Artificial intelligence identification method and system for abnormal state of production line

Abstract

The invention provides a production line abnormal state artificial intelligence identification method and system, and the method comprises the steps: collecting joint torsion moment data and ball track offset data of a humanoid robot joint at different rotation angles, and generating an evaluation result of a bearing installation state; determining the abnormal state of bearing installation according to the evaluation result of the bearing installation state; according to the dynamically adjusted judgment threshold value, the joint torsional moment data collected in real time are compared, and the severity of the abnormal state is determined; according to the severity of the abnormal state, joint rotation resistance is simulated and recognized, and a predicted value of joint fit clearance change is generated; according to the predicted value of the joint fit clearance change, updating the judgment threshold value of the dynamic adjustment; after the dynamically adjusted judgment threshold value is updated, joint assembly control configuration is generated; and generating a joint assembly control strategy according to the joint assembly control configuration.

Description

Technical Field

[0001] This invention relates to the field of information technology, and in particular to an artificial intelligence-based method and system for identifying abnormal conditions in a production line. Background Technology

[0002] As a core technology in intelligent manufacturing and automation, the precision and reliability of the joint system of humanoid robots directly determine the robot's motion performance and service life. In humanoid robot joint assembly lines, the installation quality of bearings is a key factor affecting the smoothness and durability of joint rotation. High-quality bearing assembly ensures that the joint maintains stable mechanical properties during complex movements, while even minor deviations in assembly can lead to serious performance problems. Currently, the detection of bearing assembly anomalies mainly relies on traditional manual inspection or automated inspection equipment based on fixed thresholds. These methods have significant limitations in actual production. Manual inspection is limited by the operator's experience and fatigue level, making it difficult to guarantee consistency and efficiency. While automated inspection equipment with fixed thresholds improves efficiency to some extent, it cannot adapt to dynamic fluctuations caused by material differences, equipment wear, or environmental changes during production. This static inspection method often struggles to accurately capture anomalies in complex and ever-changing assembly scenarios, leading to frequent misjudgments or missed detections. When bearings are installed with tilt or eccentricity, such as uneven pressure from assembly equipment leading to bearing insertion angle deviation, uneven machined surfaces of the joint housing causing bearing housing tilt, thermal deformation of parts during assembly causing positional displacement, or excessive assembly speed preventing the bearing from fully positioning, irregular frictional torques will be generated during joint rotation. This uneven torque distribution not only increases rotational resistance but also causes the movement trajectory of the internal bearing balls to deviate from the normal raceway operation. This deviation further exacerbates the variation in the fit clearance between the joint housing and the bearing housing, resulting in significant differences in rotational resistance at different angles. For example, when the bearing is installed with a slight tilt, the joint may exhibit a significant increase in resistance at some angles, while at other angles it may be close to normal. This dynamically changing torque characteristic is difficult to accurately judge using a single testing standard. Simultaneously, the deviation in the ball movement trajectory also leads to additional wear, further amplifying the unevenness of the torque distribution. Summary of the Invention

[0003] This invention provides an artificial intelligence-based method for identifying abnormal conditions on a production line, mainly comprising: Data on joint torsional torque and ball bearing trajectory offset of a humanoid robot joint at different rotation angles are collected to generate an evaluation result of the bearing installation status. Based on this evaluation result, an abnormal bearing installation status is determined. A dynamically adjusted threshold is generated by analyzing the correlation between the fluctuation amplitude of the joint torsional torque data at different rotation angles and the ball bearing trajectory offset data. Based on this threshold, the severity of the abnormal status is determined by comparing it with the real-time collected joint torsional torque data. Based on the severity of the abnormal status, joint rotation resistance is simulated and identified to generate a predicted value for changes in joint clearance. The dynamically adjusted threshold is updated based on the predicted value for changes in joint clearance. After updating the threshold, a joint assembly control configuration is generated. Finally, a joint assembly control strategy is generated based on this configuration.

[0004] Furthermore, the acquisition of joint torsional torque data and ball bearing trajectory offset data at different rotation angles of the humanoid robot joints generates an evaluation result of the bearing installation status, including: The raw sequence of the joint torsional torque data is collected, and the torque change rate between adjacent sampling points is calculated; the ball trajectory offset data is collected, the periodic change of the offset data is calculated, and the distribution characteristics of the offset trajectory are determined; based on the torque change rate and the distribution characteristics of the offset trajectory, a torque distribution map is constructed, the feature parameters of the map are extracted, and a quantitative evaluation result of the bearing installation status is generated.

[0005] Furthermore, determining the abnormal state of bearing installation based on the evaluation results of the bearing installation status includes: Based on the evaluation results of the bearing installation status, a clustering algorithm is used to group the joint torsional torque data and the ball trajectory offset data to generate preliminary status grouping results; the time-domain features and distribution density of each group in the preliminary status grouping results are extracted to determine the bearing installation anomaly; by calculating the distance metric between the abnormal group and the normal group, the degree of anomaly is quantified, and an abnormal status diagnosis result is generated.

[0006] Furthermore, the step of generating a dynamically adjusted discrimination threshold by analyzing the correlation between the fluctuation amplitude of the joint torsional torque data at different rotation angles and the ball trajectory offset data includes: The fluctuation amplitude of the joint torsional moment data is calculated in segments to generate a smoothed fluctuation amplitude sequence; the local strength index of the ball trajectory offset data is extracted, and the correlation coefficient between the fluctuation amplitude sequence and the local strength index is calculated; a dynamically adjusted discrimination threshold is generated based on the correlation coefficient.

[0007] Furthermore, the step of determining the severity of the abnormal state by comparing the dynamically adjusted discrimination threshold with the real-time collected joint torsional torque data includes: The difference between the joint torsional torque data and the dynamically adjusted discrimination threshold is calculated using the dynamically adjusted discrimination threshold to generate a fluctuation deviation rate; the abnormality level is marked according to the ball trajectory offset data and the fluctuation deviation rate; the severity of the abnormal state is generated according to the abnormality level and the angular distribution of the over-limit points.

[0008] Furthermore, the step of simulating and identifying joint rotational resistance and generating predicted values ​​for changes in joint fit clearance based on the severity of the abnormal state includes: The time delay relationship between the peak sequence of the joint torsional torque data and the maximum offset sequence of the ball trajectory offset data is calculated to construct a simulation calculation model of the joint rotation resistance and generate a resistance distribution curve. The abrupt change point of the resistance distribution curve is identified by gradient calculation to generate the cumulative clearance change. The clearance change rate is fitted according to the cumulative clearance change to generate a predicted value of the joint fit clearance change.

[0009] Furthermore, updating the dynamically adjusted discrimination threshold based on the predicted value of the joint fit clearance change includes: The deviation rate is calculated by comparing the predicted value of the joint fit gap change with a preset range; the dynamic adjustment discrimination threshold is updated based on the deviation rate.

[0010] Furthermore, the generation of the joint assembly control configuration includes: Extract the rate of change of the ball trajectory offset data and the joint torsional torque data, calculate the correlation coefficient, and generate a matching degree evaluation result; generate a resistance correction value sequence based on the matching degree evaluation result; adjust the drive motor control parameters through the resistance correction value sequence to generate a servo control parameter group; map the assembly process parameters through the servo control parameter group to generate the joint assembly control configuration.

[0011] Furthermore, the step of generating a control strategy based on the joint assembly control configuration includes: The joint assembly control configuration monitors the joint torsional torque data, adjusts the drive current and speed, and generates compensation records. The cumulative compensation amount is calculated using the compensation records, the changing trend of the ball trajectory offset data is verified, and a joint assembly control strategy for extending bearing service life is generated.

[0012] An artificial intelligence-based system for identifying abnormal conditions on a production line includes: The evaluation module is used to collect joint torsional torque data and ball trajectory offset data of humanoid robot joints at different rotation angles, and generate evaluation results of bearing installation status. The first determining module is used to determine the abnormal state of bearing installation based on the evaluation results of the bearing installation state. The discrimination module is used to generate a dynamically adjusted discrimination threshold by analyzing the correlation between the fluctuation amplitude of the joint torsional torque data at different rotation angles and the ball trajectory offset data; The second determining module is used to determine the severity of the abnormal state by comparing the dynamically adjusted discrimination threshold with the real-time collected joint torsional torque data. The prediction module is used to simulate and identify joint rotation resistance based on the severity of the abnormal state, and generate a predicted value for the change in joint fit clearance. An update module is used to update the dynamically adjusted discrimination threshold based on the predicted value of the change in the joint fit gap; The generation module is used to generate the joint assembly control configuration after updating the dynamically adjusted discrimination threshold. The control module is used to generate a joint assembly control strategy based on the joint assembly control configuration.

[0013] The technical solutions provided by the embodiments of the present invention may include the following beneficial effects: This invention discloses an artificial intelligence-based method for identifying abnormal conditions in production lines. Addressing the complex correlation between joint torsional torque and ball bearing trajectory offset data at different rotation angles, it accurately identifies bearing installation tilting or eccentricity anomalies through feature extraction and clustering. Based on dynamic correlation analysis of fluctuation amplitude and trajectory offset, this invention constructs a dynamically adjustable discrimination threshold, compares torsional torque data in real time, determines the severity of the anomaly, and predicts changes in joint fit clearance. The prediction results are used to adjust force information during assembly, optimizing the anomaly identification threshold. Furthermore, through matching analysis of trajectory offset and uneven torque distribution, a joint rotation resistance correction scheme is generated, outputting a high-precision assembly control configuration. Finally, through real-time monitoring and continuous data acquisition, trajectory offset is reduced and long-term anomaly tracking is achieved, significantly improving motion accuracy and bearing lifespan. This invention, through multi-dimensional data fusion and dynamic optimization, solves the problem of insufficient identification of complex abnormal conditions in traditional methods, achieving high-precision and long-life joint control. Attached Figure Description

[0014] Figure 1 This is a flowchart of an artificial intelligence method for identifying abnormal conditions in a production line according to the present invention.

[0015] Figure 2 This is a schematic diagram of the structure of an artificial intelligence recognition system for abnormal conditions in a production line according to the present invention. Detailed Implementation

[0016] To enable those skilled in the art to better understand the technical solutions in this specification, the technical solutions in the embodiments of this specification will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this specification, and not all embodiments. Based on the embodiments in this specification, all other embodiments obtained by those skilled in the art without creative effort should fall within the scope of protection of this specification.

[0017] like Figure 1-2 This embodiment of an artificial intelligence-based method and system for identifying abnormal conditions on a production line may specifically include: Step S101: Collect joint torsional torque data and ball trajectory offset data of the humanoid robot joints at different rotation angles to generate an evaluation result of the bearing installation status.

[0018] Raw torsional torque data is acquired from force sensors at 15° sampling points within the joint rotation range of 0° to 360°. Noise is removed from the torque data through filtering. The torque change rate between adjacent sampling points is calculated for the filtered torque sequence. Sampling points with a change rate exceeding a preset threshold are marked as potential abnormal angle intervals. Within these intervals, a laser displacement sensor detects the radial runout of the bearing outer ring relative to the joint housing. Radial runout data at different angular positions is recorded as the joint rotates at a constant angular velocity. The theoretical offset trajectory of the balls in the raceway is calculated based on the periodic variation of the runout data. When the offset trajectory exhibits an asymmetrical distribution within a specific angular range, a bearing installation tilt characteristic is identified within that interval. Based on the angular distribution and radial runout corresponding to the bearing installation tilt characteristic, a torque distribution map is constructed within the joint rotation cycle. The phase difference between the peak torque position and the maximum radial runout position is calculated. Dimensionality reduction is used to extract the characteristic parameter set of the torque distribution map, including the mean torque, peak-to-valley difference, and phase offset angle, to obtain a quantified bearing installation status assessment result.

[0019] For example, in one embodiment, a force sensor is mounted on the end of the drive shaft of a humanoid robot joint, and the acquisition frequency is set to 1 kHz to ensure that a sufficient density of torque data points is acquired during joint rotation.

[0020] Specifically, when the joint rotates at a constant angular velocity of 10° / s, 150 torque sampling points can be collected within every 15° angle interval. The original data is smoothed by a moving average filter with a filter window width of 5 sampling points, effectively eliminating high-frequency noise interference.

[0021] It should be noted that the torque change rate is calculated using the finite difference method, which divides the torque difference between adjacent 15° sampling points by the angular interval to obtain the change rate value. When this change rate exceeds a preset threshold of 0.5 N·m / 15°, the angular interval is marked as a potential abnormal interval, providing a target range for refined detection.

[0022] For example, the laser displacement sensor is vertically aligned with the outer ring surface of the bearing, achieving a measurement accuracy of 0.01 mm. During the constant-speed rotation of the joint, the sensor records the radial distance value every 1°, obtaining 360 radial runout data points over the complete rotation cycle.

[0023] Preferably, by analyzing the spectral characteristics of the jitter data through Fourier transform, when the amplitude of the fundamental frequency component exceeds 0.05 mm and there is an obvious second harmonic component, it indicates that the bearing installation is tilted.

[0024] In one possible implementation, the process of constructing the torque distribution map involves arranging torque data within a 360° range in polar coordinates, where the radial coordinates represent the magnitude of the torque and the angular coordinates correspond to the joint rotation position. By identifying the peak torque points and the maximum radial runout points in the map, the angular difference between the two is calculated as a phase offset parameter.

[0025] For example, when the peak torque occurs at 90° and the maximum radial runout occurs at 105°, the phase difference is 15°, which reflects the torque transmission lag caused by bearing tilt.

[0026] Understandably, the extraction of the feature parameter set adopts statistical methods. The mean torque is obtained by calculating the arithmetic mean of the torque data over the entire cycle. The peak-to-valley difference is the difference between the maximum torque and the minimum torque. These parameters together constitute a three-dimensional feature vector, which serves as a quantitative evaluation index of the bearing installation status.

[0027] Step S102: Determine the abnormal state of bearing installation based on the evaluation results of the bearing installation status.

[0028] Based on the three characteristic parameters—mean torque, peak-to-valley difference, and phase offset angle—from the bearing installation status assessment results, a multi-dimensional feature space is constructed. A clustering algorithm is used to initially group the data points in this feature space, setting the number of clusters to a preset number, corresponding to various installation states: normal installation, slight tilt, severe tilt, and eccentric installation. The Euclidean distance from each data point to the cluster center is calculated iteratively, data points are reassigned, and the cluster center positions are updated. The iteration terminates when the change in the cluster center position is less than a preset convergence threshold, yielding preliminary state grouping results. For these preliminary state grouping results, the temporal characteristics of the joint torsional torque data within each group are extracted, including the standard deviation of torque fluctuations, skewness coefficient, and kurtosis coefficient. Simultaneously, the distribution density of the ball trajectory offset data in different angle intervals is calculated. If the standard deviation of torque within a group exceeds a first preset threshold and the ball offset distribution density exceeds a preset proportion within a specific sector area, the bearing in that group is determined to have an abnormal tilt. If the offset distribution is uniform across the entire circumference and the standard deviation of torque exceeds a second preset threshold, it is determined to have an eccentricity anomaly. After obtaining the bearing tilt or eccentricity anomaly determination result, the distance metric between each data point in the abnormal group and the cluster center of the normal group is calculated. The distance metric is obtained by standardization through the covariance matrix of the feature vector. The degree of anomaly is quantified and scored according to the distance metric. When the distance value is in different preset intervals, it is defined as a mild, moderate, or severe anomaly level, respectively. The joint torsional torque data under the abnormal state is compensated by the correction coefficient corresponding to the anomaly level. After compensation, the clustering group verification is re-executed. When the proportion of data points returning to the normal group after compensation reaches a preset verification threshold, the diagnosis result of bearing installation tilt or eccentricity anomaly is determined.

[0029] For example, in one implementation, the construction of the multidimensional feature space is based on three key feature parameters extracted from the bearing installation status assessment results. The mean torque reflects the average resistance level of the joint during a complete rotation cycle, the peak-to-valley difference reflects the amplitude range of torque fluctuations, and the phase offset angle characterizes the time delay relationship between the peak torque and the peak radial runout. After normalization, these three feature parameters form a feature point cloud distribution in three-dimensional space. The clustering algorithm randomly selects initial cluster centers, iteratively calculates the Euclidean distance from each feature point to each cluster center, and assigns the feature points to the category of the nearest cluster center. After each iteration, the mean of all feature points in each category is recalculated as the new cluster center position. When the change in the cluster center position between two adjacent iterations is less than a preset convergence threshold, the algorithm terminates and outputs the final grouping results.

[0030] It should be noted that the extraction of time-domain features involves statistical analysis of the joint torsional torque data sequence. The standard deviation quantifies the intensity of data fluctuation by calculating the dispersion of torque values ​​relative to the mean. The skewness coefficient measures the asymmetry of the torque distribution; positive skewness indicates a shift of torque values ​​towards the higher end, while negative skewness indicates the opposite. The kurtosis coefficient describes the steepness of the torque distribution curve; a high kurtosis value indicates the presence of more extreme torque values. The distribution density of the ball bearing trajectory offset data is determined by dividing the 360° circumference into several sector regions and statistically analyzing the proportion of data points with offsets exceeding a threshold in each region. When the high offset density in a certain 90° sector region is significantly higher than in other regions, it indicates that the bearing has a tilted installation problem in that direction.

[0031] Specifically, the anomaly detection logic is based on a comprehensive analysis of temporal and spatial distribution characteristics. Bearing tilt anomaly manifests as a torque standard deviation exceeding a first preset threshold, while ball misalignment is concentrated within a specific fan-shaped area. This uneven distribution stems from increased local contact pressure caused by bearing tilt, resulting in additional radial displacement of the balls as they pass through this area. Eccentricity anomaly manifests as a torque standard deviation exceeding a higher second preset threshold, but with relatively uniform ball misalignment across the entire circumference. This is because the bearing deviates from its ideal installation position, causing changes in contact conditions in all directions.

[0032] For example, the distance metric is calculated using a standardization method that considers feature correlation. First, a covariance matrix is ​​constructed for all feature points within the normal group, reflecting the correlation between the three feature parameters. Eigenvalue decomposition of the covariance matrix yields eigenvectors and eigenvalues. This information is then used to transform the coordinates of feature points within the abnormal group, eliminating the influence of feature correlation. In the transformed feature space, the distance from the abnormal point to the normal cluster center is calculated; this distance more accurately reflects the degree of abnormality. Different distance values ​​falling within preset intervals correspond to different levels of abnormality: mild abnormalities typically require only minor adjustments to assembly parameters, moderate abnormalities require recalibrating the assembly equipment, and severe abnormalities require replacing bearings or reprocessing joint components.

[0033] In one possible implementation, the correction factor is determined based on statistical analysis of historical assembly data. By collecting a large number of joint assembly cases with known anomaly levels, a mapping relationship is established between the anomaly level and the optimal correction factor. The correction factor for minor anomalies is typically between 0.9 and 1.1, used for fine-tuning the amplitude of the torque data. The correction factor range for moderate anomalies expands to 0.7 to 1.3, requiring not only amplitude adjustment but also phase correction. The correction factor for severe anomalies may exceed 1.5, necessitating a complete reconstruction of the torque curve.

[0034] Preferably, the compensation process employs a piecewise linear interpolation method. The 360° range of joint rotation is divided into 24 15° intervals, and the torque data is linearly adjusted within each interval according to a correction coefficient. Cubic spline interpolation is used at the interval boundaries to ensure the continuity and smoothness of the compensated torque curve. The compensated data is then re-input into a clustering algorithm for validation grouping, and the proportion of data points that regress to the normal group is statistically analyzed.

[0035] For example, during post-assembly inspection of a humanoid robot's shoulder joint, it was found that 30% of the original torque data points were classified as abnormal groups after cluster analysis. Distance metric calculations determined this to be a moderate tilt anomaly, with a corresponding correction factor of 1.15. After applying compensation, the re-clustering results showed that 92% of the data points returned to normal groups, exceeding the preset verification threshold, confirming that the joint had a bearing tilt problem that could be resolved through software compensation.

[0036] Understandably, the final diagnostic results output includes multi-dimensional information such as anomaly type, anomaly level, and recommended handling measures. The assembly parameters of subsequent joints are automatically adjusted based on the diagnostic results, forming a closed-loop quality control mechanism. By continuously accumulating diagnostic data and optimizing the initial parameter settings and anomaly detection thresholds of the clustering algorithm, the assembly yield of the entire production line is improved.

[0037] Step S103: By analyzing the correlation between the fluctuation amplitude of the joint torsional torque data at different rotation angles and the ball trajectory offset data, a dynamically adjusted discrimination threshold is generated.

[0038] After identifying the abnormal state, the joint torsional torque data is segmented according to preset angle intervals. The difference between the maximum and minimum torque values ​​within each interval is calculated as the local fluctuation amplitude. A weighted average of the fluctuation amplitudes of multiple adjacent intervals is processed using a sliding window method. The weighting coefficients are preset based on the correlation between the central interval and neighboring intervals, resulting in a smoothed fluctuation amplitude sequence. The set of angle positions where the fluctuation amplitude exceeds a preset benchmark value is recorded. For each position in the set of angle positions, the ball trajectory offset data within a preset range before and after the corresponding angle is extracted. The root mean square value of the offset data is calculated as a local offset intensity index. The linear correlation between the fluctuation amplitude sequence and the local offset intensity sequence is calculated using a correlation analysis method. If the correlation exceeds a preset correlation threshold, a strong correlation is determined between the two. Based on the correlation value of the strong correlation, an adjustment mechanism for the discrimination threshold is constructed. The ratio of the fluctuation amplitude to the offset intensity at the current angular position is used as the adjustment basis. When the ratio is within a preset normal range, the original discrimination threshold remains unchanged. When the ratio is below the lower limit of the normal range, the discrimination threshold is decreased according to a preset decreasing rule. When the ratio is above the upper limit of the normal range, the discrimination threshold is increased to the set upper limit according to a preset increasing rule. The adjustment mechanism yields the local discrimination threshold corresponding to each angular position. Interpolation processing is performed on the local discrimination threshold sequence to generate a continuous threshold curve covering the complete rotation cycle, resulting in a dynamically adjusted discrimination threshold adapted to different rotation angle characteristics.

[0039] For example, in one implementation, the segmented processing of joint torsional torque data employs an equal-angle interval division method, dividing the complete 360° rotation cycle into 12 basic intervals of 30°. The number of torque sampling points contained in each interval depends on the ratio of the sensor sampling frequency to the joint rotation speed. The local fluctuation amplitude is obtained by calculating the difference between the maximum and minimum values ​​of all sampling points within the interval; this difference reflects the degree of torque variation of the joint within a specific angular range. The sliding window method selects three consecutive adjacent intervals as processing units, with the weight of the central interval set to 0.5 and the weights of the adjacent intervals on both sides each set to 0.25. This weight allocation considers the balance between angular continuity and local correlation.

[0040] It should be noted that the smoothed fluctuation amplitude sequence is calculated using a weighted average. The smoothed value at each location is equal to the sum of the products of the original fluctuation amplitude at that location and its immediate and adjacent locations, and their corresponding weights. The preset benchmark value is determined based on historical data statistics under normal installation conditions. Typically, the average normal fluctuation amplitude plus twice the standard deviation is taken as the threshold for judging anomalies. When the smoothed fluctuation amplitude at a certain angle exceeds this benchmark value, that location is recorded in the angle location set, forming an anomaly candidate region that requires focused analysis.

[0041] Specifically, the extraction range for ball trajectory offset data is set as a 30° sector area, 15° before and after the target angle position. Within this area, the deviation values ​​between the actual ball trajectory and the theoretical trajectory, obtained by a laser displacement sensor or visual inspection device, constitute the offset data sequence. The local offset intensity index is calculated using the root mean square (RMS) method. First, all offset values ​​within the area are squared, then the average is calculated, and finally the square root is taken to obtain the RMS value. This calculation method comprehensively reflects the average level and dispersion of the offset, avoiding the problem of positive and negative offsets canceling each other out. Correlation analysis is achieved by dividing the covariance of the fluctuation amplitude sequence and the offset intensity sequence by the product of their standard deviations. The resulting correlation coefficient ranges from -1 to 1, with the absolute value closer to 1 indicating a stronger linear correlation.

[0042] For example, the core of the discrimination threshold adjustment mechanism lies in establishing a mapping relationship between the ratio of fluctuation amplitude to offset intensity and the adjustment coefficient. When the ratio is in the normal range of 0.8 to 1.2, it indicates that the torque fluctuation and ball offset maintain a reasonable correspondence. At this time, the adjustment coefficient remains at 1.0, that is, the original discrimination threshold remains unchanged. When the ratio is below 0.8, it indicates that there is a large ball offset but the torque fluctuation is relatively small. In this case, the discrimination threshold is lowered to improve the detection sensitivity, and the adjustment coefficient decreases from 1.0 to the lower limit of 0.7 according to a linear function. When the ratio is above 1.2, it indicates that the torque fluctuation is abnormally large while the ball offset is relatively small. The discrimination threshold is increased to avoid false judgments. The adjustment coefficient increases from 1.0 according to a logarithmic function, but does not exceed the upper limit of 1.5. The use of the logarithmic function makes the adjustment amplitude increase rapidly in the early stage and then tend to level off in the later stage.

[0043] In one possible implementation, the local discrimination threshold for each angular position is calculated by multiplying the original threshold by the corresponding adjustment coefficient. Since the sampling points in the set of angular positions are discretely distributed, directly using this method would cause discontinuous jumps in the threshold curve, affecting the stability of the detection.

[0044] Preferably, the interpolation process employs a piecewise polynomial fitting method to construct a smooth transition curve between known local discrimination threshold angle positions. The interpolation function needs to satisfy three constraints: equal function values ​​at known points, continuous first derivative, and continuous second derivative, ensuring that the generated threshold curve is not only numerically continuous but also maintains a smooth transition in rate of change. For angle regions without anomaly markers, the interpolation function automatically fills in reasonable threshold values ​​to avoid the occurrence of detection blind spots.

[0045] For example, a significant torque fluctuation anomaly was detected in the knee joint of a humanoid robot within an angle range of 120° to 150°. The average fluctuation amplitude in this region was 2.5 N·m, while the corresponding ball bearing offset intensity was 0.08 mm. The calculated ratio was 31.25 mm / N·m, far exceeding the upper limit of the normal range. The system adjusted the discrimination threshold for this region from the standard value of 3.0 N·m to 4.2 N·m. Through interpolation, the thresholds for adjacent angle regions were also adjusted accordingly. Within the 105° to 120° range, the threshold gradually transitioned from 3.0 N·m to 4.2 N·m, and within the 150° to 165° range, it gradually decreased back to 3.0 N·m, forming a continuous and smooth dynamic threshold curve. The dynamically adjusted discrimination threshold can adapt to changes in bearing contact state at different angle positions, improving the accuracy and robustness of anomaly detection. This method is particularly suitable for scenarios where there are local defects in bearing installation or abnormal force on the joint within a specific angle range. The dynamic threshold adjustment avoids the problems of missed detections or false alarms caused by fixed thresholds.

[0046] Step S104: Based on the dynamically adjusted discrimination threshold, compare the real-time collected joint torsional torque data to determine the severity of the abnormal state.

[0047] Based on the obtained dynamic adjustment discrimination threshold, the real-time collected joint torsional torque data is compared point by point. The difference between the torque value at each sampling moment and the corresponding angle position threshold is calculated. When the difference is positive, it is recorded as an out-of-limit point. The proportion of out-of-limit points within a consecutive preset time window is counted to obtain the real-time fluctuation deviation rate. After obtaining the real-time fluctuation deviation rate, the mean value of the ball trajectory offset data within the same time window is extracted. The ratio of the fluctuation deviation rate to the mean value of the offset data is calculated. When the ratio is in the preset mild abnormality range, it is marked as mild abnormality; when the ratio is in the moderate abnormality range, it is marked as moderate abnormality; when the ratio exceeds the severe abnormality threshold, it is marked as severe abnormality. Based on the abnormality marking results and the concentration of out-of-limit points in the angle distribution, when the out-of-limit points are concentrated in a continuous range smaller than the preset angle and are mild abnormalities, it is judged as a local slight abnormality. When the out-of-limit points are distributed beyond the preset angle range or are moderate or above abnormalities, the severity of the abnormal state is determined by a weighted calculation of the distribution range and the abnormality level.

[0048] For example, in one implementation, the comparison process of real-time torque data is based on the difference analysis between a dynamically adjusted discrimination threshold and the actual acquired value. The torque value at each sampling moment is compared in real time with the threshold at its corresponding angular position. When the torque value exceeds the threshold, the sampling point is marked as an out-of-limit point. A preset time window is typically set to include the duration of a complete rotation cycle to ensure that abnormal joint performance at various angular positions can be captured. The percentage of out-of-limit points is calculated by dividing the total number of out-of-limit points by the total number of sampling points within the time window; the resulting ratio is the real-time fluctuation deviation rate.

[0049] It should be noted that the extraction of the mean value of the ball bearing trajectory offset data needs to be synchronized with the torque data. Within the same time window, the arithmetic mean of all offset data points is calculated to obtain the overall offset level for that period. The ratio of the fluctuation deviation rate to the mean value of the offset data reflects the relative relationship between torque anomaly and mechanical offset. The mild anomaly range is typically set at a ratio between 1.0 and 1.5, indicating that torque fluctuation and offset are basically proportional. The moderate anomaly range is 1.5 to 2.5, indicating that torque fluctuation is amplified relative to offset. When the ratio exceeds the severe anomaly threshold of 2.5, it indicates a serious torque imbalance.

[0050] Specifically, the concentration of out-of-limit points in terms of angular distribution is quantified by calculating the angular span of continuous out-of-limit areas. When out-of-limit points are concentrated in a continuous angular range of less than 60° and the anomaly level is minor, the system determines it as a local minor anomaly, which is usually caused by localized bearing wear or contaminant accumulation.

[0051] For example, when the distribution range of the out-of-limit points exceeds 120° or the anomaly level reaches moderate or above, a weighted calculation is required. The weighting coefficients are set considering two factors: the proportion of the angular distribution range to the entire period as the spatial weight, and the coefficient corresponding to the anomaly level as the intensity weight. The spatial weight is calculated by dividing the distribution angle by 360°, and the intensity weights are 0.3 for mild, 0.6 for moderate, and 1.0 for severe. The two weights are multiplied and compared with a baseline value. A product less than 0.2 indicates a slight overall anomaly, 0.2 to 0.5 indicates moderate to severe, and a product exceeding 0.5 indicates a severe anomaly.

[0052] Preferably, the final determination of the severity of the abnormal state also considers the duration of the abnormality. If anomalies of the same severity are detected in multiple consecutive time windows, the severity rating is automatically increased.

[0053] Step S105: Based on the severity of the abnormal state, simulate and identify the joint rotation resistance, and generate a predicted value for the change in joint fit clearance.

[0054] After determining the severity of the abnormal state, the peak torque sequence within the abnormal angle range is extracted from the joint torsional torque data, and the maximum offset sequence corresponding to the angle is extracted from the ball trajectory offset data. By calculating the phase difference between the two sequences on the time axis, the time delay relationship between torque and offset is obtained, and the actual contact state distribution inside the joint is determined based on the time delay relationship. Based on the actual contact state distribution, a numerical simulation method is used to construct a calculation model for the joint rotational resistance. The input parameters include the correction coefficient corresponding to the severity of the abnormality, the peak torque distribution, and the ball offset distribution. By solving the contact pressure distribution between the bearing raceway and the balls, when the pressure distribution change is less than a preset convergence threshold, the resistance distribution curves at each angle position are output. Based on the resistance distribution curves, the positions where the resistance change rate exceeds the preset threshold are identified as abrupt change points through gradient calculation. The resistance gradient value at the abrupt change point reflects the clearance change rate between the mating surfaces. By accumulating the change rate along the angle axis, the cumulative clearance change is obtained, and the cause of the clearance is determined based on the trend of the cumulative change. Based on the determination of the cause of the gap, a linear fitting relationship between the gap value and the joint running time is established. The slope of the fitted line is extracted as the gap change rate. Combined with the current cumulative gap value and the change rate, the predicted gap value within a preset time period is calculated to obtain the predicted value of the joint fit gap change.

[0055] For example, in one implementation, the extraction of the torque peak sequence is based on the precise location of the abnormal angle range. First, the angle range corresponding to the severity of the abnormal state is identified, typically a continuous interval from 60° to 180°. Within this range, torque data is extracted at 1° sampling intervals. The peak value at each angle position is determined by comparing the torque values ​​of the previous five sampling points, forming a discrete peak sequence. The maximum value extraction of the ball trajectory offset data uses the same angle correspondence to ensure the consistency of the two sequences in the spatial dimension. The phase difference is calculated by sliding and comparing the two sequences on the time axis. When the correlation between the two sequences reaches its maximum, the corresponding time offset is the phase difference. This phase difference reflects the response delay between torque change and mechanical offset; a longer delay indicates a more complex force transmission path within the bearing.

[0056] It should be noted that the determination of the actual contact state distribution depends on the variation of the phase difference at different angular positions. When the phase difference remains constant throughout the abnormal range, it indicates that the contact state between the bearing and the joint housing is uniform, and the contact pressure is symmetrically distributed along the circumferential direction. When the phase difference exhibits periodic fluctuations, the fluctuation period corresponds to the distribution interval of the bearing balls, and the fluctuation amplitude reflects the unevenness of the load on each ball. By mapping the phase difference value to a polar coordinate system, a distribution diagram describing the contact state is formed. In the diagram, the radial coordinate represents the contact pressure intensity, and the angular coordinate corresponds to the joint rotation position.

[0057] Specifically, the numerical simulation method employs a discretized contact mechanics model, dividing the bearing raceway into several micro-elements. The contact pressure of each element is calculated using Hertzian contact theory. Correction coefficients in the input parameters range from 0.5 to 2.0, depending on the severity of the anomaly, and are used to adjust the baseline value of the contact stiffness. The peak torque distribution is applied as a boundary condition to the outer ring nodes of the model, while the ball offset distribution is converted into the initial deviation of the ball center position. The solution process uses the Newton-Raphson iterative method, updating the contact pressure and deformation of each element in each iteration. Convergence is considered achieved when the pressure change rate between two adjacent iterations is less than 0.01%. The converged pressure distribution is then integrated to obtain the total resistance value at each angular position, forming a complete resistance distribution curve.

[0058] For example, the identification of resistance abrupt change points is achieved by calculating the first derivative of the resistance curve. The gradient value is calculated using a five-point difference formula for discrete resistance data points. When the absolute value of the gradient at a point exceeds three times the average gradient, it is marked as a sudden change point. The physical significance of a sudden change point lies in the abrupt change in the fit between the bearing and the raceway at that location, typically corresponding to the boundary position of a wear step or assembly defect.

[0059] Preferably, the clearance change rate is calculated by combining the gradient value at the abrupt change point with the bearing's geometric parameters. Dividing the gradient value by the bearing's radial stiffness coefficient yields the clearance change per unit angle. This change is accumulated over the entire angle range to form a cumulative clearance curve. When the cumulative curve exhibits a monotonically increasing trend and a relatively constant slope, it is determined to be clearance expansion caused by uniform wear. When the cumulative curve exhibits periodic fluctuations with alternating peaks and troughs, it is determined to be uneven clearance distribution caused by assembly deviations.

[0060] In one possible implementation, the linear fitting relationship is established using the least squares method. Gap measurements of the joint are collected at different operating durations, typically including data from multiple time points such as initial installation, 100 hours of operation, 500 hours, and 1000 hours. Using duration as the independent variable and gap value as the dependent variable, a linear equation of the form y = ax + b is fitted, where the slope 'a' represents the gap change rate in micrometers per hour.

[0061] For example, after 1000 hours of operation, the cumulative gap of a humanoid robot's hip joint is measured to be 45 micrometers, and the fitted gap change rate is 0.04 micrometers / hour. Based on these two parameters, the predicted gap value after 500 hours is 45 + 0.04 × 500 = 65 micrometers. When the predicted value exceeds the design-allowed upper limit of 80 micrometers, the system issues an early maintenance warning.

[0062] Understandably, the accuracy of gap prediction depends on the completeness of historical data and the applicability of the fitted model. For newly installed joints, gap changes during the initial break-in period may exhibit non-linear characteristics, requiring piecewise fitting or polynomial fitting methods to improve prediction accuracy. Accurate prediction of gap changes can be achieved by continuously updating historical data and dynamically adjusting the fitting parameters.

[0063] Step S106: Update the dynamic adjustment discrimination threshold based on the predicted value of the joint fit gap change.

[0064] Based on the predicted value of the joint clearance change, it is compared with the preset normal clearance range. The normal range is determined as the upper and lower limits based on the bearing model and assembly standard. When the predicted value is within the range, the current force sensor acquisition parameters remain unchanged. When the predicted value exceeds the range, the ratio of the difference between the predicted value and the range boundary to the range width is calculated as the deviation rate, providing a quantitative basis for feedback adjustment. The force sensor operating parameters during the robotic arm bearing assembly process are corrected using the deviation rate. If the predicted value exceeds the upper limit, the sampling frequency is linearly increased according to the deviation rate, while the trigger threshold is lowered to make the detection more sensitive. If the predicted value is below the lower limit, the sampling frequency is reduced and the trigger threshold is increased according to the deviation rate. The corrected force sensor operating parameters serve as the new monitoring benchmark. Using the new monitoring benchmark and deviation rate, the anomaly detection threshold is recalculated. The updated value is obtained by multiplying the original threshold by the compensation factor of the deviation rate. When the deviation rate is large, the compensation factor is greater than 1, which loosens the threshold; when the deviation rate is small, the compensation factor is less than 1, which tightens the threshold, thus optimizing anomaly identification.

[0065] For example, in one implementation, the normal range of joint clearance is preset according to the bearing model and assembly process requirements. For standard deep groove ball bearings, the normal clearance range is typically set to 10 to 50 micrometers, and the upper and lower limits of this range are stored in the system parameter library. When the predicted clearance value falls within this range, the bearing assembly status is determined to be normal, and the current force information acquisition strategy remains unchanged.

[0066] It should be noted that the deviation rate is calculated using a relative deviation method. When the predicted value exceeds the upper limit of the normal range, the deviation rate equals the predicted value minus the upper limit value, divided by the range width. When the predicted value is below the lower limit, the deviation rate equals the lower limit value minus the predicted value, divided by the range width. This normalization process ensures that the deviation rate is always positive, facilitating subsequent parameter adjustments.

[0067] Specifically, the correction of the force sensor's operating parameters follows a segmented adjustment principle. When the predicted gap value exceeds the upper limit, it indicates accelerated bearing wear, requiring increased monitoring accuracy. In this case, the sampling frequency is increased by multiplying the baseline value by 1 and adding the deviation rate. For example, if the deviation rate is 0.3, the sampling frequency is increased to 1.3 times the original value. Simultaneously, the abnormal trigger threshold is reduced to the original value divided by 1 plus the deviation rate, making the system more sensitive to minor anomalies. Conversely, when the predicted value is below the lower limit, it indicates that the bearing is too tight, and the sampling frequency and trigger threshold are adjusted in the opposite direction.

[0068] For example, the compensation factor is determined based on a nonlinear mapping relationship of the deviation rate. When the deviation rate is less than 0.2, the compensation factor adopts a linear relationship, equal to 1 plus 0.5 times the deviation rate. When the deviation rate is between 0.2 and 0.5, the compensation factor increases logarithmically to ensure a smooth transition in adjustment. When the deviation rate exceeds 0.5, the compensation factor is fixed at an upper limit of 1.5 to avoid over-adjustment leading to system instability.

[0069] Preferably, the anomaly detection threshold is updated using a gradual adjustment strategy. The new threshold is equal to the original threshold multiplied by a compensation factor, but the adjustment amount does not exceed 50% of the original value each time. This limitation ensures that the system maintains basic anomaly detection capability while adapting to interval changes.

[0070] For example, the predicted clearance of a knee joint bearing is 65 micrometers, exceeding the upper limit of 50 micrometers, with a calculated deviation rate of 0.375. By increasing the sampling frequency from 1000Hz to 1375Hz, decreasing the trigger threshold from 3 N·m to 2.18 N·m, and adjusting the anomaly discrimination threshold to 115% of its original value using a compensation factor of 1.15, dynamic optimization of the sensitivity for anomaly detection is achieved.

[0071] Step S107: After updating the dynamically adjusted discrimination threshold, generate the joint assembly control configuration.

[0072] After obtaining the optimized anomaly identification results, the offset values ​​of the ball trajectory offset data at each angular position are extracted, and the offset change rate within adjacent angular intervals is statistically analyzed. Simultaneously, the torque change rate of the joint torsional torque data at the corresponding angle is extracted. By calculating the correlation coefficient between the two sets of change rate sequences, a high matching degree is determined when the correlation coefficient is greater than a preset threshold; otherwise, a low matching degree is determined, resulting in a matching degree evaluation result. Based on the matching degree evaluation result, a compensation relationship for joint rotational resistance is constructed. The compensation value in the high matching degree region is equal to the product of the torque deviation and the preset compensation coefficient. The low matching degree region is divided into compensation intervals based on the peak position of the offset data, with each interval having an independent compensation coefficient, forming a resistance correction value sequence. The control parameters of the joint drive motor are adjusted using the resistance correction value sequence. If the correction value is positive and exceeds the first threshold, the position loop gain is increased and the speed loop bandwidth is decreased. If the correction value is negative and below the second threshold, the position loop gain is decreased and the speed loop bandwidth is increased, resulting in an optimized servo control parameter set. The servo control parameter set is used to generate a joint assembly control configuration. The configuration includes the drive current limit, acceleration constraint conditions and position error tolerance range for each angular position. By mapping the control parameter set to the assembly process parameters, a joint assembly control configuration with improved motion accuracy is obtained.

[0073] For example, in one implementation, the ball trajectory offset data is extracted based on the marked key angle positions in the optimized anomaly identification results. Offset values ​​are densely sampled at each marked position and within a 5° range before and after it, forming a local offset sequence. The offset change rate is obtained by differential calculation of adjacent sampling points, reflecting the velocity variation characteristics of the ball moving in the raceway. The change rate of the joint torsional torque is calculated using the same differential method, but the torque data needs to be low-pass filtered first to eliminate high-frequency noise interference. The correlation coefficient between the two sets of change rate sequences is obtained by the inner product operation after standardization; the coefficient range is between -1 and 1, with a larger absolute value indicating a stronger correlation.

[0074] It should be noted that the threshold setting for the matching degree evaluation takes into account the bearing type and assembly precision level. For precision bearings, the correlation coefficient threshold is usually set to 0.7. When the calculated coefficient is greater than this value, it indicates that the torque change and trajectory deviation show a synchronous response, which is judged as a high matching degree. This high matching degree usually occurs when the bearing wear is uniform or the assembly is good. A low matching degree indicates that there is an asynchronous phenomenon between the two, which may be caused by local defects or foreign object intrusion.

[0075] Specifically, the compensation relationship is constructed using a piecewise linear mapping method. In the high-matching region, since the response of torque and offset has a good linear relationship, the compensation value is directly equal to the torque deviation multiplied by a preset compensation coefficient, which is typically between 0.8 and 1.2, determined according to the bearing specifications and load conditions. The processing of the low-matching region is more complex. First, local peak points in the offset data are identified; these peak points correspond to the locations where the balls have passed through damaged or deformed areas. The entire angle range is divided into multiple compensation intervals using the peak points as boundaries, with each interval typically ranging from 15° to 30° in width. The compensation coefficient within an interval is dynamically determined based on the ratio of the peak offset to the average offset within that interval; the larger the ratio, the higher the compensation coefficient, up to a maximum of 2.0. This piecewise compensation strategy allows for appropriate corrections to be applied to different regions based on their degree of anomaly.

[0076] For example, the application of the resistance correction value sequence is reflected in the fine-tuning of drive motor control parameters. The position loop gain determines the joint's response speed to position deviations, while the speed loop bandwidth affects the system's dynamic tracking capability. When the resistance correction value is positive and exceeds the first threshold of 0.5 N·m, it indicates that there is significant motion resistance at that angular position. Increasing the position loop gain by 20% to 30% allows the motor to output a larger driving torque to overcome the resistance. Simultaneously, reducing the speed loop bandwidth by 15% to 25% prevents system oscillations. When the correction value is negative and below the second threshold of -0.3 N·m, it indicates that the resistance at that position is too low, potentially indicating excessive clearance. In this case, reducing the position loop gain and increasing the speed loop bandwidth improves the system's response speed and stability.

[0077] Preferably, the optimization of the servo control parameter set also includes adjusting the current loop response time. In the high-resistance region, the current loop response time is appropriately extended to 3 to 5 milliseconds to allow the motor sufficient time to build up the required drive current. In the low-resistance region, the response time is shortened to 1 to 2 milliseconds to improve the dynamic response capability of the system.

[0078] In one possible implementation, the mapping from control parameter sets to assembly process parameters is achieved through a lookup table. A pre-established table mapping control parameters to process parameters is used, containing drive current limits, acceleration constraints, and permissible position error ranges for different parameter combinations. The drive current limit is set based on the motor's heat capacity and heat dissipation conditions, typically 1.2 to 1.5 times the rated current. The acceleration constraint considers the strength limit of the mechanical structure to prevent structural damage caused by excessive acceleration or deceleration.

[0079] For example, if a shoulder joint detects high matching degree and a positive resistance correction value of 1.2 N·m at a 90° position, the system sets the drive current limit for that position to 8 A, the acceleration constraint to 500° / s², and the position error tolerance to ±0.5°. Conversely, if a low matching degree and a negative correction value of -0.8 N·m are detected at a 180° position, the corresponding parameters are adjusted to a current limit of 6 A, an acceleration constraint to 700° / s², and a position error tolerance to ±0.3°. The implementation of this assembly control configuration significantly improves the joint's motion accuracy. By differentiating control for different angular positions, the system compensates for motion unevenness caused by bearing defects, ensuring stable motion characteristics throughout the joint's full stroke range, and improving position control accuracy to within ±0.2°.

[0080] Step S108: Generate a joint assembly control strategy based on the joint assembly control configuration.

[0081] Based on the obtained joint assembly control configuration, the torsional torque data of each joint on the production line is monitored in real time. When the torque value exceeds the threshold set in the configuration, the drive current is adjusted according to the threshold deviation ratio. The larger the deviation, the larger the current adjustment. At the same time, the rotational speed is reduced to a preset safety value to bring the torque value back to the normal range. The torque change before and after the adjustment is recorded as a compensation record. Through the compensation record, the cumulative compensation amount in each monitoring cycle is calculated. If the cumulative amount shows a decreasing trend and the ball bearing trajectory offset data decreases synchronously, the current adjustment is determined to be effective, and the existing control parameters are maintained. If the offset data does not decrease or increases in the opposite direction, the compensation coefficient is increased by a preset increment value and the control configuration is updated. The updated control configuration is used to continuously collect joint operation data, establish a time series database to store torque and offset information, and extract the trend by calculating the difference between the average values ​​of data in adjacent time periods. When the trend slope exceeds a preset degradation threshold, a maintenance warning is generated, resulting in a control strategy output to extend the service life of the joint bearing.

[0082] For example, in one implementation, real-time monitoring is performed based on a preset torque threshold range in the joint assembly control configuration. Each joint is equipped with an independent monitoring unit that acquires torsional torque data at a frequency of 100Hz. When a torque value exceeding the upper limit threshold is detected, the excess ratio is calculated, which is the difference between the actual torque value and the threshold divided by the threshold. The adjustment amount of the drive current is equal to this ratio multiplied by 20% of the rated current, ensuring that the adjustment range is proportional to the degree of abnormality. At the same time, the rotational speed is reduced to 60% of the rated value to prevent high-speed operation from exacerbating the abnormality.

[0083] It should be noted that the compensation record includes four elements: adjustment time, original torque value, adjusted torque value, and adjustment parameters. The monitoring period is hourly, and the sum of torque changes across all compensation records within the statistical period is used as the cumulative compensation. Trend judgment employs a linear regression method, fitting the cumulative compensation over three consecutive periods; a negative slope indicates a decreasing trend.

[0084] Specifically, the synchronization of ball bearing trajectory offset data is achieved through timestamp alignment. The direction of change of the offset data within the same time period is compared with the direction of change of the cumulative compensation amount; if both change in the same direction, the adjustment is considered effective. If they change in opposite directions, it indicates that the current compensation strategy has failed to improve the bearing condition, and the compensation coefficient needs to be adjusted. The increment of the compensation coefficient is usually set to 10% of the original value to avoid over-adjustment that could cause system oscillation.

[0085] For example, the time-series database employs a circular buffer structure to store historical data for the most recent 72 hours. Trend extraction is achieved by calculating the mean difference between adjacent 6-hour data windows. The degradation threshold is set based on the bearing model and operating environment, typically 1.5 times the normal wear rate. When the trend slope exceeds this threshold, the bearing is determined to have entered an accelerated degradation phase.

[0086] Preferably, maintenance warnings are divided into three levels: a yellow warning indicates that monitoring frequency needs to be increased, an orange warning suggests scheduling a maintenance plan, and a red warning requires immediate shutdown for repair. The control strategy output includes three items: warning level, recommended maintenance time, and estimated remaining lifespan.

[0087] For example, after 1000 hours of continuous use, the cumulative compensation of a certain ankle joint increased from an initial 0.5 N·m·h to 2.8 N·m·h, with a trend slope reaching 0.003 N·m·h / h, exceeding the preset degradation threshold of 0.002 N·m·h / h. An orange warning was issued, recommending maintenance within 200 hours, with an estimated remaining service life of 800 hours. Through this preventative maintenance strategy, the actual service life of the joint bearing can be extended by more than 30%.

[0088] This invention provides an artificial intelligence-based system for identifying abnormal conditions on a production line, mainly comprising: The evaluation module is used to collect joint torsional torque data and ball trajectory offset data of humanoid robot joints at different rotation angles, and generate evaluation results of bearing installation status. The first determining module is used to determine the abnormal state of bearing installation based on the evaluation results of the bearing installation state. The discrimination module is used to generate a dynamically adjusted discrimination threshold by analyzing the correlation between the fluctuation amplitude of the joint torsional torque data at different rotation angles and the ball trajectory offset data; The second determining module is used to determine the severity of the abnormal state by comparing the dynamically adjusted discrimination threshold with the real-time collected joint torsional torque data. The prediction module is used to simulate and identify joint rotation resistance based on the severity of the abnormal state, and generate a predicted value for the change in joint fit clearance. An update module is used to update the dynamically adjusted discrimination threshold based on the predicted value of the change in the joint fit gap; The generation module is used to generate the joint assembly control configuration after updating the dynamically adjusted discrimination threshold. The control module is used to generate a joint assembly control strategy based on the joint assembly control configuration.

[0089] The preferred embodiments of the present invention disclosed above are merely illustrative of the invention. These preferred embodiments do not exhaustively describe all details, nor do they limit the invention to any specific implementation. Clearly, many modifications and variations can be made based on the content of this specification. This specification selects and specifically describes these embodiments to better explain the principles and practical applications of the invention, thereby enabling those skilled in the art to better understand and utilize the invention. The invention is limited only by the claims and their full scope and equivalents.

Claims

1. An artificial intelligence method for identifying abnormal conditions in a production line, characterized in that, include: Collect joint torsional torque data and ball trajectory offset data of humanoid robot joints at different rotation angles to generate an evaluation result of bearing installation status; Based on the evaluation results of the bearing installation status, an abnormal state of bearing installation is determined; by analyzing the correlation between the fluctuation amplitude of the joint torsional torque data at different rotation angles and the ball trajectory offset data, a dynamically adjusted discrimination threshold is generated; based on the dynamically adjusted discrimination threshold, the severity of the abnormal state is determined by comparing the real-time collected joint torsional torque data. Based on the severity of the abnormal state, the joint rotation resistance is simulated and identified, and a predicted value of the change in joint fit clearance is generated. Based on the predicted value of the change in the joint fit gap, the discrimination threshold of the dynamic adjustment is updated; after updating the discrimination threshold of the dynamic adjustment, the joint assembly control configuration is generated. A joint assembly control strategy is generated based on the joint assembly control configuration.

2. The artificial intelligence identification method for abnormal conditions of a production line according to claim 1, characterized in that, The process involves collecting joint torsional torque data and ball bearing trajectory offset data from the humanoid robot's joints at different rotation angles to generate an evaluation result of the bearing installation status, including: The raw sequence of the joint torsional torque data is collected, and the torque change rate between adjacent sampling points is calculated; the ball trajectory offset data is collected, the periodic change of the offset data is calculated, and the distribution characteristics of the offset trajectory are determined; a torque distribution map is constructed based on the torque change rate and the distribution characteristics of the offset trajectory, the feature parameters of the map are extracted, and a quantitative evaluation result of the bearing installation status is generated.

3. The artificial intelligence identification method for abnormal conditions of a production line according to claim 1, characterized in that, The step of determining the abnormal state of bearing installation based on the evaluation results of the bearing installation status includes: Based on the evaluation results of the bearing installation status, a clustering algorithm is used to group the joint torsional torque data and the ball trajectory offset data to generate preliminary status grouping results; the time-domain features and distribution density of each group in the preliminary status grouping results are extracted to determine the bearing installation anomaly; by calculating the distance metric between the abnormal group and the normal group, the degree of anomaly is quantified, and an abnormal status diagnosis result is generated.

4. The artificial intelligence identification method for abnormal conditions of a production line according to claim 1, characterized in that, The step of generating a dynamically adjusted discrimination threshold by analyzing the correlation between the fluctuation amplitude of the joint torsional torque data at different rotation angles and the ball trajectory offset data includes: The fluctuation amplitude of the joint torsional moment data is calculated in segments to generate a smoothed fluctuation amplitude sequence; the local strength index of the ball trajectory offset data is extracted, and the correlation coefficient between the fluctuation amplitude sequence and the local strength index is calculated; a dynamically adjusted discrimination threshold is generated based on the correlation coefficient.

5. The artificial intelligence identification method for abnormal conditions of a production line according to claim 1, characterized in that, The step of determining the severity of the abnormal state by comparing the dynamically adjusted discrimination threshold with the real-time collected joint torsional torque data includes: The difference between the joint torsional torque data and the dynamically adjusted discrimination threshold is calculated using the dynamically adjusted discrimination threshold to generate a fluctuation deviation rate; the abnormality level is marked according to the ball trajectory offset data and the fluctuation deviation rate, and the severity of the abnormal state is generated.

6. The artificial intelligence identification method for abnormal conditions of a production line according to claim 1, characterized in that, The step of simulating and identifying joint rotation resistance and generating predicted values ​​for changes in joint clearance based on the severity of the abnormal state includes: The time delay relationship between the peak sequence of the joint torsional torque data and the maximum offset sequence of the ball trajectory offset data is calculated to construct a simulation calculation model of the joint rotation resistance and generate a resistance distribution curve. The abrupt change point of the resistance distribution curve is identified by gradient calculation to generate the cumulative clearance change. The clearance change rate is fitted according to the cumulative clearance change to generate a predicted value of the joint fit clearance change.

7. The artificial intelligence identification method for abnormal conditions of a production line according to claim 1, characterized in that, The step of updating the dynamically adjusted discrimination threshold based on the predicted value of the change in the joint fit clearance includes: The deviation rate is calculated by comparing the predicted value of the joint fit gap change with a preset range; the dynamic adjustment discrimination threshold is updated based on the deviation rate.

8. The artificial intelligence identification method for abnormal conditions of a production line according to claim 1, characterized in that, The generated joint assembly control configuration includes: Extract the rate of change of the ball trajectory offset data and the joint torsional torque data, calculate the correlation coefficient, and generate a matching degree evaluation result; generate a resistance correction value sequence based on the matching degree evaluation result; adjust the drive motor control parameters through the resistance correction value sequence to generate a servo control parameter group; map the assembly process parameters through the servo control parameter group to generate the joint assembly control configuration.

9. The artificial intelligence identification method for abnormal conditions of a production line according to claim 1, characterized in that, The step of generating a control strategy based on the joint assembly control configuration includes: The joint assembly control configuration monitors the joint torsional torque data, adjusts the drive current and speed, and generates compensation records. The cumulative compensation amount is calculated using the compensation records, the changing trend of the ball trajectory offset data is verified, and a joint assembly control strategy for extending bearing service life is generated.

10. An artificial intelligence-based system for identifying abnormal conditions on a production line, characterized in that, include: The evaluation module is used to collect joint torsional torque data and ball trajectory offset data of humanoid robot joints at different rotation angles, and generate evaluation results of bearing installation status. The first determining module is used to determine the abnormal state of bearing installation based on the evaluation results of the bearing installation state. The discrimination module is used to generate a dynamically adjusted discrimination threshold by analyzing the correlation between the fluctuation amplitude of the joint torsional torque data at different rotation angles and the ball trajectory offset data; The second determining module is used to determine the severity of the abnormal state by comparing the dynamically adjusted discrimination threshold with the real-time collected joint torsional torque data. The prediction module is used to simulate and identify joint rotation resistance based on the severity of the abnormal state, and generate a predicted value for the change in joint fit clearance. An update module is used to update the dynamically adjusted discrimination threshold based on the predicted value of the change in the joint fit gap; The generation module is used to generate the joint assembly control configuration after updating the dynamically adjusted discrimination threshold. The control module is used to generate a joint assembly control strategy based on the joint assembly control configuration.

Images