Bearing full-size shape and position tolerance detection method and system based on multi-sensor fusion

Abstract

The application discloses a bearing full-size shape and position tolerance detection method and system based on multi-sensor fusion, and belongs to the technical field of bearing detection. The method comprises the following steps: acquiring original measurement data of a bearing under a multi-source sensor network, and attaching a credibility label containing a real-time health score of a sensor to each data point; inputting the original measurement data carrying the credibility label into a Bayesian deep learning fusion model for probabilistic fusion processing, generating a bearing cognitive digital twin with full-model measurement uncertainty information and an uncertainty cloud map; performing shape and position tolerance evaluation on the bearing cognitive digital twin, and outputting process optimization suggestions; and dynamically adjusting the detection strategy through a multi-objective optimization algorithm to form a quality feedback closed loop and a health maintenance closed loop. Thus, the balance between sub-micron level precision and minute level efficiency can be achieved in a batch online detection scene, and the system has process root cause tracing and self-evolution capabilities.

Description

Technical Field

[0001] This invention relates to a method for detecting full-size form and position tolerances of bearings based on multi-sensor fusion, belonging to the field of bearing inspection technology. Background Technology

[0002] As a core component of mechanical equipment, bearings directly determine the working performance of the host machine through their precision, performance, lifespan, and reliability. In high-end equipment fields such as aerospace, high-speed rail, precision machine tools, and new energy vehicles, the requirements for the dimensional accuracy and geometric tolerances of bearings are becoming increasingly stringent. Micrometer-level errors can lead to increased vibration, noise, shortened lifespan, or even safety accidents.

[0003] Traditional bearing inspection methods mainly rely on manual sampling combined with offline coordinate measuring machines (CMMs) or dedicated bearing measuring instruments, which suffers from low inspection efficiency, limited sample coverage, and the inability to achieve full-size online inspection. With the increasing automation of bearing production lines, the demand for online full inspection is becoming increasingly urgent. In recent years, online inspection technologies based on the fusion of multiple sensors, such as machine vision and laser scanning, have been gradually applied to the bearing industry, attempting to acquire complete geometric information of bearings in a short time by integrating various sensors.

[0004] However, existing technologies still face many challenges in practical batch online inspection scenarios. First, sensor performance gradually degrades over long-term operation, leading to system drift in measurement data. Conventional calibration cycles are insufficient to meet real-time accuracy requirements, especially for high-precision features such as bearing inner diameter and raceway curvature, where even minute drifts can cause misjudgments. Second, bearings exhibit residual thermal deformation after grinding, and environmental temperature fluctuations introduce measurement errors. Traditional temperature compensation models are mostly calibrated offline and cannot adapt to dynamic operating conditions. Third, multi-source data fusion typically employs deterministic weighted averaging, failing to effectively quantify the uncertainties of each sensor and measurement process, resulting in a lack of reliability assessment for the inspection results and making it difficult to meet the stringent reliability requirements of the bearing industry. Fourth, most inspection systems only output a pass / fail conclusion, unable to trace the root causes of out-of-tolerance bearing processes (such as abnormal grinding parameters, wheel wear, uneven cooling, etc.), making closed-loop quality control difficult. Fifth, in batch inspection scenarios, accuracy and efficiency are mutually constrained. Existing systems lack the ability to adaptively adjust inspection strategies and cannot achieve global optimization under dynamic constraints.

[0005] Therefore, there is an urgent need for an online bearing inspection method that can integrate sensor health management, dynamic compensation, probabilistic fusion, intelligent traceability and adaptive optimization to overcome the above-mentioned defects and achieve high-precision, high-efficiency and high-reliability batch full-size form and position tolerance inspection. Summary of the Invention

[0006] To address the aforementioned technical problems, this invention provides a method and system for inspecting the full-size form and position tolerances of bearings based on multi-sensor fusion. This system can achieve a balance between sub-micron level accuracy and minute-level efficiency in batch online inspection scenarios, and also possesses the ability to trace the root causes of processes and the system's self-evolution capabilities.

[0007] To achieve the above objectives, this invention provides a method for full-dimensional form and position tolerance inspection of bearings based on multi-sensor fusion, comprising the following steps: The acquisition step involves obtaining the raw measurement data of the bearing under a multi-source sensor network, and attaching a real-time credibility label representing the credibility of the data source to each data point in the raw measurement data. The real-time credibility label includes at least the sensor real-time health score. The construction step involves inputting the raw measurement data carrying the real-time credibility label into a pre-built Bayesian deep learning fusion model, performing probabilistic fusion processing, and generating a bearing cognitive digital twin with full model measurement uncertainty information. The determination step involves performing form and position tolerance evaluation on the bearing cognitive digital twin, generating a digital deviation field, and using a generative AI model to reverse-engineer the process root causes of the deviation based on the digital deviation field and a pre-set bearing process knowledge base, generating structured process optimization suggestions. The optimization step involves dynamically adjusting the detection strategy for the current or subsequent detection tasks based on the process optimization suggestions and real-time operating information using a multi-objective optimization algorithm, and generating and executing optimized detection control commands.

[0008] In one embodiment, S100 specifically includes: based on the CAD model of the bearing under test and the historical process knowledge base, using a generative AI model to automatically generate an initial multi-sensor collaborative measurement path and point layout, wherein the measurement path preferentially covers key quality characteristic parts of the bearing's inner ring, outer ring raceway, and rolling element contact area; according to the accuracy weight and efficiency weight dynamically generated by the initialization neural network based on real-time working condition information, controlling the multi-source sensor network to perform time-sequential peak-shifting data acquisition according to the measurement path; for the acquired raw measurement data, sequentially performing the first health compensation based on the real-time health score of the sensors, and the second thermal deformation compensation based on the temperature field distribution of the workpiece surface, and using the compensated data as the raw measurement data.

[0009] In one embodiment, S200 specifically includes: inputting the raw measurement data carrying the real-time credibility label into the Bayesian deep learning fusion model, which constructs uncertainty models based on the physical characteristics of different sensor data; fusing multi-source data and their uncertainties within a probabilistic framework using a variational inference algorithm to generate a super-resolution 3D model serving as the bearing cognitive digital twin, and simultaneously generating an uncertainty cloud map on the bearing cognitive digital twin to visualize the distribution of measurement uncertainty of the entire model, with the uncertainty cloud map highlighting the uncertainty levels of the bearing raceway, groove bottom, and sealing groove areas; triggering an attention mechanism when the measurement uncertainty of the bearing's key areas in the uncertainty cloud map exceeds a preset threshold, scheduling high-precision sensors to perform localized encrypted retesting of the key areas; simultaneously processing the multi-source data fusion, starting a parallel background verification thread to perform integrity verification on the data stream during the fusion process; and using a generative adversarial network or diffusion model to virtually complete the data gaps in the fused bearing cognitive digital twin caused by occlusion or surface reflection, in conjunction with the CAD model and the physical laws of geometric continuity.

[0010] In one embodiment, S300 specifically includes: automatically performing full-size form and position tolerance assessment on the bearing cognitive digital twin according to the ISO 1101 standard to generate geometric deviation data as the digital deviation field. The assessment items include at least the bearing inner diameter, outer diameter, width, roundness, cylindricity, radial runout, and end face runout. The digital deviation field is spatiotemporally aligned with historical process parameter time-series data and detection link spatiotemporal trajectory data. The influence weight of different factors on the final out-of-tolerance result is quantified through a contribution analysis model. The digital deviation field is input into the generative AI model that integrates material, process, and finite element simulation knowledge. Through reverse physics deduction, one or more virtual process parameter deviation combinations that cause the deviation field are generated. Combined with the influence weights, the most likely process root cause is identified, and structured process optimization suggestions are output.

[0011] In one embodiment, S400 specifically includes: encoding the detection strategy of the current detection task into a multi-dimensional vector containing sensor combination, scanning path, sampling density, and fusion algorithm parameters; using accuracy, efficiency, resource, and reliability objectives as objective functions, searching for a Pareto front solution set in the multi-dimensional strategy space using a multi-objective genetic algorithm; adjusting the accuracy and efficiency weights online based on the dynamic risk coefficient calculated by the real-time risk assessment model, and dynamically selecting the optimal detection strategy matching the adjusted weights from the Pareto front solution set, generating and executing the corresponding detection control command; and converting the generated... The process optimization instructions are fed back to the preceding grinding or ultra-precision machining unit in real time, and the subsequent bearing inspection results are compared with the compensation expectations to verify the optimization effect and form a quality feedback closed loop. Based on the real-time health score of the sensor, the remaining life of the sensor is predicted, triggering predictive maintenance instructions, and the maintenance effect data is used to optimize the health prediction model, forming a health maintenance closed loop. The entire process data generated by a single inspection task is stored in the central knowledge base, and the entire process data is used to jointly fine-tune the initialization neural network, the risk assessment model, the multi-objective optimization algorithm, and the generative AI model to drive the system's self-evolution.

[0012] This invention also provides a bearing full-size form and position tolerance inspection system based on multi-sensor fusion, used to implement the aforementioned bearing full-size form and position tolerance inspection method based on multi-sensor fusion, comprising: A trusted data acquisition module is used to acquire the raw measurement data of the bearing under a multi-source sensor network, and to attach a real-time trustworthiness label representing the trustworthiness of the data source to each data point in the raw measurement data. The real-time trustworthiness label includes at least the sensor real-time health score. The cognitive twin construction module is used to input the raw measurement data carrying the real-time credibility label into the pre-built Bayesian deep learning fusion model, perform probabilistic fusion processing, and generate a bearing cognitive digital twin with full model measurement uncertainty information; The intelligent judgment and optimization module is used to perform form and position tolerance evaluation on the bearing cognitive digital twin, generate a digital deviation field, and based on the digital deviation field and the preset bearing process knowledge base, use a generative AI model to reverse deduce the process root causes of the deviation and generate structured process optimization suggestions. The closed-loop control execution module is used to dynamically adjust the detection strategy of the current or subsequent detection tasks based on the process optimization suggestions and real-time operating information through a multi-objective optimization algorithm, and generate and execute the optimized detection control instructions.

[0013] In one embodiment, the trusted data acquisition module includes: a generative planning unit, used to automatically generate an initial multi-sensor collaborative measurement path and point layout based on the CAD model and historical process knowledge base of the bearing under test, using a generative AI model, wherein the measurement path preferentially covers key quality characteristic parts of the bearing's inner ring, outer ring raceway, and rolling element contact area; an adaptive acquisition control unit, used to control the multi-source sensor network to perform time-sequential peak-shifting data acquisition according to the measurement path based on the accuracy weights and efficiency weights dynamically generated by the initialized neural network according to real-time operating condition information; and a sequential compensation unit, used to sequentially perform initial health compensation based on the real-time health score of the sensors and secondary thermal deformation compensation based on the temperature field distribution of the workpiece surface on the acquired raw measurement data, and use the compensated data as the raw measurement data.

[0014] In one embodiment, the cognitive twin construction module includes: a probabilistic fusion unit, used to input raw measurement data carrying the real-time confidence label into the Bayesian deep learning fusion model, which constructs uncertainty models based on the physical characteristics of different sensor data; and to fuse multi-source data and their uncertainties within a probabilistic framework using a variational inference algorithm to generate a super-resolution 3D model serving as the bearing cognitive digital twin; and an uncertainty quantification unit, used to simultaneously generate an uncertainty cloud map on the bearing cognitive digital twin that visualizes the distribution of measurement uncertainty across the entire model, wherein the uncertainty cloud map highlights the bearing raceway, etc. The uncertainty level of the groove bottom and sealing groove area; the attention triggering unit, used to trigger the attention mechanism and schedule high-precision sensors to perform local encrypted retesting of the key area when the measurement uncertainty of the bearing key area in the uncertainty cloud map exceeds a preset threshold; the parallel verification unit, used to start a parallel background verification thread to perform integrity verification of the data stream during the fusion process while performing fusion processing of multi-source data; the virtual completion unit, used to use generative adversarial networks or diffusion models to virtually complete the data gaps caused by occlusion or surface reflection in the fused bearing cognitive digital twin, combined with the CAD model and the physical laws of geometric continuity.

[0015] In one embodiment, the intelligent judgment and optimization module includes: a geometric tolerance assessment unit, used to automatically perform full-size geometric tolerance assessment on the bearing cognitive digital twin according to the ISO 1101 standard, generating geometric deviation data as the digital deviation field, wherein the assessment items include at least the bearing inner diameter, outer diameter, width, roundness, cylindricity, radial runout, and end face runout; a contribution analysis unit, used to spatiotemporally align the digital deviation field with historical process parameter time-series data and detection link spatiotemporal trajectory data, and quantify the influence weight of different factors on the final out-of-tolerance result through a contribution analysis model; and a generative root cause inference unit, used to input the digital deviation field into the generative AI model that integrates material, process, and finite element simulation knowledge, generate one or more virtual process parameter deviation combinations that cause the deviation field through reverse physics inference, and, combined with the influence weights, lock in the most likely process root cause and output structured process optimization suggestions.

[0016] In one embodiment, the closed-loop control execution module includes: a strategy encoding unit, used to encode the detection strategy of the current detection task into a multi-dimensional vector containing sensor combination, scanning path, sampling density, and fusion algorithm parameters; a multi-objective optimization unit, used to search for a Pareto front solution set in the multi-dimensional strategy space using a multi-objective genetic algorithm with accuracy, efficiency, resource, and reliability objectives as objective functions; a dynamic strategy selection unit, used to adjust the accuracy weight and efficiency weight online according to the dynamic risk coefficient calculated by the real-time risk assessment model, and dynamically select the optimal detection strategy matching the adjusted weights from the Pareto front solution set, generating and executing the corresponding detection control command; and a quality feedback closed-loop control module. The loop unit is used to feed back the generated process optimization instructions to the preceding grinding or ultra-precision machining unit in real time, and compare the subsequent bearing inspection results with the compensation expectations to verify the optimization effect and form a quality feedback closed loop. The health maintenance closed loop unit is used to predict the remaining life of the sensor based on the real-time health score of the sensor, trigger predictive maintenance instructions, and use the maintenance effect data to optimize the health prediction model to form a health maintenance closed loop. The joint fine-tuning unit is used to store the entire process data generated by a single inspection task into the central knowledge base, and use the entire process data to jointly fine-tune the initialization neural network, the risk assessment model, the multi-objective optimization algorithm, and the generative AI model to drive the system self-evolution.

[0017] Compared with the prior art, the beneficial effects of the present invention are as follows: By attaching a confidence label containing the sensor's real-time health score to the raw measurement data and sequentially performing health compensation and thermal deformation compensation, the system errors introduced by sensor attenuation and environmental thermal interference are effectively eliminated. This allows the system to maintain sub-micron level detection accuracy even when sensor performance degrades by 20%, while reducing unplanned downtime by 35%. By employing a Bayesian deep learning framework for probabilistic fusion of multi-source data, a bearing cognitive digital twin with an uncertainty cloud map is generated, enabling the system to have self-awareness capabilities. It can identify areas where measurement uncertainty exceeds the standard and trigger an attention mechanism for encrypted retesting, significantly improving the reliability and interpretability of bearing detection results. By combining generative AI models with contribution analysis, the root causes of processes are deduced from the digital deviation field, and structured process optimization suggestions are output. This achieves a leap from detection to diagnosis, transforming the detection system from a passive measurement to an intelligent decision-making body for proactive process optimization. Attached Figure Description

[0018] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the accompanying drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on the structures shown in these drawings without creative effort.

[0019] Figure 1 This is a flowchart illustrating the bearing full-size form and position tolerance detection method based on multi-sensor fusion according to the present invention. Figure 2 This is a structural block diagram of the bearing full-size form and position tolerance detection system based on multi-sensor fusion according to the present invention; Figure 3 This is a schematic diagram illustrating the process of constructing a bearing cognitive digital twin in this invention; Figure 4 This is a schematic diagram of the dual closed-loop optimization and self-evolution mechanism in this invention. Detailed Implementation

[0020] 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.

[0021] This invention provides a method for full-dimensional form and position tolerance inspection of precision bearings based on multi-sensor fusion, such as... Figure 1As shown, the process includes the following steps: data acquisition step S100, construction step S200, judgment step S300, and optimization step S400. The following detailed description of each step is provided in conjunction with specific embodiments.

[0022] Step S100 involves acquiring raw measurement data of the bearing from a multi-source sensor network and attaching a real-time reliability label characterizing the reliability of the data source to each data point. This label includes at least a real-time health score of the sensor. The specific implementation process is as follows: S101. Based on the CAD model of the bearing under test and a historical process knowledge base, the system automatically generates an initial multi-sensor collaborative measurement path and point layout using a generative AI model. In one embodiment, the CAD model of the deep groove ball bearing is input into a pre-trained diffusion model. This model automatically identifies key quality characteristic areas of the bearing based on historical testing experience, including the inner ring raceway, outer ring raceway, groove bottom, sealing groove, and end face, and outputs the scanning trajectory, angle, and triggering sequence of each sensor. For the inner ring raceway, the system plans a line laser sensor to scan along the circumference in a helical trajectory; for areas with drastic changes in groove bottom curvature, a structured light camera is planned to project at multiple angles; for end face runout measurement, a contact probe is planned to sample at specific phase points to ensure that key features are fully covered and that there is no mutual interference between sensors.

[0023] S102. Based on real-time operating information, the system dynamically generates the accuracy weights for this detection task by initializing the neural network. and efficiency weight Real-time operating condition information includes the current bearing batch, ambient temperature, and current sensor health status. This initialization neural network is a lightweight multilayer perceptron; its input layer receives operating condition feature vectors, the hidden layer uses the ReLU activation function, and the output layer outputs data through the softmax function. and And satisfy In one embodiment, when inspecting the first P4-grade high-precision bearing, It was set to 0.85. The value is 0.15, when the production line is in steady state of mass production of P0-grade ordinary bearings. Adjusted to 0.25. It is 0.75.

[0024] S103. The control unit coordinates the multi-source sensor network to perform time-series staggered data acquisition based on the aforementioned weights and measurement paths. The multi-source sensor network includes a line laser profilometer, a structured light camera, an infrared temperature array, and a contact probe. The line laser profilometer is used for measuring the inner diameter, outer diameter, and raceway profile; the structured light camera is used for measuring complex curved surfaces such as groove bottoms and chamfers; the infrared temperature array is used to acquire the bearing surface temperature field; and the contact probe is used for high-precision feature verification such as runout.

[0025] Furthermore, the data acquisition sequence is designed as follows: First, the infrared thermography array is activated to acquire the bearing surface temperature field distribution data within 0.1 seconds; then, the linear laser profilometer is triggered to scan along the planned spiral path, taking approximately 3 seconds; next, the structured light camera projects data from multiple angles onto areas such as the bottom of the raceway groove, taking approximately 1.5 seconds; finally, a contact probe is invoked at key phase points for high-precision verification, taking approximately 0.5 seconds. The raw data streams acquired by each sensor are transmitted to the data preprocessing unit in real time.

[0026] S104. In the data preprocessing unit, attach a triple confidence label to each data point: Sensor real-time health score The score is calculated in real time by an independent health prediction model based on the sensor’s historical operating data. Specifically, a Long Short-Term Memory (LSTM) network is used to analyze the sensor’s output stability and noise level changes over the past 24 hours, predict its current performance degradation trend, and output a score of 0 to 100, where 100 indicates a brand-new state, 80 indicates that attention is needed, and 60 indicates that maintenance is needed. The dynamic fusion weight based on sensor historical consistency is calculated by the consistency evaluation algorithm based on the difference between the current batch data and the historical statistical characteristics. Specifically, if the difference between the current data of the line laser and the historical data of the same model exceeds the threshold, its weight will be dynamically reduced. In one embodiment, the current operating condition label includes ambient temperature of 22°C, humidity of 45%, bearing material GCr15, and batch number B240301.

[0027] S105. Perform sequential dual compensation on the original measurement data carrying the confidence label. The sequential dual compensation includes: Health compensation, the system is based on the health status of each sensor. The score is used to query a pre-generated health score and compensation mapping table, which is obtained through experimental calibration. In one embodiment, for a line laser sensor, a standard ring gauge is measured at a standard temperature, and the measurement deviations corresponding to different health scores are recorded to establish a correspondence. Specifically, when... When the corresponding compensation amount is found to be +0.3μm by looking up the table, the system adds 0.3μm to each point cloud coordinate value currently output by the sensor for the first correction. Thermal deformation compensation is performed based on the bearing surface temperature field distribution obtained from an infrared thermometer array. Combined with a database of the linear expansion coefficient of GCr15 bearing steel, a finite element model is used to calculate the theoretical thermal deformation of various parts of the bearing under the current temperature field. The finite element model discretizes the bearing into a finite element mesh, inputs the temperature field distribution, and outputs the thermal deformation displacement field of each node relative to a standard temperature of 20℃. The system subtracts this thermal deformation displacement from the initially corrected measured values, unifying all measured values ​​to the theoretical values ​​at the standard temperature of 20℃. The data after this sequential compensation is the original measurement data with a confidence label, which is then output to the next construction step.

[0028] Construction step S200 involves inputting the raw measurement data carrying confidence labels into a pre-built Bayesian deep learning fusion model, and generating a bearing cognitive digital twin with full model measurement uncertainty information through probabilistic fusion processing. Specifically, this includes: S201. The system inputs data from different sensors into the corresponding branch networks in the Bayesian deep learning fusion model. In one embodiment, taking line laser data as an example, its branch network includes an uncertainty estimation submodule. This submodule constructs an uncertainty model based on the physical characteristics of the laser sensor. The sources of uncertainty of the laser sensor include: speckle noise caused by surface roughness, intensity attenuation caused by changes in the incident angle, and ambient light interference.

[0029] Furthermore, based on historical calibration data, the submodule learns the mapping relationship between these factors and measurement uncertainty, and outputs the prior uncertainty of each data point, expressed in the form of Gaussian variance. Specifically, for points with a large incident angle on the raceway sidewall, the output variance value is larger, indicating that the measurement confidence at that point is lower.

[0030] S202. The outputs of all branch networks are fed into a global fusion layer. This fusion layer uses a variational inference algorithm to jointly optimize multi-source data and their uncertainties within a probabilistic framework. Specifically, the fusion layer uses the data from each sensor as the observed values ​​and the actual bearing geometry as the latent variable to construct a probabilistic graphical model. The probabilistic graphical model assumes that the observed data is generated by transforming the actual geometry through sensor characteristics and adding noise. The variance of the noise is the prior uncertainty of the output of each branch network. By minimizing the KL divergence between the observed values ​​and the model prediction values, the posterior distribution of the actual geometry is iteratively solved.

[0031] Furthermore, the fusion process dynamically adjusts the weights of corresponding observations using the confidence labels of each data point: data with low health automatically receive lower weights in the objective function optimization, thereby reducing their impact on the fusion results; data with high health have higher weights and dominate the fusion results.

[0032] S203. After fusion, the system generates a super-resolution 3D model, namely the bearing cognitive digital twin. This super-resolution 3D model not only contains a high-density geometric coordinate point cloud, but each coordinate point also includes a posterior uncertainty obtained by Bayesian inference, expressed in standard deviation.

[0033] Furthermore, based on this, the system maps posterior uncertainty to color values, simultaneously generating a visual cloud map on the surface of the 3D model, namely the uncertainty cloud map. The uncertainty cloud map pays special attention to the key functional areas of the bearing, such as the bottom of the raceway groove and the sealing groove, which are highlighted in the cloud map. Specifically, the mapping rules are as follows: areas with uncertainty below 0.3μm are displayed in dark green, 0.3–0.5μm in light green, 0.5–0.8μm in yellow, 0.8–1.2μm in orange, and above 1.2μm in red.

[0034] S204. During the fusion process, the system starts a parallel background verification thread, which runs independently of the main fusion calculation thread and continuously performs cyclic redundancy checks and consistency checks on the fused data stream.

[0035] In one embodiment, a CRC32 checksum is calculated for the point cloud data packets in transmission and compared with the checksum attached by the sending end; simultaneously, the continuity of data timestamps is checked, and anomalies are recorded when packet loss or out-of-order delivery is detected. Once a corrupted data packet or timing anomaly is detected, an alarm signal is immediately sent to the main thread, triggering data retransmission or re-acquisition to ensure that the fusion process is based on complete and correct data.

[0036] S205. For data gaps in the fused bearing cognitive digital twin caused by workpiece geometric occlusion, high surface reflectivity, etc., the system calls a pre-trained generative adversarial network (GAN) for virtual completion. In one embodiment, taking a gap caused by occlusion at the bottom of the raceway groove as an example: the local point cloud region containing the gap and its neighborhood data are taken as input. The GAN generator generates a completed point cloud that is geometrically consistent with the surrounding area and conforms to the arc law, based on the prior geometric shape and physical continuity law provided by the CAD model. The discriminator then judges whether the generated point cloud is realistic and credible. After adversarial training, the generator can output high-fidelity completed data. Specifically, for the rolling element contact area data missing due to cage occlusion, the model can reasonably complete it according to symmetry and kinematic laws.

[0037] Step S300 performs form and position tolerance evaluation on the generated bearing cognitive digital twin, and based on the evaluation results, reverse-engineers the root causes of process problems, outputting structured process optimization suggestions. Specifically, this includes: S301. The system inputs the bearing cognitive digital twin into the geometrical tolerance assessment engine. Specifically, this engine incorporates assessment algorithms for all geometrical tolerance items specified in the ISO 1101 standard and optimizes them for bearing characteristics. The assessment engine automatically identifies geometric features on the twin, namely, identifying the inner ring cylindrical surface, outer ring cylindrical surface, raceway surface, end face plane, chamfered surface, etc., through curvature analysis.

[0038] Furthermore, for the inner cylindrical surface, the least squares method is used to fit the cylinder based on the point cloud data to calculate the actual diameter and cylindricity error; for the raceway surface, a toroidal surface is fitted to calculate the groove bottom diameter, groove curvature radius, and roundness error; for the end face, a plane is fitted to calculate the flatness and end face runout relative to the axis.

[0039] In one embodiment, taking the deep groove ball bearing 6204 as an example, the evaluation items include: inner diameter d=20mm, outer diameter D=47mm, width B=14mm, inner ring raceway roundness ≤1.5μm, inner ring radial runout ≤3μm, end face runout ≤4μm, etc. All the above evaluation results constitute a digital deviation field, which indicates the degree of deviation of each part of the bearing from the nominal geometry in three-dimensional space.

[0040] S302. The system invokes the contribution analysis model to perform in-depth analysis of the deviation field. Specifically, the model first aligns the deviation field data with historical process parameter time-series data and the spatiotemporal trajectory data of the current inspection link. After alignment, an interpretability analysis method based on SHAP values ​​is used to calculate the contribution weight of each potential factor to the final out-of-tolerance result. The process parameter time-series data includes: grinding spindle current curve, coolant temperature curve, grinding wheel dressing interval, feed rate, etc. These data correspond one-to-one with the bearing and are stored by timestamp. In one embodiment, for a bearing with an inner diameter 2.5μm larger than the tolerance, the model analysis shows that: the spindle current fluctuation in the final stage of grinding contributes 55%, the coolant temperature rise of 3℃ contributes 25%, the measurement uncertainty introduced by the laser sensor health score of 78 during inspection contributes 12%, and other factors contribute 8%. The model outputs a structured report, clearly indicating the dominant factors.

[0041] Furthermore, the spatiotemporal trajectory data of the detection link includes: the acquisition time of each sensor, its spatial location, and its health score at that time.

[0042] S303. Input the digitized deviation field into a generative AI model that integrates material, process, and finite element simulation knowledge. Specifically, this model is a conditional variational autoencoder, whose training data includes a large number of historical deviation field samples and their corresponding process parameter combinations. The model uses the deviation field as input conditions and generates various virtual process parameter deviation combinations that may lead to the deviation field through back-calculation.

[0043] In one embodiment, for a three-lobed waveform exhibiting out-of-tolerance raceway roundness, the model may generate two possibilities: third harmonic vibration caused by spindle imbalance; and grinding wheel dressing shape error. Combining the weights derived from contribution analysis, the system uses a voting algorithm to identify the most likely root cause as unbalanced vibration caused by wear of the grinding wheel spindle bearing, and outputs structured process optimization suggestions: the primary cause is excessive vibration of the inner ring grinding spindle, suggesting checking the spindle bearing and rebalancing; the secondary cause is large fluctuations in coolant temperature, suggesting checking the temperature control unit; and the system recommends prioritizing spindle dynamic balancing, which is expected to reduce roundness error by 40%.

[0044] S304. When the uncertainty in a critical area shown in the uncertainty cloud map exceeds a preset threshold, the system automatically triggers the attention mechanism. Specifically, taking the inner raceway of a deep groove ball bearing as an example, if the cloud map shows that the uncertainty of a certain section of the raceway reaches 1.0 μm, where the threshold is set to 0.8 μm, the system pauses the current fusion process and schedules idle high-precision contact probes to perform localized and intensified retesting of the area at a specific axial position and a specific circumferential angle.

[0045] Furthermore, the probe scans along the raceway busbar at a higher density to acquire accurate data, which is then re-injected into the fusion model to update the posterior distribution and uncertainty contour map of the region. This process forms a microscopic measurement, verification, and supplementary measurement closed loop, ensuring that key features obtain sufficient data quality.

[0046] In optimization step S400, based on process optimization suggestions and real-time operating information, the detection strategy is dynamically adjusted using a multi-objective optimization algorithm, forming a dual-closed-loop drive system for self-evolution. The specific implementation process is as follows: S401. The system encodes the detection strategy of the current detection task into a multi-dimensional vector, which includes variables such as sensor combination, scanning path, sampling density, and fusion algorithm parameters.

[0047] S402. Start the multi-objective optimization engine. This engine uses four objectives as optimization functions: accuracy, efficiency, resource, and reliability. A multi-objective genetic algorithm is used to search for a Pareto front solution set in the multi-dimensional policy space. In one embodiment, the algorithm may find a set of solutions: solution A (accuracy 0.5 μm, time 9.5 seconds), solution B (accuracy 0.7 μm, time 6.8 seconds), and solution C (accuracy 0.9 μm, time 5.2 seconds). These solutions are non-dominant and constitute a Pareto front.

[0048] S403. During the detection process, a real-time risk assessment model continuously runs. This model takes the current data stream characteristics as input and outputs a dynamic risk coefficient. Data stream characteristics include: real-time noise level, sudden changes in sensor health, and the rate of change of ambient temperature. The model uses a lightweight gradient boosting tree and outputs a risk coefficient ranging from 0 to 1. When the risk coefficient increases, the system adjusts the accuracy weights online. With efficiency weight In one embodiment, the following will be used: It increased from 0.6 to 0.8. The weight is reduced from 0.4 to 0.2, and then the optimal strategy matching the current weight is selected from the Pareto front solution set: if Wp = 0.8, then the solution A with higher accuracy is selected; if the risk is extremely high, i.e. If the value is greater than 0.8, the system may trigger a safety mode, switching to the most conservative sensor combination and the highest sampling density. The selected strategy generates corresponding detection control commands.

[0049] Furthermore, the optimization process also includes two closed-loop execution mechanisms: The system employs a closed-loop quality feedback mechanism, sending generated process optimization commands in real-time to the CNC systems of the preceding grinding or ultra-precision machining units via the OPC UA interface. When subsequent bearings complete machining and re-enter the inspection system, their inspection results are automatically compared with the expected compensation. Specifically, the expected reduction in roundness error after dynamic balancing was 40%, and the actual inspection results showed that the roundness decreased from 2.5μm to 1.6μm, close to the expectation. The system records this optimization case, storing the data on spindle imbalance, dynamic balancing, and a 36% improvement in roundness into a process and quality-related knowledge base to enhance the reasoning capabilities of the generative AI model.

[0050] A closed-loop health maintenance system based on real-time health scores from sensors. The system uses a Wiener process-based remaining lifetime prediction model to predict the remaining lifespan of each sensor based on time series data. In one embodiment, the line laser sensor currently has a health score of 82 and a predicted remaining effective lifespan of 720 hours. When the predicted value falls below a threshold, a predictive maintenance command is automatically triggered, a spare parts replacement work order is generated, and an online self-calibration procedure is executed during the inspection interval. After maintenance is completed, the maintenance effect data is fed back to the health prediction model to optimize its prediction accuracy.

[0051] S404. All data generated from a single inspection task, including the bearing cognitive digital twin, uncertainty cloud map, judgment report, diagnostic report, optimization strategy log, and closed-loop execution results, are structured and stored in the central knowledge base. The system periodically, or during idle periods, uses this newly added data to jointly fine-tune the internal parameters of the initial neural network, risk assessment model, multi-objective optimization algorithm, and generative AI model. Specifically, transfer learning is used to update the weights of the generative AI model, enabling it to more accurately map the deviation field to the process root cause; reinforcement learning is used to adjust the preference parameters of the multi-objective optimization algorithm, making it more adaptable to the current production line conditions. Each completed inspection provides experience for improving the next inspection, forming a positive cycle and achieving continuous self-evolution of the entire inspection system.

[0052] Corresponding to the above method, this embodiment also provides a bearing full-size form and position tolerance inspection system based on multi-sensor fusion, such as... Figure 2 As shown, it includes a trusted data acquisition module, a cognitive twin construction module, an intelligent judgment and optimization module, and a closed-loop control execution module.

[0053] Furthermore, the reliable data acquisition module includes a generative planning unit, an adaptive acquisition control unit, and a sequential compensation unit. The generative planning unit, based on the CAD model of the bearing under test and a historical process knowledge base, automatically generates an initial multi-sensor collaborative measurement path and point layout using a generative AI model. The measurement path prioritizes covering key quality characteristic areas of the bearing's inner and outer raceways and rolling element contact areas. The adaptive acquisition control unit, based on the accuracy and efficiency weights dynamically generated by the initial neural network according to real-time operating conditions, controls the multi-source sensor network to perform time-sequential, staggered data acquisition along the measurement path. The sequential compensation unit sequentially performs initial health compensation based on the real-time sensor health scores and secondary thermal deformation compensation based on the workpiece surface temperature field distribution on the acquired raw measurement data, outputting the compensated data as the raw measurement data.

[0054] Furthermore, the cognitive twin construction module includes a probabilistic fusion unit, an uncertainty quantification unit, an attention triggering unit, a parallel verification unit, and a virtual completion unit. The probabilistic fusion unit inputs raw measurement data carrying real-time credibility labels into a Bayesian deep learning fusion model, constructing uncertainty models based on the physical characteristics of different sensor data, and fusing them within a probabilistic framework using variational inference algorithms to generate a bearing cognitive digital twin. The uncertainty quantification unit simultaneously generates an uncertainty cloud map on the bearing cognitive digital twin, highlighting the uncertainty levels of the bearing raceway, groove bottom, and sealing groove areas. The attention triggering unit monitors the uncertainty cloud map; when the uncertainty in key areas exceeds a threshold, it triggers an attention mechanism to schedule encrypted retesting of sensors. The parallel verification unit, while processing multi-source data fusion, initiates a background verification thread for integrity verification. The virtual completion unit uses generative adversarial networks or diffusion models to complete data gaps in accordance with physical laws.

[0055] Furthermore, the intelligent judgment and optimization module includes a geometrical tolerance assessment unit, a contribution analysis unit, and a generative root cause analysis unit. The geometrical tolerance assessment unit performs full-size geometrical tolerance assessment on the bearing cognitive digital twin according to the ISO 1101 standard, generating a digital deviation field. The assessment items include at least the bearing's inner diameter, outer diameter, width, roundness, cylindricity, radial runout, and axial runout. The contribution analysis unit aligns the deviation field with historical process parameters and inspection data, quantifying the influence weight of each factor through a contribution analysis model. The generative root cause analysis unit inputs the deviation field into a generative AI model, reverse-engineers the process root causes, and outputs structured process optimization suggestions.

[0056] Furthermore, the closed-loop control execution module includes a strategy encoding unit, a multi-objective optimization unit, a dynamic strategy selection unit, a quality feedback closed-loop unit, a health maintenance closed-loop unit, and a joint fine-tuning unit. The strategy encoding unit encodes the detection strategy into a multi-dimensional vector. The multi-objective optimization unit, with accuracy, efficiency, resource efficiency, and reliability as objectives, uses a multi-objective genetic algorithm to search for the Pareto front solution set. The dynamic strategy selection unit adjusts weights based on real-time risk coefficients and selects the optimal strategy, generating detection control commands. The quality feedback closed-loop unit feeds back the process optimization commands to the upstream grinding or ultra-precision machining unit to verify the optimization effect. The health maintenance closed-loop unit predicts the remaining lifespan based on sensor health and triggers predictive maintenance. The joint fine-tuning unit uses full-process data to jointly fine-tune multiple AI models in the system, driving the system's self-evolution.

[0057] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A method for inspecting the full-dimensional form and position tolerances of bearings based on multi-sensor fusion, characterized in that, Includes the following steps: S100. Obtain the raw measurement data of the bearing under a multi-source sensor network, and attach a real-time credibility label representing the credibility of the data source to each data point in the raw measurement data. The real-time credibility label includes at least the sensor real-time health score. S200. Input the raw measurement data carrying the real-time credibility label into the pre-built Bayesian deep learning fusion model, perform probabilistic fusion processing, and generate a bearing cognitive digital twin with full model measurement uncertainty information. S300. Perform form and position tolerance evaluation on the bearing cognitive digital twin, generate a digital deviation field, and based on the digital deviation field and the preset bearing process knowledge base, use a generative AI model to reverse-engineer the process root causes of the deviation and generate structured process optimization suggestions. S400. Based on the process optimization suggestions and real-time operating information, dynamically adjust the detection strategy of the current or subsequent detection tasks through a multi-objective optimization algorithm, and generate and execute the optimized detection control instructions.

2. The method for detecting full-dimensional form and position tolerances of bearings based on multi-sensor fusion according to claim 1, characterized in that, Specifically, S100 includes: Based on the CAD model of the bearing under test and the historical process knowledge base, the initial multi-sensor collaborative measurement path and point layout are automatically generated using a generative AI model. The measurement path prioritizes covering the key quality characteristic parts of the bearing's inner ring, outer ring raceway, and rolling element contact area. Based on the accuracy weights and efficiency weights dynamically generated by the initial neural network according to real-time operating conditions, the multi-source sensor network is controlled to perform time-series staggered data acquisition according to the measurement path. For the collected raw measurement data, the first health compensation based on the real-time health score of the sensor and the second thermal deformation compensation based on the temperature field distribution of the workpiece surface are performed sequentially, and the compensated data is used as the raw measurement data.

3. The method for detecting full-dimensional form and position tolerances of bearings based on multi-sensor fusion according to claim 1, characterized in that, Specifically, S200 includes: The raw measurement data carrying the real-time credibility label is input into the Bayesian deep learning fusion model, which constructs uncertainty models based on the physical characteristics of different sensor data. Using variational inference algorithms, multi-source data and their uncertainties are fused within a probabilistic framework to generate a super-resolution 3D model that serves as the bearing cognitive digital twin. Simultaneously, an uncertainty cloud map is generated on the bearing cognitive digital twin to visualize the distribution of measurement uncertainty of the entire model. The uncertainty cloud map highlights the uncertainty levels of the bearing raceway, groove bottom, and sealing groove regions. When the measurement uncertainty of the critical area of ​​the bearing in the uncertainty cloud map exceeds a preset threshold, the attention mechanism is triggered, and a high-precision sensor is scheduled to perform localized and encrypted retesting of the critical area. While fusing multi-source data, a parallel background verification thread is started to perform integrity verification on the data stream during the fusion process; By using generative adversarial networks or diffusion models, virtual completion is performed on the data gaps in the fused bearing cognitive digital twin caused by occlusion or surface reflection, in conjunction with the CAD model and the physical laws of geometric continuity.

4. The method for detecting full-dimensional form and position tolerances of bearings based on multi-sensor fusion according to claim 1, characterized in that, Specifically, S300 includes: On the bearing cognitive digital twin, full-size form and position tolerance evaluation is automatically performed according to ISO 1101 standard to generate geometric deviation data as the digital deviation field. The evaluation items include at least the bearing inner diameter, outer diameter, width, roundness, cylindricity, radial runout, and end face runout. The digital deviation field is spatiotemporally aligned with historical process parameter time-series data and detection link spatiotemporal trajectory data, and the influence weight of different factors on the final deviation result is quantified through a contribution analysis model. The digital deviation field is input into the generative AI model that integrates material, process, and finite element simulation knowledge. Through reverse physics deduction, one or more virtual process parameter deviation combinations that cause the deviation field are generated. Combined with the influence weights, the most likely process root cause is identified and the structured process optimization suggestions are output.

5. The method for detecting full-dimensional form and position tolerances of bearings based on multi-sensor fusion according to claim 2, characterized in that, Specifically, S400 includes: The detection strategy of the current detection task is encoded as a multi-dimensional vector containing sensor combination, scanning path, sampling density and fusion algorithm parameters; Using accuracy, efficiency, resource, and reliability as objective functions, a multi-objective genetic algorithm is employed to search for Pareto front solutions in a multi-dimensional policy space. Based on the dynamic risk coefficient calculated by the real-time risk assessment model, the accuracy weight and efficiency weight are adjusted online, and the optimal detection strategy matching the adjusted weights is dynamically selected from the Pareto front solution set, and the corresponding detection control command is generated and executed. The generated process optimization instructions are fed back to the preceding grinding or ultra-precision machining unit in real time, and the subsequent bearing test results are compared with the compensation expectations to verify the optimization effect and form a quality feedback closed loop. Based on the real-time health score of the sensor, the remaining lifespan of the sensor is predicted, a predictive maintenance command is triggered, and the maintenance effect data is used to optimize the health prediction model, forming a health maintenance closed loop. The entire process data generated by a single detection task is stored in a central knowledge base, and the entire process data is used to jointly fine-tune the initialization neural network, the risk assessment model, the multi-objective optimization algorithm, and the generative AI model, driving the system to self-evolve.

6. A bearing full-size form and position tolerance inspection system based on multi-sensor fusion, characterized in that, The method for implementing the bearing full-dimensional form and position tolerance inspection method based on multi-sensor fusion as described in any one of claims 1-5 includes: A trusted data acquisition module is used to acquire the raw measurement data of the bearing under a multi-source sensor network, and to attach a real-time trustworthiness label representing the trustworthiness of the data source to each data point in the raw measurement data. The real-time trustworthiness label includes at least the sensor real-time health score. The cognitive twin construction module is used to input the raw measurement data carrying the real-time credibility label into the pre-built Bayesian deep learning fusion model, perform probabilistic fusion processing, and generate a bearing cognitive digital twin with full model measurement uncertainty information; The intelligent judgment and optimization module is used to perform form and position tolerance evaluation on the bearing cognitive digital twin, generate a digital deviation field, and based on the digital deviation field and the preset bearing process knowledge base, use a generative AI model to reverse deduce the process root causes of the deviation and generate structured process optimization suggestions. The closed-loop control execution module is used to dynamically adjust the detection strategy of the current or subsequent detection tasks based on the process optimization suggestions and real-time operating information through a multi-objective optimization algorithm, and generate and execute the optimized detection control instructions.

7. The bearing full-size form and position tolerance inspection system based on multi-sensor fusion according to claim 6, characterized in that, The trusted data acquisition module includes: Generative planning unit is used to automatically generate an initial multi-sensor collaborative measurement path and point layout based on the CAD model and historical process knowledge base of the bearing under test and the generative AI model. The measurement path prioritizes covering the key quality characteristic parts of the bearing inner ring, outer ring raceway and rolling element contact area. An adaptive acquisition and control unit is used to control the multi-source sensor network to perform time-series staggered data acquisition according to the measurement path based on the accuracy weights and efficiency weights dynamically generated by the initial neural network based on real-time operating conditions. The sequential compensation unit is used to sequentially perform initial health compensation based on the real-time health score of the sensor and secondary thermal deformation compensation based on the temperature field distribution of the workpiece surface on the collected raw measurement data, and use the compensated data as the original measurement data.

8. The bearing full-size form and position tolerance inspection system based on multi-sensor fusion according to claim 6, characterized in that, The cognitive twin construction module includes: The probabilistic fusion unit is used to input the raw measurement data carrying the real-time credibility label into the Bayesian deep learning fusion model, which constructs uncertainty models based on the physical characteristics of different sensor data respectively; through variational inference algorithm, multi-source data and their uncertainties are fused in a probabilistic framework to generate a super-resolution three-dimensional model as the bearing cognitive digital twin. An uncertainty quantification unit is used to synchronously generate an uncertainty cloud map on the bearing cognitive digital twin, which visualizes the distribution of measurement uncertainty of the entire model. The uncertainty cloud map highlights the uncertainty level of the bearing raceway, groove bottom and sealing groove regions. The attention triggering unit is used to trigger the attention mechanism when the measurement uncertainty of the key area of ​​the bearing in the uncertainty cloud map exceeds a preset threshold, and to schedule a high-precision sensor to perform local encrypted retesting of the key area. The parallel verification unit is used to start a parallel background verification thread to perform integrity verification on the data stream during the fusion process while performing fusion processing on multi-source data. The virtual completion unit is used to virtually complete the data gaps in the fused bearing cognitive digital twin caused by occlusion or surface reflection by using generative adversarial networks or diffusion models, combined with the CAD model and the physical laws of geometric continuity.

9. The bearing full-size form and position tolerance inspection system based on multi-sensor fusion according to claim 6, characterized in that, The intelligent judgment and optimization module includes: The form and position tolerance evaluation unit is used to automatically perform full-size form and position tolerance evaluation on the bearing cognitive digital twin according to the ISO 1101 standard, and generate geometric deviation data as the digital deviation field. The evaluation items include at least the bearing inner diameter, outer diameter, width, roundness, cylindricity, radial runout and end face runout. The contribution analysis unit is used to align the digital deviation field with historical process parameter time series data and detection link spatiotemporal trajectory data in a spatiotemporal manner, and quantify the influence weight of different factors on the final deviation result through the contribution analysis model. The generative root cause inference unit is used to input the digital deviation field into the generative AI model that integrates material, process and finite element simulation knowledge. Through reverse physics inference, it generates one or more virtual process parameter deviation combinations that cause the deviation field, and combines the influence weights to lock in the most likely process root cause and output the structured process optimization suggestions.

10. The bearing full-size form and position tolerance inspection system based on multi-sensor fusion according to claim 6, characterized in that, The closed-loop control execution module includes: The strategy encoding unit is used to encode the detection strategy of the current detection task into a multi-dimensional vector containing sensor combination, scanning path, sampling density and fusion algorithm parameters; The multi-objective optimization unit is used to search for Pareto front solutions in a multi-dimensional policy space using a multi-objective genetic algorithm with accuracy, efficiency, resource, and reliability objectives as objective functions. The dynamic strategy selection unit is used to adjust the accuracy weight and efficiency weight online based on the dynamic risk coefficient calculated by the real-time risk assessment model, and to dynamically select the optimal detection strategy that matches the adjusted weight from the Pareto front solution set, and generate and execute the corresponding detection control command. The quality feedback closed-loop unit is used to feed back the generated process optimization instructions to the preceding grinding or ultra-precision machining unit in real time, and compare the subsequent bearing test results with the compensation expectation to verify the optimization effect and form a quality feedback closed loop. The health maintenance closed-loop unit is used to predict the remaining lifespan of the sensor based on the real-time health score of the sensor, trigger predictive maintenance instructions, and use the maintenance effect data to optimize the health prediction model to form a health maintenance closed loop. The joint fine-tuning unit is used to store the entire process data generated by a single detection task into the central knowledge base, and use the entire process data to jointly fine-tune the initialization neural network, the risk assessment model, the multi-objective optimization algorithm and the generative AI model, thereby driving the system to self-evolve.

Images