An industrial motor health degree AI predictive maintenance and closed-loop management system

Abstract

This invention discloses an AI-based predictive maintenance and closed-loop control system for industrial motor health, belonging to the field of motor monitoring and predictive maintenance. The system includes: a multi-dimensional high-frequency synchronous sensing unit for collecting motor operating parameters; a dynamic operating condition feature decoupling unit to eliminate the influence of speed fluctuations on signals under variable frequency drive through order analysis; an electromagnetic-mechanical coupling feature fusion unit to achieve deep fusion of current and vibration signals and construct a composite fault feature map; an artificial intelligence health prediction engine to output quantitative health assessment results based on a spatiotemporal sequence model; and an adaptive closed-loop control execution unit to automatically adjust motor operating parameters according to the assessment results, forming a complete link from sensing and prediction to closed-loop control. This invention solves the problems of high false alarm rate and difficulty in early fault identification under variable frequency operating conditions, achieving accurate assessment and automated closed-loop control of motor health status, significantly improving equipment reliability and intelligent operation and maintenance levels.

Description

Technical Field

[0001] This invention belongs to the field of motor monitoring and predictive maintenance, specifically relating to an AI-based predictive maintenance and closed-loop management system for industrial motor health. Background Technology

[0002] In the field of industrial automation, motors, as core power equipment, directly affect the overall efficiency of the production system due to their operational stability. Artificial intelligence-based predictive maintenance technology, by collecting operating parameters in real time and assessing equipment health status, can effectively reduce the risk of unplanned downtime and has become an important development direction for current industrial operations and maintenance.

[0003] However, existing technologies still have significant limitations when dealing with complex operating conditions. First, in inverter-driven scenarios, frequent fluctuations in motor speed can cause non-fault-related spectral shifts in vibration and current signals, making traditional alarm methods based on fixed thresholds prone to false alarms and missed alarms. Second, existing systems often process vibration and current signals independently, lacking deep fusion analysis across physical domains. This makes it difficult to identify complex faults caused by electromagnetic and mechanical interactions, such as implicit rotor imbalance caused by abnormal electromagnetic forces, resulting in a high rate of missed early fault diagnoses. Third, most monitoring systems only provide early warnings and fail to automatically integrate prediction results into the control loop, making it impossible to adaptively adjust motor operating parameters and delay the evolution of faults.

[0004] Therefore, how to eliminate signal characteristic interference under varying operating conditions, achieve deep coupling analysis of electromagnetic mechanical characteristics, and transform artificial intelligence predictions into safe closed-loop control commands are technical challenges that urgently need to be solved in this field. Summary of the Invention

[0005] The purpose of this invention is to provide an AI-based predictive maintenance and closed-loop management system for the health status of industrial motors, in order to solve the technical problems of high false alarm rate of fixed threshold alarms caused by frequent fluctuations in speed and load of industrial motors under variable frequency drive conditions, and low accuracy of fault diagnosis in complex electromagnetic environments due to lack of deep integration of multi-source heterogeneous data, thereby achieving accurate assessment of motor health status and real-time automated closed-loop management.

[0006] The technical solution of this invention is to provide an AI-based predictive maintenance and closed-loop management system for industrial motor health, comprising: The multi-dimensional high-frequency synchronous sensing unit collects the operating parameters of the motor under variable frequency drive, including at least current signals and vibration signals. The dynamic operating condition feature decoupling unit receives the operating parameters and maps the current signal and vibration signal in the time domain to the angle domain through instantaneous frequency estimation and angle domain resampling technology to generate an order feature vector, so as to eliminate the frequency modulation interference of speed fluctuation on signal characteristics under variable frequency drive. The electromagnetic-mechanical coupling feature fusion unit receives the order feature vector and performs semantic-level fusion of the current signal and vibration signal by constructing a cross-domain mapping matrix to generate a composite fault feature map characterizing the electromagnetic-mechanical coupling relationship. The artificial intelligence health prediction engine receives the composite fault feature map, performs in-depth analysis through a spatiotemporal sequence model, and outputs a quantitative assessment result of the motor's health at the current moment. The adaptive closed-loop control execution unit receives the health quantification assessment results and automatically generates and sends control commands to the motor driver to adjust the motor operating parameters based on the health quantification assessment results, forming a complete link from perception and prediction to closed-loop control.

[0007] Furthermore, the dynamic operating condition characteristic decoupling unit is specifically used for: The fundamental component is extracted from the current signal, and the Hilbert transform is applied to the fundamental component to extract the instantaneous frequency. Based on the instantaneous frequency and the number of pole pairs of the motor, the instantaneous mechanical speed is calculated, and the instantaneous mechanical speed is integrated to obtain the rotation angle information; Based on the preset number of sampling points per revolution, a time sequence corresponding to equal angular intervals is determined, and based on the time sequence, the original vibration signal and the current signal are resampled in the angular domain through interpolation to generate the order feature vector decoupled from the rotational speed fluctuation.

[0008] Furthermore, the electromagnetic-mechanical coupling feature fusion unit is specifically used for: Multi-scale wavelet packet decomposition is performed on the order feature vectors of the input current signal and vibration signal respectively to obtain wavelet packet coefficients of multiple frequency bands; The mutual information between the current signal frequency band and the vibration signal frequency band is calculated by using sparse coding technology guided by cross-domain mutual information matrix, and the wavelet packet coefficients are sparsified based on the mutual information to enhance the characteristics of weak composite faults. Feature vectors are extracted from the wavelet packet coefficients of each frequency band after sparsification, and the Kronecker product is used to fuse the feature vectors of the current signal and the vibration signal to construct a composite fault feature map.

[0009] Furthermore, the electromagnetic-mechanical coupling feature fusion unit is also used for: Using gradient information from a pre-trained AI health prediction engine, the contribution weight of each pixel in the composite fault feature map to the final health prediction result is calculated, and a weight mask is generated. We use a weighted mask to enhance the composite fault feature map to highlight the coupling region related to the fault, generate an enhanced composite fault feature map, and use it as input to an artificial intelligence health prediction engine.

[0010] Furthermore, the multi-dimensional high-frequency synchronous sensing unit includes an analog-to-digital converter, a signal conditioning circuit, and a field-programmable gate array logic control module; The field-programmable gate array (FPGA) logic control module is used to generate synchronous trigger pulses. The hardware-triggered synchronization mechanism controls the precise alignment of the sampling times of the current signal and the vibration signal, ensuring that the sampling synchronization error is within 10 microseconds, thus providing a time reference for subsequent cross-domain correlation analysis.

[0011] Furthermore, the AI ​​health prediction engine includes a combined architecture of residual networks and bidirectional gated recurrent units; The residual network is used to extract spatial coupling features from the composite fault feature map, and the bidirectional gated recurrent unit is used to model the time sequence composed of spatial coupling features to learn the evolution law of health status and output the quantitative assessment results of health and the predicted value of remaining life.

[0012] Furthermore, the AI-powered health prediction engine also includes a built-in uncertainty assessment module; The uncertainty assessment module implements Bayesian inference based on the Monte Carlo Dropout method, which is used to calculate the confidence level of the health quantification assessment results during the inference phase; When the confidence level is lower than the preset threshold, the system triggers a data re-verification process and uses the data integrity indicators of the data quality monitoring layer to check the sensor status.

[0013] Furthermore, the adaptive closed-loop control execution unit has a built-in multi-objective optimization controller; The multi-objective optimization controller is used to optimize motor operating parameters based on the results of health quantification assessment and current production process requirements, with the goal of maximizing motor operating efficiency and minimizing mechanical losses. Before generating and issuing control commands, the controller also performs resonance range avoidance verification. By comparing the target speed with the pre-stored list of mechanical resonance ranges, it ensures that the adjusted operating parameters avoid the mechanical resonance range.

[0014] Furthermore, it also includes a data quality monitoring layer, deployed in the edge computing gateway, for real-time monitoring of the operating status and signal quality of all sensors; The data quality monitoring layer determines the electrical connection status by detecting the DC bias voltage of the sensor output signal, and determines whether clipping distortion exists by calculating whether the signal amplitude reaches the full-scale range of the analog-to-digital converter. When a sensor malfunction or signal quality degradation is detected, the system automatically triggers a redundancy switching mechanism, prioritizing switching to redundant sensors of the same type or switching to the current observer data integrated inside the motor driver to maintain basic system operation.

[0015] Furthermore, it also includes an augmented reality maintenance assistance module; The augmented reality maintenance assistance module is activated when the health measurement assessment results reach the preset maintenance threshold. It is used to retrieve matching maintenance strategies from the fault knowledge base and push the maintenance strategies and the three-dimensional digital twin model of the motor to the mobile terminal of the maintenance personnel. The mobile terminal uses augmented reality technology to overlay a 3D digital twin model onto the actual motor image in the correct orientation, and renders colors according to the health scores of each key part to intuitively guide maintenance personnel to perform precise maintenance.

[0016] In summary, this application includes at least one of the following beneficial technical effects: 1. This invention achieves signal mapping from the time domain to the angle domain through a dynamic operating condition feature decoupling unit. By utilizing instantaneous frequency estimation and resampling technology, it fundamentally eliminates the frequency modulation interference of speed fluctuations on the spectral characteristics under variable frequency drive. Compared with the traditional fixed threshold monitoring method, this solution can accurately extract order features, effectively avoid false alarms and missed alarms caused by speed drift, and greatly improve the stability and reliability of the system in complex variable operating conditions.

[0017] 2. This invention introduces an electromagnetic-mechanical coupling feature fusion unit, which uses cross-domain mapping matrix and semantic-level fusion technology to perform deep correlation analysis between current features and vibration features. This design overcomes the limitations of monitoring in a single physical dimension and can identify potential composite faults caused by electromagnetic and mechanical interactions, such as hidden rotor imbalance caused by uneven distribution of electromagnetic force. Through cross-physical domain correlation parameter monitoring, the system can identify weak early fault signs earlier than traditional methods, significantly extending the warning window and providing ample time for operation and maintenance decisions.

[0018] 3. This invention constructs a complete closed-loop process from real-time perception to artificial intelligence prediction, and then to adaptive closed-loop control. The artificial intelligence health prediction engine achieves high-precision quantitative assessment of health through spatiotemporal sequence modeling, while the adaptive closed-loop control execution unit can automatically adjust motor operating parameters based on the assessment results. This automated closed-loop logic transforms traditional passive maintenance into proactive predictive maintenance and has the ability to automatically reduce load to delay fault evolution. Combined with the low-latency response characteristics of edge computing architecture, this invention can effectively prevent major production accidents, significantly reduce unplanned downtime and maintenance costs, and improve the overall efficiency and intelligent management level of industrial assets.

[0019] 4. This invention enhances the robustness of the system in harsh industrial environments by integrating an uncertainty assessment module and a sensor redundancy mechanism. When the data quality of the perception layer is impaired or the confidence of model inference decreases, the system can ensure the safety and reliability of decision-making through data verification and redundancy switching, avoiding the risk of making erroneous control decisions based on noisy data. At the same time, the deep collaboration mechanism with the manufacturing execution system enables the optimal balance between equipment maintenance and production scheduling, providing solid technical support for building a smart factory. Attached Figure Description

[0020] Figure 1 This is an overall schematic diagram of an AI-based predictive maintenance and closed-loop management system for industrial motor health. Figure 2 This is a schematic diagram illustrating the principle of dynamic working condition feature decoupling and order feature generation; Figure 3 This is a flowchart of the fusion of electromagnetic-mechanical coupling characteristics and the construction of a composite fault map; Figure 4 This is a schematic diagram illustrating the interaction and data flow between the AI ​​health prediction engine and the adaptive closed-loop control execution unit. Detailed Implementation

[0021] Please refer to the attached document. Figure 1 This embodiment provides an AI-based predictive maintenance and closed-loop management system for the health of industrial motors. The system is built on a distributed industrial edge computing architecture and aims to solve the technical problems of high false alarm rate of fixed threshold alarms caused by frequent fluctuations in speed and load of industrial motors under variable frequency drive conditions, and low fault diagnosis accuracy in complex electromagnetic environments due to lack of deep integration of multi-source heterogeneous data.

[0022] The entire system achieves intelligent control across the entire chain, from the acquisition of underlying physical signals to the issuance of high-level decision commands, through the collaborative interaction of multi-dimensional high-frequency synchronous sensing units, dynamic operating condition feature decoupling units, electromagnetic-mechanical coupling feature fusion units, artificial intelligence health prediction engines, and adaptive closed-loop control execution units.

[0023] The multi-dimensional high-frequency synchronous sensing unit, serving as the input to the entire system, is deployed near the industrial motor and is responsible for acquiring real-time physical characteristic signals of the motor during operation across all dimensions. These signals include three-phase AC current signals, three-phase AC voltage signals, triaxial vibration acceleration signals, stator winding temperature signals, and speed pulse signals.

[0024] To ensure the accuracy of subsequent algorithms in extracting subtle fault features, this unit employs high-specification hardware configuration. The sampling frequency of the three-phase AC current signal is set to no less than 20 kHz to capture high-frequency harmonic components; the sampling frequency of the three-axis vibration acceleration signal is set to no less than 50 kHz to cover the high-frequency vibration characteristics of the motor bearings and rotor.

[0025] During data acquisition, all physical quantity signals are precisely aligned at the sampling time through a hardware-triggered synchronization mechanism, with the sampling synchronization error strictly controlled within 10 microseconds. This high-precision time base provides a reliable guarantee for subsequent cross-domain correlation analysis of electromagnetic and mechanical signals. The multi-dimensional high-frequency synchronous sensing unit integrates an analog-to-digital converter, signal conditioning circuitry, and a field-programmable gate array (FPGA) logic control module.

[0026] Among them, the field-programmable gate array is responsible for managing the synchronous trigger pulse, ensuring that the timestamps of current, voltage, vibration and speed data packets are completely consistent in each sampling period.

[0027] The collected raw data stream is transmitted in real time to the edge computing gateway via a high-speed industrial Ethernet interface for processing by the subsequent dynamic operating condition feature decoupling unit.

[0028] Combined with appendix Figure 2 The dynamic operating condition characteristic decoupling unit receives the raw high-frequency data stream from the multi-dimensional high-frequency synchronous sensing unit. Its core task is to eliminate the influence of non-stationary operating conditions caused by frequency converter drive on signal characteristics. During frequency converter regulation, the motor speed and load change frequently, rendering traditional spectrum analysis methods based on Fast Fourier Transform ineffective because the spectrum components drift with the speed, resulting in a tailing phenomenon. To solve this problem, this unit adopts a decoupling method based on order analysis, the specific process of which is as follows.

[0029] First, one phase current from the three-phase AC current signal is selected as the analysis object because the fundamental frequencies of the three phase currents are the same. This phase current signal is then bandpass filtered to extract the fundamental component. The filter's center frequency is dynamically tracked based on the motor's rated frequency or an adaptive algorithm is used; the bandwidth should be sufficient to cover the speed variation range to ensure the removal of harmonics and noise interference. The filtered fundamental current signal is denoted as... .

[0030] Next, the filtered fundamental current signal By applying the Hilbert transform, an analytic signal is constructed, thereby extracting the instantaneous frequency. The formula for calculating the instantaneous frequency is: in, This represents the calculated instantaneous electrical frequency value, in Hertz (Hz). This is the filtered fundamental current signal; Indicates to The imaginary part of the signal obtained after performing a Hilbert transform. By taking the time derivative of the phase angle of the arctangent function, the fundamental frequency of the motor at any instant can be accurately obtained.

[0031] According to the number of pole pairs in the motor nameplate parameters Converting instantaneous electrical frequency into instantaneous mechanical speed The transformation relationship is as follows: in The unit is revolutions per minute. It is an extreme logarithm.

[0032] Obtain instantaneous mechanical speed Next, it needs to be converted into angle information for angle domain resampling. First, the rotational speed is integrated to obtain the steering angle. : The rotation angle is measured in radians. The number of sampling points per rotation cycle is set. (i.e., order resolution), usually determined based on the analysis frequency range and the highest order of interest, for example, 1024 points per revolution. From this, the time series corresponding to equal angular intervals can be calculated. ,satisfy: That is, at fixed angular increments Corresponding to a sampling time These moments are usually not the original sampling points and need to be interpolated to obtain the corresponding values ​​from the original time-domain signal. Interpolation operations are performed on the original vibration acceleration signal and the current signal (here, the current signal can be either the original current or a filtered fundamental signal, depending on the subsequent analysis requirements, but generally the original broadband signal is retained to preserve fault characteristics). Linear interpolation or cubic spline interpolation can be used to obtain the values ​​at time [time value missing]. The signal value is used to generate a signal sampled at equal angular intervals in the angular domain.

[0033] After the above resampling process, the vibration and current signals that originally varied with time are mapped to the angular domain. The influence of speed fluctuations is eliminated, and the frequency components in the signal are transformed into stable order components. For example, the vibration component with the same frequency as the speed is represented as a first-order component in the angular domain, and the fault features related to the speed harmonics correspond to integer orders. The final output order feature vector is the signal sequence sampled at equal angular intervals. It provides a normalized feature representation of the operating conditions, providing a stable and reliable data foundation for the subsequent electromagnetic-mechanical coupling feature fusion unit.

[0034] Through this series of steps, the dynamic operating condition characteristic decoupling unit successfully eliminates the interference of speed fluctuations on the spectrum characteristics under variable frequency drive, ensures that the positions of each order component remain constant during the frequency converter adjustment process, eliminates false spectrum offsets, and enables subsequent fault diagnosis to be carried out in a stable feature space.

[0035] Please refer to the attached document. Figure 3 The electromagnetic-mechanical coupling feature fusion unit receives order feature vectors from the dynamic operating condition feature decoupling unit. Its core task is to perform deep semantic fusion of current and vibration signals to reveal the composite fault characteristics generated by electromagnetic and mechanical interactions. This unit constructs a composite fault feature map rich in coupling information through steps such as multi-scale decomposition, feature quantization, attention weighting, and explicit association extraction. The specific implementation process is as follows.

[0036] First, multi-scale wavelet packet decomposition is performed on the input current order signal and vibration order signal respectively. In this embodiment, the Daubechies4 wavelet basis is selected, and the decomposition level is set to 4 levels. Through wavelet packet decomposition, the original signal is divided into 16 equal-width frequency bands, and the wavelet packet coefficients of each frequency band are obtained. For the current signal, the wavelet packet coefficient of the i-th frequency band is denoted as... Where i = 1, 2, ..., 16, and j is the coefficient index; for vibration signals, the wavelet packet coefficients corresponding to the frequency band are denoted as... .

[0037] The number of decomposition layers and frequency bands can be adjusted according to the sampling frequency and the range of fault frequencies of interest to ensure coverage of the typical fault characteristic frequency bands of the motor.

[0038] To address the issue of sparse and easily noise-saturated composite fault features in a single physical domain, this unit introduces a cross-domain mutual information-guided sparse coding technique to enhance the extraction of weak fault features. Specifically, after obtaining the wavelet packet coefficients, a cross-domain mutual information matrix is ​​first constructed. Its elements Defined as the first The current frequency band and the first Mutual information between vibration frequency bands: in For the first A sequence of coefficient amplitudes for each current frequency band. For the first A sequence of coefficient amplitudes for each vibration frequency band. Joint probability density. Mutual information. The magnitude of this value directly reflects the statistical dependence between the two frequency bands. For early composite faults, although the energy change in a single frequency band may be weak, the statistical dependence between the electromagnetic and mechanical frequency bands will be significantly enhanced.

[0039] Based on the mutual information matrix, this unit implements sparse coding: only mutual information values ​​higher than the global threshold are retained. The frequency band combination sets the coefficients of frequency bands below a threshold to zero, thereby eliminating background noise and redundant information unrelated to the fault. This embodiment uses a threshold value. The wavelet packet coefficients after sparse coding can improve the signal-to-noise ratio by about 8-12dB, making early weak fault features stand out from the background noise.

[0040] After wavelet packet decomposition and sparse coding, feature extraction is performed on the coefficients of each frequency band, mainly extracting energy entropy, singular value features, and permutation entropy. The energy entropy is calculated as follows: First, the energy of each frequency band is calculated. ,in The wavelet packet coefficients for this frequency band are used; then the total energy of all frequency bands is calculated. This allows us to obtain the energy percentage of each frequency band. Finally, the energy entropy of this frequency band. The energy entropy vector of the current signal is obtained by calculating it sequentially for all frequency bands. and the energy entropy vector of the vibration signal .

[0041] Singular value features are extracted by constructing a Hankel matrix for the wavelet packet coefficients of each frequency band and performing singular value decomposition. Specifically, for a coefficient sequence of length L, a Hankel matrix with m rows (e.g., m=10) is constructed, with matrix elements filled sequentially by the coefficient sequence. Singular value decomposition is then performed on this matrix to obtain a series of singular values. In this embodiment, the first three largest singular values ​​are selected. These are the singular value features of this frequency band. Combining the singular values ​​of all frequency bands forms the singular value feature vector of the current signal. (16 frequency bands × 3 singular values) and singular value eigenvectors of vibration signals .

[0042] To further capture the nonlinear dynamic abrupt changes within the signal, this unit introduces multi-scale permutation entropy features. Permutation entropy, by comparing the arrangement patterns of adjacent values ​​in the time series, can effectively detect nonlinear dynamic changes in the signal. For each frequency band's wavelet packet coefficient sequence... First, phase space reconstruction is performed, and the length is extracted as follows: (Embedded dimension, taken in this embodiment) Given the permutations of , calculate the probability of each permutation occurring. Then the permutation entropy is defined as: The permutation entropy vector of the current signal is obtained by calculating sequentially for all frequency bands. and the permutation entropy vector of the vibration signal Each vector has a dimension of 16.

[0043] The energy entropy vector of the current signal is concatenated with the singular value eigenvector to obtain the complete current eigenvector. (16+48); similarly, the vibration characteristic vector is obtained. Next, an attention mechanism is used to dynamically weight and fuse current and vibration characteristics to highlight the cross-dimensional coupling relationships related to the fault. This embodiment employs a scaled dot product attention structure, firstly... and Concatenate into a joint feature vector Through three trainable linear transformation matrices. Generate the query matrix Q, key matrix K, and value matrix V respectively: in , As the dimension of the key, this embodiment takes The attention weight matrix A is calculated by the following formula: A is a A matrix whose elements These represent the attention weights of current characteristics on themselves, current on vibration, vibration on current, and vibration on themselves, respectively. The weighted feature vector is: This attention mechanism is jointly trained end-to-end with the neural network in the subsequent AI health prediction engine. The specific training data construction, loss function design and optimization objectives will be detailed in the AI ​​health prediction engine section.

[0044] Obtain the weighted current eigenvector and vibration eigenvectors Subsequently, to match the size of the pre-set composite fault feature map, its dimension was reduced from 80 to 64 using a fully connected layer, resulting in the dimensionality-reduced feature vector. and The parameters of this fully connected layer are optimized along with the rest of the network during end-to-end training. A fused feature space is then constructed using the Kronecker product.

[0045] The Kronecker product is defined as: Soon Each element and Multiplying each element together yields a 64×64 matrix.

[0046] This matrix is ​​a composite fault feature map, where each element represents the coupling strength between current characteristics and vibration characteristics in a specific dimension, and can intuitively reflect the correlation between electromagnetic anomalies and mechanical responses.

[0047] To further enhance the sensitivity and interpretability of the graph to early, minor faults, this unit introduces a graph augmentation technique based on class activation mapping. Specifically, gradient information from a pre-trained AI health prediction engine (detailed below) is used to calculate the current graph. A weight mask is generated by assigning a weight to each pixel in the final health prediction result. The weight mask is obtained by inputting the current graph into a trained health prediction network and then forward-propagating to obtain the predicted health values. Then backpropagation calculation For each pixel in the image, the gradient is calculated, and global average pooling is performed on the gradient to obtain the importance weight of each pixel. Then, the original image is weighted and augmented. in For element-wise multiplication, For the enhancement coefficient (in this embodiment, we take...) Enhanced map The fault-related coupling regions are highlighted, while irrelevant regions are suppressed, thus providing more discriminative inputs for subsequent convolutional neural networks.

[0048] To address the issue of aliasing of multiple fault modes, this unit generates a fault decoupling identifier vector along with the output enhancement map. This vector is formed by performing global max pooling and global average pooling on the aforementioned enhancement map and then concatenating the two: This identifier vector is designed to give different composite faults a larger inter-class distance and a smaller intra-class distance in the feature space, thereby facilitating accurate classification and quantification by the final health prediction engine.

[0049] In addition to the aforementioned graphs, this unit also extracts a set of explicit correlation features as supplementary data, including the phase difference between the negative-sequence current component and the vibration harmonic component, and the linear correlation coefficient between voltage fluctuation rate and temperature rise rate. The negative-sequence current component is calculated from the instantaneous values ​​of the three-phase current using the symmetrical component method. Let the three-phase current be... Then the negative order component The calculation formula is: in , Since it is a complex number, take its amplitude. and phase The vibration harmonic component is obtained by order tracking of the vibration order signal, extracting the Fourier coefficients of the first harmonic, and thus obtaining the amplitude and phase of this frequency component. The phase of the first harmonic is denoted as... Then the phase difference .

[0050] Voltage fluctuation rate is defined as the effective value of voltage. The rate of change per unit time, i.e. The rate of temperature rise is defined as the rate of change of the stator winding temperature T per unit time, i.e. Select a measurement sequence over a continuous time period and calculate the voltage fluctuation rate sequence. and temperature rise rate sequence The linear correlation coefficient between the two is calculated using the Pearson correlation coefficient formula: in and denoted as the mean of the two sequences, and n is the sequence length.

[0051] The electromagnetic-mechanical coupling feature fusion unit will ultimately enhance the composite fault feature map. (64×64 matrix), fault decoupling identification vector and the aforementioned explicit correlation features (including phase difference) Correlation coefficient (Each component, including the enhanced graph, is packaged together and passed as output to the AI ​​health prediction engine. The enhanced graph serves as the primary input to the subsequent residual network, while the fault decoupling identifier vector is used as an auxiliary feature and fused with the temporal features extracted from the graph. This unit, through multi-dimensional and multi-level fusion processing, transforms the originally isolated current and vibration signals into feature representations rich in electromagnetic-mechanical coupling information. This significantly enhances the early identification capability of complex faults such as rotor cracks, stator inter-turn short circuits, and bearing latent wear, providing a high-quality data foundation for subsequent quantitative health assessment.)

[0052] Combined with appendix Figure 4 The AI-powered health prediction engine is the core logic layer of the system, deployed on a workshop-level edge server. It is responsible for deep analysis of complex fault feature maps and outputting a quantitative assessment of the motor's health status. This engine employs a combined architecture based on residual networks and bidirectional gated recurrent units, while also integrating an uncertainty assessment module to ensure the reliability of the prediction results. The following provides a detailed explanation of the engine's network structure, training process, and operational logic.

[0053] The input to the AI ​​health prediction engine consists of two parts: an enhanced composite fault feature map generated by the electromagnetic-mechanical coupling feature fusion unit. Its size is a 64×64 two-dimensional matrix, representing the coupling relationship between current and vibration characteristics; and a fault decoupling identification vector. The network structure consists of three main parts: a residual network feature extraction layer, a bidirectional gated recurrent unit temporal modeling layer, and a multi-task output layer.

[0054] First, the enhanced map is processed by a residual network to extract spatial features, resulting in a 512-dimensional spatial feature vector. This vector is concatenated with the fault decoupling identifier vector at the current time step to form a fused feature vector. Subsequently, a time series is constructed using a sliding window approach: the fused feature vector at the current time step t is stacked with the fused feature vectors from the past 9 time steps (t-9 to t-1) to form a sequence of length 10, which serves as the input to a bidirectional gated recurrent unit.

[0055] The residual network consists of four stacked residual blocks. Each residual block contains two convolutional layers with 3×3 kernels and a stride of 1. SAME padding is used to ensure the feature map size remains constant. The first residual block outputs 64 feature maps, the second 128, the third 256, and the fourth 512. Each convolutional layer is followed by a batch normalization layer and a ReLU activation function. Residual connections are implemented through skip-addition, directly adding the input of each residual block to its output. The last residual block is followed by a global average pooling layer, compressing the 512 feature maps into a 512-dimensional feature vector, which aggregates the spatial structure information in the feature maps.

[0056] The 512-dimensional feature vector is then fed into a bidirectional gated recurrent unit for time series modeling. Considering that the evolution of the health state is a continuous process, this engine uses a sliding window approach to construct the time series: the feature vector of the current time t is stacked with the feature vectors of the past 9 times (t-9 to t-1) to form a sequence of length 10, which serves as the input to the bidirectional gated recurrent unit. The time interval between adjacent times is fixed at 1 hour, but can be adjusted according to the data acquisition frequency in practical applications. The bidirectional gated recurrent unit contains two layers, each with 256 hidden units. The forward and backward hidden states are concatenated to obtain a 512-dimensional time series feature vector.

[0057] The temporal feature vector is input to the multi-task output layer. The output layer consists of two branches: a health regression branch and a remaining lifespan regression branch. The health regression branch is composed of a fully connected layer containing one neuron, with the sigmoid activation function, and the output value is between 0 and 1, denoted as . Where 1 represents perfect health and 0 represents catastrophic failure. The remaining lifespan prediction branch also consists of a fully connected layer containing one neuron, using ReLU as the activation function, and outputting a non-negative real number, denoted as . The unit is hours, which indicates the estimated time the motor can continue to run under its current condition.

[0058] The training data comes from two sources. One is a historical maintenance database, which stores the full lifecycle data of multiple motors of the same type over many years, including raw signals collected by the multi-dimensional high-frequency synchronous sensing unit, as well as corresponding maintenance records and fault occurrence times. These raw signals are processed sequentially by a dynamic operating condition feature decoupling unit and an electromagnetic-mechanical coupling feature fusion unit to generate corresponding composite fault feature maps, and the health score and remaining lifespan before the fault are marked according to the maintenance records.

[0059] The true label of health score Calibrated by an expert system. The expert system calculates the value based on a weighted average of vibration RMS, current harmonic distortion rate, winding temperature rise rate, and cumulative operating time. The specific formula is as follows: in The health factor is the normalized vibration RMS value. The normalization reference value is the statistical median of the vibration RMS values ​​of motors of the same model under healthy conditions. When the measured vibration RMS value equals the reference value, The higher the measured value, the better. Linearly decrease to 0; The normalized health factor is the harmonic distortion rate of the current. The normalization benchmark is the median of the harmonic distortion rate under healthy conditions. The health factor is the normalized winding temperature rise rate, which decreases rapidly when the temperature rise rate exceeds the threshold. The health factor, defined as the normalized ratio of cumulative uptime to mean time between failures, is: ,in The cumulative operating hours are represented by MTBF, which is the mean time between failures (MTBF) for motors of the same model. Weighting coefficients are also used. The value is determined by referring to a table based on the motor type and rated power, and it meets the following requirements. The normalized baseline values ​​and weight tables for each health factor are pre-set and embedded in the system by the equipment manufacturer based on a large amount of historical data.

[0060] The true label of remaining lifespan Defined as the time interval from the current moment to the actual occurrence of a fault or downtime for maintenance, in hours. For historical maintenance data, if the sample collection time is [missing information] from the actual fault time... Hours, For laboratory simulation data, samples at different degradation stages are obtained through accelerated life testing. The test time is converted into equivalent actual operating time according to the acceleration factor, and then labeled according to the above rules.

[0061] All samples were divided into training, validation, and test sets in a 7:2:1 ratio. The training set was used for model parameter optimization, the validation set was used for hyperparameter tuning and early stopping detection, and the test set was used for final performance evaluation.

[0062] The AI ​​health prediction engine employs a multi-task joint training method, with a total loss function. The weighted sum of the health regression loss and the remaining life regression loss: Among them, health regression loss Mean square error is used: This represents the batch sample size. For the first The predicted health score for each sample. This corresponds to the actual health score.

[0063] Remaining life regression loss Mean squared logarithmic error is used to balance the accuracy of early and late predictions: in and The first The predicted and actual remaining lifespan values ​​for each sample are added by 1 to avoid the logarithm being zero.

[0064] Weighting coefficient and Initially set to 0.5 and 0.5, the values ​​can be fine-tuned based on the performance of the validation set to make the loss magnitudes of the two tasks comparable.

[0065] Training employed the Adam optimizer with an initial learning rate of 0.001. The batch size was set to 32, and the number of training epochs was 200. An early stopping strategy was used, halting training when the validation set loss did not decrease within 20 consecutive epochs. Simultaneously, a learning rate decay was employed, multiplying the learning rate by 0.5 every 10 epochs if the validation loss did not decrease. The training objective was to minimize the total loss, thereby obtaining model parameters capable of accurately predicting health scores and remaining lifespan. The linear transformation matrix in the attention mechanism is also included. It also participates in end-to-end optimization along with other network parameters during this process.

[0066] To ensure the reliability of the prediction results, the engine incorporates an uncertainty assessment module. This module implements Bayesian inference based on the Monte Carlo Dropout method. During training, a Dropout layer is added after the fully connected layers of the residual network, with a dropout rate of 0.5. During the inference phase, the Dropout layer is retained and forward propagation is performed multiple times (e.g., 30 times) to obtain a set of samples with health scores, and their mean is calculated. and standard deviation The confidence interval is defined as follows: The confidence level is 95%. The confidence level can be quantified as... An approximation of the value, which is simplified in this embodiment to when The prediction is considered reliable when the confidence level is higher than 0.85.

[0067] When predicting the standard deviation of health score When the confidence level exceeds 0.075 (i.e., below 0.85), the engine determines that the current input data quality is insufficient or the model confidence level is too low. In this case, no specific score is output; instead, a data re-verification process is triggered. The re-verification logic calls the data integrity indicators of the data quality monitoring layer to check for sensor anomalies. The data quality monitoring layer monitors the electrical connection status and signal integrity of each sensor in real time, such as whether the DC bias of the vibration signal exceeds the preset range or whether the current signal exhibits clipping distortion. If a sensor fault is confirmed, the system automatically switches to redundancy mode, using the current observer data integrated within the motor driver to replace the external high-precision sensor data; if it is only a transient interference, it waits for the data to be re-evaluated at the next moment. This mechanism effectively avoids making erroneous control decisions based on unreliable data.

[0068] The AI-powered health prediction engine, through a joint architecture of residual networks and bidirectional gated recurrent units, extracts spatially coupled features from a composite fault feature map and models temporal evolution patterns, ultimately outputting a health score and remaining life prediction. Combining a multi-task loss function and uncertainty assessment, this engine not only provides quantitative evaluation but also possesses self-questioning and data verification capabilities, offering a reliable basis for subsequent adaptive closed-loop management.

[0069] The adaptive closed-loop control execution unit automatically executes closed-loop control logic based on the health score and remaining life prediction results output by the AI ​​health prediction engine, combined with the real-time requirements of the current production process. This unit incorporates a multi-objective optimization controller, whose optimization objective is to maximize motor operating efficiency and minimize mechanical losses while ensuring the completion of production tasks. To achieve this objective, the controller uses a weighted summation method to construct a comprehensive performance index: in The overall performance index that needs to be maximized; This represents the current output power of the motor. Rated power; This is an estimated value for mechanical loss. This refers to mechanical losses under rated operating conditions. This represents the current production task completion rate (e.g., output per unit time). To achieve the highest possible level of completion; The weighting coefficient is dynamically adjusted according to production priority and satisfies the following conditions: The controller maximizes output frequency or current limits by optimizing them. The optimization results are then sent to the frequency converter as instructions.

[0070] The calculation model for the health quantification score integrates neural network predictions, prior knowledge of inherent equipment reliability, and environmental compensation factors. The specific formula is as follows: in The final health measurement score ranges from 0 to 1. The AI ​​health prediction engine is based on the current composite fault feature map. Output health prediction value; Let the prior value of the inherent reliability of the device based on the Weibull distribution model be defined as follows: In the formula The cumulative running time of the motor. For scale parameters, For shape parameters, both are pre-calibrated and fixed in the system using maximum likelihood estimation based on historical life data of the same type of motor; The environmental compensation coefficient is calculated by the data quality monitoring layer based on comprehensive data from on-site environmental sensors. ,in For ambient temperature, For reference temperature, For ambient humidity, For reference humidity, This is an empirical coefficient, ranging from 0.1 to 0.3, to ensure... Between 0.9 and 1.1; and This is a dynamic weighting coefficient, whose value is adjusted in real time based on the motor's historical maintenance records and current operating conditions. The adjustment rule is as follows: , ,in For the baseline value (e.g.) ), This is the current load factor (the ratio of actual power to rated power). The normalized value of the historical fault frequency of the same model motor, ranging from 0 to 1, and satisfying the following conditions: .

[0071] The control strategy of the adaptive closed-loop control execution unit exhibits a clear hierarchical characteristic. When the health metric score is in the warning range of 0.6 to 0.8, the system determines that the motor has slight deterioration. At this time, the controller generates a derating operation command, which modifies the output current limit of the frequency converter through an industrial Ethernet protocol (such as PROFINET or EtherNet / IP), reducing the motor operating load by 20% to 30%, thereby reducing winding heat generation and bearing load, and achieving the purpose of suppressing the speed of fault evolution.

[0072] When the health metric score falls below 0.4, the system determines that the motor faces a high-risk fault. The controller immediately sends an emergency shutdown request to the factory's manufacturing execution system and, upon system confirmation, instructs the braking unit to execute a controlled shutdown procedure. Before sending any commands that affect the motor's operating status, the system must verify whether the motor is currently in manual adjustment mode and ensure that the issued frequency adjustment parameters avoid the mechanical resonance range. The system pre-stores a list of critical speed resonance ranges for the motor under no-load and full-load conditions, such as 1200 to 1500 rpm and 2800 to 3100 rpm. When the target speed falls into either resonance range, the controller automatically adjusts the adjustment range to quickly pass through the resonance range or stabilize at a safe speed, avoiding new mechanical resonance caused by closed-loop adjustment. This dual verification mechanism ensures the safety of closed-loop control.

[0073] The adaptive closed-loop control execution unit transforms artificial intelligence prediction results into specific equipment control actions through the aforementioned multi-objective optimization, integrated health model, and hierarchical safety strategy, realizing a complete closed loop from perception to decision-making to execution, effectively delaying fault development and preventing sudden shutdowns.

[0074] The system also integrates an augmented reality maintenance assistance module. This module is activated when the health metric score reaches a preset maintenance threshold, which can be set according to the importance of the equipment; in this embodiment, it is set to 0.6. When the adaptive closed-loop management execution unit detects that the health score is below 0.6, it automatically triggers the maintenance assistance process, retrieving a maintenance strategy matching the current fault characteristics from the fault knowledge base.

[0075] A pre-built fault knowledge base stores historical fault cases of motors of the same model. Each case record includes a composite fault feature map at the time of the fault, the root cause localization conclusion, a list of recommended replacement spare parts, and standardized maintenance operation instructions. The knowledge base uses a vectorized storage method, extracting feature vectors from the fault feature map of each case using a convolutional neural network and creating an index. During retrieval and matching, the system also extracts feature vectors from the currently generated composite fault feature map using the same convolutional neural network, calculates the cosine similarity between this vector and the vectors of all cases in the knowledge base, selects the case with the highest similarity as the matching result, and outputs the corresponding maintenance information.

[0076] The retrieved maintenance information is pushed in real time to the mobile terminals of on-site maintenance personnel via a wireless network. These mobile terminals can be smartphones or AR glasses with augmented reality capabilities. The information push uses the MQTT protocol to ensure low latency and high reliability.

[0077] Upon arrival at the site, maintenance personnel use a mobile device to scan the QR code or RFID tag affixed to the motor casing. The QR code or RFID tag encodes the motor's unique device ID. After scanning, the mobile device retrieves the motor's 3D digital twin model and current health scores for key components from an edge server or cloud via 5G or WiFi network. The 3D digital twin model is built based on the motor's CAD design drawings and pre-defines key areas such as bearings, end windings, and rotor supports.

[0078] The mobile terminal utilizes augmented reality technology to overlay a 3D digital twin model onto a real-time image of the motor captured by a camera, ensuring the model is in the correct orientation and position. During overlay, different parts of the model are color-coded based on their health scores: parts with a health score above 0.8 are displayed in green, indicating good condition; scores between 0.6 and 0.8 are displayed in yellow, indicating attention is needed; and scores below 0.6 are displayed in red, indicating a serious risk of failure. Maintenance personnel can visually observe the health status distribution of the motor's internal structure on the screen.

[0079] When maintenance personnel click on the area displayed in red on the screen, the mobile terminal sends a request to the edge server to retrieve in-depth spectrum analysis charts and fault evolution prediction curves for that part. The spectrum analysis charts display the order spectrum of the vibration or current signal of that part, and the fault evolution prediction curves show the trend of the health score of that part over a period of time and its predicted future trend. These charts are generated based on historical data and prediction results stored in the artificial intelligence health prediction engine, helping maintenance personnel to deeply understand the causes and development trends of faults, thereby making accurate maintenance decisions.

[0080] The augmented reality maintenance assistance module transforms abstract fault data into intuitive visual information through the above steps, and combines it with the real-world scene, greatly improving maintenance efficiency and accuracy. This module works in conjunction with the adaptive closed-loop management execution unit. Once maintenance is complete, maintenance personnel can confirm completion on their mobile devices, and the system automatically updates the fault knowledge base and resets the health score, forming a complete maintenance closed loop.

[0081] In terms of underlying support, this system has a comprehensive data quality monitoring layer. As the underlying support module of the system, the data quality monitoring layer is responsible for monitoring the operating status and signal quality of all sensors in real time, ensuring the reliability and integrity of the input data. This layer is deployed in the edge computing gateway and maintains continuous communication with the multi-dimensional high-frequency synchronous sensing unit, acquiring the status information of each channel at a fixed period, such as once per second.

[0082] The monitoring includes two aspects: electrical connection status and signal integrity. Electrical connection status is detected by monitoring the physical link between the sensor and the acquisition module. For current and vibration sensors, the system continuously monitors the DC bias voltage of their output signals. Each sensor has a specific DC bias range under normal operating conditions. For example, IEPE accelerometers typically have a DC bias of 2.5V or 3.3V. When the bias voltage deviates from the normal value by more than ±0.5V, the system determines that the sensor has a loose connection or a broken wire. For temperature sensors, the connection status is determined by detecting whether their output resistance or current value exceeds the lower limit of the effective range.

[0083] Signal integrity monitoring focuses on the quality of the analog signal output by the sensor. For vibration acceleration signals, the system calculates the amplitude at each sampling point and compares it with the full-scale range of the analog-to-digital converter. When multiple consecutive sampling points reach or approach the upper limit of the full-scale range, clipping distortion is detected. In this embodiment, a clipping alarm is triggered when the amplitude at 10 consecutive sampling points exceeds 98% of the full-scale range. For current signals, the system monitors for abnormal truncation or abrupt changes in the waveform, identifying this by calculating the first-order difference of the signal and comparing it with historical statistical thresholds. Simultaneously, the signal-to-noise ratio (SNR) is monitored; when the noise floor significantly increases beyond a preset threshold, the sensor is deemed to be experiencing severe electromagnetic interference.

[0084] The data quality monitoring layer is also responsible for assessing the quality of the speed pulse signal. By measuring the uniformity of the pulse interval and calculating the coefficient of variation of the continuous pulse interval, when the coefficient of variation exceeds 5%, it is determined that the speed signal quality has deteriorated, which may indicate an encoder malfunction or loose installation.

[0085] Once a hardware fault or signal quality issue is detected at the sensor level, the system automatically triggers a redundancy switching mechanism. The switching logic follows a priority principle: data from redundant sensors of the same type is used first; if no redundant sensors are available, the system switches to data from the current observer integrated within the motor driver. The driver current observer reconstructs the three-phase current based on the inverter DC bus current and switch status. Although its accuracy is lower than that of external high-precision current sensors, it can maintain basic system operation even if external sensors fail. For vibration signals, if no redundant sensors are available, the system temporarily freezes vibration-related fault diagnosis functions, performs a health assessment based solely on current and temperature data, and displays a message on the interface indicating that vibration data is unavailable.

[0086] The redundancy switching process employs a seamless switching design to ensure uninterrupted data flow. Specifically, the system simultaneously receives data from both the primary sensor and the backup data source in the background, comparing their differences in real time. When the primary sensor fails, the output data source automatically switches to the backup channel, with the switching time controlled within 100 milliseconds, unnoticed by the upper-layer algorithm. After the switch is complete, the system generates an alarm record and pushes it to the maintenance terminal, prompting timely repair of the faulty sensor.

[0087] Through the aforementioned mechanisms, the data quality monitoring layer ensures the reliability of input data, providing a reliable data foundation for the AI ​​health prediction engine and the adaptive closed-loop control execution unit. This avoids misjudgments or malfunctions caused by sensor failures and ensures the continuous and stable operation of the system in harsh industrial environments.

[0088] In summary, the industrial motor health AI-based predictive maintenance and closed-loop management system constructed in this embodiment achieves precise control over the motor's operating status under complex conditions through full-stack optimization from the perception layer to the execution layer. It is not merely a monitoring system, but an intelligent agent with self-diagnosis, self-adjustment, and self-protection capabilities, providing a solid and reliable power guarantee for building modern smart factories.

[0089] It will be apparent to those skilled in the art that the present invention is not limited to the details of the exemplary embodiments described above, and that the present invention can be implemented in other specific forms without departing from the spirit or essential characteristics of the present invention. Therefore, the embodiments should be regarded as exemplary and non-limiting in all respects.

[0090] Furthermore, it should be understood that although this specification describes embodiments, not every embodiment contains only one independent technical solution. This narrative style is merely for clarity. Those skilled in the art should consider the specification as a whole, and the technical solutions in each embodiment can also be appropriately combined to form other embodiments that can be understood by those skilled in the art.

Claims

1. An AI-based predictive maintenance and closed-loop management system for industrial motor health, characterized in that, include: The multi-dimensional high-frequency synchronous sensing unit collects the operating parameters of the motor under variable frequency drive, including at least current signals and vibration signals. The dynamic operating condition feature decoupling unit receives the operating parameters and maps the current signal and vibration signal in the time domain to the angle domain through instantaneous frequency estimation and angle domain resampling technology to generate an order feature vector, so as to eliminate the frequency modulation interference of speed fluctuation on signal characteristics under variable frequency drive. The electromagnetic-mechanical coupling feature fusion unit receives the order feature vector and performs semantic-level fusion of the current signal and vibration signal by constructing a cross-domain mapping matrix to generate a composite fault feature map characterizing the electromagnetic-mechanical coupling relationship. The artificial intelligence health prediction engine receives the composite fault feature map, performs in-depth analysis through a spatiotemporal sequence model, and outputs a quantitative assessment result of the motor's health at the current moment. The adaptive closed-loop control execution unit receives the health quantification assessment results and automatically generates and sends control commands to the motor driver to adjust the motor operating parameters based on the health quantification assessment results, forming a complete link from perception and prediction to closed-loop control.

2. The industrial motor health AI predictive maintenance and closed-loop management system according to claim 1, characterized in that, The dynamic operating condition characteristic decoupling unit is specifically used for: The fundamental component is extracted from the current signal, and the Hilbert transform is applied to the fundamental component to extract the instantaneous frequency. Based on the instantaneous frequency and the number of pole pairs of the motor, the instantaneous mechanical speed is calculated, and the instantaneous mechanical speed is integrated to obtain the rotation angle information; Based on the preset number of sampling points per revolution, a time sequence corresponding to equal angular intervals is determined, and based on the time sequence, the original vibration signal and the current signal are resampled in the angular domain through interpolation to generate the order feature vector decoupled from the rotational speed fluctuation.

3. The industrial motor health AI predictive maintenance and closed-loop management system according to claim 1, characterized in that, The electromagnetic-mechanical coupling feature fusion unit is specifically used for: Multi-scale wavelet packet decomposition is performed on the order feature vectors of the input current signal and vibration signal respectively to obtain wavelet packet coefficients of multiple frequency bands; The mutual information between the current signal frequency band and the vibration signal frequency band is calculated by using sparse coding technology guided by cross-domain mutual information matrix, and the wavelet packet coefficients are sparsified based on the mutual information to enhance the characteristics of weak composite faults. Feature vectors are extracted from the wavelet packet coefficients of each frequency band after sparsification, and the Kronecker product is used to fuse the feature vectors of the current signal and the vibration signal to construct a composite fault feature map.

4. The industrial motor health AI predictive maintenance and closed-loop management system according to claim 3, characterized in that, The electromagnetic-mechanical coupling feature fusion unit is also used for: Using gradient information from a pre-trained AI health prediction engine, the contribution weight of each pixel in the composite fault feature map to the final health prediction result is calculated, and a weight mask is generated. We use a weighted mask to enhance the composite fault feature map to highlight the coupling region related to the fault, generate an enhanced composite fault feature map, and use it as input to an artificial intelligence health prediction engine.

5. The industrial motor health AI predictive maintenance and closed-loop management system according to claim 1, characterized in that, The multi-dimensional high-frequency synchronous sensing unit includes an analog-to-digital converter, a signal conditioning circuit, and a field-programmable gate array logic control module; The field-programmable gate array (FPGA) logic control module is used to generate synchronous trigger pulses. The hardware-triggered synchronization mechanism controls the precise alignment of the sampling times of the current signal and the vibration signal, ensuring that the sampling synchronization error is within 10 microseconds, thus providing a time reference for subsequent cross-domain correlation analysis.

6. The industrial motor health AI predictive maintenance and closed-loop management system according to claim 1, characterized in that, The AI ​​health prediction engine includes a combined architecture of residual networks and bidirectional gated recurrent units; The residual network is used to extract spatial coupling features from the composite fault feature map, and the bidirectional gated recurrent unit is used to model the time sequence composed of spatial coupling features to learn the evolution law of health status and output the quantitative assessment results of health and the predicted value of remaining life.

7. The industrial motor health AI predictive maintenance and closed-loop management system according to claim 1 or 6, characterized in that, The AI-powered health prediction engine also has a built-in uncertainty assessment module; The uncertainty assessment module implements Bayesian inference based on the Monte Carlo Dropout method, which is used to calculate the confidence level of the health quantification assessment results during the inference phase; When the confidence level is lower than the preset threshold, the system triggers a data re-verification process and uses the data integrity indicators of the data quality monitoring layer to check the sensor status.

8. The industrial motor health AI predictive maintenance and closed-loop management system according to claim 1, characterized in that, The adaptive closed-loop control execution unit has a built-in multi-objective optimization controller; The multi-objective optimization controller is used to optimize motor operating parameters based on the results of health quantification assessment and current production process requirements, with the goal of maximizing motor operating efficiency and minimizing mechanical losses. Before generating and issuing control commands, the controller also performs resonance range avoidance verification. By comparing the target speed with the pre-stored list of mechanical resonance ranges, it ensures that the adjusted operating parameters avoid the mechanical resonance range.

9. The industrial motor health AI predictive maintenance and closed-loop management system according to claim 1, characterized in that, It also includes a data quality monitoring layer, deployed in the edge computing gateway, used to monitor the operating status and signal quality of all sensors in real time; The data quality monitoring layer determines the electrical connection status by detecting the DC bias voltage of the sensor output signal, and determines whether clipping distortion exists by calculating whether the signal amplitude reaches the full-scale range of the analog-to-digital converter. When a sensor malfunction or signal quality degradation is detected, the system automatically triggers a redundancy switching mechanism, prioritizing switching to redundant sensors of the same type or switching to the current observer data integrated inside the motor driver to maintain basic system operation.

10. The industrial motor health AI predictive maintenance and closed-loop management system according to claim 1, characterized in that, It also includes an augmented reality maintenance assistance module; The augmented reality maintenance assistance module is activated when the health measurement assessment results reach the preset maintenance threshold. It is used to retrieve matching maintenance strategies from the fault knowledge base and push the maintenance strategies and the three-dimensional digital twin model of the motor to the mobile terminal of the maintenance personnel. The mobile terminal uses augmented reality technology to overlay a 3D digital twin model onto the actual motor image in the correct orientation, and renders colors according to the health scores of each key part to intuitively guide maintenance personnel to perform precise maintenance.

Images