Edge buffer adaptive arbitration sampling method and device for industrial robot

Abstract

The application discloses an industrial robot edge cache adaptive arbitration sampling method and equipment, and belongs to the technical field of industrial robot state monitoring. In view of the problems of fixed window data redundancy, high false alarm rate and low review value when the edge end and the cloud end diagnosis results conflict, the method comprises the following steps: constructing a high-frequency data cycle cache at the edge end, taking edge risk, cloud risk and remaining service life prediction results as multi-source diagnosis evidence, and calculating evidence conflict entropy score; based on change point detection, dynamically generating an adaptive sampling window, dividing candidate segments in the window, quantifying segment review value by information gain, and uploading key segments under bandwidth constraints; outputting health determination results through cloud review fusion, and online correcting algorithm parameters through a feedback reward mechanism. The application significantly reduces data upload volume and working condition false alarm, improves the reliability of diagnosis in conflict scenarios, and is suitable for industrial robot production line collaborative health monitoring.

Description

Technical Field

[0001] This application relates to the field of industrial robot condition monitoring technology, and in particular to an adaptive arbitration sampling method and device for edge buffering of industrial robots. Background Technology

[0002] As the number of industrial robots deployed in smart manufacturing production lines continues to grow, the operational status of core components such as joint motors, reducers, and transmission mechanisms directly affects production line cycle time and product quality. However, unplanned robot downtime can lead to a decrease in overall line efficiency and an increase in maintenance costs.

[0003] Currently, predictive maintenance of industrial robots mainly adopts two technical approaches: one is to continuously upload raw high-frequency data such as vibration and current to the cloud for unified diagnosis and analysis. Although this method can ensure data integrity, it has high bandwidth consumption, storage costs, and cloud computing pressure. The other approach is to perform initial screening at the edge using threshold rules or feature summaries, and only upload abnormal alarms or feature values. This method has lower implementation costs, but it is easily affected by operating condition disturbances such as load switching, fixture replacement, and trajectory changes, leading to frequent false alarms or missed alarms.

[0004] In actual production line operation, inconsistencies frequently arise between edge-side rule-based judgments and cloud-based model judgments. Directly trusting one side's judgment may lead to erroneous maintenance decisions; uploading all cached raw data results in a large amount of invalid transmission. Existing methods typically employ fixed threshold triggers and fixed time windows for uploading when conflicts occur, making it difficult to distinguish between short-term fluctuations caused by operating condition changes and persistent anomalies caused by component degradation. Furthermore, the selection of cached segments often relies on single indicators such as anomaly amplitude or temporal proximity, lacking a quantitative evaluation of the segment's verification value. This easily leads to uploading large amounts of data that contribute little to diagnostic conclusions, wasting valuable uplink bandwidth resources. Summary of the Invention

[0005] To address the technical problems existing in the background art, this application provides an adaptive arbitration sampling method and device for edge caching of industrial robots. The method includes: constructing a circular cache of original high-frequency data at the edge of the industrial robot, and simultaneously acquiring the edge risk probability distribution and edge confidence level output from the edge, as well as the cloud risk probability distribution, cloud confidence level, predicted remaining useful life, and confidence interval width of the remaining useful life output from the cloud; calculating the difference between the edge risk probability distribution and the cloud risk probability distribution, and using the degradation risk probability distribution, edge risk probability distribution, and cloud risk probability distribution generated by mapping the predicted remaining useful life and the confidence interval width of the remaining useful life as multi-source diagnostic evidence, calculating the diagnostic uncertainty entropy after fusing the multi-source diagnostic evidence; weighting the difference, diagnostic uncertainty entropy, cloud confidence level, predicted remaining useful life confidence interval width, short-term degradation trend slope, and load mutation factor to obtain an evidence conflict entropy score; when the evidence conflict entropy score is greater than a preset trigger threshold, performing change point detection on the feature sequence corresponding to the original high-frequency data to obtain the anomaly start point and anomaly recovery point, determining the retention time before and after the anomaly based on the evidence conflict entropy score, and then... An adaptive sampling window is determined by considering the normal start point, the abnormal recovery point, the retention time before the abnormality, and the retention time after the abnormality. Within this adaptive sampling window, the original high-frequency data is cyclically cached and divided into several candidate segments according to a preset time period. The information gain of each candidate segment is calculated. The information gain is the difference in diagnostic uncertainty before and after the candidate segment is uploaded and participates in the review. The abnormal amplitude, spectral abnormality degree, temporal proximity to the time of the abnormality, operating condition matching degree, data volume cost, and the information gain of each candidate segment are weighted to determine the comprehensive score of each candidate segment. Within preset bandwidth constraints and the number of uploaded segments... Under certain conditions, candidate segments with high comprehensive scores are selected and uploaded to the cloud. Based on the cloud's review and diagnosis of the uploaded candidate segments, the probability distribution of review risks and the confidence level of review are obtained. The fusion weight is determined based on the review confidence level, the edge confidence level, and the cloud confidence level. Based on the fusion weight, the probability distributions of edge risks, cloud risks, and review risks are weighted and fused to obtain the health judgment result. A feedback reward function is constructed based on the health judgment result and the manual maintenance confirmation result. Based on the feedback reward function, the weight of evidence conflict entropy calculation, the weight of candidate segment comprehensive score, and the trigger threshold are corrected.

[0006] In one example, calculating the difference between the edge risk probability distribution and the cloud risk probability distribution, and calculating the diagnostic uncertainty entropy after fusing the multi-source diagnostic evidence, specifically includes: calculating the difference between the edge risk probability distribution and the cloud risk probability distribution using Jensen-Shannon divergence; weighting the edge risk probability distribution, cloud risk probability distribution, and degradation risk probability distribution according to their respective confidence levels to obtain a fused risk probability distribution; determining the confidence level corresponding to the degradation risk probability distribution by normalizing the remaining useful life confidence interval width; calculating the Shannon information entropy based on the fused risk probability distribution, and using the Shannon information entropy as the diagnostic uncertainty entropy.

[0007] In one example, the evidence conflict entropy score is obtained by weighting the difference, diagnostic uncertainty entropy, cloud confidence, remaining useful life prediction confidence interval width, short-term degradation trend slope, and load mutation factor. Specifically, this includes: obtaining a sequence of historical remaining useful life prediction values ​​within a preset time window; obtaining the trend slope through linear fitting, which is used as the short-term degradation trend slope; calculating the magnitude of the change in operating conditions between the current moment and the previous steady-state moment based on industrial robot operating condition load data collected at the edge, thus obtaining the load mutation factor; and weighting the remaining useful life prediction confidence interval... The width, short-term degradation trend slope, and load mutation factor are normalized to obtain their respective normalized values. The cloud confidence level is converted into a confidence deficiency term, where the confidence deficiency term is the difference between 1 and the cloud confidence level. The difference degree, diagnostic uncertainty entropy, confidence deficiency term, normalized remaining useful life prediction confidence interval width, and normalized short-term degradation trend slope are used as positive weighting terms, and the normalized load mutation factor is used as a suppression weighting term. The weighted terms are summed using preset weighting coefficients to obtain the evidence conflict entropy score.

[0008] In one example, when the evidence conflict entropy score is greater than a preset trigger threshold, change point detection is performed on the feature sequence corresponding to the original high-frequency data to obtain the anomaly start point and anomaly recovery point. The retention time before and after the anomaly is determined based on the evidence conflict entropy score. An adaptive sampling window is determined based on the anomaly start point, anomaly recovery point, and the retention times before and after the anomaly. Specifically, this includes: when the evidence conflict entropy score is greater than the preset trigger threshold, calculating the moving mean and moving standard deviation of the feature sequence; constructing a standardized deviation based on the difference between the current feature value and the moving mean, combined with the moving standard deviation; constructing a cumulative change statistic based on the standardized deviation, and introducing a drift suppression parameter to offset normal small-amplitude random fluctuations. Cumulative impact; when the cumulative change statistic is greater than a preset anomaly determination threshold, the corresponding time is determined as the anomaly start point; when the cumulative change statistic falls back to below a preset recovery threshold, the corresponding time is determined as the anomaly recovery point; the evidence conflict entropy score is linearly calculated to obtain the retention time before and after the anomaly, and a preset anomaly duration adjustment component is superimposed on the retention time after the anomaly; preset upper and lower limit constraints are applied to the retention time before and after the anomaly, respectively, and the retention time is clipped to the corresponding minimum and maximum limit ranges; the time when the retention time before the anomaly is shifted forward from the anomaly start point is taken as the window start time, and the time when the retention time after the anomaly is shifted backward from the anomaly recovery point is taken as the window end time, thus obtaining an adaptive sampling window.

[0009] In one example, within the adaptive sampling window, the original high-frequency data cyclic cache is divided into several candidate segments according to a preset time period, and the information gain of each candidate segment is calculated. Specifically, this includes: within the adaptive sampling window, the original high-frequency data in the original high-frequency data cyclic cache is segmented according to a preset time period to obtain several candidate segments; the system health status conditional entropy based solely on existing diagnostic evidence from the edge and cloud is used as the diagnostic uncertainty before the candidate segment participates in the review; the system health status conditional entropy after the original data with the corresponding candidate segment is added to participate in the review diagnosis is used as the diagnostic uncertainty after the candidate segment participates in the review; the difference between the diagnostic uncertainty before the review and the diagnostic uncertainty after the review is determined as the information gain of the candidate segment.

[0010] In one example, the comprehensive score for each candidate segment is determined by weighting the abnormal amplitude, spectral abnormality, temporal proximity to the time of the abnormality, operating condition matching degree, and data volume cost, along with the information gain. Specifically, this includes: normalizing the abnormal amplitude, spectral abnormality, temporal proximity, operating condition matching degree, and data volume cost of the candidate segment to a unified numerical range; the abnormal amplitude is the deviation of the signal peak value within the candidate segment from the normal baseline; the spectral abnormality is the amplitude of harmonic energy change in the characteristic frequency band within the candidate segment; the temporal proximity is the proximity of the candidate segment to the time of the abnormality; the operating condition matching degree is the degree of matching between the operating condition corresponding to the candidate segment and the current review operating condition; the information gain, normalized abnormal amplitude, spectral abnormality, temporal proximity, and operating condition matching degree are used as positive scoring items, and the normalized data volume cost is used as a cost suppression item. These are then weighted and summed using preset corresponding weight coefficients to obtain the comprehensive score for each candidate segment.

[0011] In one example, the fusion weights are determined based on the verification confidence, edge confidence, and cloud confidence. Then, based on these fusion weights, the edge risk probability distribution, cloud risk probability distribution, and verification risk probability distribution are weighted and fused to obtain a health assessment result. Specifically, this includes: normalizing the edge confidence, cloud confidence, and verification confidence, and calculating the edge fusion weight, cloud fusion weight, and verification fusion weight respectively; using the edge fusion weight, cloud fusion weight, and verification fusion weight respectively, weighting and summing the edge risk probability distribution, cloud risk probability distribution, and verification risk probability distribution to obtain a final risk probability distribution; and outputting the corresponding health assessment result based on the probability values ​​of various health states in the final risk probability distribution; the health states include normal, suspected abnormal, and abnormal.

[0012] In one example, a feedback reward function is constructed based on the health assessment result and the manual maintenance confirmation result. Based on the feedback reward function, the weights for calculating the evidence conflict entropy, the weights for the comprehensive score of candidate segments, and the trigger threshold are corrected. Specifically, this includes: setting the diagnostic accuracy benefit as a positive reward item, and setting the cost of uploading data, the cost of review latency, the penalty for false positives, and the penalty for false negatives as negative deduction items. The feedback reward function is constructed by weighting each item with corresponding preset weight coefficients. Based on the output value of the feedback reward function and the gradient direction of the segment scoring function relative to the parameter to be corrected, the weights for calculating the evidence conflict entropy, the weights for the comprehensive score of candidate segments, and the trigger threshold are initially iteratively updated. The segment scoring function is a weighted process for calculating the comprehensive score of candidate segments. A preset forgetting coefficient is introduced, and the parameters after the initial update and the historically verified parameter values ​​are subjected to exponential sliding smoothing updates to obtain the final corrected parameters, so as to suppress the drastic fluctuations in parameters caused by a single abnormal sample.

[0013] In one example, when selecting a candidate segment with a high comprehensive score as the upload segment to the cloud, the method further includes: when the conflict of the candidate segment comes from vibration spectrum anomalies, priority is given to uploading vibration waveform candidate segments with high comprehensive scores; when the conflict of the candidate segment comes from current harmonic anomalies, priority is given to uploading current waveform candidate segments with high comprehensive scores; when the conflict of the candidate segment comes from repetitive positioning errors, priority is given to uploading trajectory error sequence candidate segments with high comprehensive scores.

[0014] On the other hand, embodiments of this application provide an adaptive arbitration sampling device for industrial robot edge cache, comprising: at least one processor; and a memory communicatively connected to the at least one processor; wherein the memory stores instructions executable by the at least one processor, the instructions being executed by the at least one processor to enable the at least one processor to perform any of the above-mentioned adaptive arbitration sampling methods for industrial robot edge cache.

[0015] On the other hand, embodiments of this application provide an industrial robot edge cache adaptive arbitration sampling non-volatile computer storage medium storing computer-executable instructions, which can execute any of the above-mentioned industrial robot edge cache adaptive arbitration sampling methods.

[0016] The above-described technical solutions adopted in the embodiments of this application can achieve the following beneficial effects: This solution only uploads critical raw data segments as needed when conflicts arise between edge and cloud diagnostics or when cloud confidence is insufficient. This significantly reduces the amount of raw high-frequency data uploaded, effectively alleviating the burden on network bandwidth, cloud storage, and computing power, and adapting to the limited edge-cloud transmission conditions in industrial settings. By dynamically generating sampling windows through change point detection, the verification value of cached segments is quantified using information gain. Under bandwidth constraints, high-value data is selectively uploaded, avoiding redundant invalid data from fixed-window uploads while ensuring no missing critical fault signals, significantly improving the efficiency and relevance of verification and diagnosis. Furthermore, the solution relies on a feedback reward mechanism to online correct algorithm weights and threshold parameters, adapting to different robot models, production cycles, and load conditions without requiring repeated manual parameter adjustments. It can continuously iterate and optimize as equipment ages and the on-site environment changes. Attached Figure Description

[0017] To more clearly illustrate the technical solution of this application, some embodiments of this application will be described in detail below with reference to the accompanying drawings, in which: Figure 1 A flowchart illustrating an adaptive arbitration sampling method for edge buffering of an industrial robot provided in this application embodiment; Figure 2A schematic diagram illustrating the deployment of an algorithm for an adaptive arbitration sampling method for edge caching of industrial robots, provided in an embodiment of this application, in an edge-cloud system; Figure 3 This is a schematic diagram of the structure of an adaptive arbitration sampling device for edge buffering of an industrial robot, provided in an embodiment of this application. Detailed Implementation

[0018] To make the objectives, technical solutions, and advantages of this application clearer, the technical solutions of this application will be clearly and completely described below in conjunction with specific embodiments and corresponding drawings. Obviously, the described embodiments are only a part of the embodiments of this application, and not all of them. Based on the embodiments in this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0019] Some embodiments of this application will now be described in detail with reference to the accompanying drawings.

[0020] Figure 1 This is a flowchart illustrating an adaptive arbitration sampling method for edge buffering in industrial robots, provided as an embodiment of this application. This method can be applied to various business domains. Certain input parameters or intermediate results in this process can be manually adjusted to help improve accuracy.

[0021] The analysis method involved in the embodiments of this application can be implemented by a terminal device or a server, and this application does not impose any special limitations on it. For ease of understanding and description, the following embodiments are all described in detail using a server as an example.

[0022] Based on this Figure 1 The process may include the following steps: S101: Build a raw high-frequency data loop cache at the edge of the industrial robot and simultaneously acquire the edge risk probability distribution and edge confidence level output by the edge, as well as the cloud risk probability distribution, cloud confidence level, remaining service life prediction value and remaining service life confidence interval width output by the cloud.

[0023] In some embodiments of this application, data caching and synchronous acquisition of multi-source information are achieved by relying on edge hardware units deployed on-site at the industrial robot site.

[0024] Specifically, the edge device connects to various acquisition units of the robot body via an industrial bus and communication interface, including vibration sensors, current acquisition modules, encoder trajectory acquisition units, and controller task log interfaces. It collects raw high-frequency operating data such as vibration signals, current signals, trajectory errors, and task logs in real time, and builds a raw high-frequency data loop cache inside the edge device.

[0025] The cache adopts a first-in-first-out circular storage mechanism, continuously writing the latest collected raw high-frequency data. When the cache capacity reaches the preset limit, it automatically overwrites the oldest historical data, always retaining the complete raw high-frequency data of the most recent fixed duration, providing a full data foundation for subsequent arbitration sampling.

[0026] At the same time, the edge end synchronously acquires the edge risk probability distribution and edge confidence level output by its own edge risk judgment module, which characterizes the preliminary diagnostic conclusions and credibility obtained by the edge side based on feature summaries; there are also diagnostic and prediction results output by the cloud side, specifically including the cloud risk probability distribution and cloud confidence level output by the cloud diagnostic model, as well as the remaining useful life prediction value and remaining useful life confidence interval width output by the cloud RUL prediction module. The above cloud results are periodically synchronized to the edge end through the edge-cloud communication link, keeping the timestamp aligned with the original data cached locally on the edge end, ensuring the time consistency of subsequent multi-source evidence fusion.

[0027] S102: Calculate the difference between the edge risk probability distribution and the cloud risk probability distribution, and use the degradation risk probability distribution, edge risk probability distribution and cloud risk probability distribution generated by mapping the remaining useful life prediction value with the remaining useful life confidence interval width as multi-source diagnostic evidence, and calculate the diagnostic uncertainty entropy after the fusion of the multi-source diagnostic evidence.

[0028] In some embodiments of this application, the degree of discrepancy between edge and cloud-based diagnostic results is first quantified, and the edge risk probability distribution is calculated using Jensen-Shannon divergence. Risk probability distribution in the cloud Difference The formula is: .

[0029] The greater the difference between the two probability distributions, the higher the divergence value, which can intuitively reflect the degree of inconsistency between the diagnostic conclusions at the edge and in the cloud.

[0030] Subsequently, using the predicted remaining useful life as the distribution center and the width of the remaining useful life confidence interval as the distribution span, a probability mapping algorithm is used to map the predicted remaining useful life to the width of the remaining useful life confidence interval to generate a degradation risk probability distribution, which represents the health risk conclusions obtained based on the life degradation trend.

[0031] Based on this, the probability distributions of edge risk, cloud risk, and degradation risk are used together as multi-source diagnostic evidence, and a multi-evidence fusion operation is performed. Specifically, the probability distributions of edge risk, cloud risk, and degradation risk are weighted according to their respective confidence levels to obtain the fused risk probability distribution. The confidence level corresponding to the probability distribution of degradation risk is determined by normalizing the confidence interval width of the remaining useful life. The larger the confidence interval width, the lower the corresponding confidence level and the smaller the weight ratio in the fusion process.

[0032] Ultimately based on the fusion risk probability distribution Calculate information entropy and use it as the diagnostic uncertainty entropy. The formula is: .

[0033] A higher entropy value indicates greater uncertainty in the overall diagnostic result after fusion of multi-source evidence, making it more necessary to upload the original data for further review and diagnosis.

[0034] S103: The difference degree, diagnostic uncertainty entropy, cloud confidence degree, remaining useful life prediction confidence interval width, short-term degradation trend slope and load mutation factor are weighted to obtain the evidence conflict entropy score.

[0035] In some embodiments of this application, a sequence of historical remaining useful life predictions within a preset time window is first obtained, and the trend slope is obtained through univariate linear fitting, which serves as the short-term degradation trend slope. The slope reflects the recent rate of equipment degradation. A positive slope and a larger value indicate an accelerating trend of equipment degradation and a higher probability of diagnosing a real fault.

[0036] Simultaneously, based on the industrial robot's operating load data collected at the edge, the magnitude of the operating condition change between the current moment and the previous steady-state moment is calculated to obtain the load mutation factor. The operating data covers dimensions such as robot joint load commands, speed settings, and workpiece model switching signals, and is used to distinguish between operating condition switching disturbances and actual component degradation.

[0037] After feature extraction, the remaining useful life prediction confidence interval width, short-term degradation trend slope, and load mutation factor are normalized to obtain their respective normalized values. All input features are mapped to a unified numerical range of 0 to 1, eliminating the influence of dimensional differences on the weighted calculation. Simultaneously, the cloud-based confidence score is... It is converted into a confidence deficiency item, where the confidence deficiency item is the difference between 1 and the cloud confidence level. The lower the cloud confidence level, the larger the value of this item, and the stronger its positive driving effect on conflict entropy.

[0038] Ultimately based on the degree of difference Diagnostic uncertainty entropy Items with insufficient confidence Confidence interval width for normalized remaining useful life prediction Normalized short-term degradation trend slope As a positive weighting term, the normalized load mutation factor As a suppression weighting term, by presetting weight coefficients The evidence conflict entropy score is obtained by summing the weighted terms. The formula is: .

[0039] A higher value for the positive weighting term indicates a more significant diagnosis of the conflict and a stronger necessity for review. A higher value for the suppression weighting term indicates that the conflict is more likely to be caused by operating condition disturbances, and the conflict entropy score is lowered accordingly to reduce false triggering.

[0040] S104: When the evidence conflict entropy score is greater than the preset trigger threshold, change point detection is performed on the feature sequence corresponding to the original high-frequency data to obtain the abnormal start point and abnormal recovery point. The retention time before the abnormality and the retention time after the abnormality are determined according to the evidence conflict entropy score. The adaptive sampling window is determined according to the abnormal start point, the abnormal recovery point, the retention time before the abnormality and the retention time after the abnormality.

[0041] In some embodiments of this application, a trigger determination is first performed, and the evidence conflict entropy score calculated in real time is applied. The evidence conflict entropy score is compared with a preset trigger threshold. When the value exceeds the preset trigger threshold, the current diagnostic conflict level is determined to meet the review criteria, and the adaptive sampling process is initiated.

[0042] Subsequently, change point detection is performed to accurately locate anomaly boundaries. The moving mean and moving standard deviation of the feature sequence corresponding to the original high-frequency data are calculated, based on the feature values ​​at the current time. With moving average The difference, combined with the sliding standard deviation Constructing standardized deviation To eliminate the effects of dimensional differences and baseline fluctuations in different characteristics; the formula is: .

[0043] Construct a cumulative change statistic based on standardized deviation. Furthermore, a drift suppression parameter v is introduced to counteract the cumulative effect of normal small-amplitude random fluctuations, thus avoiding false detections caused by random noise during normal equipment operation; the formula is: .

[0044] When the cumulative change statistic exceeds the preset anomaly detection threshold, the corresponding time point is determined to be the start point of the anomaly. When the cumulative change statistic falls below the preset recovery threshold, the corresponding time point is determined to be an abnormal recovery point. This allows for precise identification of the start and end times of the abnormal process.

[0045] After anomaly identification is completed, the evidence conflict entropy score is linearly calculated using a preset constant to obtain the retention time before and after the anomaly. The higher the conflict entropy score, the longer the retention time, so as to cover a more complete anomaly evolution process. At the same time, a preset anomaly duration adjustment component is superimposed on the retention time after the anomaly. The longer the anomaly duration, the longer the retention time after the anomaly, ensuring that the entire anomaly recovery process is fully recorded.

[0046] Subsequently, preset upper and lower limits were applied to the retention time before and after the anomaly, respectively, and the data was trimmed to the corresponding minimum and maximum limits to avoid missing key information due to excessively short retention time or invalid data redundancy due to excessively long retention time.

[0047] Finally, the adaptive sampling window is obtained by shifting the retention time before the anomaly starting point forward to the anomaly start point as the window start time, and by shifting the retention time after the anomaly recovery point backward to the window end time. This defines the range of subsequent candidate segments for selection; the formula is: .

[0048] S105: Within the adaptive sampling window, the original high-frequency data is cyclically cached and divided into several candidate segments according to a preset time period, and the information gain of each candidate segment is calculated; the information gain is the difference in diagnostic uncertainty before and after the candidate segment is uploaded and participates in the review.

[0049] In some embodiments of this application, the candidate segments are first divided. Within the time range corresponding to the adaptive sampling window, the original high-frequency data in the original high-frequency data cyclic buffer is segmented according to a preset time period to obtain several candidate segments. The segmentation adopts a continuous segmentation method with equal duration, and the preset time period can be flexibly configured according to the on-site sampling frequency and bandwidth constraints.

[0050] Subsequently, information gain is calculated segment by segment to quantify its review value. Specifically, the system health conditional entropy based solely on existing diagnostic evidence from the edge and cloud is used as the diagnostic uncertainty before a candidate segment participates in the review, representing the overall uncertainty of the system under the current diagnostic information. The system health conditional entropy after incorporating the original data of the corresponding candidate segment into the review diagnosis is used as the diagnostic uncertainty after the candidate segment participates in the review, representing the change in the system uncertainty after supplementing the original data of that segment. The diagnostic uncertainty before review is then calculated. Uncertainty in diagnosis after review The difference is determined as the information gain of the candidate segment. The formula is: .

[0051] The larger the information gain value, the stronger the contribution of the original data segment to reducing diagnostic uncertainty and clarifying fault conclusions, and the higher its verification value after uploading.

[0052] S106: Weight the abnormal amplitude, spectral abnormality degree, temporal proximity to the time of the abnormality, operating condition matching degree, data volume cost, and information gain of each candidate segment to determine the comprehensive score of each candidate segment.

[0053] In some embodiments of this application, multi-dimensional evaluation features of each candidate segment are first extracted, and normalization preprocessing is performed: that is, abnormal amplitudes of candidate segments are analyzed. Spectral anomaly degree Temporal proximity Operating condition matching degree Data volume cost Normalization is performed separately to map to a unified numerical range, eliminating dimensional differences across dimensions. Specifically, the abnormal amplitude is the deviation of the signal peak within a candidate segment from the normal baseline, which can also be characterized by the proportion of samples exceeding the threshold within the candidate segment, reflecting the severity of the time-domain signal anomaly; the spectral anomaly is the amplitude of harmonic energy changes in characteristic frequency bands within the candidate segment, obtained by extracting frequency band energy through Fast Fourier Transform and comparing it with the normal operating baseline, reflecting the intensity of abnormal features in the frequency domain; temporal proximity is the proximity of the candidate segment to the time of the anomaly occurrence, with higher proximity values ​​indicating stronger diagnostic reference value; operating condition matching is the degree of matching between the operating condition corresponding to the candidate segment and the current review operating condition, with greater consistency in operating conditions leading to stronger comparability and reference value; and data volume cost is the original data size corresponding to the candidate segment, reflecting the bandwidth resources required to upload the segment.

[0054] After feature preprocessing, information gain, normalized abnormal amplitude, spectral abnormality degree, temporal proximity, and working condition matching degree are used as positive scoring items; the higher the value, the higher the comprehensive diagnostic value of the segment. Normalized data cost is used as a cost suppression item; the larger the data volume, the stronger the negative impact on the comprehensive score. Through preset corresponding weight coefficients ( , , , , , The positive scoring items and cost suppression items are weighted and summed to obtain the comprehensive score for each candidate segment. A higher overall score indicates a higher cost-effectiveness of uploading the segment under limited bandwidth constraints. The formula is: .

[0055] S107: Under the preset bandwidth constraints and number of upload segments, select the candidate segments with the highest comprehensive scores as upload segments and upload them to the cloud.

[0056] In some embodiments of this application, candidate segment selection is modeled as a constrained optimization problem. Under preset bandwidth constraints and the number of uploaded segments, the goal is to maximize the sum of the comprehensive scores of the selected segments to obtain the optimal combination of uploaded segments. In specific implementation, this selection process can be equivalent to a 0-1 knapsack optimization problem, with constraints that the total data volume of the selected segments does not exceed the current maximum allowed uploaded data volume and the total number of segments does not exceed a preset upper limit. The optimization objective is to maximize the sum of the comprehensive scores of the selected segments. The optimal combination of segments is obtained by solving a dynamic programming algorithm or a greedy algorithm, ensuring that the uploaded data has the highest comprehensive verification value under limited uplink bandwidth resources.

[0057] Meanwhile, a data type priority matching mechanism is introduced during the screening process to further enhance the relevance of uploaded data: when the conflict of candidate segments originates from vibration spectrum anomalies, vibration waveform candidate segments with higher comprehensive scores are uploaded first; when the conflict of candidate segments originates from current harmonic anomalies, current waveform candidate segments with higher comprehensive scores are uploaded first; when the conflict of candidate segments originates from repetitive positioning errors, trajectory error sequence candidate segments with higher comprehensive scores are uploaded first. By combining type priority matching with comprehensive score ranking, efficient use of bandwidth resources is ensured, while also ensuring a high degree of matching between uploaded data and the root cause of conflicts, thereby improving the efficiency and accuracy of cloud-based review.

[0058] S108: Based on the cloud-based review and diagnosis of the uploaded candidate segments, the review risk probability distribution and review confidence are obtained. The fusion weight is determined based on the review confidence, edge confidence, and cloud confidence. Based on the fusion weight, the edge risk probability distribution, cloud risk probability distribution, and review risk probability distribution are weighted and fused to obtain the health judgment result.

[0059] In some embodiments of this application, after the cloud receives the candidate segments uploaded by the edge device, the review and diagnosis module performs refined review and diagnosis on the uploaded original high-frequency data segments. Through deep feature extraction and high-precision fault identification model calculation, it outputs the review risk probability distribution. With review confidence level The verification result is calculated based on the full amount of original high-frequency data, and has higher accuracy and confidence compared with edge-side feature-level diagnosis.

[0060] Subsequently, a weighted fusion of multi-source diagnostic results is performed: the edge confidence, cloud confidence, and verification confidence are normalized, and the edge fusion weights are calculated separately. Cloud-based integrated weights , review and integration weight The sum of the three weights is 1, and the higher the confidence value, the greater the corresponding fusion weight is assigned, ensuring that the diagnostic results with high confidence account for a higher weight in the final conclusion.

[0061] The edge risk probability distribution, cloud risk probability distribution, and verification risk probability distribution are weighted and summed by using edge fusion weight, cloud fusion weight, and verification fusion weight respectively, to obtain the final risk probability distribution.

[0062] Finally, based on the probability values ​​of various health states in the final risk probability distribution, the corresponding health judgment results are output. Health states include three categories: normal, suspected abnormal, and abnormal. When the probability of an abnormal state exceeds the preset judgment threshold, an abnormal alarm is output. When the probability is in the transition range, a suspected abnormal prompt is output. When the probability of a normal state is absolutely dominant, a normal state conclusion is output, providing a quantitative basis for subsequent operation and maintenance decisions.

[0063] S109: Construct a feedback reward function based on the health assessment result and the manual maintenance confirmation result. Based on the feedback reward function, correct the weight of the evidence conflict entropy calculation, the weight of the candidate fragment comprehensive score, and the trigger threshold.

[0064] In some embodiments of this application, after obtaining the health assessment result and subsequent manual maintenance confirmation result, a feedback reward function is constructed: the diagnostic accuracy benefit is... This is set as a positive reward item; a reward is given when the health assessment result matches the manual maintenance confirmation result, with a higher reward value for greater consistency. The cost of uploading data will also be considered. Review delay costs False alarm penalty items Penalties for Failure to Report These are designated as negative deduction items, with higher deduction values ​​for larger uploaded data volumes and longer review delays. Additional penalty points are applied for false positives or missed positives. A feedback reward function is constructed by weighting each item using corresponding preset weight coefficients to comprehensively measure the overall benefits and costs of this arbitration sampling process. The function is as follows: .

[0065] Subsequently, based on the output value of the feedback reward function and the gradient direction of the fragment scoring function relative to the parameters to be corrected, the weights for calculating the evidence conflict entropy, the weights for the comprehensive score of candidate fragments, and the trigger threshold are initially iteratively updated. The fragment scoring function is a weighted process for calculating the comprehensive score of candidate fragments. The parameter updates are performed along the gradient direction of reward enhancement, gradually improving the overall benefit of the algorithm. The formula is as follows: .

[0066] Where θ represents the set of parameters to be updated. Indicates the learning rate. This represents the gradient direction of the segment scoring function with respect to the parameter θ.

[0067] To avoid drastic parameter fluctuations caused by a single outlier sample, a preset forgetting coefficient is introduced. For the initially updated parameters The parameters are updated using an exponential sliding smoothing method based on historically validated values ​​to obtain the final corrected parameters, as shown in the formula: .

[0068] By smoothing the parameters to suppress drastic fluctuations caused by a single abnormal sample, the parameters can be continuously optimized adaptively as the equipment operating conditions change and the aging process progresses, so that they can be continuously adapted to the field operating environment without repeated manual parameter adjustments.

[0069] It should be noted that, although the embodiments in this application are based on... Figure 1 Steps S101 to S109 will be described sequentially, but this does not mean that steps S101 and S109 must be performed in a strict order. The reason this embodiment follows this order is... Figure 1 The order in which steps S101 to S109 are described is provided to facilitate understanding of the technical solutions of the embodiments of this application by those skilled in the art. In other words, in the embodiments of this application, the order of steps S101 to S109 can be appropriately adjusted according to actual needs.

[0070] Figure 2 This is a schematic diagram illustrating the deployment of an algorithm for an adaptive arbitration sampling method for edge caching of industrial robots, provided in an embodiment of this application, in an edge-cloud system.

[0071] exist Figure 2The paper demonstrates the layered deployment structure of the proposed method under the edge-cloud collaborative architecture, which is divided into four layers from bottom to top: device side, edge side, cloud side, and decision feedback side. The device side deploys the industrial robot body and its supporting sensing and acquisition units, collecting raw operational data such as vibration signals, current signals, trajectory errors, and task logs in real time and transmitting them to the edge side. The edge side integrates a data preprocessing module, an edge risk judgment module, and an edge circular cache and arbitration sampling module. On the one hand, it extracts feature summaries, uploads them to the cloud, and outputs edge risk judgment results; on the other hand, it retains raw high-frequency data through circular caching, performs adaptive arbitration sampling based on edge-cloud diagnostic conflicts, and selects key raw fragments to upload to the cloud. The cloud side sets up a cloud diagnostic model, a RUL prediction module, a verification diagnostic module, and a fusion judgment module. Based on feature summaries, it outputs routine diagnostic and lifespan prediction results, performs verification diagnostics based on the uploaded raw fragments, and completes multi-source result fusion. The decision feedback side includes an alarm and maintenance decision module and a parameter feedback correction module, outputting the final health judgment result and generating maintenance decisions. Simultaneously, it corrects the algorithm weights and threshold parameters of the edge side based on the feedback results, forming a closed-loop optimization chain.

[0072] Figure 3 A schematic diagram of an adaptive arbitration sampling device for edge buffering of an industrial robot provided in this application embodiment includes: At least one processor; and, A memory that is communicatively connected to at least one processor; wherein, An industrial robot edge cache adaptive arbitration sampling method is described, wherein the memory stores instructions that can be executed by at least one processor, and the instructions are executed by at least one processor to enable at least one processor to perform any of the above-mentioned actions.

[0073] Some embodiments of this application provide an industrial robot edge cache adaptive arbitration sampling non-volatile computer storage medium storing computer-executable instructions that can execute any of the above-described industrial robot edge cache adaptive arbitration sampling methods.

[0074] The various embodiments in this application are described in a progressive manner. Similar or identical parts between embodiments can be referred to mutually. Each embodiment focuses on describing the differences from other embodiments. In particular, the device and medium embodiments are basically similar to the method embodiments, so the description is relatively simple; relevant parts can be referred to the description of the method embodiments.

[0075] The devices and media provided in this application are one-to-one with the methods. Therefore, the devices and media also have similar beneficial technical effects as their corresponding methods. Since the beneficial technical effects of the methods have been described in detail above, the beneficial technical effects of the devices and media will not be repeated here.

[0076] Those skilled in the art will understand that embodiments of the present invention can be provided as methods, systems, or computer program products. Therefore, the present invention can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present invention can take the form of a computer program product embodied on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0077] This invention is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of the invention. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart illustrations and / or block diagrams. Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0078] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0079] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0080] In a typical configuration, a computing device includes one or more processors (CPU), input / output interfaces, network interfaces, and memory.

[0081] Memory may include non-persistent storage in computer-readable media, random access memory (RAM), and non-volatile memory such as read-only memory (ROM) or flash RAM. Memory is an example of computer-readable media.

[0082] Computer-readable media includes both permanent and non-permanent, removable and non-removable media that can store information using any method or technology. Information can be computer-readable instructions, data structures, modules of programs, or other data. Examples of computer storage media include, but are not limited to, phase-change memory (PRAM), static random access memory (SRAM), dynamic random access memory (DRAM), other types of random access memory (RAM), read-only memory (ROM), electrically erasable programmable read-only memory (EEPROM), flash memory or other memory technologies, CD-ROM, digital versatile optical disc (DVD) or other optical storage, magnetic tape, magnetic disk storage or other magnetic storage devices, or any other non-transferable medium that can be used to store information accessible by a computing device. As defined herein, computer-readable media does not include transient computer-readable media, such as modulated data signals and carrier waves.

[0083] It should also be noted that the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, method, article, or apparatus. Unless otherwise specified, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes that element.

[0084] The above are merely embodiments of this application and are not intended to limit this application. Various modifications and variations can be made to this application by those skilled in the art. Any modifications, equivalent substitutions, or improvements made within the technical principles of this application should fall within the protection scope of this application.

Claims

1. An adaptive arbitration sampling method for edge buffering in industrial robots, characterized in that, The method includes: A raw high-frequency data loop cache is built at the edge of the industrial robot, and the edge risk probability distribution and edge confidence level output by the edge are obtained simultaneously, as well as the cloud risk probability distribution, cloud confidence level, remaining service life prediction value and remaining service life confidence interval width output by the cloud. Calculate the difference between the edge risk probability distribution and the cloud risk probability distribution, and use the degradation risk probability distribution, edge risk probability distribution and cloud risk probability distribution generated by mapping the remaining useful life prediction value to the remaining useful life confidence interval width as multi-source diagnostic evidence, and calculate the diagnostic uncertainty entropy after the fusion of the multi-source diagnostic evidence; The evidence conflict entropy score is obtained by weighting the difference degree, diagnostic uncertainty entropy, cloud confidence degree, remaining useful life prediction confidence interval width, short-term degradation trend slope and load mutation factor. When the evidence conflict entropy score is greater than the preset trigger threshold, change point detection is performed on the feature sequence corresponding to the original high-frequency data to obtain the abnormal start point and abnormal recovery point. The retention time before the abnormality and the retention time after the abnormality are determined according to the evidence conflict entropy score. The adaptive sampling window is determined according to the abnormal start point, abnormal recovery point, retention time before the abnormality and retention time after the abnormality. Within the adaptive sampling window, the original high-frequency data is cyclically cached and divided into several candidate segments according to a preset time period, and the information gain of each candidate segment is calculated; the information gain is the difference in diagnostic uncertainty before and after the candidate segment is uploaded and participates in the review. The comprehensive score of each candidate segment is determined by weighting the abnormal amplitude, spectral abnormality degree, temporal proximity to the time of the abnormality, operating condition matching degree, data volume cost, and information gain. Under the preset bandwidth constraints and the number of uploaded segments, candidate segments with high comprehensive scores are selected as uploaded segments and uploaded to the cloud; Based on the cloud-based review and diagnosis of uploaded candidate segments, the review risk probability distribution and review confidence are obtained. The fusion weight is determined based on the review confidence, edge confidence, and cloud confidence. Based on the fusion weight, the edge risk probability distribution, cloud risk probability distribution, and review risk probability distribution are weighted and fused to obtain the health judgment result. A feedback reward function is constructed based on the health assessment results and the manual maintenance confirmation results. Based on the feedback reward function, the weights for calculating the evidence conflict entropy, the weights for the comprehensive score of candidate fragments, and the trigger threshold are adjusted.

2. The method according to claim 1, characterized in that, Calculating the difference between the edge risk probability distribution and the cloud risk probability distribution, and calculating the diagnostic uncertainty entropy after the fusion of multi-source diagnostic evidence, specifically includes: The difference between the edge risk probability distribution and the cloud risk probability distribution is calculated using Jensen-Shannon divergence. The edge risk probability distribution, cloud risk probability distribution, and degradation risk probability distribution are weighted according to their respective confidence levels to obtain a fused risk probability distribution; the confidence level corresponding to the degradation risk probability distribution is determined by normalizing the remaining useful life confidence interval width. Shannon information entropy is calculated based on the fusion risk probability distribution, and the Shannon information entropy is used as the diagnostic uncertainty entropy.

3. The method according to claim 1, characterized in that, The evidence conflict entropy score is obtained by weighting the difference degree, diagnostic uncertainty entropy, cloud confidence degree, remaining useful life prediction confidence interval width, short-term degradation trend slope, and load mutation factor. Specifically, it includes: Obtain the historical remaining useful life prediction value sequence within a preset time window, and obtain the trend slope through linear fitting, which is used as the short-term degradation trend slope; Based on the industrial robot operating load data collected at the edge, the magnitude of the operating condition change between the current moment and the previous steady-state moment is calculated to obtain the load mutation factor. The remaining useful life prediction confidence interval width, short-term degradation trend slope, and load mutation factor were normalized to obtain their respective normalized values. The cloud confidence level is converted into a confidence-deficient item; the confidence-deficient item is the difference between 1 and the cloud confidence level. The difference degree, diagnostic uncertainty entropy, insufficient confidence term, normalized remaining useful life prediction confidence interval width, and normalized short-term degradation trend slope are used as positive weighting terms, and the normalized load mutation factor is used as a suppression weighting term. The evidence conflict entropy score is obtained by weighting and summing each weighting term through preset weight coefficients.

4. The method according to claim 1, characterized in that, When the evidence conflict entropy score is greater than a preset trigger threshold, change point detection is performed on the feature sequence corresponding to the original high-frequency data to obtain the anomaly start point and anomaly recovery point. The pre-anomaly retention time and post-anomaly retention time are determined based on the evidence conflict entropy score. An adaptive sampling window is determined based on the anomaly start point, anomaly recovery point, and the pre-anomaly retention time and post-anomaly retention time. Specifically, this includes: When the evidence conflict entropy score is greater than the preset trigger threshold, the moving mean and moving standard deviation of the feature sequence are calculated, and a standardized deviation is constructed based on the difference between the feature value and the moving mean at the current time and the moving standard deviation. Based on the standardized deviation, a cumulative change statistic is constructed, and a drift suppression parameter is introduced to counteract the cumulative effect of normal small-amplitude random fluctuations. When the cumulative change statistics exceed the preset anomaly determination threshold, the corresponding time is determined to be the start point of the anomaly. When the cumulative change statistics fall below the preset recovery threshold, the corresponding time is determined to be an abnormal recovery point; Linear calculation is performed on the evidence conflict entropy score to obtain the retention time before and after the anomaly, and a preset anomaly duration adjustment component is superimposed on the retention time after the anomaly. Preset upper and lower limit constraints are applied to the retention time before and after the anomaly, respectively, and the retention time is trimmed to the corresponding minimum and maximum limit ranges. The adaptive sampling window is obtained by shifting the retention time before the anomaly starts forward from the anomaly start point as the window start time, and shifting the retention time after the anomaly recovers backward from the anomaly recovery point as the window end time.

5. The method according to claim 1, characterized in that, Within the adaptive sampling window, the original high-frequency data is cyclically buffered and divided into several candidate segments according to a preset time period, and the information gain of each candidate segment is calculated. Specifically, this includes: Within the adaptive sampling window, the original high-frequency data in the original high-frequency data cyclic buffer is segmented according to a preset time period to obtain several candidate segments; The system health status conditional entropy based solely on existing diagnostic evidence from the edge and cloud is used as a candidate fragment to participate in the diagnostic uncertainty before review. The system health status conditional entropy after the original data with corresponding candidate fragments is used to participate in the review and diagnosis is used as the diagnostic uncertainty after the candidate fragments participate in the review. The difference between the diagnostic uncertainty before and after review is determined as the information gain of the candidate segment.

6. The method according to claim 1, characterized in that, The comprehensive score for each candidate segment is determined by weighting the abnormal amplitude, spectral abnormality degree, temporal proximity to the time of the abnormality, operating condition matching degree, data volume cost, and information gain. Specifically, this includes: The abnormal amplitude, spectral anomaly degree, temporal proximity, operating condition matching degree, and data volume cost of the candidate segments are normalized and mapped to a unified numerical range. The abnormal amplitude is the degree of deviation of the signal peak value in the candidate segment from the normal baseline. The spectral anomaly degree is the amplitude of harmonic energy change in the characteristic frequency band in the candidate segment. The temporal proximity is the time close to the time when the candidate segment occurs. The operating condition matching degree is the degree of matching between the operating condition corresponding to the candidate segment and the current review condition. Information gain, normalized abnormal amplitude, spectral abnormality degree, temporal proximity, and working condition matching degree are used as positive scoring items, and normalized data volume cost is used as a cost suppression item. The comprehensive score of each candidate segment is obtained by weighting and summing the data using preset corresponding weight coefficients.

7. The method according to claim 1, characterized in that, The fusion weights are determined based on the verification confidence, edge confidence, and cloud confidence. Then, based on these fusion weights, the edge risk probability distribution, cloud risk probability distribution, and verification risk probability distribution are weighted and fused to obtain a health assessment result, specifically including: The edge confidence, cloud confidence, and verification confidence are normalized, and the edge fusion weight, cloud fusion weight, and verification fusion weight are calculated respectively. The edge risk probability distribution, cloud risk probability distribution, and verification risk probability distribution are weighted and summed by using edge fusion weight, cloud fusion weight, and verification fusion weight respectively to obtain the final risk probability distribution. Based on the probability values ​​of various health states in the final risk probability distribution, the corresponding health judgment results are output; the health states include normal, suspected abnormal, and abnormal.

8. The method according to claim 1, characterized in that, The step of constructing a feedback reward function based on the health assessment result and the manual maintenance confirmation result, and correcting the weights for calculating the evidence conflict entropy, the comprehensive score weights for candidate fragments, and the trigger threshold based on the feedback reward function, specifically includes: The benefits of diagnostic accuracy are set as positive reward items, while the costs of uploading data, review latency, false alarm penalties, and missed alarm penalties are set as negative deduction items. The feedback reward function is constructed by weighting each item with corresponding preset weight coefficients. Based on the output value of the feedback reward function and the gradient direction of the fragment scoring function relative to the parameter to be corrected, the weights for calculating the evidence conflict entropy, the weights for the comprehensive score of candidate fragments, and the trigger threshold are initially iteratively updated; the fragment scoring function is a weighted process for calculating the comprehensive score of candidate fragments. By introducing a preset forgetting coefficient, an exponential sliding smoothing update is performed on the initially updated parameters and the historically verified parameter values ​​to obtain the final corrected parameters, thereby suppressing drastic parameter fluctuations caused by a single abnormal sample.

9. The method according to claim 1, characterized in that, When selecting candidate segments with high overall scores as upload segments to the cloud, the method further includes: When the conflict between candidate segments stems from abnormal vibration spectra, the candidate segment with the higher overall score is uploaded first. When the conflict between candidate segments comes from current harmonic anomalies, the current waveform candidate segment with the higher comprehensive score is uploaded first. When the conflict of candidate segments comes from repeated positioning errors, the candidate segments of trajectory error sequences with higher comprehensive scores are uploaded first.

10. An adaptive arbitration sampling device for edge buffering of an industrial robot, characterized in that, include: At least one processor; as well as, A memory communicatively connected to the at least one processor; wherein, The memory stores instructions that can be executed by the at least one processor, which, when executed by the at least one processor, enables the at least one processor to perform the industrial robot edge buffer adaptive arbitration sampling method according to any one of claims 1-8.

Images