequency band for shaft eccentricity is 1.8-4.2 times the motor fundamental frequency multiplied by the pole pair number (corresponding to harmonics of mass imbalance forces), and the characteristic frequency band for gear tooth breakage is ±3 times the meshing fundamental frequency (including sideband modulation energy). The third level calculates the energy entropy and kurtosis of each subband signal: an energy entropy below 0.7 indicates concentrated energy, while a kurtosis greater than 3.5 indicates impact characteristics. Subband nodes that meet both low entropy and high kurtosis are selected, and the time domain signals of these nodes are reconstructed. The reconstruction process uses an inverse wavelet packet transform, and the fused signal is generated by superposition.

[0057] The fault-related frequency band signal presents a periodic impact sequence in the time domain: bearing damage is manifested as a narrow pulse (pulse width 50-200 microseconds) appearing at a fixed phase per revolution, shaft eccentricity is displayed as a wide-amplitude oscillation (duration 1-3 milliseconds) that is an integer multiple of the rotation frequency, and gear abnormality is manifested as a pulse group modulated by the meshing frequency (group duration 2-5 milliseconds).

[0058] The fault-related frequency band signal is input into a Hilbert transformer to generate an analytical signal. The Hilbert transform is implemented using a 90-degree phase-shift network consisting of a 601-order finite-length unit impulse response filter with a passband ripple of less than 0.1 dB. The modulus of the analytical signal is extracted as the envelope signal to eliminate the influence of the high-frequency carrier. A 1024-point fast Fourier transform is performed on the envelope signal with a frequency resolution of 0.5 Hz. Significant peaks are detected in the envelope spectrum. Peaks are identified when the amplitude exceeds three standard deviations of the background noise mean and the half-power bandwidth is less than 2 Hz.

[0059] Three types of periodic features are extracted: the fundamental frequency corresponds to the bearing ball passing frequency or the gear meshing fundamental frequency; the harmonic family feature calculates the cumulative amplitude sum of the 2nd to 10th harmonics of the fundamental frequency; and the sideband feature, targeting gear faults, calculates the ratio of the sideband amplitudes centered on the meshing frequency and ± the rotational frequency. The output is a 12-dimensional impact periodic feature vector, including parameters such as the fundamental frequency, harmonic energy fraction, and sideband modulation depth, fully characterizing the periodicity of the impact event.

[0060] A preset mechanical parameter database is used: bearing geometry parameters include ball diameter, pitch diameter, and contact angle; gear parameters include number of teeth, module, and pressure angle; and shafting parameters include rotor mass and support stiffness. Physical models are used to convert impact periodicity characteristics: bearing fault energy is calculated based on Hertzian contact theory, converting the integral of the harmonic energy of the ball passing frequency into an equivalent damage area. The damage location is determined by the frequency component (the characteristic frequency of inner race faults is the ball passing frequency multiplied by 0.6, and the outer race by 0.4).

[0061] Shaft eccentricity energy is calculated by converting the amplitude of the rotational harmonics into an equivalent imbalance (in grams per millimeter) based on a mass-stiffness model. Gear fault energy is calculated by mapping the sideband modulation depth to a percentage of the root crack depth based on the tooth root bending stress formula. This output is a three-dimensional fault impact energy distribution vector: the first dimension represents the bearing damage index (0-100%), the second dimension represents the shaft eccentricity magnitude (in microns), and the third dimension represents the gear fault intensity (a dimensionless coefficient between 0 and 1.0). This distribution quantifies the extent of physical damage to the mechanical component.

[0062] A three-state hidden Markov model was established: the hidden states consisted of a healthy state, an early fault state, and a severe fault state. The observation sequence consisted of a 12-dimensional impulse periodicity feature vector concatenated with a 3-dimensional fault impulse energy distribution vector. The state transition probability matrix was trained using historical fault data: the Baum-Welch algorithm was used to iteratively optimize the transition probabilities, with the number of iterations set to 1000 or a convergence threshold of 0.001. The Viterbi algorithm was used to decode the optimal state sequence: the most likely state transition path was calculated under the current observation sequence.

[0063] The judgment rule is: when the early fault state lasts for more than 60% of the total observation time and the probability of transitioning to a severe fault state is greater than 0.3, a latent fault flag is triggered. The output latent fault vibration signature includes three elements: fault type code (bearing wear code 1, shaft eccentricity code 2, gear anomaly code 3), fault development stage (0 healthy / 1 early / 2 severe), and confidence level (a 0-100% value calculated based on the state transition probability). This feature enables quantitative assessment and graded early warning of mechanical component failures.

[0064] This embodiment uses adaptive noise cancellation of a dynamic noise baseline to accurately separate circuit switching noise from mechanical shock responses, eliminate false fault indications caused by environmental coupling interference, and ensure that the vibration signal only reflects the true status of components such as bearings and gears. Based on multi-scale resonant frequency band extraction and physical characteristic frequency band preset, fault-sensitive frequency bands of bearing wear, shaft eccentricity, and gear meshing anomalies are captured in a targeted manner, the signal-to-noise ratio of weak shock features is enhanced, and feature dilution caused by broadband analysis is avoided. Combining envelope spectrum analysis with physical modeling of mechanical parameters, the impact periodic characteristics are converted into quantifiable engineering parameters such as bearing damage area and shaft eccentricity magnitude, realizing physical meaning mapping of the fault degree.

[0065] In one embodiment, component association mapping is performed on the fault precursor current characteristics and the latent fault vibration characteristics to obtain an internal fault component identifier, including: The core of physical characteristic frequency band analysis lies in converting the time domain signal into a frequency domain representation that can reflect the fault characteristics. The fault precursor current characteristics are usually manifested as harmonic distortion or amplitude abnormality in a specific frequency band, while the latent fault vibration characteristics are mostly concentrated in the mechanical resonance frequency band or near the bearing defect characteristic frequency. The analysis process uses a multi-scale signal decomposition method to decompose the original current signal into fundamental components, harmonic components, and high-frequency noise components. At the same time, the vibration signal is envelope demodulated to extract the fault characteristics modulated on the carrier frequency. The frequency band division of the current signal is determined by the motor type and power supply frequency. For example, for an asynchronous motor with a 50Hz power supply, the focus is on the sideband components in the range of 100-1500Hz; the frequency band analysis of the vibration signal combines the bearing geometric parameters to calculate the fault characteristic frequency and extract the energy distribution in the range of 1-5 times the rotation frequency.

[0066] The construction of the joint feature vector of current and vibration signals relies on the cross-integration of time-frequency domain features. Correlation is calculated in the frequency domain between the Park vector modulus analysis results of the current signal and the order ratio analysis results of the vibration signal to extract co-occurrence features of the two under the same fault mode. For example, a winding short-circuit fault may manifest as a third harmonic increase in the current spectrum while simultaneously triggering an increase in electromagnetic vibration at twice the line frequency in the vibration spectrum. This cross-modal feature is encoded as one dimension of the joint feature vector. Other dimensions of the feature vector include cross-domain statistics such as the ratio of the total harmonic distortion of the current to the total effective value of the vibration and the difference in energy entropy in specific frequency bands. Dimensional optimization of the joint current-vibration feature is achieved through principal component analysis, retaining principal components with a cumulative contribution exceeding 85% as the final feature representation.

[0067] The fault component mapping rules are based on the topological relationships of the motor's physical structure, converting the coupling relationship between electrical and mechanical features into graph constraints. Each fault component node in the mapping relationship library contains three sets of attributes: an electrical signature, a mechanical signature, and the connection weights between components. For example, the electrical signature of a bearing component may contain high-frequency components of the shaft current, while the mechanical signature contains energy in a specific frequency band of radial vibration acceleration. The topological matching process uses a bidirectional propagation algorithm, starting from both the electrical component nodes triggered by the current signature and the mechanical component nodes triggered by the vibration signature, and then spreading feature similarity along the connecting edges.

[0068] The generation of candidate sets follows a dual verification mechanism. The primary screening, based on the feature amplitude threshold, selects components associated with dimensions in the joint feature vector that exceed the normal range by 3σ. The secondary screening imposes a topological consistency constraint, requiring that candidate components must appear in the intersection of both the electrical and mechanical propagation paths. For permanent magnet synchronous motors, candidate components for rotor demagnetization faults must simultaneously meet the conditions of excessive current harmonic amplitude and torque pulsation frequency matching. The mapping rule also introduces a time decay factor to perform a weighted average of the matching degrees of historical fault features to avoid mismatches caused by transient interference. The final output candidate fault component set is sorted by matching confidence, and the top K components are retained for the next stage of verification.

[0069] The essence of frequency band distribution verification is to verify the physical feasibility of fault characteristics. Each component in the candidate component set must meet the requirement for the consistency between its theoretical fault characteristic frequency and the actual detection frequency band. For components with bearing outer race faults, it is necessary to verify the presence of the theoretically calculated outer race fault frequency and its harmonics in the vibration spectrum, and to check whether the corresponding slip modulation sideband appears in the current signal. The verification process uses frequency band energy ratio analysis to calculate the relative change in actual signal energy compared to the baseline energy within the theoretical fault frequency band. Components whose change exceeds a threshold and persists for more than three sampling cycles are marked as valid.

[0070] The consistency verification of multiple physical quantities is the core link. Valid components must exhibit coupling anomalies in both current and vibration characteristics. For example, insulation-aged components must simultaneously meet the conditions of excessive current leakage components and enhanced local discharge ultrasonic vibration. The frequency band coherence index is introduced into the verification process to calculate the cross-power spectral density of the current specific harmonics and the vibration characteristic frequency band, and components with a coherence coefficient greater than 0.7 are retained. For the open-circuit fault of the power device IGBT, the time difference between the collector current drop moment and the radiator vibration impact signal is required to be within the microsecond range. The final set of valid fault components is screened by constructing a fault feature decision tree. Each branch node corresponds to a frequency band verification condition. Only components that pass all necessary nodes can enter the final set.

[0071] Primary-secondary conflict resolution addresses the issue of feature competition when multiple faults occur simultaneously. The mapping relationship library predefines the fault propagation paths and causal relationship matrix between components. For example, bearing wear can lead to shaft misalignment, which in turn causes rotor eccentricity. The resolution algorithm utilizes an evidence theory framework to convert the feature matching degree of each valid component into a basic probability distribution and calculate the joint confidence using the Dempster combination rule. For component pairs with a causal relationship, the confidence of the predecessor component influences the final score of the successor component with a weight of 0.7.

[0072] Conflict detection is based on the principle of fault signature exclusivity. Conflict detection is triggered when the theoretical fault frequency band overlap of two components exceeds 30%. The origin of the signature is determined by tracing back the time-frequency distribution of the original signal. For example, a stator winding short circuit and a rotor bar breakage both cause a 2sf frequency component in the current, but the former is accompanied by distortion of the air gap flux waveform, while the latter exhibits torque pulsation characteristics. A fault development time series model is introduced during the resolution process, prioritizing the component that first exhibits abnormal characteristics as the primary fault source. The final set of primary fault component identifiers meets two conditions: the difference in confidence between components within the set is greater than 20%, and the difference in confidence between any component and components outside the set is greater than 35%.

[0073] Reverse mapping establishes a precise correlation from macroscopic features to microscopic defects. The component fault mapping relationship library stores the failure mode fingerprint library of each major fault component. For example, the fingerprint of the inner ring peeling of the bearing includes a nonlinear relationship curve between the peeling area and the vibration impact amplitude. The mapping process adopts a hierarchical matching strategy. The first layer matches the general category of fault type (such as electrical fault, mechanical fault), and the second layer matches the specific failure mechanism (such as insulation thermal aging, mechanical fatigue fracture). For the permanent magnet motor rotor system, it is necessary to distinguish between the two modes of local demagnetization of permanent magnets and rotor eccentricity. The former manifests as prominent fractional harmonics in the current harmonics, while the latter causes the growth of integer multiple harmonics of the rotation frequency.

[0074] Internal localization of faulty components is achieved through characteristic propagation path analysis. For stator winding interturn short circuits, the fault phase is determined based on the phase difference between the negative-sequence current component and the axial vibration component. For bearing faults, the circumferential position of the damage point relative to the sensor is determined by calculating the vibration signal propagation delay. The final output of the internal fault component identification contains three levels of information: component type (e.g., deep groove ball bearing), fault nature (e.g., outer ring spalling), and damage severity (e.g., spalling area as a percentage of the contact area). This identification information is linked to the coordinates of the device's 3D model, accurately marking the fault location in the visual interface.

[0075] This embodiment fuses the fault features of current and vibration signals into a joint feature vector through physical feature band analysis, realizes the collaborative characterization of multi-physical quantity fault features, and significantly improves the detection sensitivity of early weak faults. The component mapping rules based on topological constraints establish an association network of electrical and mechanical features, so that the screening process of candidate fault components is both physically reasonable and computationally efficient. The frequency band distribution verification link effectively eliminates the misjudgment caused by interference signals through multi-modal feature consistency verification, ensuring the accuracy of fault location. The primary and secondary conflict resolution mechanism combined with the fault propagation path analysis can distinguish the root components and derivative components in concurrent faults, providing an accurate basis for maintenance decisions. Ultimately, the accurate inversion from macroscopic features to microscopic defects is achieved through reverse mapping of fault modes, enabling maintenance personnel to directly identify the specific damage type and location, significantly shortening the fault diagnosis and disposal time.

[0076] In one embodiment, an alarm analysis is performed based on the fault component identification, the fault precursor current characteristics, and the latent fault vibration characteristics to obtain a hierarchical alarm decision, including: Fault component identification is one of the core inputs of intelligent monitoring and alarm methods. Its essence is to mark the real-time status of the internal components of the motor controller through sensors or diagnostic modules. The fault type parsing process uses the preset component type information to deeply deconstruct the semantics of the identification, and converts the abstract fault code or signal characteristics into a classifiable fault knowledge system. Electrical fault type information usually covers typical problems such as winding insulation deterioration, power device breakdown, and contactor contact erosion. Mechanical fault type information includes physical damage such as bearing wear, rotor eccentricity, and coupling looseness. During the parsing process, the fault component identification is matched with the predefined fault knowledge graph. The knowledge graph organizes the fault types in a tree structure. The root node distinguishes between electrical and mechanical categories, and the sub-nodes refine the specific fault mode.

[0077] The analysis of electrical faults requires consideration of the historical degradation trajectory of component electrical parameters. For example, the exponential decay of insulation resistance may indicate dampness in the windings, while a sudden change in the on-state voltage drop of an IGBT module may indicate thermal fatigue in the chip. The analysis of mechanical faults relies on the correlation of time-frequency characteristics of vibration signals. For example, the appearance of characteristic frequencies of bearing faults is often accompanied by an increase in the energy of the high-frequency resonance band. Fuzzy inference mechanisms are introduced into the analysis process to address fault identifiers with unclear boundaries, and the uncertainty of the fault type is quantified through membership functions. The correlation between electrical and mechanical faults is of particular concern. For example, rotor eccentricity may simultaneously cause increased current harmonics and excessive radial vibration. In this case, the analysis results generate a composite fault type label. Fault type information is output in a structured data format, containing metadata such as fault code, confidence level, and timestamp, providing standardized input for subsequent multimodal feature fusion.

[0078] Pre-fault current signatures reflect early electrical anomalies in motor controllers before complete failure, manifesting as current waveform distortion, harmonic component variation, or asymmetry in the dq-axis currents. Hidden fault vibration signatures capture the sub-health of mechanical structures, such as the impact vibration envelope modulation caused by microcracks in the inner ring of a bearing. The core of multimodal feature fusion lies in establishing a cross-domain correlation model between electrical and mechanical signals, overcoming the limitations of traditional single-modal diagnosis. Current feature extraction focuses on time-frequency indicators such as the Park vector trajectory ellipticity and negative-sequence current ratio of the three-phase current. Vibration signatures extract fault-sensitive components from the envelope spectrum and wavelet energy entropy of the acceleration signal.

[0079] Feature fusion is performed at three levels: the data layer aligns the time scales of current and vibration signals, ensuring synchronization through a sliding time window. The feature layer uses an attention mechanism to dynamically weight the contributions of different modal features, for example, increasing the weight when current harmonics and vibration sidebands are coupled. The decision layer uses a graph neural network to construct topological relationships between features, with nodes representing individual features and edge weights reflecting causal relationships between features. An adversarial autoencoder is introduced during the fusion process to eliminate modal differences caused by sensor noise while retaining common features that are strongly correlated with the fault. Manifold learning is used to reduce the dimensionality of the integrated fault feature vector, projecting high-dimensional features into a low-dimensional fault representation space while maintaining intra-class compactness and inter-class separability. The resulting vector contains temporal features (such as the vibration response delay at the current mutation point), statistical features (such as the covariance between vibration kurtosis and current peaks), and topological features (such as the phase coupling matrix of the multi-sensor signals).

[0080] Hazard diffusion assessment simulates the propagation path and impact range of faults in the motor controller system. The comprehensive fault feature vector provides a quantitative description of the current state, while the fault type information defines the potential propagation direction. Hazard assessment for electrical faults follows the circuit topology. For example, an open-circuit IGBT fault can cause DC bus voltage oscillation, and the degree of hazard is strongly correlated with the location of the bridge arm. Mechanical fault assessment is based on the mechanical impedance network model. Bearing faults can transmit vibration energy through the shaft system, causing encoder signal distortion. The assessment process uses a dynamic Bayesian network to model fault propagation probability. Nodes represent subsystem states, and edge weights are dynamically updated based on the timing correlation parameters in the comprehensive feature vector.

[0081] The hazard level score is calculated using three dimensions: the temporal dimension assesses the fault development rate, using the current characteristic slope combined with the vibration energy accumulation rate to predict the remaining useful life. The spatial dimension analyzes the extent of the fault's impact; localized overheating in the stator winding can affect the insulation of adjacent phases through heat conduction. The system dimension quantifies the risk of functional failure. For example, a position sensor failure causing a closed-loop control breakdown is more severe than a cooling fan failure. A scoring mechanism processes uncertain information, with a library representing the fault state and transition rules defined by an expert knowledge base. The final score is based on a 0-100 scale, with scores below 30 indicating tolerable degradation and scores above 70 requiring immediate shutdown. The score results are accompanied by a description of the propagation path, such as the chain reaction chain of "1. Bearing wear, 2. Rotor dynamic imbalance, 3. Increased current harmonics, 4. Bus capacitor overheating."

[0082] The pre-configured alarm policy library is a hybrid decision-making system based on rules and machine learning. It incorporates industry standards (such as ISO 10816 vibration thresholds), equipment manufacturer maintenance procedures, and a library of historical fault resolution cases. The decision mapping process discretizes continuous hazard scores into graded alarm actions, typically categorized as early warning (monitoring level), alarm (maintenance level), and emergency alarm (shutdown level). The policy library employs a multi-level decision tree structure, with the root node determining whether the fault is self-recoverable, and sub-nodes evaluating contextual factors such as the availability of maintenance resources and the impact on production plans.

[0083] A reinforcement learning mechanism is introduced to generate alarm decisions, and the reward function considers the balance between false alarm rate and missed alarm rate. For example, when the false positive rate of vibration characteristics is high, the alarm threshold is appropriately increased. The dynamic weight adjustment module optimizes the strategy according to the equipment operation stage (such as the running-in period and the stabilization period). The running-in period is more sensitive to minor anomalies. The classification results output a structured instruction set, including the alarm level, recommended disposal measures (such as "replace the bearing within 48 hours"), related parameter adjustments (such as derating to 10% power), and verification methods (such as infrared thermal imager re-inspection). The alarm decision is linked to the equipment management system to automatically generate work orders or trigger the spare parts procurement process. At the same time, the decision basis is pushed to the system for production scheduling optimization. The decision feedback loop updates the strategy library in real time, and the strategy self-optimization process is triggered when the same fault is misjudged multiple times.

[0084] This embodiment analyzes the types of electrical faults and mechanical faults, which can clearly distinguish different fault modes and their potential impacts, and avoid misjudgments or missed judgments caused by traditional single diagnostic methods. Multimodal feature fusion technology effectively integrates the complementary information of current and vibration signals, significantly improving the comprehensiveness and reliability of fault feature extraction, and is particularly suitable for hidden fault detection under complex working conditions. Hazard diffusion assessment combined with a dynamic propagation model can quantify the potential hazard level of a fault, provide an objective basis for maintenance decisions, and reduce the risk of over-maintenance or under-maintenance. The intelligent mapping mechanism of the hierarchical alarm strategy library can dynamically adjust the alarm threshold according to the actual operating status, optimize the allocation of operation and maintenance resources while ensuring equipment safety, and reduce the probability of unplanned downtime.

[0085] In one embodiment, a hazard diffusion assessment is performed on the electrical fault type information and the mechanical fault type information based on the comprehensive fault feature vector to obtain a fault hazard level score, including: The comprehensive fault feature vector is a high-dimensional data representation derived from the fusion of multimodal features, encompassing coupled information about electrical and mechanical faults. The goal of fault parameter extraction is to convert it into a quantifiable global fault intensity index, which reflects the severity and development trend of the fault. The extraction process combines feature dimensionality reduction with key indicator aggregation. Principal component analysis (PCA) or autoencoders is first used to eliminate redundant features, retaining the core dimensions that are sensitive to the fault. Electrical feature components such as current harmonic distortion and negative sequence current ratio are normalized to a range of 0–1, while mechanical feature components such as vibration envelope amplitude and shock pulse count are normalized according to ISO standards.

[0086] The calculation of the global fault severity index integrates time series and statistical characteristics. The time series component evaluates the evolution rate of fault features, such as the growth rate of current waveform distortion over the past 5 minutes. The statistical component calculates the degree of deviation of features from the historical baseline, such as the probability of the effective vibration value exceeding the 3σ range under normal operating conditions. The index synthesis uses a weighted geometric mean rather than an arithmetic mean to prevent outliers in a single feature from excessively influencing the overall assessment. Weights are dynamically adjusted based on the correlation between the feature and the fault type; for example, vibration features in a bearing fault are weighted higher than current features. The resulting global fault severity index is a unitless scalar ranging from 0 to 100, with larger values ​​indicating a greater threat to the overall system performance. This index is accompanied by a confidence metric that reflects the uncertainty in data quality during the feature extraction process.

[0087] The dependencies among motor controller components form a complex network of physical and functional coupling. Electrical faults (such as IGBT overheating) can affect the mechanical structure through thermal conduction (such as heat sink deformation), and mechanical faults (such as bearing wear) can also cause electrical parameter anomalies (such as back EMF distortion). Graph modeling uses components as nodes and energy flows, signal flows, or physical connections as edges to construct a directed weighted graph. Electrical dependencies include power topology (such as the DC bus, inverter, and motor windings) and control loops (such as current sampling and PWM modulation). Mechanical dependencies cover shaft system dynamics (such as bearings, rotors, and couplings) and structural vibration transmission paths (such as the housing and mounting base).

[0088] The modeling process incorporates multi-layer graph theory. The bottom layer is a physical connection graph (e.g., the relationship between PCB traces and mechanical assembly), the middle layer is a functional dependency graph (e.g., the cooling capacity of the heat dissipation system for the power module), and the upper layer is a logical impact graph (e.g., the control algorithm failure caused by a sensor failure). Edge weights quantify the probability of fault propagation. For example, the probability of a short circuit caused by aging winding insulation is higher than the probability of electromagnetic interference on the encoder signal. A dynamic graph update mechanism adjusts weights in real time. When a temperature sensor detects localized overheating, the outgoing edge weight of the corresponding node is increased. The output of the fault propagation path network is stored in a graph database, supporting SPARQL query language for fast retrieval of multi-hop propagation paths, such as the complete chain of "1. stator winding overheating, 2. permanent magnet demagnetization, 3. torque fluctuation, 4. mechanical resonance."

[0089] Path propagation simulation models the dynamic propagation of faults through a component network, using a global fault intensity index as the initial stimulus injected into the source node. The simulation engine utilizes a discrete event-driven model, evaluating the fault propagation intensity along each edge at each time step, taking into account attenuation factors (such as distance attenuation and damping effects) and amplification factors (such as positive feedback loops). Electrical path simulation relies on circuit theory, where the distribution of fault current within the impedance network determines the degree of impact on each branch. Mechanical path simulation utilizes the vibration transfer function, calculating the attenuation rate of impact energy across different structures.

[0090] The simulation output includes three key parameters: The spatial dimension generates the fault impact domain, identifying the affected components and their contamination levels (e.g., "busbar capacitor temperature rise +8K"); the temporal dimension predicts the fault progression stage, estimating the time window from the current state to functional failure (e.g., "remaining life 72 ± 15 hours"); and the system dimension calculates the chain reaction risk value, quantifying the likelihood of triggering a secondary fault (e.g., "the probability of encoder failure due to bearing wear is 37%"). The parameter calculation incorporates the Monte Carlo method to address uncertainty, generating a probability distribution through thousands of random samplings. The fault propagation assessment parameters are ultimately encapsulated as a multidimensional vector, containing the boundary values ​​for the worst-case, most likely, and optimal scenarios, providing probabilistic input for the subsequent hazard level scoring.

[0091] The historical fault case database is a structured database of similar equipment failure events. Each record contains fault characteristics, diffusion paths, remediation measures, and consequence severity labels. The assessment process uses case-based reasoning (CBR) to compare the current fault diffusion assessment parameters with historical cases for similarity. Similarity is calculated by covering the Euclidean distance of parameter values, graph isomorphism of propagation paths, and dynamic time warping (DTW) distance of time evolution. The matching results retrieve the top-K most similar cases, and the weighted average of their hazard levels is used as the baseline score.

[0092] Expert rules are introduced in the scoring calibration process to correct deviations. For example, when the diffusion parameters show that the fault involves safety-critical components (such as brake resistors), the score will automatically increase by 20%. The final hazard level score adopts a five-level system: Level 1 (minor, can be delayed), Level 2 (medium, requires planned maintenance), Level 3 (serious, recommended shutdown for maintenance), Level 4 (critical, must be shut down immediately), and Level 5 (catastrophic, may cause equipment damage). The scoring report is accompanied by historical case references, such as "The current fault is 82% similar to Case #47 in 2023. This case caused the motor to burn out due to untimely processing." The dynamic learning mechanism automatically stores newly occurring fault cases in the database and continuously optimizes the accuracy of the assessment. The fault hazard level score is associated with the maintenance knowledge graph, and the corresponding emergency plan document is directly pushed to the operation and maintenance terminal. This embodiment can accurately quantify the severity and development trend of the fault by extracting the global fault intensity index of the comprehensive fault feature vector, avoid misjudgment of a single feature, and improve the reliability of fault assessment. Graph structure modeling based on internal component dependencies can clearly depict the coupled propagation paths of electrical and mechanical faults, visualize the fault impact range, and facilitate accurate location of risk sources. Path propagation simulation combined with dynamic attenuation and amplification factors can simulate the spatiotemporal evolution of fault diffusion, predict remaining life and chain reaction risks, and provide a decision-making basis for proactive maintenance. Matching of historical fault case libraries and expert rule calibration can optimize hazard level scoring based on past experience, ensuring that the assessment results are consistent with both theoretical models and actual operation and maintenance needs. The resulting five-level hazard scoring system can intuitively reflect the urgency of the fault and associate it with emergency plans, significantly improving fault response efficiency and equipment safety.

[0093] Reference Figure 2 As shown, the present invention also provides an intelligent monitoring and alarm system for a motor controller, which is applied to any of the above-mentioned intelligent monitoring and alarm methods for a motor controller, comprising: The acquisition module is used to obtain the internal circuit signal and body vibration signal of the motor controller, and perform joint noise analysis to generate a dynamic noise baseline; An analysis module is used to identify component faults based on a dynamic noise baseline and internal circuit signals to obtain pre-fault current characteristics; A correlation module is used to separate the component impact response based on the dynamic noise baseline and the body vibration signal to obtain the vibration characteristics of the hidden fault; A processing module is used to perform component association mapping on the fault precursor current characteristics and the latent fault vibration characteristics to obtain an internal fault component identifier; The control module is used to perform alarm analysis based on the fault component identification, fault precursor current characteristics and hidden fault vibration characteristics to obtain a graded alarm decision.

[0094] The present invention provides an intelligent monitoring and alarm system for a motor controller. It uses a dynamic noise baseline to perform joint noise analysis on internal circuit signals and body vibration signals, overcoming the limitations of traditional methods that rely on only a single signal. It can more keenly capture early fault characteristics and reduce false alarms and missed alarms. Component fault identification and impulse response separation are performed based on the dynamic noise baseline, effectively distinguishing normal operating noise from fault characteristics, reducing the impact of environmental interference, and improving the detection rate of hidden faults (such as aging of high-frequency switching devices, mechanical microcracks, etc.). Utilizing a component fault mapping relationship library to perform correlation analysis on current characteristics and vibration characteristics, it is possible to accurately identify faulty components, avoid misjudgments caused by signal segmentation analysis in traditional methods, and provide a reliable basis for predictive maintenance. Intelligent alarm analysis is performed in combination with fault component identification, current characteristics, and vibration characteristics to achieve graded warnings from minor anomalies to serious faults, making operation and maintenance responses more targeted, reducing unnecessary downtime and maintenance, and improving equipment operation efficiency.

[0095] It should be noted that, those skilled in the art will clearly understand that, for the sake of convenience and brevity of description, the specific working processes of the above-described system and each module can refer to the corresponding processes in the aforementioned method embodiments and will not be repeated here.

[0096] The above description is only a preferred embodiment of the present invention and does not limit the patent scope of the present invention. Any equivalent structure or equivalent process transformation made by using the contents of the present invention description and drawings, or directly or indirectly applied in other related technical fields, are also included in the patent protection scope of the present invention.

Claims

1. An intelligent monitoring and alarm method for a motor controller, characterized in that: include: Obtain the internal circuit signal and body vibration signal of the motor controller, perform joint noise analysis, and generate a dynamic noise baseline; performing component fault identification based on the dynamic noise baseline and the internal circuit signal to obtain a fault precursor current characteristic; Separating component impact responses based on the dynamic noise baseline and the body vibration signal to obtain a hidden fault vibration signature; Performing component association mapping on the fault precursor current characteristics and the hidden fault vibration characteristics to obtain an internal fault component identifier; An alarm analysis is performed based on the fault component identifier, the fault precursor current characteristics and the hidden fault vibration characteristics to obtain a graded alarm decision.

2. The intelligent monitoring and alarm method for a motor controller according to claim 1, characterized in that: The obtaining of the internal circuit signal and the body vibration signal of the motor controller and performing joint noise reduction to generate a dynamic noise baseline includes: collecting current from a DC bus circuit of the motor controller to obtain the internal circuit signal; Performing triaxial vibration acquisition on the power device installation area of ​​the motor controller housing to obtain the body vibration signal; performing bandpass filtering on the internal circuit signal to obtain a current waveform in a target frequency band; Cross-correlation analysis and blind source separation processing are performed on the target frequency band current waveform and the body vibration signal to obtain the dynamic noise baseline.

3. The intelligent monitoring and alarm method for a motor controller according to claim 1, characterized in that: The performing component fault identification based on the dynamic noise baseline and the internal circuit signal to obtain a fault precursor current feature includes: Performing multi-scale time-frequency decomposition on the internal circuit signal to obtain a decomposition signal set; Performing frequency band noise suppression on the decomposed signal set according to the dynamic noise baseline to obtain a target frequency band signal; Performing waveform segment interception processing on the target frequency band signal to obtain candidate fault waveform segments; Performing multi-dimensional feature matching on the candidate fault waveform segment based on a preset fault feature library to obtain a multi-dimensional feature vector; Perform similarity matching on the multidimensional feature vector according to the fault template of the preset fault feature library, and output a probability distribution of the fault type; The fault type probability distribution is screened using a confidence threshold to obtain the fault precursor current characteristics.

4. The intelligent monitoring and alarm method for a motor controller according to claim 1, characterized in that: The separating of component impact responses based on the dynamic noise baseline and the body vibration signal to obtain a hidden fault vibration feature includes: Suppressing environmental coupling noise on the body vibration signal according to the dynamic noise baseline to obtain a denoised vibration signal; Performing multi-scale resonance frequency band extraction on the denoised vibration signal to obtain fault-related frequency band signals including bearing wear, shaft eccentricity and gear meshing abnormality; Performing envelope spectrum analysis on the fault-related frequency band signal to obtain impact periodicity characteristics; Performing physical characteristic frequency band analysis on the impact periodicity characteristics based on preset mechanical component structural parameters to obtain fault impact energy distribution; The hidden fault vibration characteristics are obtained by performing a probabilistic judgment based on the impact periodic characteristics and the fault impact energy distribution.

5. The intelligent monitoring and alarm method for a motor controller according to claim 1, characterized in that: The performing component association mapping on the fault precursor current feature and the hidden fault vibration feature to obtain an internal fault component identifier includes: Performing physical characteristic frequency band analysis on the fault precursor current characteristics and the latent fault vibration characteristics to obtain a current vibration joint characteristic; Performing topological constraint matching on the current vibration joint features according to a preset component fault mapping relationship library to obtain a set of candidate fault components; Performing fault characteristic frequency band distribution verification on the candidate fault component set to obtain a valid fault component set; Based on the component failure mapping relationship library, primary and secondary conflicts are resolved on the valid fault component set to obtain a primary fault component identification set; Fault mode reverse mapping is performed on the main fault component identifier set to obtain the internal fault component identifier.

6. The intelligent monitoring and alarm method for a motor controller according to claim 1, characterized in that: The alarm analysis based on the fault component identifier, the fault precursor current characteristics and the hidden fault vibration characteristics to obtain a graded alarm decision includes: Performing fault type analysis based on the preset component type information and the fault component identifier to obtain electrical fault type information and mechanical fault type information; Performing multi-modal feature fusion on the fault precursor current feature and the hidden fault vibration feature to obtain a comprehensive fault feature vector; Performing a hazard diffusion assessment on the electrical fault type information and the mechanical fault type information according to the comprehensive fault feature vector to obtain a fault hazard level score; The fault hazard level score is subjected to decision mapping based on a preset alarm strategy library to obtain the graded alarm decision.

7. The intelligent monitoring and alarm method for a motor controller according to claim 6, characterized in that: The performing a hazard diffusion assessment on the electrical fault type information and the mechanical fault type information according to the comprehensive fault feature vector to obtain a fault hazard level score includes: Extracting fault parameters from the comprehensive fault feature vector to obtain a global fault intensity index; Performing graph structure modeling on the gas fault type information and the mechanical fault type information based on a preset internal component dependency relationship to obtain a fault propagation path network; Performing path propagation simulation on the global fault intensity index based on the fault propagation path network to obtain a fault diffusion evaluation parameter; A fault evaluation is performed on the fault diffusion evaluation parameters according to a preset historical fault case library to obtain the fault hazard level score.

8. An intelligent monitoring and alarm system for a motor controller, characterized in that: The intelligent monitoring and alarm method for the motor controller according to any one of claims 1 to 7 comprises: An acquisition module is used to acquire internal circuit signals and body vibration signals of the motor controller, and perform joint noise analysis to generate a dynamic noise baseline; an analysis module, configured to identify component faults based on the dynamic noise baseline and the internal circuit signal to obtain a pre-fault current characteristic; a correlation module, configured to separate component impact responses based on the dynamic noise baseline and the body vibration signal to obtain a hidden fault vibration signature; a processing module configured to perform component association mapping on the precursory current characteristics and the latent fault vibration characteristics to obtain an internal fault component identifier; A control module is used to perform alarm analysis based on the fault component identifier, the fault precursor current characteristics and the hidden fault vibration characteristics to obtain a hierarchical alarm decision.

Images