Contact window detection method and device based on dynamic density clustering decision
The contact window detection method based on dynamic density clustering decision-making solves the problems of high false detection rate and large response delay in contact window detection by using dynamic calibration and multi-source signal fusion, combined with signal-to-noise ratio and clustering algorithm, and achieves efficient contact window recognition.
Patent Information
- Application Number
- CN202511596256.8
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-11-04
- Publication Date
- 2026-01-16
AI Technical Summary
Existing contact window detection methods in industrial automation scenarios suffer from high false detection rates, large response delays, and high data annotation costs, making them difficult to adapt to dynamically changing contact scenarios.
A contact window detection method based on dynamic density clustering decision is adopted. The compensation coefficient is calculated by dynamic calibration method, and the multi-source physical signals are weighted and summed to generate a synthetic signal. The contact segment is identified by combining the signal-to-noise ratio correlation coefficient and dynamic density threshold with Grubbs outlier detection and DBSCAN clustering algorithm.
It achieves real-time and accurate identification of contact windows, reduces the false detection rate to 2.2%, controls the response latency within 8ms, adapts to dynamic environments, and reduces the dependence on data annotation.
Smart Images
Figure CN121350672A_ABST
Abstract
Description
TECHNICAL FIELD
[0001] The present application relates to the technical field of robot perception, and in particular to a contact window detection method and device based on dynamic density clustering decision. BACKGROUND
[0002] In the industrial automation scene, the contact window detection method generally adopts static threshold detection method, frequency domain analysis method and traditional machine learning, but the traditional static threshold detection method is difficult to adapt to the dynamic contact scene, the static threshold cannot adapt to the change of signal characteristics, resulting in residual inertial interference components in the synthesized signal, in the scene of robot collaborative work or minimally invasive surgery which requires precise force control, when the load fluctuation is greater than 15%, the false detection rate is as high as 40%; the frequency domain analysis method has significant response delay, and the processing has a time delay of more than 30ms, resulting in obvious lag of the haptic feedback of the electronic device; the traditional machine learning has strong data dependence, and needs 10000+ labeled samples, the data labeling cost is high and the labeling quality is difficult to guarantee. The above defects directly restrict the intelligent upgrading process of high-end equipment. SUMMARY
[0003] In order to solve the above technical problems, the present application proposes a contact window detection method based on dynamic density clustering decision. In the method and device, the robustness bottleneck of the static threshold method in the dynamic environment can be broken, the response delay defect of the frequency domain analysis method can be eliminated, the strong dependence of the traditional machine learning on labeled data and parameters can be overcome, and real-time and accurate identification of the contact window can be realized.
[0004] In order to achieve the above purpose, the technical scheme of the present application is as follows:
[0005] The contact window detection method based on dynamic density clustering decision comprises the following steps:
[0006] Collecting multiple source physical signals; calculating the compensation coefficients corresponding to the multiple source physical signals by a dynamic calibration method, weighting and summing the multiple source physical signals based on the compensation coefficients to obtain a synthesized T_syn signal;
[0007] Dividing the continuous synthesized T_syn signal into several equal-width intervals and generating a histogram; extracting the maximum count value from the histogram and multiplying it by a signal-to-noise ratio correlation coefficient k as a main density threshold; comparing the main density threshold with the 0.05 quantile value of the synthesized T_syn signal, and taking the larger one as a dynamic density threshold; extracting a segment in the continuous synthesized T_syn signal that meets the dynamic density threshold and has a duration of more than M as a feature sequence;
[0008] The number of samples of the signal amplitude in the characteristic sequence is determined, and if the number of samples is not greater than 5, the processing is performed based on a statistical mode: an abnormal point in the amplitude set is identified as a contact section using a Grubbs outlier detection method; if the number of samples is greater than 5, the processing is switched to a clustering mode: a neighborhood scale parameter is dynamically calculated, and a DBSCAN clustering algorithm is used to cluster the amplitude set, and the cluster center of each cluster is taken as a contact section.
[0009] Preferably, the multi-source physical signal includes a mechanical torque signal, an angular acceleration signal, and a deformation displacement signal.
[0010] Preferably, the dynamic calibration method is a dynamic calibration method based on environmental parameters through a calibrate_coeff function to generate compensation coefficients α, β, and γ corresponding to the mechanical torque signal, the angular acceleration signal, and the deformation displacement signal.
[0011] Preferably, the M is an arbitrary value between 15 ms and 25 ms.
[0012] Preferably, the signal-to-noise ratio correlation coefficient k is obtained in the following steps:
[0013] The signal-to-noise ratio of the current window is calculated.
[0014] If the signal-to-noise ratio is greater than N, the signal-to-noise ratio correlation coefficient k is 0.15; if the signal-to-noise ratio is less than or equal to N, the signal-to-noise ratio correlation coefficient k is 0.3.
[0015] Preferably, the N is an arbitrary value between 10 and 20.
[0016] Preferably, the amplitude set is clustered using a DBSCAN clustering algorithm, including the following steps:
[0017] The neighborhood scale parameter is dynamically calculated, and the neighborhood scale parameter is the minimum value between the square root of the number of samples and 5;
[0018] The neighborhood distance parameter is calculated using the K-neighbor distance median method with the neighborhood scale parameter;
[0019] The amplitude set is clustered using the DBSCAN clustering algorithm with eps as the neighborhood distance parameter and the minimum sample number set to 1, and the cluster center is output as a contact section.
[0020] Preferably, the following steps are further included:
[0021] Mapping the abnormal points or cluster centers back to the original signal time axis to obtain time stamps; performing intensity division based on the number of abnormal points or cluster centers to determine intensity grading; obtaining a reliability score based on a weighted sum of signal smoothness and clustering profile coefficient in a clustering mode or obtaining a reliability score based on a weighted sum of signal smoothness and abnormal point statistical significance in a statistical mode;
[0022] Outputting a structured contact event data package including the time stamps, the intensity grading, and the reliability score.
[0023] Based on the above, the application further discloses a contact window detection device based on dynamic density clustering decision, comprising,
[0024] The acquisition and conversion module is configured to acquire multi-source physical signals, calculate compensation coefficients corresponding to the multi-source physical signals by using a dynamic calibration method, and perform weighted summation on the multi-source physical signals based on the compensation coefficients to obtain a synthesized T_syn signal.
[0025] The dynamic density perception and extraction module is configured to divide the continuous synthesized T_syn signal into a plurality of equal-width intervals and generate a histogram, extract a maximum count value from the histogram and multiply the maximum count value by a signal-to-noise ratio correlation coefficient k to obtain a main density threshold, compare the main density threshold with a 0.05 quantile value of the synthesized T_syn signal and take the larger one as a dynamic density threshold, and extract a section that meets the dynamic density threshold and has a time length greater than M from the continuous synthesized T_syn signal as a feature sequence.
[0026] The intelligent decision module is configured to judge a sample number of signal amplitudes in the feature sequence, perform processing based on a statistical mode if the sample number is not greater than 5, identify abnormal points in an amplitude set as contact sections by using a Grubbs outlier detection method, switch to a clustering mode for processing if the sample number is greater than 5, dynamically calculate a neighborhood scale parameter, and perform clustering on the amplitude set by using a DBSCAN clustering algorithm to take a cluster center of each cluster as a contact section.
[0027] Based on the above technical scheme, the beneficial effects of the present application are: the present application discloses a contact window detection method and device based on dynamic density clustering decision. The method collects multi-source physical signals, performs weighted summation on the multi-source physical signals based on a compensation coefficient, and obtains a synthesized T_syn signal; the continuous synthesized T_syn signal is divided into several equal-width intervals and a histogram is generated; the maximum count value is extracted from the histogram and multiplied by a signal-to-noise ratio correlation coefficient k as a main density threshold; the main density threshold is compared with the 0.05 quantile value of the synthesized T_syn signal, and the larger one is taken as the dynamic density threshold; the section in the continuous synthesized T_syn signal that meets the dynamic density threshold and has a duration longer than M is extracted as a feature sequence; the sample number of the signal amplitude in the feature sequence is judged, and if the sample number is not greater than 5, a statistical mode processing is used to identify the contact section; if the sample number is greater than 5, a clustering mode processing is used to identify the contact section. The dynamic calibration method of the present application calculates the compensation coefficient, and the closed-loop adaptive mechanism of the compensation coefficient ensures the stability of signal fusion under variable working conditions; the dynamic density threshold is obtained, the signal distribution characteristics are adaptively adapted, the threshold failure caused by noise is avoided, and real-time accurate identification of the contact window is realized; the clustering-statistical dual-mode decision mechanism triggered by the sample number is adaptively switched to obtain the detection result. Through experimental verification, the false detection rate can be reduced to 2.2%, and the response delay can be controlled within 8ms. BRIEF DESCRIPTION OF DRAWINGS
[0028] Fig. 1 It is an embodiment of a contact window detection method based on dynamic density clustering decision flow diagram;
[0029] Fig. 2 It is an embodiment of the positioning mark of the contact window in the real-time torque waveform (for comparison of detection effect);
[0030] Fig. 3 It is an embodiment of a torque signal density histogram partition diagram (showing low-density area identification). DETAILED DESCRIPTION
[0031] The technical solutions in the embodiments of the present application will be described clearly and completely below with reference to the drawings in the embodiments of the present application.
[0032] As shown in Figs. 1 to 3 , the present embodiment provides a contact window detection method based on dynamic density clustering decision, comprising the following steps:
[0033] Step 1, collect multi-source physical signals; calculate the compensation coefficient by a dynamic calibration method, perform weighted summation on the multi-source physical signals based on the compensation coefficient, and obtain a synthesized T_syn signal.
[0034] In this embodiment, mechanical torque (T_raw) and angular acceleration (T_raw) are acquired through a multi-physical quantity sensor array (12-axis torque sensor acquisition (1kHz sampling), MEMS accelerometer, deformation displacement sensor). The original deformation displacement (δ_deform) signal is used. Considering that static coefficient calibration (i.e., fixing the values of α, β, and γ) for multi-source signal fusion results in poor adaptability to various operating conditions, this embodiment calculates compensation coefficients using a dynamic calibration method. Based on these compensation coefficients, the multi-source physical signals are weighted and summed to obtain the synthesized T_syn signal.
[0035] Step 2: Divide the continuous synthetic T_syn signal into several equal-width intervals and generate histograms; extract the maximum count value from the histogram and multiply it by the signal-to-noise ratio correlation coefficient k as the principal density threshold; compare the principal density threshold with the 0.05 quantile value of the synthetic T_syn signal and take the larger one as the dynamic density threshold; extract the segments in the continuous synthetic T_syn signal that meet the dynamic density threshold and have a duration of more than M as feature sequences.
[0036] In this embodiment, firstly, the continuous T_syn signal is divided into 50 to 200 equal-width intervals (preferably 100) and a histogram is generated to analyze its amplitude distribution. For example... Fig. 3 As shown, the histogram clearly distinguishes between high-density areas (background noise) and low-density areas (potential contact events). Next, the maximum count value is extracted from the histogram and multiplied by the signal-to-noise ratio (SNR) correlation coefficient k to obtain the principal density threshold. The coefficient k is determined by the SNR of the current sliding window (e.g., 100ms): if SNR > 10⁻²⁰ (preferably 15), k = 0.15 (high SNR, more sensitive threshold); if SNR ≤ this threshold, k = 0.3 (low SNR, more conservative threshold, better interference resistance). To ensure threshold robustness, this principal threshold is compared with the 0.05 quantile of the T_syn signal, and the larger value is taken as the final dynamic density threshold.
[0037] Output feature sequences (low-density continuous segments that meet the dynamic density threshold and have a duration of more than 15ms to 25ms (preferably 20ms)) to extract feature data streams of suspected contact events for processing by the decision engine.
[0038] Step 3: Determine the number of signal amplitude samples in the feature sequence. If the number of samples is no more than 5, process based on statistical mode: use the Grubbs outlier detection method to identify outliers in the amplitude set as contact segments; if the number of samples is greater than 5, switch to clustering mode: dynamically calculate the neighborhood size parameter, use the DBSCAN clustering algorithm to cluster the amplitude set, and take the cluster center of each cluster as the contact segment.
[0039] In this embodiment, the clustering quality: evaluate the reliability of the clustering results after clustering is completed, calculate the clustering silhouette coefficient (S). The coefficient evaluates the compactness and separation degree of the clustering cluster, and the calculation process includes: for each sample point in the cluster, calculate the average distance (a) of the point with other points in the same cluster and the average distance (b) with the nearest different cluster, the silhouette coefficient s(i) of the point = (b-a) / max(a, b); finally, the silhouette coefficients of all non-noise points are averaged to obtain the overall silhouette coefficient S. The value of S is closer to 1, indicating that the clustering effect is better, and the distribution of the contact event in the feature space is clearer.
[0040] Output contact section: take the cluster center (such as the amplitude average value of all core points in the cluster) of each clustering cluster as the representative amplitude of the contact section.
[0041] This dual-mode switching mechanism ensures that the statistical method is preferred in small samples, and the density clustering is optimized in large samples.
[0042] In this embodiment, the adaptive switching of the clustering-statistics dual-mode decision mechanism triggered by the number of samples. First, the number of samples of the signal amplitude in the feature sequence is judged (i.e. the set of all original data points of the synthesized signal T_syn contained in each feature sequence segment that meets the condition extracted in step 2.). If the number of samples does not exceed 5, the statistical mode is used for processing, and the Grubbs outlier detection method is used to identify the abnormal points in the amplitude set as the contact section (the process of using the Grubbs outlier detection method includes: calculating the mean (μ) and standard deviation (σ) of the amplitude set, finding the data point with the largest absolute difference from the mean, and calculating its G value (G = |max value - μ| / σ). Compare this G value with the Grubbs test critical value table (according to the number of samples and the significance level, such as α = 0.05), if the G value exceeds the critical value, it is determined that the point is an abnormal point (i.e. the contact section).); If the number of samples exceeds 5, switch to the clustering mode for processing: dynamically calculate the neighborhood size parameter (adaptive_k), which is the minimum value between the square root of the number of samples and 5; Use this parameter to calculate the neighborhood distance parameter (eps) by the K-neighbor distance median method (the process of using the neighborhood size parameter to calculate the neighborhood distance parameter (eps) by the K-neighbor distance median method includes: for each sample point in the amplitude set, calculate the distance to the Kth (K = neighborhood size parameter) nearest neighbor point. Then, sort all sample points by K-neighbor distance in ascending order, and take the median as the final eps value. This method can adapt to the density of data distribution and avoid the subjectivity of manual parameter setting.), to resist noise interference; Finally, apply the DBSCAN clustering algorithm (with eps as the neighborhood distance parameter, and the minimum sample number is set to 1) to cluster the amplitude set, and output the cluster center as the contact section. This dual-mode switching mechanism ensures that the statistical method is preferred when the sample is small, and the density clustering is optimized for accuracy when the sample is large.
[0043] In the contact window detection method based on dynamic density clustering decision of one embodiment, a specific process of the dynamic calibration method for obtaining the compensation coefficient is also provided, including the following steps:
[0044] 1) Current harmonic interference compensation
[0045] Obtain motor current samples, calculate the total harmonic distortion ratio (harmonic_ratio) based on FFT spectrum analysis;
[0046] The calculation formula of the compensation coefficient α is as follows:
[0047] α = 0.85 - 0.15 * sigmoid(harmonic_ratio * 5)
[0048] 2) Load state self-adaptation
[0049] The active power and apparent power of the AC circuit are obtained, the power factor (power_factor) = active power (active_power) / apparent power (apparent_power) is calculated, when the power factor > 0.9, β = 0.25, the no-load state is strengthened and the acceleration compensation is carried out; when 0.7 <= power factor <= 0.9, β = 0.25 + 0.1 * (0.85 - power_factor), linear transition is carried out; when the power factor <0.7, β = 0.30, the heavy load state is weakened and the dynamic compensation is carried out.
[0050] 3) Transmission back gap compensation
[0051] The actual angle and target angle of the target are obtained, and the displacement integral error (displacement_error) = actual angle (actual_theta) - target angle (target_theta) is calculated.
[0052] The calculation formula of the compensation coefficient γ is as follows:
[0053] γ = 0.12 + 0.05 * tanh(displacement_error* 20)
[0054] The physical quantity fusion is performed to obtain a synthesized T_syn signal, and the formula is as follows:
[0055]
[0056] In the formula, T_raw is the mechanical torque signal data; is the angular acceleration signal data; δ_deform is the deformation displacement signal data, and α, β and γ are compensation coefficients corresponding to the mechanical torque signal, the angular acceleration signal and the deformation displacement signal respectively. The essence is to establish a closed-loop adaptive mechanism of equipment operating state -> signal interference characteristics -> compensation coefficient, so as to maintain the signal fusion stability under variable working conditions.
[0057] In an embodiment of the contact window detection method based on dynamic density clustering decision, a signal-to-noise ratio correlation coefficient k acquisition method is provided, including the following steps:
[0058] The signal-to-noise ratio of the current window is calculated.
[0059] If the signal-to-noise ratio is greater than N, the signal-to-noise ratio correlation coefficient k is 0.15; if the signal-to-noise ratio is less than or equal to N, the signal-to-noise ratio correlation coefficient k is 0.3.
[0060] The current window refers to a fixed-length signal data segment dynamically sliding for real-time analysis. The technical features are that the system sequentially slides on the continuous synthetic signal T_syn in fixed time intervals (for example, every 10 ms) or fixed data point quantities (for example, a window containing 100 sampling points). After each sliding, the signal-to-noise ratio of the latest data in the window is calculated. This ensures that the evaluation of the signal-to-noise ratio can reflect the latest changes in signal quality in real time, providing immediate and accurate basis for the subsequent adaptive adjustment of the dynamic density threshold.
[0061] The signal-to-noise ratio (SNR) is used to quantify the strength relationship between the effective components and noise in the signal. Based on the signal-to-noise ratio, the standard working condition and the high-noise environment are distinguished, and the signal-to-noise ratio is used to drive the threshold adjustment (N is an arbitrary value between 10 and 20, preferably 15). Different signal-to-noise ratio correlation coefficients k are provided to adapt to the subsequent dynamic density threshold determination in complex working conditions and obtain the effective section.
[0062] In an embodiment of the contact window detection method based on dynamic density clustering decision, a process of clustering the amplitude set using the DBSCAN clustering algorithm is provided, which includes the following steps:
[0063] The neighborhood size parameter is dynamically calculated, and the neighborhood size parameter is the minimum value between the square root of the sample number and 5;
[0064] The neighborhood distance parameter is calculated by the K-neighbor distance median method using the neighborhood size parameter;
[0065] The DBSCAN clustering algorithm is applied to cluster the amplitude set with eps as the neighborhood distance parameter and the minimum sample number set to 1, and the cluster center is output as the contact section.
[0066] In an embodiment of the contact window detection method based on dynamic density clustering decision, a process of generating and outputting the contact window is also provided, which includes the following steps:
[0067] Map the abnormal points or cluster centers back to the original signal time axis to obtain the time stamp; perform intensity division based on the number of abnormal points or cluster centers to determine the intensity classification; obtain the credibility score based on the weighted sum of signal smoothness and clustering profile coefficient in clustering mode or obtain the credibility score based on the weighted sum of signal smoothness and abnormal point statistical significance in statistical mode;
[0068] Output the structured contact event data packet including the time stamp, intensity classification, and credibility score.
[0069] In this embodiment, 1) timestamp mapping: the position of the amplitude point corresponding to the abnormal point or cluster center in the feature sequence is mapped back to the original time base (ms level precision); 2) intensity classification: level 1 is a single abnormal point or the smallest cluster (≤3 points); level 2 is a medium-sized cluster (4-10 points); and level 3 is a large cluster (>10 points).
[0070] The determination method of the credibility score in the clustering mode is as follows:
[0071] C = 0.4*F + 0.6*S
[0072] In the formula, F is the signal smoothness, and S is the clustering profile coefficient.
[0073] The determination method of the credibility score in the statistical mode: when the number of samples in the feature sequence does not exceed 5 (small sample scenario), the system switches to the statistical mode. In this mode, due to too few data points, effective clustering analysis cannot be performed, and therefore the clustering profile coefficient (S) cannot be calculated. Therefore, the application designs an alternative credibility score scheme, which is determined by the signal smoothness (F) and the abnormal point statistical significance (G_ratio). The calculation formula of the credibility score C is as follows:
[0074] C = 0.7*F + 0.3*G_ratio
[0075] In the formula, F is the signal smoothness, and G_ratio is the abnormal point statistical significance.
[0076] Specifically, the calculation method of the signal smoothness F is exactly the same as in the clustering mode (F = 1 / (1 + mean(|diff (feature sequence)|)), which continues to be used as a basic index for evaluating signal quality. In the statistical mode, its weight is increased to 0.7, and it becomes the dominant factor because the quality of the signal itself is more critical in the small sample scenario.
[0077] The abnormal point statistical significance (G_ratio) is used to quantify the degree of significance of the abnormal point (i.e., the contact point) identified by the Grubbs test. The core idea is that the abnormal value is not just "big" or "small", but "significantly" deviates from other data points in a statistical sense. The calculation formula is: G_ratio = G / G_critical, wherein:
[0078] Calculate G value: calculate the G statistical value of the abnormal point through the Grubbs test formula. Specifically, let the sample set be X, the mean be μ, and the standard deviation be σ, then the G value of the abnormal point X* is: G = |X* - μ| / σ.
[0079] Determine the critical value: According to the sample size n and the selected significance level (such as α = 0.05), the critical value G_critical is obtained by looking up the Grubbs test critical value table.
[0080] Quantify significance: When G_ratio > 1, it indicates that the point is statistically significant outliers, not random fluctuations, and the contact event is highly reliable; Adaptive threshold: The ratio automatically adapts to the statistical critical value under different sample sizes, which is more universal and scientific than directly using G value; Supplementary information: When the cluster quality evaluation cannot be used, G_ratio provides a strong and reliable evidence based on statistical theory.
[0081] Technical effect summary in statistical mode: By shifting the evaluation focus from "group mode" (clustering) to "individual significance" (statistical test), the present application ensures that in the small sample scenario, the reliability score still has a solid mathematical foundation. The weight distribution (0.7 for signal quality and 0.3 for statistical significance) reflects the design idea of prioritizing the reliability of the signal itself while considering the statistical significance of outliers in this mode, ensuring the rationality and effectiveness of the scoring mechanism.
[0082] It should be understood that although each step in the above flowchart is displayed in sequence according to the direction of the arrow, these steps are not necessarily executed in the order indicated by the arrow. Unless explicitly stated herein, there is no strict order restriction on the execution of these steps, and these steps can be executed in other orders. Moreover, at least part of the steps in the above flowchart can include multiple sub-steps or multiple stages, which are not necessarily executed at the same time, but can be executed at different times, and the execution order of these sub-steps or stages is not necessarily sequential, but can be executed in rotation or alternation with at least part of other steps or sub-steps or stages of other steps.
[0083] Based on the same inventive concept, the embodiments of the present application also provide a dynamic density clustering decision-based contact window detection device for implementing the dynamic density clustering decision-based contact window detection method described above. The problem-solving implementation scheme provided by the system is similar to the implementation scheme described in the above method, so the specific limitations in one or more dynamic density clustering decision-based contact window detection device embodiments provided below can refer to the limitations of the dynamic density clustering decision-based contact window detection method described above, which will not be repeated here.
[0084] In one embodiment, a dynamic density clustering decision-based contact window detection device is provided, comprising,
[0085] The acquisition and conversion module is used to acquire multi-source physical signals; it calculates the compensation coefficients corresponding to the multi-source physical signals through a dynamic calibration method, and performs weighted summation of the multi-source physical signals based on the compensation coefficients to obtain the synthesized T_syn signal;
[0086] The dynamic density sensing and extraction module is used to divide the continuous synthetic T_syn signal into several equal-width intervals and generate a histogram; extract the maximum count value from the histogram and multiply it by the signal-to-noise ratio correlation coefficient k as the principal density threshold; compare the principal density threshold with the 0.05 quantile value of the synthetic T_syn signal and take the larger one as the dynamic density threshold; extract the segments in the continuous synthetic T_syn signal that meet the dynamic density threshold and have a duration of more than M as feature sequences;
[0087] The intelligent decision-making module is used to determine the number of signal amplitude samples in the feature sequence. If the number of samples is no more than 5, it processes based on the statistical mode: using the Grubbs outlier detection method to identify outliers in the amplitude set as contact segments; if the number of samples is greater than 5, it switches to the clustering mode: dynamically calculating the neighborhood size parameter, using the DBSCAN clustering algorithm to cluster the amplitude set, and taking the cluster center of each cluster as the contact segment.
[0088] The above are merely preferred embodiments of the present application and are not intended to limit the embodiments of the present application. For those skilled in the art, the embodiments of the present application can have various modifications and variations. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the embodiments of the present application should be included within the protection scope of the embodiments of the present application.
Claims
1. A contact window detection method based on dynamic density clustering decision, characterized in that, The method comprises the following steps: Collecting multi-source physical signals; calculating compensation coefficients corresponding to the multi-source physical signals by a dynamic calibration method, and performing weighted summation on the multi-source physical signals based on the compensation coefficients to obtain a synthesized T_syn signal; Dividing the continuous synthesized T_syn signal into several equal-width intervals and generating a histogram; Extracting a maximum count value from the histogram and multiplying the maximum count value by a signal-to-noise ratio correlation coefficient k to obtain a main density threshold; comparing the main density threshold with a 0.05 quantile value of the synthesized T_syn signal, and taking the larger one as a dynamic density threshold; extracting a section in the continuous synthesized T_syn signal that satisfies the dynamic density threshold and has a time length greater than M as a feature sequence; Judging the sample number of signal amplitudes in the feature sequence; if the sample number is not greater than 5, performing processing based on a statistical mode: identifying an abnormal point in an amplitude set as a contact segment by using a Grubbs outlier detection method; if the sample number is greater than 5, switching to a clustering mode for processing: dynamically calculating a neighborhood scale parameter, and clustering the amplitude set by using a DBSCAN clustering algorithm, and taking a cluster center of each cluster as a contact segment.
2. The contact window detection method based on dynamic density clustering decision according to claim 1, characterized in that, The multi-source physical signals comprise a mechanical torque signal, an angular acceleration signal and a deformation displacement signal.
3. The contact window detection method based on dynamic density clustering decision according to claim 2, characterized in that, The dynamic calibration method is a dynamic calibration method for generating compensation coefficients α, β and γ corresponding to the mechanical torque signal, the angular acceleration signal and the deformation displacement signal based on an environment parameter by using a calibrate_coeff function.
4. The contact window detection method based on dynamic density clustering decision according to claim 1, characterized in that, The M is an arbitrary value between 15 ms and 25 ms.
5. The method of claim 1, wherein the method further comprises: The signal-to-noise ratio correlation coefficient k is obtained by the following steps: Calculating a current window signal-to-noise ratio; If the signal-to-noise ratio is greater than N, the signal-to-noise ratio correlation coefficient k is 0.15; if the signal-to-noise ratio is less than or equal to N, the signal-to-noise ratio correlation coefficient k is 0.
3.
6. The method of claim 1, wherein the method further comprises: The N is an arbitrary value between 10 and 20.
7. The method of claim 1, wherein the method is based on dynamic density clustering decision. The clustering of the amplitude set by using the DBSCAN clustering algorithm comprises the following steps: Dynamically calculating a neighborhood scale parameter, wherein the neighborhood scale parameter is the minimum value between a square root of the sample number and 5; Calculating a neighborhood distance parameter by using the neighborhood scale parameter through a K-nearest neighbor distance median method; Applying the DBSCAN clustering algorithm to the amplitude set by taking eps as the neighborhood distance parameter and taking 1 as the minimum sample number, and outputting a cluster center as a contact segment.
8. The method of claim 1, wherein the method is based on dynamic density clustering decision. The method further comprises the following steps: Mapping the abnormal point or the cluster center back to an original signal time axis to obtain a timestamp; performing intensity division based on the number of abnormal points or cluster centers to determine an intensity classification; obtaining a credibility score based on a weighted summation of signal smoothness and a clustering profile coefficient in the clustering mode or based on a weighted summation of signal smoothness and abnormal point statistical significance in the statistical mode; Outputting a structured contact event data packet comprising the timestamp, the intensity classification and the credibility score.
9. A contact window detection apparatus based on dynamic density clustering decision, characterized in that, The method comprises the following steps: Collecting multi-source physical signals by using a collecting and converting module; calculating compensation coefficients corresponding to the multi-source physical signals by a dynamic calibration method, and performing weighted summation on the multi-source physical signals based on the compensation coefficients to obtain a synthesized T_syn signal; A dynamic density perception and extraction module is configured to divide the continuous synthetic T_syn signal into several equal-width intervals and generate a histogram; extract the maximum count value from the histogram and multiply it by a signal-to-noise ratio correlation coefficient k to obtain a main density threshold; compare the main density threshold with a 0.05 quantile value of the synthetic T_syn signal, and take the larger one as a dynamic density threshold; and extract a section of the continuous synthetic T_syn signal that satisfies the dynamic density threshold and has a time length greater than M as a feature sequence; An intelligent decision module is configured to judge the sample quantity of the signal amplitude in the feature sequence; if the sample quantity is not greater than 5, a statistical mode is used for processing: using a Grubbs outlier detection method to identify an abnormal point in the amplitude set as a contact section; if the sample quantity is greater than 5, a clustering mode is switched for processing: dynamically calculating a neighborhood scale parameter, using a DBSCAN clustering algorithm to cluster the amplitude set, and taking the cluster center of each cluster as a contact section.