Intelligent detection method for comprehensive quality of wheel assembly based on vibration analysis

Abstract

The invention discloses a wheel assembly comprehensive quality intelligent detection method based on vibration analysis, and the method comprises the steps: arranging a high-precision vibration sensor and a photoelectric encoder near a hub and a bearing seat, and synchronously collecting multi-channel vibration signals and rotating speed information; de-noising is carried out by adopting wavelet transform or EMD, equal-angle resampling is realized based on order tracking, and the problem of frequency spectrum fuzziness under the variable speed working condition is eliminated; aiming at quality dimensions such as dynamic balance, bearing state, bolt looseness and tire uniformity, extracting targeted characteristics such as 1 * / 2 * rotation frequency amplitude, fault characteristic frequency envelope spectrum, kurtosis and impact pulse factor; training a multi-task intelligent diagnosis model in combination with the labeled data set, synchronously outputting defect types, severity and 0-100 score comprehensive quality scores, and introducing a confidence coefficient mechanism to trigger manual reinspection; the linkage production line realizes automatic rejection of unqualified products, forms a'detection-judgment-disposal 'closed loop, and remarkably improves the detection efficiency and the quality control level.

Description

Technical Field

[0001] This invention relates to the field of wheel assembly quality inspection technology, and more specifically, to an intelligent method for comprehensive quality inspection of wheel assemblies based on vibration analysis. Background Technology

[0002] Traditional dynamic balancing machines only detect imbalance and cannot identify bearing damage, loose bolts, or tire uniformity issues. Bearing inspection often relies on offline disassembly or listening rods, which are inefficient and subjective. They lack the ability to simultaneously and comprehensively identify multiple defect types. Offline inspection disrupts production cycles, requiring most quality inspections to be conducted at dedicated workstations, making it impossible to integrate into high-speed assembly lines. Defective products are detected late, resulting in high rework costs and even the products ending up in the next process or market. Under variable speed conditions, signal distortion is severe. During wheel start-stop or speed changes, traditional FFT spectra exhibit "spectral smear" due to speed fluctuations, making it difficult to accurately extract fault characteristic frequencies. The lack of effective equal-angle resampling (order analysis) methods leads to a high false alarm rate. High noise interference and poor feature extraction robustness are also issues. Environmental vibrations, electromagnetic interference, and installation errors introduce a large amount of unrelated noise. Traditional filtering methods (such as bandpass filtering) easily filter out weak early fault impact signals. There is a lack of adaptive denoising and multi-scale feature fusion mechanisms. The lack of a structured feature system oriented towards the "quality dimension" means that existing methods mostly use general statistical features (such as RMS and kurtosis), which cannot distinguish between defects with different mechanisms, such as "dynamic imbalance" and "bearing spalling." The upgrade to intelligent manufacturing places new demands on inspection systems, requiring a closed-loop capability of "inspection-judgment-handling." Industry 4.0 requires quality inspection systems to automatically trigger rejection, rework, or alarms, rather than simply providing data. Currently, most systems remain at the "data acquisition + manual judgment" stage, failing to achieve a truly intelligent closed loop. This invention achieves a technological leap from "passive inspection" to "proactive quality control" through high-precision synchronous acquisition, order-domain signal processing, mechanism-oriented feature engineering, intelligent multi-task diagnostics, and production line-level closed-loop linkage. Summary of the Invention

[0003] A smart detection method for the overall quality of a wheel assembly based on vibration analysis includes the following steps; S1. Detection system setup and sensor placement: Install high-precision vibration sensors (such as accelerometers) at the detection station near the wheel hub or bearing housing of the wheel assembly. Use speed sensors (such as photoelectric encoders) to synchronously acquire the rotation frequency. Build a data acquisition system (DAQ) to ensure that the sampling frequency meets the Nyquist sampling theorem (usually ≥10 times the highest frequency of interest). S2. Wheel assembly excitation and data acquisition: The wheel assembly is rotated within a set speed range (e.g., 300–3000 rpm) by a drive device (e.g., a motor). Multi-channel vibration signals (time domain data) are acquired in real time during steady state or speed change, and operating parameters such as speed, temperature, and load are recorded simultaneously. S3. Signal preprocessing: Wavelet transform, EMD (Empirical Mode Decomposition) or filters (such as bandpass / lowpass) are used to remove environmental noise, convert the time domain signal into equal angle sampling (order analysis) to eliminate the influence of speed fluctuations, and segment the continuous signal according to the period or time window to facilitate subsequent feature extraction; S4. Feature Extraction: Feature extraction oriented towards the quality dimension, including: Dynamic balancing / eccentricity: Extracting 1× and 2× frequency amplitudes and order spectral energy; Bearing condition: Calculate the characteristic frequency amplitude and envelope spectrum peak value of faults such as BPFO / BPFI; Loose bolts: Calculate kurtosis, impact pulse factor, and low-frequency modulation sideband; Tire uniformity: Analysis of vibration components related to first- and second-order harmonic forces; Simultaneously extract general features of energy entropy in the time domain (RMS, peak-to-peak value) and frequency domain; S5. Intelligent Diagnosis and Quality Assessment: Based on the labeled dataset (including labels such as "qualified", "imbalance", "bearing damage", "loose", etc.), a training dataset is constructed to train a multi-classification / multi-output model (such as 1D-CNN+Attention or random forest). The model synchronously outputs: defect type (such as "mild dynamic imbalance" "early wear of bearing outer ring"), comprehensive quality score (0-100 points), and introduces a confidence threshold to trigger manual re-inspection for uncertain samples; S6. Results Visualization and Feedback: Displays inspection conclusions, defect locations, and recommendations on the HMI interface; automatically generates quality inspection reports; automatically marks non-conforming products and links them to the production line for rejection or rework, achieving a closed loop of "inspection-judgment-disposal".

[0004] Furthermore, an intelligent detection method for the overall quality of wheel assemblies based on vibration analysis is proposed. In step S1, a speed sensor is used: an optical encoder is used to synchronously acquire the rotation frequency, and the specific synchronization strategy is as follows; Synchronization mechanism design: Strict time alignment between vibration and rotation speed signals Hardware synchronization: The same data acquisition card is used to simultaneously acquire vibration (analog input) and encoder (digital counting / quadrature decoding) signals. The acquisition system supports hardware triggering: the encoder Z-phase signal serves as the start marker for each revolution. Software alignment: Equal-angle resampling based on encoder pulses: Record the moment of each encoder pulse and use interpolation: splines and sinc to convert the vibration signal from the time domain to the angular domain: one sampling point every 0.5° and 1°; eliminate the influence of speed fluctuations on spectrum analysis; Timestamp alignment: All channels: Vibration + rotation speed share high-precision system clock: PXI platform, synchronization error <1 μs.

[0005] Furthermore, an intelligent detection method for the overall quality of wheel assemblies based on vibration analysis is proposed. In step S2, multi-channel vibration signals are acquired in real time during steady-state and variable-speed processes, as detailed below; Speed ​​range setting during variable speed data acquisition: Lower limit: 500 rpm: higher than the minimum stable speed of the drive system, to ensure sufficient encoder pulses (e.g., 500 rpm = 8.3 Hz, 500 PPR → 4167 pulses per second, to meet sampling requirements), and to avoid excessively low signal-to-noise ratio (small vibration amplitude) at low speeds. Upper limit: 3000 rpm: close to the common high-speed operating conditions of vehicles (such as 120 km / h corresponding to about 2500–3000 rpm, depending on tire size), covering the frequency band where the bearing failure characteristic frequency is located (the higher the speed, the higher the BPFO / BPFI frequency, and the easier it is to separate from noise), and not exceeding the safe limit speed of the wheel assembly (usually ≥5000 rpm, with a margin). Typical settings: 500–3000 rpm, covering the main operating range from urban roads to highways; And equivalent simulation of key dynamic loads: Rotational inertial load: Simulates acceleration / cruising / deceleration by setting a rotational speed curve; Centrifugal force effect: naturally generated during high-speed rotation, affecting dynamic balance performance; Thermal effect: Monitoring the impact of temperature changes on bearing clearance after long-term operation.

[0006] Furthermore, an intelligent detection method for the overall quality of wheel assemblies based on vibration analysis is proposed. In step S3, the time-domain signal is converted into an equal-angle sample to eliminate the influence of speed fluctuations, and the continuous signal is segmented according to the period and time window, as follows; Equal-angle sampling effectively eliminates the influence of rotational speed fluctuations by using high-precision encoder timestamps and spline / VKF interpolation. Combined with an intelligent segmentation strategy based on rotation period, window length and overlap rate can be configured as needed to improve the repeatability, robustness, and diagnostic sensitivity of subsequent features such as order amplitude and impact index.

[0007] Furthermore, an intelligent detection method for the overall quality of wheel assemblies based on vibration analysis is proposed. In step S5, a training dataset is constructed based on the labeled dataset: qualified, unbalanced, bearing damaged, and loose. The specific steps for constructing the dataset are as follows: When constructing the dataset for wheel assembly vibration testing, the annotation was based on authoritative standards and combined with engineering practice. Dynamic balance defects are classified into acceptable, minor, and severe levels according to the amount of imbalance, in accordance with ISO 1940 standard. The condition of bearings is classified into normal, early wear, moderate damage and severe failure according to ISO 15243 or the company's internal control standards, and is confirmed by disassembly and re-inspection. Bolt loosening is defined based on the preload loss ratio and marked by low-frequency modulation and impact characteristics. Tire uniformity is indirectly assessed by the amplitude of first- and second-order harmonic vibrations, referring to relevant ASTM or SAE specifications. To improve the model's generalization ability, a variety of data augmentation strategies are adopted: controllable Gaussian noise and power frequency interference are added to the original signal to simulate the field environment; By applying virtual rotational speed perturbation to the angular domain signal, equivalent samples at different rotational speeds are generated. Signal interpolation is performed between samples with similar defects to expand the rare fault categories. Time offset and amplitude scaling operations are also used to enhance the model's robustness to sensor differences and phase changes. All enhancements are only used during the training phase to ensure that the test results are true and reliable.

[0008] Furthermore, an intelligent detection method for the overall quality of wheel assemblies based on vibration analysis is proposed. In step S5, a multi-class / multi-output model is trained. The model outputs the following in sync: defect type: mild dynamic imbalance, early wear of bearing outer ring. The model output is combined to give a comprehensive quality score of 0-100 points. A confidence threshold is introduced to trigger manual re-inspection for uncertain samples. The specific steps are as follows. S51. Construct a structured labeled dataset: collect vibration signals from a large number of wheel assemblies under the S1–S4 process, and have experts and destructive testing confirm their true state; Label definition: Each sample corresponds to the following multi-dimensional labels: Defect type labels: normal, mild dynamic imbalance, moderate dynamic imbalance, early wear of bearing outer ring, damage of bearing inner ring, loose bolts, uneven tires; Overall quality rating labels (continuous values ​​from 0 to 100): Experts score the defects based on their type, severity, and safety margin: Normal = 95–100, Slight imbalance = 80–89, Early bearing wear = 70–79, Severe looseness = 30–50. S52. Design a multi-output model architecture: adopt a multi-task learning framework, share the underlying feature extraction network, and branch outputs results for different tasks; Loss function design: Total loss = α × Classification loss + β × Regression loss; The classification loss is cross-entropy, the regression loss is mean squared error and smoothing L1 loss, and α and β are hyperparameters: α=1.0, β=0.5, tuned using the validation set; S53. Model Training and Validation: Train the model using the training set, validate the model's performance using the validation set, and test the model's accuracy using the test set; S54. Inference Phase – Synchronous Output of Defect Type and Quality Score: Perform the following inference process on new samples: Input preprocessed features → Model forward propagation, obtaining two outputs: Defect type probability vector: p=[p1,p2,...,pK], where ∑pi=1; Quality score: s ∈[0,100], determine the main defect type: select the class with the highest probability: y =argmax(p), example output: "slight dynamic imbalance" (corresponding to pimbalance_light = 0.92); S55. Introduce a confidence threshold to trigger manual review: Confidence definition: Take the maximum predicted probability value conf=max(p), and set a threshold: τ (e.g., τ=0.85): When conf≥τ: Accept the model's judgment and automatically output the result; When conf < τ: it is determined to be an "uncertain sample" and a manual re-inspection process is triggered; A re-inspection linkage mechanism was designed, which highlights the "re-inspection required" status on the HMI interface, automatically saves the original vibration data, feature maps, and intermediate model outputs for engineers to analyze, pauses the automatic rejection action, and waits for manual confirmation before further processing.

[0009] The beneficial effects of this invention are as follows: By integrating multi-dimensional features such as dynamic balance, bearing condition, bolt connection, and tire uniformity, it can simultaneously diagnose multiple typical quality defects (such as imbalance, bearing outer ring wear, and loose bolts) in one go, avoiding the drawbacks of traditional methods that require multiple sets of equipment and multiple tests. Specialized features (such as order spectrum energy, envelope spectrum peak value, and impact pulse factor) are designed for different fault mechanisms, significantly improving the defect identification rate and classification accuracy, with a measured accuracy of over 95%. The introduction of equal-angle resampling (order tracking) technology converts the time-domain vibration signal into an order-domain signal synchronized with the rotation angle, completely eliminating spectral ambiguity caused by speed fluctuations. Even under non-steady-state conditions such as start-stop, acceleration, and deceleration, it can still stably extract 1× and 2× rotational frequencies and bearing fault characteristic frequencies, greatly improving the adaptability and robustness of online detection. It not only outputs the defect type, but also generates a comprehensive quality score of 0-100, which intuitively reflects the overall health status of the wheel assembly. The quality score is based on multi-dimensional feature weighted fusion and supports flexible configuration of weights according to OEM standards to meet the customized needs of different customers. The model output includes a confidence index, which automatically triggers manual re-inspection for low-confidence samples, balancing automation efficiency and judgment reliability. Attached Figure Description

[0010] Figure 1 This is a flowchart of an intelligent detection method for the overall quality of wheel assemblies based on vibration analysis; Detailed Implementation

[0011] A smart detection method for the overall quality of a wheel assembly based on vibration analysis includes the following steps; S1. Detection system setup and sensor placement: Install high-precision vibration sensors (such as accelerometers) at the detection station near the wheel hub or bearing housing of the wheel assembly. Use speed sensors (such as photoelectric encoders) to synchronously acquire the rotation frequency. Build a data acquisition system (DAQ) to ensure that the sampling frequency meets the Nyquist sampling theorem (usually ≥10 times the highest frequency of interest). S2. Wheel assembly excitation and data acquisition: The wheel assembly is rotated within a set speed range (e.g., 300–3000 rpm) by a drive device (e.g., a motor). Multi-channel vibration signals (time domain data) are acquired in real time during steady state or speed change, and operating parameters such as speed, temperature, and load are recorded simultaneously. S3. Signal preprocessing: Wavelet transform, EMD (Empirical Mode Decomposition) or filters (such as bandpass / lowpass) are used to remove environmental noise, convert the time domain signal into equal angle sampling (order analysis) to eliminate the influence of speed fluctuations, and segment the continuous signal according to the period or time window to facilitate subsequent feature extraction; S4. Feature Extraction: Feature extraction oriented towards the quality dimension, including: Dynamic balancing / eccentricity: Extracting 1× and 2× frequency amplitudes and order spectral energy; Bearing condition: Calculate the characteristic frequency amplitude and envelope spectrum peak value of faults such as BPFO / BPFI; Loose bolts: Calculate kurtosis, impact pulse factor, and low-frequency modulation sideband; Tire uniformity: Analysis of vibration components related to first- and second-order harmonic forces; Simultaneously extract general features of energy entropy in the time domain (RMS, peak-to-peak value) and frequency domain; S5. Intelligent Diagnosis and Quality Assessment: Based on the labeled dataset (including labels such as "qualified", "imbalance", "bearing damage", "loose", etc.), a training dataset is constructed to train a multi-classification / multi-output model (such as 1D-CNN+Attention or random forest). The model synchronously outputs: defect type (such as "mild dynamic imbalance" "early wear of bearing outer ring"), comprehensive quality score (0-100 points), and introduces a confidence threshold to trigger manual re-inspection for uncertain samples; S6. Results Visualization and Feedback: Displays inspection conclusions, defect locations, and recommendations on the HMI interface; automatically generates quality inspection reports; automatically marks non-conforming products and links them to the production line for rejection or rework, achieving a closed loop of "inspection-judgment-disposal".

[0012] Furthermore, an intelligent detection method for the overall quality of wheel assemblies based on vibration analysis is proposed. In step S1, a speed sensor is used: an optical encoder is used to synchronously acquire the rotation frequency, and the specific synchronization strategy is as follows; Synchronization mechanism design: Strict time alignment between vibration and rotation speed signals Hardware synchronization: The same data acquisition card is used to simultaneously acquire vibration (analog input) and encoder (digital counting / quadrature decoding) signals. The acquisition system supports hardware triggering: the encoder Z-phase signal serves as the start marker for each revolution. Software alignment: Equal-angle resampling based on encoder pulses: Record the moment of each encoder pulse and use interpolation: splines and sinc to convert the vibration signal from the time domain to the angular domain: one sampling point every 0.5° and 1°; eliminate the influence of speed fluctuations on spectrum analysis; Timestamp alignment: All channels: Vibration + rotation speed share high-precision system clock: PXI platform, synchronization error <1 μs.

[0013] Furthermore, an intelligent detection method for the overall quality of wheel assemblies based on vibration analysis is proposed. In step S2, multi-channel vibration signals are acquired in real time during steady-state and variable-speed processes, as detailed below; Speed ​​range setting during variable speed data acquisition: Lower limit: 500 rpm: higher than the minimum stable speed of the drive system, to ensure sufficient encoder pulses (e.g., 500 rpm = 8.3 Hz, 500 PPR → 4167 pulses per second, to meet sampling requirements), and to avoid excessively low signal-to-noise ratio (small vibration amplitude) at low speeds. Upper limit: 3000 rpm: close to the common high-speed operating conditions of vehicles (such as 120 km / h corresponding to about 2500–3000 rpm, depending on tire size), covering the frequency band where the bearing failure characteristic frequency is located (the higher the speed, the higher the BPFO / BPFI frequency, and the easier it is to separate from noise), and not exceeding the safe limit speed of the wheel assembly (usually ≥5000 rpm, with a margin). Typical settings: 500–3000 rpm, covering the main operating range from urban roads to highways; And equivalent simulation of key dynamic loads: Rotational inertial load: Simulates acceleration / cruising / deceleration by setting a rotational speed curve; Centrifugal force effect: naturally generated during high-speed rotation, affecting dynamic balance performance; Thermal effect: Monitoring the impact of temperature changes on bearing clearance after long-term operation.

[0014] Furthermore, an intelligent detection method for the overall quality of wheel assemblies based on vibration analysis is proposed. In step S3, the time-domain signal is converted into an equal-angle sample to eliminate the influence of speed fluctuations, and the continuous signal is segmented according to the period and time window, as follows; Equal-angle sampling effectively eliminates the influence of rotational speed fluctuations by using high-precision encoder timestamps and spline / VKF interpolation. Combined with an intelligent segmentation strategy based on rotation period, window length and overlap rate can be configured as needed to improve the repeatability, robustness, and diagnostic sensitivity of subsequent features such as order amplitude and impact index.

[0015] Furthermore, an intelligent detection method for the overall quality of wheel assemblies based on vibration analysis is proposed. In step S5, a training dataset is constructed based on the labeled dataset: qualified, unbalanced, bearing damaged, and loose. The specific steps for constructing the dataset are as follows: When constructing the dataset for wheel assembly vibration testing, the annotation was based on authoritative standards and combined with engineering practice. Dynamic balance defects are classified into acceptable, minor, and severe levels according to the amount of imbalance, in accordance with ISO 1940 standard. The condition of bearings is classified into normal, early wear, moderate damage and severe failure according to ISO 15243 or the company's internal control standards, and is confirmed by disassembly and re-inspection. Bolt loosening is defined based on the preload loss ratio and marked by low-frequency modulation and impact characteristics. Tire uniformity is indirectly assessed by the amplitude of first- and second-order harmonic vibrations, referring to relevant ASTM or SAE specifications. To improve the model's generalization ability, a variety of data augmentation strategies are adopted: controllable Gaussian noise and power frequency interference are added to the original signal to simulate the field environment; By applying virtual rotational speed perturbation to the angular domain signal, equivalent samples at different rotational speeds are generated. Signal interpolation is performed between samples with similar defects to expand the rare fault categories. Time offset and amplitude scaling operations are also used to enhance the model's robustness to sensor differences and phase changes. All enhancements are only used during the training phase to ensure that the test results are true and reliable.

[0016] Furthermore, an intelligent detection method for the overall quality of wheel assemblies based on vibration analysis is proposed. In step S5, a multi-class / multi-output model is trained. The model outputs the following in sync: defect type: mild dynamic imbalance, early wear of bearing outer ring. The model output is combined to give a comprehensive quality score of 0-100 points. A confidence threshold is introduced to trigger manual re-inspection for uncertain samples. The specific steps are as follows. S51. Construct a structured labeled dataset: collect vibration signals from a large number of wheel assemblies under the S1–S4 process, and have experts and destructive testing confirm their true state; Label definition: Each sample corresponds to the following multi-dimensional labels: Defect type labels: normal, mild dynamic imbalance, moderate dynamic imbalance, early wear of bearing outer ring, damage of bearing inner ring, loose bolts, uneven tires; Overall quality rating labels (continuous values ​​from 0 to 100): Experts score the defects based on their type, severity, and safety margin: Normal = 95–100, Slight imbalance = 80–89, Early bearing wear = 70–79, Severe looseness = 30–50. S52. Design a multi-output model architecture: adopt a multi-task learning framework, share the underlying feature extraction network, and branch outputs results for different tasks; Loss function design: Total loss = α × Classification loss + β × Regression loss; The classification loss is cross-entropy, the regression loss is mean squared error and smoothing L1 loss, and α and β are hyperparameters: α=1.0, β=0.5, tuned using the validation set; S53. Model Training and Validation: Train the model using the training set, validate the model's performance using the validation set, and test the model's accuracy using the test set; S54. Inference Phase – Synchronous Output of Defect Type and Quality Score: Perform the following inference process on new samples: Input preprocessed features → Model forward propagation, obtaining two outputs: Defect type probability vector: p=[p1,p2,...,pK], where ∑pi=1; Quality score: s ∈[0,100], determine the main defect type: select the class with the highest probability: y =argmax(p), example output: "slight dynamic imbalance" (corresponding to pimbalance_light = 0.92); S55. Introduce a confidence threshold to trigger manual review: Confidence definition: Take the maximum predicted probability value conf=max(p), and set a threshold: τ (e.g., τ=0.85): When conf≥τ: Accept the model's judgment and automatically output the result; When conf <τ：判定为"不确定样本”，触发人工复检流程； A re-inspection linkage mechanism was designed, which highlights the "re-inspection required" status on the HMI interface, automatically saves the original vibration data, feature maps, and intermediate model outputs for engineers to analyze, pauses the automatic rejection action, and waits for manual confirmation before further processing.< / τ：判定为"不确定样本”，触发人工复检流程；

Claims

1. A smart detection method for the overall quality of a wheel assembly based on vibration analysis, characterized in that, Includes the following steps; S1. Detection system setup and sensor placement: Install high-precision vibration sensors near the wheel hub and bearing housing of the wheel assembly at the detection station. Use speed sensors and photoelectric encoders to synchronously acquire the rotation frequency. Build a data acquisition system so that the sampling frequency satisfies the Nyquist sampling theorem: usually ≥10 times the highest frequency of interest. S2. Wheel assembly excitation and data acquisition: The wheel assembly is rotated within a set speed range by the drive device. Multi-channel vibration signals are collected in real time during steady state and speed change, and speed, temperature and load parameters are recorded simultaneously. S3. Signal preprocessing: Wavelet transform and EMD are used to remove environmental noise, the time domain signal is converted into equal-angle sampling to eliminate the influence of speed fluctuation, and the continuous signal is segmented by period and time window to facilitate subsequent feature extraction; S4. Feature Extraction: Feature extraction oriented towards the quality dimension, including: Dynamic balancing / eccentricity: Extracting 1× and 2× frequency amplitudes and order spectral energy; Bearing condition: Calculate the characteristic frequency amplitude and envelope spectrum peak value of faults such as BPFO / BPFI; Loose bolts: Calculate kurtosis, impact pulse factor, and low-frequency modulation sideband; Tire uniformity: Analysis of vibration components related to first- and second-order harmonic forces; Simultaneously extract time-domain features: RMS, peak-to-peak value, and frequency-domain energy entropy; S5. Intelligent Diagnosis and Quality Assessment: Based on the labeled dataset: qualified, unbalanced, bearing damaged, loose, a training dataset is constructed to train a multi-classification / multi-output model. The model synchronously outputs: defect type: mild dynamic imbalance, early wear of bearing outer ring. Combined with the model output, a comprehensive quality score of 0-100 is given, and a confidence threshold is introduced to trigger manual re-inspection for uncertain samples. S6. Results Visualization and Feedback: Displays inspection conclusions, defect locations, and recommendations on the HMI interface; automatically generates quality inspection reports; automatically marks non-conforming products and links them to the production line for rejection or rework, achieving a closed loop of "inspection-judgment-disposal".

2. The intelligent detection method for the comprehensive quality of a wheel assembly based on vibration analysis as described in claim 1, characterized in that, In step S1, a speed sensor is used: an optical encoder is used to synchronously acquire the rotation frequency, and the specific synchronization strategy is as follows; Synchronization mechanism design: Strict time alignment between vibration and rotation speed signals Hardware synchronization: The same data acquisition card is used to simultaneously acquire vibration and encoder signals. The acquisition system supports hardware triggering: the encoder Z-phase signal is used as the start mark for each revolution. Software alignment: Equal-angle resampling based on encoder pulses: Record the moment of each encoder pulse and use interpolation: splines and sinc to convert the vibration signal from the time domain to the angular domain: one sampling point every 0.5° and 1°; eliminate the influence of speed fluctuations on spectrum analysis; Timestamp alignment: All channels: Vibration + rotation speed share high-precision system clock: PXI platform, synchronization error < 1 μs.

3. The intelligent detection method for the comprehensive quality of a wheel assembly based on vibration analysis as described in claim 1, characterized in that, In step S2, multi-channel vibration signals are acquired in real time during steady-state and variable-speed processes, as detailed below; Speed ​​range setting during the speed change test data acquisition process: Lower limit: 500 rpm: higher than the minimum stable speed of the drive system, to ensure sufficient encoder pulses to meet sampling requirements and avoid excessively low signal-to-noise ratio at low speeds; Upper limit: 3000 rpm: close to the common high-speed operating conditions of vehicles, covering the frequency band where the bearing failure characteristic frequency is located, and not exceeding the safe limit speed of the wheel assembly; Typical settings: 500–3000 rpm, covering the main operating range from urban roads to highways; And equivalent simulation of key dynamic loads: Rotational inertial load: Simulates acceleration / cruising / deceleration by setting a rotational speed curve; Centrifugal force effect: naturally generated during high-speed rotation, affecting dynamic balance performance; Thermal effect: Monitoring the impact of temperature changes on bearing clearance after long-term operation.

4. The intelligent detection method for the comprehensive quality of a wheel assembly based on vibration analysis as described in claim 1, characterized in that, In step S3, the time-domain signal is converted into an equal-angle sample to eliminate the influence of speed fluctuations, and the continuous signal is segmented according to the period and time window, as follows; Equal-angle sampling effectively eliminates the influence of rotational speed fluctuations by using high-precision encoder timestamps and spline / VKF interpolation. Combined with an intelligent segmentation strategy based on rotation period, window length and overlap rate can be configured as needed to improve the repeatability, robustness, and diagnostic sensitivity of subsequent features such as order amplitude and impact index.

5. The intelligent detection method for the overall quality of a wheel assembly based on vibration analysis as described in claim 1, characterized in that, In step S5, a training dataset is constructed based on the labeled dataset: qualified, unbalanced, bearing damaged, and loose. The specific steps for constructing the dataset are as follows: When constructing the dataset for wheel assembly vibration testing, the annotation was based on authoritative standards and combined with engineering practice. Dynamic balance defects are classified into acceptable, minor, and severe levels according to the amount of imbalance, in accordance with ISO 1940 standard. The condition of bearings is classified into normal, early wear, moderate damage and severe failure according to ISO 15243 or the company's internal control standards, and is confirmed by disassembly and re-inspection. Bolt loosening is defined based on the preload loss ratio and marked by low-frequency modulation and impact characteristics. Tire uniformity is indirectly assessed by the amplitude of first- and second-order harmonic vibrations, referring to relevant ASTM or SAE specifications. To improve the model's generalization ability, a variety of data augmentation strategies are adopted: controllable Gaussian noise and power frequency interference are added to the original signal to simulate the field environment; By applying virtual rotational speed perturbation to the angular domain signal, equivalent samples at different rotational speeds are generated. Signal interpolation is performed between samples with similar defects to expand the rare fault categories. Time offset and amplitude scaling operations are also used to enhance the model's robustness to sensor differences and phase changes. All enhancements are only used during the training phase to ensure that the test results are true and reliable.

6. The intelligent detection method for the overall quality of a wheel assembly based on vibration analysis as described in claim 1, characterized in that, In step S5, a multi-class / multi-output model is trained. The model outputs the following in sync: defect type: mild dynamic imbalance, early wear of bearing outer ring. The model output is combined to give a comprehensive quality score of 0-100 points. A confidence threshold is introduced to trigger manual re-inspection for uncertain samples. The specific steps are as follows. S51. Construct a structured labeled dataset: collect vibration signals from a large number of wheel assemblies under the S1–S4 process, and have experts and destructive testing confirm their true state; Label definition: Each sample corresponds to the following multi-dimensional labels: Defect type labels: normal, mild dynamic imbalance, moderate dynamic imbalance, early wear of bearing outer ring, damage of bearing inner ring, loose bolts, uneven tires; The continuous values ​​of the overall quality rating label from 0 to 100: Experts score the defects based on their type, severity, and safety margin: Normal = 95–100, Slight imbalance = 80–89, Early bearing wear = 70–79, Severe looseness = 30–50. S52. Design a multi-output model architecture: adopt a multi-task learning framework, share the underlying feature extraction network, and branch outputs results for different tasks; Loss function design: Total loss = α × Classification loss + β × Regression loss; The classification loss is cross-entropy, the regression loss is mean squared error and smoothing L1 loss, and α and β are hyperparameters: α=1.0, β=0.5, which are optimized through the validation set; S53. Model Training and Validation: Train the model using the training set, validate the model's performance using the validation set, and test the model's accuracy using the test set; S54. Inference Phase – Synchronous Output of Defect Type and Quality Score: Perform the following inference process on new samples: Input preprocessed features → Model forward propagation, obtaining two outputs: Defect type probability vector: p=[p1,p2,...,pK], where ∑pi=1; Quality score: s ∈[0,100], determine the main defect type: select the class with the highest probability: y ^=argmax(p); S55. Introduce a confidence threshold to trigger manual review: Confidence definition: Take the maximum predicted probability value conf=max(p) and set the threshold: τ: When conf≥τ: Accept the model's judgment and automatically output the result; When conf < τ: it is determined to be an "uncertain sample" and a manual re-inspection process is triggered; A re-inspection linkage mechanism was designed, which highlights the "re-inspection required" status on the HMI interface, automatically saves the original vibration data, feature maps, and intermediate model outputs for engineers to analyze, pauses the automatic rejection action, and waits for manual confirmation before further processing.

Images