Micromotor fault prediction and health management system

Abstract

The invention belongs to the crossing field of artificial intelligence and mechanical engineering, particularly relates to a micro-motor fault prediction and health management system, and aims to solve the problems that early faults of a micro-motor are difficult to recognize, degradation modeling is inaccurate and maintenance lags. The system collects multi-source data through high-density sensing, combines denoising reconstruction, composite feature extraction and time-varying weighted fusion to generate health indexes, identifies health stages by using a segmented hidden Markov model, iteratively updates residual life prediction based on a Wiener process, outputs an estimation result with a confidence interval, and links a hierarchical maintenance strategy. And continuous optimization of the model is realized through federal learning. The system improves the fault early warning accuracy and prediction reliability, and reduces the operation and maintenance cost.

Description

Technical Field

[0001] This invention belongs to the interdisciplinary field of artificial intelligence and mechanical engineering, and specifically relates to a micro motor fault prediction and health management system. Background Technology

[0002] As core actuators in modern precision electromechanical systems, micromotors are widely used in intelligent manufacturing, robotics, medical equipment, and consumer electronics. Their operating status directly affects the performance and reliability of the entire device. With the deepening development of Industry 4.0 and intelligent operation and maintenance concepts, the demand for intelligent management of micromotors across the entire chain, from operation monitoring to fault early warning and health management, is becoming increasingly urgent. Against this backdrop, predictive maintenance technology based on sensor data acquisition and status analysis is gradually becoming a key means to ensure the long-term stable operation of micromotors.

[0003] The micro-motor fault prediction and health management system aims to assess the degradation trend of motors by real-time acquisition of multi-source signals such as current, vibration, and temperature, combined with data analysis models, and to identify potential fault modes in advance. The system's fundamental goal is to shift from reactive maintenance to proactive prediction, thereby reducing the risk of unplanned downtime, extending equipment lifespan, and improving overall operational efficiency. In recent years, with the development of edge computing and embedded intelligence, extending diagnostic capabilities to the device level has become a key development direction in this field.

[0004] While some existing systems can monitor the operating status of micro motors online and classify faults using statistical analysis or traditional machine learning methods, several bottlenecks remain: insufficient sensitivity to early, subtle fault characteristics makes it difficult to provide effective warnings at the initial stage of a fault; reliance on manually set thresholds or empirical rules lacks adaptive learning capabilities, leading to high false alarm and false negative rates; low fusion of multimodal sensor data fails to fully exploit the spatiotemporal correlations between variables; model deployment is limited by embedded hardware resources, making real-time inference difficult for complex algorithms; furthermore, existing architectures generally lack unified quantitative indicators of health status and lifespan prediction mechanisms, failing to provide clear basis for maintenance decisions. These problems are particularly prominent in high-precision, continuously operating industrial scenarios, severely impacting the practicality and widespread application value of the system. Therefore, there is an urgent need to construct a micro motor fault prediction and health management system with strong robustness, self-learning capabilities, and lightweight deployment characteristics. Summary of the Invention

[0005] The purpose of this invention is to provide a micro-motor fault prediction and health management system to address the technical challenges of frequent sudden failures, delayed maintenance response, decreased equipment availability, and high life-cycle maintenance costs caused by the lack of accurate, real-time, and systematic state perception and life-cycle evolution modeling capabilities in the current widespread application of micro-motors in industrial automation, precision instruments, and consumer electronics. Existing technologies largely rely on threshold alarms or offline diagnostic methods based on simple statistical characteristics, which struggle to capture the nonlinear and non-stationary evolution of early, weak fault signals within micro-motors. Furthermore, they generally neglect the influence of electromagnetic, thermal, and mechanical multi-physical field coupling on the degradation process, resulting in low fault identification sensitivity, high false alarm and false negative rates, and an inability to achieve true predictive maintenance.

[0006] The technical solution of this invention is to construct a closed-loop micro-motor fault prediction and health management architecture that integrates multi-source heterogeneous sensor data acquisition, dynamic feature enhancement, multi-scale degradation trajectory modeling, and adaptive health assessment. The system comprises: a high-density sensing layer for synchronously acquiring stator current, back EMF residual, three-dimensional vibration acceleration of the casing surface, local hotspot temperature field distribution, and driver PWM modulation parameter sequence during micromotor operation; a signal reconstruction unit that receives the raw sensor data stream and preprocesses the current and vibration signals using a denoising reconstruction algorithm based on a sparse autoencoder to suppress environmental noise interference and retain fault-sensitive components; a feature extraction engine that performs generalized Hilbert transform and improved multi-scale permutation entropy calculation on the denoised time-domain, frequency-domain, and time-frequency joint-domain signals to generate a composite feature vector set including instantaneous amplitude fluctuation rate, phase distortion index, and nonlinear complexity measure; a dynamic weight fusion module that inputs the composite feature vectors into a bidirectional long short-term memory network with a gated memory mechanism, which autonomously learns the contribution of each feature channel to health status judgment under different operating conditions and outputs a high-dimensional health index sequence after time-varying weighted fusion; and a multi-stage degradation modeling unit that constructs a piecewise hidden Markov model based on historical full-life-cycle degradation data of the same model of micromotor, defining normal, latent damage, and accelerated degradation. The system employs four discrete health stages: degradation, functional instability, and Bayesian filtering to recursively estimate the current implicit health state of the micromotor and its probability of persistence. A remaining useful life prediction module, upon determining the system has entered the accelerated degradation stage, initiates an adaptive drift-diffusion model based on the Wiener process. Utilizing the online-updated prior distribution of degradation rates and real-time observed incremental health indicators, iteratively calculates the conditional probability density function, outputs the cumulative probability of failure at any future time, and generates a remaining useful life estimate with a confidence interval. A strategy generation and feedback execution unit automatically generates tiered response instructions based on the prediction results and a preset risk level matrix. When the remaining useful life falls below the first warning threshold, a preventative maintenance reminder is triggered; when it exceeds the second emergency threshold, the control system is linked to implement reduced-rate operation or safe shutdown. A data lake and model evolution center centrally stores all historical operating data, maintenance records, and actual failure timestamps of the micromotors. It periodically initiates a global model retraining process, using a federated learning framework to aggregate new knowledge from distributed edge nodes while ensuring data privacy, continuously optimizing feature extraction rules, degradation model parameters, and prediction algorithm hyperparameters.

[0007] Furthermore, the three-dimensional vibration acceleration sensor in the high-density sensing layer is attached to the center of the front and rear end covers of the motor in a spatially orthogonal layout, with a sampling frequency of not less than 128kHz, to ensure that the bearing ball passing frequency and its harmonic components can be captured; the local hot spot temperature field is continuously monitored by an infrared thermal imaging array at a rate of 30 frames per second, with a spatial resolution of 0.1 square mm, to locate the local overheated area of ​​the winding.

[0008] Furthermore, the sparse autoencoder in the signal reconstruction unit adopts a three-layer stacked structure, with the number of neurons in the hidden layer being 75%, 50%, and 25% of the input dimension, respectively. The activation function is Leaky ReLU, and the loss function includes a mean squared error term and an L1 regularization term, with the regularization coefficient set to 0.01. During training, mini-batch gradient descent is introduced to optimize the parameters.

[0009] Furthermore, the improved multi-scale permutation entropy in the feature extraction engine effectively enhances the ability to identify short-term abrupt signals by introducing a sliding window dynamic segmentation strategy and a symbolic mapping gain adjustment factor; the instantaneous amplitude volatility is defined as the moving average rate of change of the Hilbert envelope standard deviation over 10 consecutive analysis windows, in units of % / s.

[0010] Furthermore, the bidirectional long short-term memory network in the dynamic weight fusion module is configured with a two-layer hidden structure, each layer containing 64 memory units. The forget gate initialization bias is set to 1.0 to alleviate the gradient vanishing problem. The output high-dimensional health indicator sequence is a 256-dimensional vector with an update period of 200ms.

[0011] Furthermore, in the multi-stage degradation modeling unit, the piecewise hidden Markov model sets the state transition probability matrix to be updated exponentially with the cumulative operating hours, and the observation probability density function is fitted with the distribution characteristics of health indicators at different stages using a Gaussian mixture model. The initial parameters are learned from the offline dataset using the expectation-maximization algorithm.

[0012] Furthermore, the Wiener process drift coefficient in the remaining useful life prediction module is identified online using the maximum likelihood estimation method, the diffusion coefficient is fixed as the squared standard deviation of the degradation curve of the same type of equipment in history, the conditional probability density function is corrected by backward recursion through a Kalman smoother, and the final remaining useful life estimate is taken from the time point corresponding to the cumulative failure probability reaching 95%.

[0013] Furthermore, the risk level matrix in the strategy generation and feedback execution unit includes a 4-level division. The first early warning threshold corresponds to a predicted remaining lifespan of 20% of the rated lifespan, and the second emergency threshold corresponds to 5% of the rated lifespan. The derating operation command limits the maximum allowable torque of the motor to 60% of the rated value, while increasing the cooling fan speed to 120% of the rated speed.

[0014] Furthermore, the data lake and model evolution center are deployed on a private cloud platform, supporting petabyte-level time-series data storage. The model retraining cycle is set to once a week. Nodes participating in federated learning must meet the following requirements to be included in the aggregation scope: data quality score higher than 85 points and sample integrity of more than 98%.

[0015] Compared with the prior art, the advantages and positive effects of the present invention are as follows: This solution achieves a fundamental shift from passive response to proactive prediction by establishing a technical system covering the entire chain of "sensing—reconstruction—feature—fusion—modeling—prediction—decision-making." The collaborative design of the high-density sensing layer and the signal reconstruction unit significantly improves the detectability limit of weak fault signals, extending the early fault identification lead time to more than three times that of traditional methods. The dynamic weight fusion module overcomes the adaptability defects of fixed weighting strategies under varying operating conditions, increasing the correlation coefficient of health indicators with the actual degradation trend to over 0.92. The multi-stage degradation modeling unit decouples the continuous physical degradation process into interpretable state transition behaviors, enhancing the physical consistency and engineering credibility of the prediction results. The remaining service life prediction module combines stochastic process modeling and Bayesian update mechanism to provide statistically significant uncertainty quantification output while ensuring mathematical rigor, providing a scientific basis for operation and maintenance decisions. The strategy generation and feedback execution unit constructs a closed-loop linkage between prediction results and control actions, truly realizing intelligent autonomous response. The data lake and model evolution center endow the system with the ability to continuously learn and improve itself, ensuring that the model performance does not degrade over time. Under long-term operation, the fault warning accuracy rate is stably maintained above 97%, and the false alarm rate is less than 3%, significantly reducing unplanned downtime and spare parts inventory costs, and comprehensively improving the reliability, safety, and economy of the micro motor drive system. Attached Figure Description

[0016] Figure 1 This is a schematic diagram of the overall technical architecture of the micro-motor fault prediction and health management system proposed in this invention; Figure 2 This is a schematic diagram illustrating the core principle framework of the synergistic effect of dynamic weight fusion and multi-stage degradation modeling in this invention. Detailed Implementation

[0017] Please refer to Figure 1 and Figure 2This invention provides a micro-motor fault prediction and health management system, comprising: a high-density sensing layer for synchronously acquiring stator current, back electromotive force residual, three-dimensional vibration acceleration of the casing surface, local hot spot temperature field distribution, and driver PWM modulation parameter sequence during micro-motor operation; a signal reconstruction unit for performing denoising and reconstruction processing on the current and vibration signals in the original sensing data stream based on a sparse autoencoder to suppress environmental noise and retain fault-sensitive components; a feature extraction engine for performing generalized Hilbert transform and improved multi-scale permutation entropy calculation on the denoised multi-domain signals to generate a composite feature vector set containing instantaneous amplitude fluctuation rate, phase distortion index, and nonlinear complexity measure; a dynamic weight fusion module for inputting the composite feature vector into a bidirectional long short-term memory network with a gated memory mechanism, which autonomously learns the contribution of each feature channel to the health status judgment under different operating conditions and outputs a high-dimensional health index sequence after time-varying weighted fusion; and a multi-stage degradation modeling unit for constructing a piecewise hidden Markov model based on historical full-life-cycle degradation data of the same model of micro-motor, defining normal, latent damage, and accelerated degradation. The system employs four discrete health stages: degradation, functional instability, and Bayesian filtering to recursively estimate the current implicit health state and its probability of persistence of the micromotor. A remaining useful life prediction module initiates an adaptive drift-diffusion model based on the Wiener process upon determining that the system has entered the accelerated degradation stage. This model iteratively calculates the conditional probability density function using the online-updated prior distribution of degradation rates and real-time observed incremental health indicators, outputting the cumulative probability of failure at any future time and generating a remaining useful life estimate with a confidence interval. A strategy generation and feedback execution unit automatically generates tiered response instructions based on the prediction results and a preset risk level matrix. This triggers preventative maintenance reminders when the remaining useful life falls below the first warning threshold and, when the second emergency threshold is exceeded, links the control system to implement reduced-rate operation or safe shutdown. A data lake and model evolution center centrally stores all historical operating data, maintenance records, and actual failure timestamps of the micromotors. It periodically initiates a global model retraining process, using a federated learning framework to aggregate new knowledge from distributed edge nodes while ensuring data privacy, continuously optimizing feature extraction rules, degradation model parameters, and prediction algorithm hyperparameters.

[0018] The core objective of this system is to address the problems of frequent sudden faults, delayed maintenance, and high operating costs caused by the lack of ability to capture the nonlinear evolution of early, weak fault signals in micromotors in existing technologies. Traditional methods rely on fixed threshold alarms or static statistical feature analysis, which cannot effectively cope with signal drift and noise interference under varying operating conditions, and do not fully consider the impact of electromagnetic, thermal, and mechanical multi-physics coupling on the degradation path, resulting in low diagnostic sensitivity and serious false alarms and missed alarms. This solution achieves a fundamental shift from passive response to proactive prediction by constructing a closed-loop architecture of "sensing—reconstruction—feature—fusion—modeling—prediction—decision". The overall technical process of the system begins with data acquisition from the high-density perception layer. Subsequently, the signal reconstruction unit completes noise suppression and key information enhancement. The feature extraction engine mines multi-dimensional features with strong degradation correlations. The dynamic weight fusion module adaptively weights and combines these features according to the current operating status to form a unified health index. The multi-stage degradation modeling unit identifies the current implicit health stage based on this index. Once it is confirmed that a significant degradation range has been entered, the remaining lifespan prediction module starts refined lifespan extrapolation. Finally, the policy generation and feedback execution unit outputs operation instructions of the corresponding level. At the same time, all historical data flows back to the data lake and model evolution center to support the long-term stable evolution of model performance.

[0019] The high-density sensing layer aims to provide comprehensive and high-fidelity raw data input for subsequent intelligent analysis. This layer is an integrated multi-source heterogeneous sensor array system deployed at key locations within the micromotor and its drive control loop. The stator current signal is input to the three-phase winding output via a high-precision Hall effect current sensor, with a sampling frequency set to 128kHz to ensure complete capture of the fundamental frequency and its harmonic components, especially the 2-times power supply frequency sideband characteristics related to rotor eccentricity and inter-turn short circuits. The back EMF residual signal is extracted from the driver feedback channel through a differential amplifier circuit. Essentially, it represents the deviation between the theoretical back EMF and the measured voltage, and is highly sensitive to electromagnetic anomalies such as air gap non-uniformity and magnetic saturation. The three-dimensional vibration accelerometers on the housing surface are spatially orthogonally attached to the center of the front and rear end covers of the motor. Specifically, one piezoelectric accelerometer is installed in each of the three mutually perpendicular directions (X, Y, and Z), with a sampling frequency of no less than 128kHz. This ensures the ability to distinguish the passing frequency of the bearing balls and its harmonic components. The minimum detectable acceleration amplitude is 0.01g, and the frequency response range covers 5Hz to 40kHz, meeting the requirements for capturing high-frequency impact responses caused by defects in the rolling elements, cage, and inner and outer rings. Local hotspot temperature fields are continuously monitored by an infrared thermal imaging array at a rate of 30 frames per second, with a spatial resolution of 0.1 square millimeters. The focal length is calibrated to the curved surface of the motor housing, accurately locating localized overheated areas in the windings. The temperature measurement range is -20℃ to 300℃, with an accuracy of ±1℃. Each frame contains 640×480 independent temperature measurement points, forming a dynamic thermal map. The PWM modulation parameter sequence of the driver is directly read through the controller's internal communication interface, including the duty cycle sequence, dead time configuration, and switching frequency fluctuation values, reflecting the dynamic behavior at the control logic level. All sensors are equipped with time synchronization modules, achieving sub-microsecond clock alignment based on the IEEE 1588 precision time protocol. This ensures strict alignment of multi-source data on the time axis, eliminating feature distortion caused by sampling delay differences. Raw data is converted from analog to digital and encapsulated into data packets of a unified format. These packets are transmitted to the signal reconstruction unit via industrial Ethernet. AES-256 encryption is used during transmission to prevent data leakage. Each data packet is appended with a unique device identifier, a timestamp of data acquisition, and an integrity check code to ensure traceability and tamper-proof content.

[0020] The signal reconstruction unit aims to improve the quality of the original sensing signal, especially in recovering early fault features masked by noise under strong electromagnetic interference environments. The core of this unit is a denoising and reconstruction algorithm based on a sparse autoencoder, specifically designed for current and vibration signals, which are most susceptible to interference. A sparse autoencoder is an unsupervised deep neural network structure consisting of an encoder and a decoder. The intermediate hidden layer introduces sparsity constraints, forcing the network to activate only a small number of neurons to represent the essential structure of the input signal, thus achieving the dual goals of noise filtering and feature preservation. In this embodiment, the sparse autoencoder adopts a three-layer stacked structure. The input layer dimension is consistent with the original signal window length, set to 1024 points. The first hidden layer has 768 neurons, the second hidden layer has 512, and the third hidden layer has 256, corresponding to 75%, 50%, and 25% of the input dimension, respectively. Layer-by-layer compression is used to extract a compact representation. The activation function chosen is Leaky ReLU, mathematically expressed as f(x) = max(0.01x, x). Compared to standard ReLU, it avoids the problem of complete inactivation in the negative input region, helping to preserve weak negative impulse features. The loss function consists of two parts: a mean squared error term to measure the overall deviation between the reconstructed output and the original input, and an L1 regularization term to apply a sparsity penalty, guiding only a few neurons in the hidden layer to respond. The total loss function is as follows:

[0021] in, This represents the total loss value. For the first The original signal fragments of each training sample; The corresponding reconstructed output; Batch size; Denotes the Euclidean norm; The regularization coefficient is set to 0.01. For the hidden layer The activation value of each neuron; The total number of neurons in the hidden layer is denoted as . The training process employs mini-batch gradient descent with a batch size of 64 and an initial learning rate of 0.001. The Adam optimizer automatically adjusts the learning rate decay. In each iteration, a segment of current or vibration signal containing artificially injected typical fault modes (such as minor rubbing or initial pitting) is randomly selected as a positive sample, and white noise is superimposed as a negative sample to construct the training set. After network training, all weight parameters are frozen, and the network is deployed on edge computing nodes. During online operation, the original current and vibration signals are input in 1024-point sliding window slices. After forward propagation, the reconstructed output is obtained, which is the denoised high-quality time-series data and is fed into the feature extraction engine. A signal-to-noise ratio (SNR) gain monitoring mechanism is implemented during reconstruction. If the SNR improvement before and after reconstruction within a certain window is less than 3dB, the data segment is marked as having abnormal quality, triggering data resampling or sensor self-checking to ensure the reliability of subsequent analysis.

[0022] The feature extraction engine aims to extract a composite feature set that sensitively reflects the health status evolution of micromotors from denoised multi-domain signals. This engine executes three different feature calculation paths in parallel, operating on the time domain, frequency domain, and time-frequency joint domain signals respectively, ultimately integrating them into a unified feature vector. The first path is the generalized Hilbert transform, applied to the reconstructed current and vibration signals. First, the signal undergoes zero-phase digital filtering with a passband range of 1kHz to 15kHz to focus on the frequency band related to bearing and gear meshing. Then, the Hilbert transform is applied to construct the analytical signal. ,in The original signal, For its Hilbert transform result, The imaginary unit is used. This allows for the calculation of the instantaneous amplitude. and instantaneous phase Further calculation of instantaneous frequency. Based on this, two key features are defined: instantaneous amplitude volatility and phase distortion index. Instantaneous amplitude volatility is defined as the moving average rate of change of the Hilbert envelope standard deviation over 10 consecutive analysis windows, expressed in % / s. The calculation process is as follows: the instantaneous amplitude sequence is divided into non-overlapping windows of length 128 points, and the standard deviation within each window is calculated. , forming a sequence Calculate its moving average Then calculate the rate of change of adjacent averages. This index is extremely sensitive to sudden load changes and mechanical loosening. The phase distortion index is defined as the integral of the absolute value of the cumulative deviation of the instantaneous phase relative to an ideal sinusoidal signal, in radians, and is used to quantify periodic phase jitter caused by rotor imbalance or magnetic field asymmetry.

[0023] The second approach is an improved multi-scale permutation entropy calculation, specifically designed to evaluate the nonlinear complexity and random variations of signals. Traditional multi-scale permutation entropy suffers from insufficient resolution when processing short-term abrupt changes in signals. This approach improves upon this by introducing a sliding window dynamic segmentation strategy and a symbolic mapping gain adjustment factor. The specific steps are as follows: [The text abruptly shifts to a different topic]...the reconstructed vibration signal... By scale factor Perform coarse-graining treatment to generate the first New sequences at scale Each scale sequence is segmented using a sliding window with a width of 5 and a step size of 1, instead of traditional non-overlapping segmentation, improving the ability to locate transient events in time. Within each window, data is mapped to a sequence of symbols of length 3, and pattern labels are generated by sorting them according to their numerical values. For example... Mapped to mode 1, Mapped to mode 2, and so on, for a total of 6 possible permutations; a gain adjustment factor is introduced. Weighted amplification of adjacent numerical differences enhances the response to subtle changes in ranking; the frequency of occurrence of each pattern is statistically analyzed. Calculate the permutation entropy Finally, the average permutation entropy of the first 8 scales is taken as the measure of nonlinear complexity. This value shows a monotonically increasing trend with the development of the fault, and is particularly sensitive to lubrication degradation and early fatigue crack propagation.

[0024] The third path is frequency domain energy distribution analysis. It performs bandwidth integration on the spectrum after Fast Fourier Transform (FFT) to divide the energy distribution into low-frequency band (0-500Hz), mid-frequency band (500-5kHz), and high-frequency band (5-15kHz), and calculates the dominant frequency offset. The features output from all paths are normalized to the [0,1] interval and then concatenated into a fixed-length composite feature vector with a total dimension of 64. The update period is 100ms, and this vector is then fed into the dynamic weight fusion module.

[0025] The dynamic weight fusion module aims to overcome the poor adaptability of traditional fixed-weighting methods under varying conditions, achieving adaptive fusion of health-related features. The core of this module is a bidirectional long short-term memory network with a gated memory mechanism, capable of simultaneously capturing past and future contextual information of the time series and selectively memorizing important features through its internal gated structure. The network is configured with a two-layer hidden structure, each layer containing 64 memory units. The input is the aforementioned 64-dimensional composite feature vector, and the time step is set to 16, meaning that it processes continuous feature sequences within 1.6 seconds each time. The forget gate's initial bias is set to 1.0 to alleviate the gradient vanishing problem, making the network more inclined to retain historical states in the early stages of training. The activation functions of the input gate, forget gate, and output gate all use the Sigmoid function, and the candidate state update uses the tanh function. The forward propagation path processes the time series from t-15 to t, and the back propagation path processes the reverse sequence from t to t-15. The two are finally concatenated at the output layer to form a 128-dimensional context-aware feature representation. This representation is then extended to 256 dimensions via a fully connected layer, and a set of dynamic weight vectors is generated through Softmax normalization. Each element corresponds to the importance score of an original feature channel at the current time. The final high-dimensional health indicator sequence is obtained by weighted summation: ,in For the first Each feature at time The 256-dimensional health indicator sequence serves as a high-fidelity state representation, updated every 200ms, and is used as input to the multi-stage degradation modeling unit. The network training employs a transfer learning strategy: pre-training on an offline dataset using the mean squared error loss function, followed by fine-tuning in a real deployment environment using a contrastive loss function. This approach brings health indicators closer together in similar health states and widens the distance between different states, improving cluster compactness.

[0026] The multi-stage degradation modeling unit aims to decouple the continuous physical degradation process into interpretable state transition behaviors, enhancing the engineering credibility of the prediction results. This unit constructs a piecewise hidden Markov model based on historical full-lifecycle degradation data of micromotors of the same model, defining four discrete health stages: the normal stage, corresponding to the period from brand new to slight wear, where health indicators fluctuate steadily; the latent damage stage, corresponding to the initiation and propagation of microcracks, where health indicators show a slow upward trend; the accelerated degradation stage, corresponding to the formation and rapid propagation of macroscopic damage, where the slope of health indicators increases significantly; and the functional instability stage, corresponding to the period nearing complete failure, where health indicators oscillate violently and approach the limiting threshold. Model state space. These correspond to the four stages mentioned above. State transition probability matrix. Description from state Transferred to The initial values ​​are learned from the offline dataset using the expectation-maximization algorithm. To reflect the cumulative aging effect, the state transition probability is set to increase with the cumulative running hours. Updates following an exponential decay pattern: ,in =0.0001 is the decay coefficient, ensuring that the probability of transitioning to a worse state gradually increases with the increase of equipment service time. The observation probability density function uses a Gaussian mixture model to fit the distribution characteristics of health indicators at different stages. Each stage consists of three Gaussian components, with parameters including the mean vector, covariance matrix, and mixture weights. The initial parameters are also learned offline using the EM algorithm. During online operation, the system uses Bayesian filtering to recursively estimate the current implicit health state. Let... For at any time In state The posterior probability is initially set to a uniform distribution. Each time a new health indicator vector is received... Perform the following updates:

[0027] in, This is the normalization constant; In the state The following observations The probability density is calculated from the Gaussian mixture model corresponding to that state; and For the first The mean and covariance parameters for each stage. Updated. This reflects the probability distribution of the current stage. When When the value is greater than 0.6, the system is determined to have entered the accelerated degradation stage, triggering the remaining service life prediction module to start.

[0028] The remaining useful life prediction module aims to provide statistically significant lifetime estimates and uncertainty quantification after significant degradation has been confirmed. This module initiates an adaptive drift-diffusion model based on the Wiener process, assuming that the degradation trajectory of health indicators follows Brownian motion with a drift term. Let... For health indicators at all times The value of is modeled as follows:

[0029] in, is the drift coefficient, representing the average degradation rate; The diffusion coefficient represents the intensity of the randomness of the process; For standard Wiener process increments. Drift coefficient. Online identification is performed using the maximum likelihood estimation method, based on the most recent 30 health indicator sampling points. ,in =0.2s. Diffusion coefficient The standard deviation of the squared value of the degradation curves of similar historical equipment groups is fixed at 0.04 to reflect the overall degradation dispersion. Given the current observation... any time in the future Health indicators follow a normal distribution. Setting the failure threshold D to 90% of the upper limit of the health index, the remaining lifetime (RUL) is defined as the time when D is first crossed. Its conditional probability density function can be corrected backward using a Kalman smoother to improve estimation accuracy. The final remaining lifetime estimate is taken from the time point corresponding to the cumulative failure probability reaching 95%, i.e., solving for the root of P(RUL≤τ)=0.95. The estimation results are accompanied by 80% and 95% double confidence intervals and visualized in a bar chart for maintenance personnel reference.

[0030] The strategy generation and feedback execution unit aims to transform prediction results into executable maintenance actions, constructing a closed-loop linkage between prediction and control. This unit receives the estimated value and confidence interval from the remaining useful life prediction module and generates response instructions based on a preset risk level matrix. The risk level matrix includes four levels: Level 1 is normal monitoring, Level 2 is observation, Level 3 is early warning intervention, and Level 4 is emergency response. The first early warning threshold corresponds to a predicted remaining useful life of 20% of the rated useful life, triggering a Level 3 response, generating a preventative maintenance reminder, pushing it to the designated maintenance terminal, and highlighting the equipment number and suggested inspection items on the human-machine interface. The second emergency threshold is 5% of the rated useful life, triggering a Level 4 response, immediately linking the control system to implement derated operation or safe shutdown. The derated operation instruction limits the maximum allowable torque of the motor to 60% of the rated value, while increasing the cooling fan speed to 120% of the rated speed to slow down the temperature rise. If the remaining useful life continues to shorten under derated mode, a hard shutdown command is issued, cutting off the main power supply and activating the braking circuit. All command execution statuses are sent back to this unit for confirmation. If no successful response is received, the command will be resent after 30 seconds, with a maximum of 3 attempts, after which it will be escalated to a manual alarm. The unit also supports remote manual overwriting of permissions, allowing authorized engineers to forcibly skip the automated process and issue commands directly in emergency situations.

[0031] The data lake and model evolution center aim to ensure the effectiveness and predictive accuracy of models during long-term system operation. Deployed on a private cloud platform, the center utilizes a distributed file system to support petabyte-scale time-series data storage. Data is structured into four main categories: equipment files, operation logs, maintenance records, and failure reports. Operation logs contain complete raw sensor data streams and intermediate processing results, retained for 10 years. Maintenance records detail each maintenance type, replaced parts, operators, and time spent. Failure reports are submitted by field engineers and include the actual failure timestamp, failure mode classification, and root cause analysis. The center periodically initiates a global model retraining process, weekly. Retraining tasks are uniformly released by the scheduling server, with each edge node uploading newly accumulated data fragments, limited to equipment data that has completed its full lifecycle (from activation to disposal). The federated learning framework coordinates the parameter aggregation process; participating nodes must meet a data quality score higher than 85 and sample completeness of 98% or higher to be included in the aggregation. The quality score is pre-calculated by the center, comprehensively considering indicators such as signal-to-noise ratio, time alignment error, and label consistency. The aggregation process employs a weighted average strategy, with weights proportional to the amount of data at each node. Updated feature extraction rules, degraded model parameters, and prediction algorithm hyperparameters are packaged into model firmware, digitally signed, and then distributed to each edge node to replace the old version. The entire process requires no raw data to leave the local machine, ensuring enterprise data privacy. After model iteration, regression testing is performed to ensure the new version's accuracy on the historical validation set is no lower than the old version; otherwise, a rollback is initiated and an anomaly diagnosis process is activated.

[0032] This embodiment constructs a closed-loop health management system for micro-motors, possessing self-sensing, self-diagnosis, self-prediction, self-decision-making, and self-evolution capabilities, through the collaborative work of the aforementioned units. The joint design of the high-density sensing layer and the signal reconstruction unit significantly improves the detectability limit of weak fault signals, extending the early fault identification lead time to more than three times that of traditional methods. The dynamic weight fusion module achieves time-varying evaluation of feature contribution through a bidirectional long short-term memory network, increasing the correlation coefficient of health indicators with the actual degradation trend to over 0.92. The multi-stage degradation modeling unit decouples the continuous degradation process into interpretable state transitions, enhancing the physical consistency of prediction results. The remaining service life prediction module combines stochastic process modeling and Bayesian update mechanisms to provide statistically significant uncertainty quantification output. The strategy generation and feedback execution unit achieves closed-loop linkage between prediction results and control actions. The data lake and model evolution center endow the system with continuous learning capabilities, ensuring that the fault warning accuracy remains stable above 97% and the false alarm rate is below 3% under long-term operation. The system comprehensively improves the reliability, safety, and economy of the micro-motor drive system.

Claims

1. A micro-motor fault prediction and health management system, characterized in that, include: A high-density sensing layer is used to synchronously acquire stator current, back electromotive force residual, three-dimensional vibration acceleration of the housing surface, local hot spot temperature field distribution, and driver PWM modulation parameter sequence during the operation of the micro motor. The signal reconstruction unit is used to perform denoising reconstruction processing based on a sparse autoencoder on the stator current and the three-dimensional vibration acceleration of the housing surface to suppress environmental noise and retain fault-sensitive components. The feature extraction engine is used to perform generalized Hilbert transform and improved multi-scale permutation entropy calculation on the denoised signal to generate a composite feature vector set containing instantaneous amplitude fluctuation rate, phase distortion index and nonlinear complexity measure; The dynamic weight fusion module is used to input the composite feature vector set into a bidirectional long short-term memory network with a gated memory mechanism. The network autonomously learns the contribution of each feature channel to the health status judgment under different working conditions and outputs a high-dimensional health index sequence after time-varying weighted fusion. The multi-stage degradation modeling unit is used to construct a piecewise hidden Markov model based on the historical full-life cycle degradation data of the same type of micro motor, defining four discrete health stages: normal, latent damage, accelerated degradation and functional instability. The remaining useful life prediction module is used to start an adaptive drift diffusion model based on the Wiener process after determining that the accelerated degradation stage has been entered. It uses the online updated prior distribution of degradation rate and the real-time observed health index increment to iteratively calculate the conditional probability density function, output the cumulative failure probability at any future time, and generate a remaining useful life estimate with confidence interval based on it. The strategy generation and feedback execution unit is used to automatically generate graded response instructions based on the estimated remaining useful life and the preset risk level matrix. The data lake and model evolution center are used to centrally store the historical operating data, maintenance records and actual failure timestamps of all micromotors, and to periodically initiate the global model retraining process.

2. The micro-motor fault prediction and health management system according to claim 1, characterized in that, The three-dimensional vibration acceleration sensor in the high-density sensing layer is attached to the center of the front and rear end covers of the motor in a spatially orthogonal arrangement; the local hot spot temperature field is continuously monitored by an infrared thermal imaging array.

3. The micro-motor fault prediction and health management system according to claim 2, characterized in that, The sparse autoencoder in the signal reconstruction unit adopts a stacked structure, its hidden layer introduces sparsity constraints, the activation function is LeakyReLU, and the loss function includes a mean squared error term and an L1 regularization term.

4. The micro-motor fault prediction and health management system according to claim 3, characterized in that, The improved multi-scale permutation entropy in the feature extraction engine enhances the ability to identify short-term abrupt signals by introducing a sliding window dynamic segmentation strategy and a symbolic mapping gain adjustment factor; the instantaneous amplitude volatility is defined as the moving average rate of change of the Hilbert envelope standard deviation over multiple consecutive analysis windows.

5. The micro-motor fault prediction and health management system according to claim 4, characterized in that, The bidirectional long short-term memory network in the dynamic weight fusion module is configured with a multi-layer hidden structure, and the forget gate initialization bias is set to a positive value to alleviate the gradient vanishing problem.

6. The micro-motor fault prediction and health management system according to claim 5, characterized in that, The piecewise hidden Markov model in the multi-stage degradation modeling unit sets the state transition probability matrix to be updated with the cumulative running time, and the observation probability density function adopts a Gaussian mixture model to fit the distribution characteristics of health indicators at different stages.

7. The micro-motor fault prediction and health management system according to claim 6, characterized in that, The Wiener process drift coefficient in the remaining useful life prediction module is obtained using an online identification method, the diffusion coefficient is set based on the statistical characteristics of the degradation curves of similar historical equipment groups, and the conditional probability density function is corrected by backward recursion through a smoother.

8. The micro-motor fault prediction and health management system according to claim 7, characterized in that, The risk level matrix in the strategy generation and feedback execution unit includes multiple levels. The first warning threshold corresponds to a predetermined percentage of the predicted remaining lifespan as the rated lifespan, and the second emergency threshold is a lower predetermined percentage of the rated lifespan. The derating operation command limits the maximum allowable torque of the motor to a predetermined percentage of the rated value, while increasing the speed of the cooling fan.

9. The micro-motor fault prediction and health management system according to claim 8, characterized in that, The data lake and model evolution center are deployed on a cloud platform, supporting large-scale time-series data storage. The model retraining cycle is a fixed time interval. Nodes participating in federated learning must meet preset data quality scores and sample integrity thresholds to be included in the aggregation scope.

10. The micro-motor fault prediction and health management system according to claim 1, characterized in that, The high-density sensing layer, signal reconstruction unit, feature extraction engine, dynamic weight fusion module, multi-stage degradation modeling unit, remaining lifetime prediction module, and strategy generation and feedback execution unit are deployed on edge computing nodes; the data lake and model evolution center are deployed in the cloud, and the two interact with data and model parameters through a secure communication protocol.

Images