A condition monitoring method and system for a high-pressure blower type shaftless external rotor motor

Abstract

This invention provides a condition monitoring method and system for a high-pressure fan-type shaftless external rotor motor, relating to the field of motor condition monitoring and fault diagnosis technology. Specific steps include: First, uniformly arranging four sensors on each of the upper and lower guide rails of the shaftless external rotor motor to construct a 360° circular monitoring array; second, acquiring eight channels of vibration data, and improving the signal-to-noise ratio through adaptive multi-scale filtering and wavelet envelope analysis; then, using a spatial vector decomposition and fusion algorithm to compress the eight-dimensional signal into a three-dimensional equivalent vector, and constructing a five-dimensional feature vector by combining speed normalization and temperature compensation; next, inputting the data into a lightweight Lite-CNN model for fault identification and severity quantification, and achieving fault location through confidence weighting; finally, triggering online incremental learning and updating the model using knowledge distillation. This invention has significant advantages in terms of recognition rate, positioning accuracy, and power consumption.

Description

Technical Field

[0001] This invention belongs to the field of motor condition monitoring and fault diagnosis technology, specifically relating to a condition monitoring method and system for a high-pressure fan-type external rotor motor without a central shaft, which is particularly suitable for online condition monitoring and predictive maintenance of external rotor motors without a central shaft under operating conditions of 800V high voltage and 1500-3000rpm. Background Technology

[0002] High-pressure fan-type shaftless external rotor motors are widely used in new energy vehicle thermal management, industrial ventilation, and data center heat dissipation due to their advantages such as compact structure, high power density, and low wind resistance. However, the inherent characteristics of this type of motor cause traditional monitoring technologies to fail: the shaftless structure makes vibration energy mainly radial and tangential modes, with extremely weak axial signals. Traditional axial sensor-based solutions have more than 30% missing monitoring dimensions, making it impossible to effectively identify typical faults such as imbalance and eccentricity; the traditional single-plane layout of 1 to 4 sensors has at least a 180° monitoring blind zone, and has almost no ability to locate faults with spatial directionality such as eccentricity and local wear; the fan operating environment has complex noise, with a typical signal-to-noise ratio of only 3 to 5 dB. Traditional FFT and single envelope analysis methods cannot extract the weak early fault signals of the submerged bearings, and the early fault detection rate is less than 45%.

[0003] Furthermore, using independent thresholds for three types of heterogeneous data—vibration, temperature, and rotational speed—fails to capture the chain reaction of frictional heat generated by bearing wear, leading to misdiagnosis and missed diagnosis of complex faults. Traditional fixed-parameter filtering and thresholding have errors exceeding 8% over a wide rotational speed range of 1500–3000 rpm, with the fault identification rate decreasing significantly as the rotational speed deviates. Traditional industrial control computer solutions consume as much as 25–30W, making them unsuitable for miniaturized embedded devices, while simplified thresholding solutions sacrifice diagnostic accuracy. Traditional AI black-box models only output binary judgments, failing to provide decision-making basis such as fault location and severity. Offline-trained models cannot adapt to equipment aging and environmental changes, resulting in an average annual decrease in diagnostic accuracy of 3–5%, and high costs associated with regular retraining.

[0004] To address the aforementioned issues, there is currently no effective solution that can simultaneously meet the requirements of full-dimensional monitoring, high-precision positioning, low signal-to-noise ratio robustness, wide speed adaptability, and edge deployment for shaftless external rotor motors. Summary of the Invention

[0005] The purpose of this invention is to overcome the shortcomings of the prior art and provide a condition monitoring method and system for high-pressure blower type external rotor motor without central shaft. This solves the problems of traditional technology in this type of motor, such as lack of monitoring dimensions, large blind spots, low early fault detection rate, difficulty in multi-parameter fusion, poor adaptability to wide speed range, limited edge deployment, uninterpretable results, and long-term performance degradation.

[0006] To solve the above problems, the present invention is achieved through the following technology: S1. Arrange four or more vibration sensors on the upper and lower circumferential support rails of the motor to form a multi-channel circumferential monitoring array. Configure corresponding installation angle parameters and height factors for each sensor, wherein the height factors of the upper rail sensor and the lower rail sensor have opposite signs. S2. Synchronously acquire and preprocess the raw vibration signals of all channels; S3. Based on the preprocessed vibration signals of each channel, the corresponding installation angle, the sensor calibration weight coefficient and the height factor, spatial vector decomposition and fusion are performed. The three-dimensional equivalent vibration vectors in the radial, tangential and axial directions, which characterize the overall vibration state of the motor, are reconstructed using the differential signals of the upper and lower guide rail sensors. S4. Combine the three-dimensional equivalent vibration vector, the real-time temperature signal of the motor, and the real-time speed signal to construct a multi-parameter fusion feature vector; S5. Input the multi-parameter fused feature vector into the intelligent diagnostic model, and output a multi-dimensional diagnostic result including fault type, spatial location, severity level, and judgment confidence.

[0007] Preferably, the preprocessing in step S2 includes adaptive multi-scale filtering: extracting the fundamental frequency based on the FFT result of the multi-channel synthesized signal and converting it into real-time rotational speed; dynamically adjusting the cutoff frequency of the FIR bandpass filter to filter the vibration signal according to the real-time rotational speed; and extracting early bearing fault features through wavelet envelope analysis.

[0008] Preferably, the calculation method for spatial vector decomposition and fusion in step S3 is as follows: in, This represents the installation azimuth angle of the i-th sensor. This represents the calibration weighting coefficient for the i-th sensor. This represents the height factor of the i-th sensor. This represents the radial, tangential, and reconstructed axial vibration vectors obtained after fusion.

[0009] Preferably, in step S4, when constructing the multi-parameter fusion feature vector, the temperature signal is first normalized and the three-dimensional vibration vector is temperature compensated, and then the rotation speed signal is normalized to the [0,1] interval. Finally, the vector is combined to form a 5-dimensional feature vector containing the compensated vibration vector, normalized temperature, and normalized rotation speed.

[0010] Preferably, the intelligent diagnostic model is an enhanced lightweight multi-parameter fusion diagnostic model, which adopts a dual-output head structure. The first output head outputs the probability distribution of fault types through a multi-classification problem activation function, and the second output head outputs a fault severity level quantification value of 0 to 1 through a regression task activation function.

[0011] Preferably, in step S5, the fault location is determined by confidence-weighted fusion: the main vibration direction angle and the location of the sensor with the largest amplitude are calculated, and after being assigned corresponding confidence levels, they are fused to obtain the accurate fault location.

[0012] Preferably, step S5 also includes an alarm triggering step: when the confidence level of the fault type determination is greater than a preset threshold, an alarm is triggered, and the fault severity level is divided into three maintenance priorities: minor, moderate, and severe.

[0013] This invention also provides a condition monitoring system for a high-pressure blower-type shaftless external rotor motor, characterized in that it includes: A multi-channel circumferential monitoring array includes four or more vibration sensors arranged on the upper and lower circumferential support rails of the motor to form a multi-channel circumferential monitoring array. Each sensor is configured with corresponding installation angle parameters and height factors, wherein the height factors of the upper rail sensor and the lower rail sensor have opposite signs. The signal acquisition unit is electrically connected to the sensor array and is used to synchronously acquire multi-channel raw vibration signals; The spatial vector fusion unit is used to perform spatial vector decomposition and fusion on the preprocessed vibration signal to reconstruct the three-dimensional equivalent vibration vector. A multi-parameter fusion unit is used to combine three-dimensional vibration vectors, temperature and rotation speed signals to construct a multi-parameter fusion feature vector. The intelligent diagnostic unit is used to input multi-parameter feature vectors and output multi-dimensional diagnostic results.

[0014] Preferably, four or more vibration sensors are arranged on the upper and lower circumferential support rails of the motor, all evenly arranged along the circumference. The sensors on the lower rail and the upper rail are symmetrically distributed at 180° in the circumferential direction, forming a 360° monitoring coverage without blind spots.

[0015] Preferably, the system further includes an online incremental learning unit, which is used to fine-tune the intelligent diagnostic model by knowledge distillation after collecting a preset number of samples, and to perform replacement only when the performance of the new model is improved through a version management mechanism.

[0016] Beneficial effects of this invention: 1. Full-dimensional blind-spot-free monitoring: The dual-rail eight-channel layout combined with height factor differential reconstruction restores the monitoring dimension coverage to over 99%, eliminating the 180° traditional monitoring blind spot, and achieving fault location accuracy of ±10°.

[0017] 2. Strong robustness under low signal-to-noise ratio: Adaptive multi-scale filtering combined with wavelet envelope analysis reduces the lowest identifiable signal-to-noise ratio from 10dB to 3dB, and the early fault detection rate reaches 98% under 3-5dB conditions.

[0018] 3. Excellent adaptability across a wide speed range: Speed ​​normalization processing ensures that the diagnostic error is ≤1% across the entire range of 1500-3000rpm, which is 8 times better than traditional solutions.

[0019] 4. High efficiency and low power consumption in edge deployment: Spatial vector fusion compresses 60% of the computation, the Lite-CNN model has less than 15KB of parameters, a single inference time of 3.2ms, and the system power consumption is only 4.8W, which can run on low-power MCUs such as STM32F4.

[0020] 5. Interpretable diagnostic results: The four-dimensional diagnostic report system provides fault type, location, severity, and confidence level, providing direct basis for operation and maintenance decisions, with a false alarm rate controlled below 0.5%.

[0021] 6. Long-term stable performance: The online incremental learning mechanism ensures that the diagnostic accuracy remains above 98.5% after the system has been running continuously for 12 months, solving the performance degradation problem of traditional models. Attached Figure Description

[0022] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used together with the embodiments of the invention to explain the invention and do not constitute a limitation thereof.

[0023] Figure 1 This is a flowchart illustrating the overall process of the condition monitoring method of the present invention. Figure 2 A comparison curve of the fault identification rate of each scheme at different speeds; Figure 3 A comparison curve of diagnostic time for different solutions under different loads; Figure 4 A comparison curve of the positioning accuracy of various solutions under different fault types; Figure 5 A power consumption comparison curve of each scheme after 24 hours of continuous operation; Figure 6 A comparison curve of false alarm rates for different schemes under different noise environments. Detailed Implementation

[0024] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0025] Example 1 S1, Data Acquisition To address the vibration characteristics of external rotor motors without a central shaft, an innovative three-dimensional layout of eight sensors along upper and lower dual guide rails is employed. Four sensors (0°, 45°, 90°, 135°) are evenly distributed along each of the upper and lower guide rails, forming a 360° ring-shaped monitoring array with no blind spots. By introducing a height factor (upper guide rail = +1, lower guide rail = -1), the axial vibration component is reconstructed using the differential signals between the sensors on the upper and lower guide rails, fundamentally solving the problem of missing axial monitoring dimensions caused by the lack of a central shaft structure.

[0026] S2, Adaptive Multi-Scale Filtering Algorithm This invention designs a speed-adaptive multi-scale filtering algorithm that dynamically adjusts the cutoff frequency of the FIR bandpass filter based on the real-time estimated speed, accurately matching the fault characteristic frequency band under the current operating conditions. Simultaneously, it integrates wavelet envelope analysis technology to extract weak fault signals from strong noise backgrounds through joint time-frequency domain demodulation. This scheme reduces the minimum identifiable signal-to-noise ratio for early faults from 10dB in traditional schemes to 3dB, while maintaining a detection rate of over 98% even under severe operating conditions of 3-5dB.

[0027] (1) Real-time speed estimation: in, This represents the original time-domain signal acquired by the i-th vibration sensor. This represents the Fast Fourier Transform operator. This indicates the frequency corresponding to the maximum amplitude of the spectrum. This represents the estimated fundamental frequency of the motor rotation (unit: Hz). This indicates the estimated motor speed (unit: revolutions per minute, RPM).

[0028] Speed ​​estimate The output is used to drive adaptive parameter adjustments in subsequent algorithms. If no obvious peak is found in the FFT (e.g., the device is stationary), then... Set the default value to 1500.

[0029] (2) Determine the filter frequency band based on the rotational speed: like : like : like : in, Indicates the lower cutoff frequency of the filter band. This indicates the upper limit cutoff frequency of the filter band.

[0030] (3) FIR filter design: in, Indicates For passband, a bandpass filter, The time series of samples is usually implicitly represented by the index number of the signal array, and there is no need to explicitly store the time value.

[0031] right Perform 4-level discrete wavelet packet decomposition (using the 'db4' wavelet basis) to obtain a set of sub-band coefficients. , where j represents the sub-band index.

[0032] Selecting high-frequency detail subband coefficients that include bearing fault characteristic frequencies .

[0033] (4) Wavelet envelope analysis (for bearing faults) Wavelet packet decomposition: in, This represents the wavelet packet coefficients of all approximations and detail subbands in the 4th layer. This represents the discrete wavelet packet decomposition operator. This represents the time-domain signal after adaptive bandpass filtering. This represents the selected Daubechies4 wavelet basis functions. This indicates the number of decomposition layers (producing a total of 2^4 = 16 sub-bands).

[0034] High-frequency detail subband selection and reconstruction: in, This represents the high-frequency detail subband coefficients selected from the fourth layer coefficients, corresponding to the bearing fault resonance frequency band. This indicates the wavelet packet reconstruction operator (which restores the coefficients to a time-domain signal of the same length as the original signal).

[0035] in, This represents the Hilbert transform operator, used to obtain analytic signals. This represents the envelope spectrum of the i-th signal.

[0036] FIR bandpass filtering adaptively adjusts the cutoff frequency based on the current rotational speed, effectively filtering out speed-related background noise and high-frequency interference. Wavelet envelope analysis is specifically designed for early-stage, subtle bearing faults. Through demodulation techniques, it extracts clear fault characteristic frequencies from the modulated signal, addressing the problem of bearing faults being masked by strong rotational speed components in the vibration signal.

[0037] S3, Spatial Vector Decomposition and Fusion (1) Vector composition calculation in, This represents the installation azimuth angle of the i-th sensor, with values ​​{0°, 45°, 90°, 135°, 180°, 225°, 270°, 315°}. This represents the calibration weighting coefficient for the i-th sensor, ranging from 0.8 to 1.2, used to compensate for sensitivity differences among sensor channels. This represents the height factor of the i-th sensor, with +1 for the upper guide rail sensor and -1 for the lower guide rail sensor. This represents the radial (X-direction), tangential (Y-direction), and reconstructed axial (Z-direction) vibration vectors obtained after fusion.

[0038] (2) Normalization in, This indicates taking the absolute value. Indicates taking The maximum absolute value within an analysis time window It is a very small constant to prevent division by zero. , Similarly, , , This represents the normalized triaxial vibration vector, with values ​​in the interval [-1, 1].

[0039] , , These features are directly used as the core input features for subsequent models, characterizing the motor's vibration state in the radial, tangential, and axial directions.

[0040] The signals from eight spatially dispersed sensors are processed through a sensor-calibrated weighting system. Height factor The vector synthesis formula is used to fuse the equivalent vibration vectors in three directions. Height factor The introduction of this technology is a key innovation – the upper guide rail takes a value of +1 and the lower guide rail takes a value of -1. By using the differential information between the upper and lower guide rails, the axial vibration component is reconstructed, which solves the problem of failure of traditional axial monitoring in motors without a central shaft.

[0041] This process compresses the data dimension from 8 dimensions to 3 dimensions (reducing the data volume by about 60%), while reconstructing the sensor point information into vector information that can characterize the overall three-dimensional vibration state of the motor.

[0042] S4. Fault Feature Extraction This invention designs a speed-adaptive multi-scale filtering algorithm that dynamically adjusts the cutoff frequency of the FIR bandpass filter based on the real-time estimated speed, accurately matching the fault characteristic frequency band under the current operating conditions. Simultaneously, it integrates wavelet envelope analysis technology to extract weak fault signals from strong noise backgrounds through joint time-frequency domain demodulation. This scheme reduces the minimum identifiable signal-to-noise ratio for early faults from 10dB in traditional schemes to 3dB, while maintaining a detection rate of over 98% even under severe operating conditions of 3-5dB.

[0043] (1) Calculate the principal direction angle in, Represents the arctangent function in the four quadrants. The principal vibration direction angle (unit: degrees) is calculated based on the fused vector.

[0044] (2) Location of maximum amplitude in, This indicates that the index that maximizes the value of the function within the parentheses is returned. This indicates the index of the sensor channel with the largest instantaneous amplitude among the 8 filtered signals. This indicates the azimuth angle of the sensor at its maximum amplitude.

[0045] (3) Calculate the confidence level in, Represents the confidence level in the principal direction, which is the modulus of the normalized radial vibration vector. This indicates the confidence level of the maximum amplitude, representing the proportion of the maximum single-channel signal energy to the total energy.

[0046] (4) Confidence-weighted fusion in, Indicates fusion and The azimuth angle of the subsequent refining fault was obtained.

[0047] Main direction angle The main source direction of vibration energy is calculated from the vibration vectors in the X and Y directions using the arctangent function. The location of the sensor with the maximum amplitude is also determined. It provides a direct physical location reference. By combining the two through confidence-weighted fusion, it leverages both the high reliability of physical sensor locations and the global perspective of vector computation.

[0048] Final Fault Location Outputting angle values ​​provides a clear positioning range. .when When the location result is reliable, the positioning result is reliable; when At that time, the vibration signal was too weak, and the positioning result was for reference only.

[0049] S5, Multi-parameter Fusion and Diagnostic Model This invention constructs an enhanced lightweight multi-parameter fusion diagnostic model (Lite-CNN), which integrates the normalized three-dimensional vibration vector, temperature, and rotational speed into a 5-dimensional feature vector. The implicit correlations between parameters are automatically learned through a deep separable convolutional network. A specially designed temperature compensation mechanism can eliminate the interference of temperature on vibration amplitude and simultaneously identify the causal chain between vibration anomalies and temperature increases, achieving accurate separation of complex faults.

[0050] (1) Temperature normalization and compensation in, This indicates the currently collected bearing temperature. and This represents the baseline and maximum values ​​for temperature normalization. This represents the normalized temperature value. , , This represents the vibration vector after temperature compensation.

[0051] (2) Rotation speed normalization in, This represents the rotational speed normalized to the [0,1] interval.

[0052] (3) Construct a 5-dimensional feature input vector in, This represents a five-dimensional column vector, which serves as the input to the diagnostic model.

[0053] (4) Lite-CNN model inference Convolutional layer 1: 8 1D convolutional kernels (size = 3), ReLU activation, BatchNorm1D: in, This indicates batch normalization, which standardizes the output of this layer (subtracting the mean and dividing by the standard deviation) to stabilize its distribution. This indicates a one-dimensional convolution performed on 5-dimensional input features, with a kernel size of 3 and 8 filters. This represents the linear rectification activation function.

[0054] Depthwise separable convolutional layer: 16 convolutional kernels in, This means that standard convolution is decomposed into two steps: channel-wise convolution (each input channel is spatially convolved using a separate kernel) and pointwise convolution (using 1x1 convolution to fuse channel information). Indicates the number of channels. This indicates the output of this layer.

[0055] Global average pooling in, This means averaging over all time steps within each channel (where each channel can be considered a spatial point), condensing each feature channel into a single scalar value. This indicates the output of this layer.

[0056] Fully connected layer 1: 32 neurons in, This represents a fully connected layer, which integrates 16-dimensional... The weighted summation is mapped to a 32-dimensional space to learn higher-order combinations and nonlinear relationships among these global features. This indicates a randomly deactivated layer, where, during training, the outputs of some neurons in this layer are randomly set to zero with a 30% probability. This indicates the output of this layer.

[0057] Fully connected layer 2: 16 neurons Output header 1: Fault type (5 categories, Softmax) in, Including: normal, unbalanced, bearing wear, eccentricity, overheating, This means converting the 5 Logits into a probability distribution.

[0058] Output Header 2: Severity Level (Sigmoid, output 0~1) in, Severity level="mild"; Severity level: "Medium"; Severity level="Severe".

[0059] Unlike simple threshold judgment, this module uses a lightweight convolutional neural network to automatically learn the complex mapping relationship between multidimensional features and fault types.

[0060] Temperature compensation mechanism: Increased temperature amplifies vibration amplitude. The temperature interference on vibration characteristics is eliminated by the compensation factor, thereby improving the robustness of the model under different working conditions.

[0061] Speed ​​normalization: The speed is unified to the [0,1] range, so that the model can adapt to diagnosis within a wide speed range (1500~3000rpm) with an error ≤1%.

[0062] Depthwise separable convolution: While ensuring feature extraction capabilities, the number of model parameters is compressed to <15KB, adapting to the storage and computing limitations of low-power MCUs.

[0063] Dual-output head design: Simultaneously outputs the probability distribution of fault types and the quantitative score of severity level, realizing "qualitative + quantitative" diagnosis.

[0064] (5) Fault type identification and alarm in, This indicates that the highest probability value is taken as the confidence level for this diagnosis. This indicates that the index corresponding to the highest probability value is used to obtain the final fault type. :Fault type="normal"; :Fault type="unbalance; Fault type="bearing wear"; :Fault type="eccentric"; Fault type="overheating".

[0065] The core of the judgment mechanism is Softmax probability maximization. The category with the highest probability in the five-element probability vector output by the model is the final fault type. An alarm is only triggered when the highest probability exceeds the 95% threshold. This mechanism effectively filters out ambiguous signals and keeps the false alarm rate below 0.5%.

[0066] (6) The complete diagnostic output format is shown in Table 1: Table 1 Diagnostic Report Form The diagnostic report presents complete information across four dimensions: "what the fault is, how serious it is, where it is located, and how certain it is," providing a directly actionable basis for predictive maintenance decisions.

[0067] Example 2 In this embodiment, the method of the present invention is compared with other traditional technologies in various aspects.

[0068] (1) A comparison of the method of the present invention with traditional techniques in multiple dimensions is shown in Table 2: Table 2 Dimensional Comparison Table The table shows the following differences between the present invention and traditional technologies: This invention employs a dual-plane, 8-sensor, 360° ring layout to achieve comprehensive three-dimensional monitoring in radial, tangential, and axial directions, with a 0° blind zone, while traditional single-plane layouts have blind zones of ≥180°. The fault identification rate of this invention reaches 99.2%, and the early fault detection rate is 98%, far exceeding the 72%-85% and 45%-65% of traditional technologies, respectively. This invention's accuracy is down to ±10°, while traditional technologies cannot locate faults or only achieve ±45°. The rotational speed error of this invention is ≤1%, superior to the traditional 5%-8%. The effective signal-to-noise ratio threshold of this invention can be as low as >3dB, while traditional methods require >10dB. The system power consumption of this invention is only 4.8W, and the model parameter size is <15KB, suitable for edge deployment; traditional solutions consume 25-30W and have large models. The accuracy of this invention remains above 98.5% after 12 months of continuous operation, while traditional methods experience an annual attenuation of 3-5%. This invention provides a four-dimensional diagnostic report, rather than the traditional black-box binary output.

[0069] (2) The present invention is now compared with traditional solutions A and B. The scheme settings of the present invention and traditional solutions A and B are shown in Table 3: Table 3 Comparison Scheme Settings Table from Figures 2-6 As can be seen from the data, compared with schemes A and B, the present invention has the highest fault identification rate under different speeds, the least time for diagnosis under different loads, the highest positioning accuracy for different fault types, the lowest power consumption in the system power consumption comparison, and the lowest false alarm rate in the comparison of false alarm rates under different noise environments.

[0070] The differences in performance indicators between the present invention and traditional solutions A and B are shown in Table 4: Table 4 Performance Comparison of Different Solutions As can be seen from the table, compared with traditional scheme A and traditional scheme B, the present invention has advantages in fault identification rate, false alarm rate, positioning accuracy, diagnostic time, system power consumption, wide speed adaptability, and noise robustness. It can improve the fault identification rate and reduce the false alarm rate while maintaining low power consumption, and also improve positioning accuracy and wide speed adaptability.

[0071] Example 3 A condition monitoring system for a high-pressure blower-type shaftless external rotor motor mainly includes: A multi-channel circumferential monitoring array includes four or more vibration sensors arranged on the upper and lower circumferential support rails of the motor to form a multi-channel circumferential monitoring array. Each sensor is configured with corresponding installation angle parameters and height factors, wherein the height factors of the upper rail sensor and the lower rail sensor have opposite signs. The signal acquisition unit is electrically connected to the sensor array and is used to synchronously acquire multi-channel raw vibration signals; The spatial vector fusion unit is used to perform spatial vector decomposition and fusion on the preprocessed vibration signal to reconstruct the three-dimensional equivalent vibration vector. A multi-parameter fusion unit is used to combine three-dimensional vibration vectors, temperature and rotation speed signals to construct a multi-parameter fusion feature vector. The intelligent diagnostic unit is used to input multi-parameter feature vectors and output multi-dimensional diagnostic results.

[0072] The system also includes an online incremental learning unit. This module integrates an online incremental learning mechanism based on knowledge distillation, automatically triggering model fine-tuning every 1000 samples. Using the old model as the teacher and the new model as the student, it adapts to new data features while preserving existing knowledge through KL divergence constraints. A version management mechanism ensures that replacement is only performed when the new model's performance improves, avoiding the performance degradation risks that online learning might introduce. Real-world testing shows that after 12 months of continuous operation, the system's diagnostic accuracy remains above 98.5%, effectively solving the long-term performance degradation problem of traditional fixed models.

[0073] Knowledge distillation fine-tuning and version management in, This represents the teacher model, i.e., the optimal model currently deployed. Represents the student model, and the new model copy to be trained. This represents the cross-entropy loss, which enables the student model to predict... Approaching the true label Y, This represents the mean squared error loss, which makes the student model output approximate the teacher model output. ("soft label") The balance coefficient, 0 < <1.

[0074] Version management logic: when When, use replace Otherwise, keep .

[0075] This module enables the algorithm to continuously evolve. The motor's operating environment changes slowly due to factors such as aging and seasonal temperature differences, and the performance of offline-trained models will degrade after long-term use.

[0076] Instead of outputting a single result, the decision to update is based on a comparison of model performance. If the old model is retained and logs are recorded, operations and maintenance personnel can use this information to analyze the reasons for performance degradation.

[0077] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some or all of the technical features therein. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method for monitoring the condition of a high-pressure blower-type shaftless external rotor motor, characterized in that, Includes the following steps: S1. Arrange four or more vibration sensors on the upper and lower circumferential support rails of the motor to form a multi-channel circumferential monitoring array. Configure corresponding installation angle parameters and height factors for each sensor, wherein the height factors of the upper rail sensor and the lower rail sensor have opposite signs. S2. Synchronously acquire and preprocess the raw vibration signals of all channels; S3. Based on the preprocessed vibration signals of each channel, the corresponding installation angle, the sensor calibration weight coefficient and the height factor, spatial vector decomposition and fusion are performed. The three-dimensional equivalent vibration vectors in the radial, tangential and axial directions, which characterize the overall vibration state of the motor, are reconstructed using the differential signals of the upper and lower guide rail sensors. S4. Combine the three-dimensional equivalent vibration vector, the real-time temperature signal of the motor, and the real-time speed signal to construct a multi-parameter fusion feature vector; S5. Input the multi-parameter fused feature vector into the intelligent diagnostic model, and output a multi-dimensional diagnostic result including fault type, spatial location, severity level, and judgment confidence.

2. The condition monitoring method for a high-pressure blower-type shaftless external rotor motor according to claim 1, characterized in that, The preprocessing in step S2 includes adaptive multi-scale filtering: the fundamental frequency is extracted from the FFT result of the multi-channel synthesized signal and converted into real-time rotational speed; the cutoff frequency of the FIR bandpass filter is dynamically adjusted according to the real-time rotational speed to filter the vibration signal; and early bearing fault characteristics are extracted through wavelet envelope analysis.

3. The condition monitoring method for a high-pressure blower-type external rotor motor without a central shaft according to claim 1, characterized in that, The calculation method for spatial vector decomposition and fusion in step S3 is as follows: in, This represents the installation azimuth angle of the i-th sensor. This represents the calibration weighting coefficient for the i-th sensor. This represents the height factor of the i-th sensor. This represents the radial, tangential, and reconstructed axial vibration vectors obtained after fusion.

4. The condition monitoring method for a high-pressure blower-type external rotor motor without a central shaft according to claim 1, characterized in that, In step S4, when constructing the multi-parameter fusion feature vector, the temperature signal is first normalized and the three-dimensional vibration vector is temperature compensated. Then, the rotation speed signal is normalized to the [0,1] interval. Finally, the vector is combined to form a 5-dimensional feature vector containing the compensated vibration vector, normalized temperature, and normalized rotation speed.

5. A method for monitoring the condition of a high-pressure blower-type external rotor motor without a central shaft, as described in claim 1 or 2, characterized in that, The intelligent diagnostic model is an enhanced lightweight multi-parameter fusion diagnostic model, which adopts a dual-output head structure. The first output head outputs the probability distribution of fault types through a multi-classification problem activation function, and the second output head outputs a fault severity level quantification value of 0 to 1 through a regression task activation function.

6. The condition monitoring method for a high-pressure blower-type shaftless external rotor motor according to claim 1, characterized in that, In step S5, the spatial location of the fault is determined by confidence-weighted fusion: the principal direction angle of vibration and the location of the sensor with the largest amplitude are calculated, and after being assigned corresponding confidence levels, they are fused to obtain the precise fault location.

7. The condition monitoring method for a high-pressure blower-type external rotor motor without a central shaft according to claim 1, characterized in that, Step S5 also includes an alarm triggering step: when the confidence level of the fault type determination is greater than a preset threshold, an alarm is triggered, and the fault severity level is divided into three maintenance priorities: minor, moderate, and severe.

8. A condition monitoring system for a high-pressure blower-type shaftless external rotor motor, characterized in that, include: A multi-channel circumferential monitoring array includes four or more vibration sensors arranged on the upper and lower circumferential support rails of the motor to form a multi-channel circumferential monitoring array. Each sensor is configured with corresponding installation angle parameters and height factors, wherein the height factors of the upper rail sensor and the lower rail sensor have opposite signs. The signal acquisition unit is electrically connected to the sensor array and is used to synchronously acquire multi-channel raw vibration signals; The spatial vector fusion unit is used to perform spatial vector decomposition and fusion on the preprocessed vibration signal to reconstruct the three-dimensional equivalent vibration vector. A multi-parameter fusion unit is used to combine three-dimensional vibration vectors, temperature and rotation speed signals to construct a multi-parameter fusion feature vector. The intelligent diagnostic unit is used to input multi-parameter feature vectors and output multi-dimensional diagnostic results.

9. A condition monitoring system for a high-pressure blower-type shaftless external rotor motor according to claim 8, characterized in that, More than four vibration sensors are arranged on the upper and lower circumferential support rails, all evenly distributed along the circumference. The sensors on the lower rail and the upper rail are symmetrically distributed at 180° in the circumferential direction, forming a 360° monitoring coverage without blind spots.

10. A condition monitoring system for a high-pressure blower-type shaftless external rotor motor according to claim 8, characterized in that, It also includes an online incremental learning unit, which is used to fine-tune the intelligent diagnostic model by knowledge distillation after collecting a preset number of samples, and replaces it only when the performance of the new model is improved through a version management mechanism.

Images