Robot tactile signal segmentation and alignment method based on period-phase optimization

Abstract

The invention discloses a period-phase optimized robot tactile signal segmentation and alignment method, which can quickly extract effective signal segments when a robot contacts an object, and solves the problems of unstable tactile signal form, baseline drift, uncertain period length and the like. The method provides a joint optimization normal form: firstly, estimating the average period of a tactile signal through self-correlation analysis; then, unified modeling is carried out on the segmentation tangency points and the fragment duration under structural constraints to form variables to be optimized, and an aggregation similarity objective function with normalized mutual measurement standards is constructed; and finally, iteratively solving the optimal segmentation point and phase of the time sequence signal in the contact process by maximizing the objective function. Compared with a traditional non-global optimization method, the method can be suitable for non-fixed-length quasi-periodic signals with unstable forms, has high robustness for amplitude drift and periodic jitter, is higher in segmentation precision and universality, and can provide accurate contact segment effective signals for robot touch recognition.

Description

Technical Field

[0001] This invention relates to the field of robot tactile perception and signal processing technology, specifically to a period-phase optimized method for robot tactile signal segmentation and alignment. Background Technology

[0002] With the rapid development of intelligent equipment such as service robots, industrial robots, and medical rehabilitation robots, utilizing tactile sensors to acquire mechanical information during contact and identify the shape, material, and surface condition of the contacted object has become an important way to improve the robot's environmental perception capabilities. When robots perform operations such as sliding contact and reciprocating sweeping, the output of tactile sensors is often a one-dimensional time-series periodic or quasi-periodic signal. Due to fluctuations in robot operating speed, changes in contact force, and interference from environmental noise, these signals are usually accompanied by uncertainties in period length, nonlinear drift in amplitude, and distortion of local morphology. Therefore, accurate extraction and phase alignment of effective periodic segments of such non-stationary signals are key preprocessing steps for achieving stable performance in subsequent feature extraction and recognition algorithms.

[0003] In the existing technology, the processing of such periodic signals is mainly divided into two typical technical routes.

[0004] The first type of method is mainly based on "local feature detection" or "fixed window truncation". These methods typically segment the period by setting a fixed amplitude threshold, detecting the zero-crossing points and peak-to-valley values ​​of the waveform, or directly using a fixed-length sliding window. However, this method requires extremely high signal stability. In robotic tactile perception scenarios, due to the irregularity of the contact surface or the jitter of the robotic arm's movement, the signal often exhibits significant baseline drift and period length variations. Relying on local feature points easily leads to incorrect or missed segments, while the fixed window method cannot adapt to the expansion and contraction of the period, resulting in phase disorder in the segmented segments, making them difficult to use for subsequent high-precision analysis.

[0005] The second category of methods is mainly based on "template matching" or "Dynamic Time Warping (DTW)". These methods typically require the pre-construction of a standard reference template (or center sequence) and the calculation of the similarity or distance between the signal under test and the template to achieve alignment or recognition. Although algorithms such as DTW can solve the problem of nonlinear distortion of the time axis to some extent, this type of method has significant limitations: First, it relies heavily on prior knowledge, that is, it must obtain a "standard template" in advance through manual screening or training with a large amount of data, and lacks adaptability for unknown materials or novel contact conditions; second, algorithms such as DTW have high computational complexity and usually require the endpoints of the signal to be known, resulting in low efficiency when processing long-term continuous stream signals.

[0006] More importantly, the aforementioned existing technologies generally adopt a sequential processing strategy of "segmentation first, alignment later." That is, the start and end points of the cycle are first determined by a certain rule (segmentation), and then the similarity of the segmented segments is calculated (alignment). This strategy severs the intrinsic connection between "segmentation" and "alignment." Once an error occurs in the first step of segmentation (such as a cut point offset), the error will be amplified in subsequent steps and cannot be corrected, resulting in poor consistency of the final extracted segment shape and seriously affecting the robustness of the tactile recognition system.

[0007] The differences compared to existing technologies are as follows:

[0008] Technical comparison with patent CN120582167A "A method and system for predicting frequency modulation commands for energy storage based on chaotic phase injection"

[0009] 1. Patent CN120582167A is positioned for power system energy storage devices, modeling and predicting the time series of power grid control such as energy storage frequency regulation commands. Its overall goal is to improve the prediction accuracy and eliminate response lag under quasi-periodic frequency regulation commands with chaotic characteristics through "periodic segmentation + chaotic phase injection + deep network prediction". It belongs to the prediction algorithm in the power system control scenario. In contrast, this patent is specifically for one-dimensional periodic or quasi-periodic tactile sensing signals in the sliding / reciprocating contact process of robots. The purpose is to automatically extract multi-periodic, phase-aligned effective contact segments from long-term tactile raw data to provide high-quality input for subsequent material / texture recognition algorithms. It belongs to the front-end signal processing solution for robot tactile perception. The two are significantly different in terms of application objects and system roles.

[0010] 2. The core mechanism of patent CN120582167A is based on "chaotic phase injection + periodic subsequence selection + sequence distance metric": First, the original energy storage frequency modulation command is segmented into candidate periods. The nonlinear time series alignment problem is handled using the constructed period distance or consistency index. Then, the chaotic phase injection (PCI) is used to control the chaotic intensity and periodic perturbation to achieve noise-resistant period detection. The reconstructed sequence is then input into a pre-trained GRU network to complete the prediction. Its essence is a time series prediction method based on dynamic system modeling and deep learning. The core mechanism of this patent is "joint optimization segmentation and phase alignment under periodic prior constraints": the average period is estimated using autocorrelation. All segmentation starting points 𝒔 and segment lengths 𝑻 are modeled as joint optimization variables, and a global objective function 𝐽(𝒔,𝑻) is constructed, consisting of normalized cross-correlation between multi-period segments, aggregated similarity, and structure regularization terms. This objective function is then applied to... A directed acyclic graph is constructed on the candidate state lattice, and the optimal (𝒔*,𝑻*) is obtained through shortest path search, realizing global optimization that integrates segmentation and alignment. One focuses on "chaotic phase + deep network prediction", while the other focuses on analytical signal segmentation based on "graph theory optimization + cross-correlation similarity". They have fundamental differences in their problem-solving approaches and mathematical tools.

[0011] 3. Patent CN120582167A requires a long history of frequency regulation command data and corresponding power grid operating conditions in terms of input and prior dependence to train deep network parameters such as GRU. Furthermore, its periodic detection and alignment steps serve specific power system control objectives, belonging to the paradigm of strong domain prior + data-driven prediction. In contrast, this patent does not rely on any pre-trained models or label data in its design. It only uses the tactile signal to be processed itself to obtain the periodic prior through autocorrelation peak search, and then performs joint optimization under this prior constraint. It can be applied to single-channel tactile signals and can also be naturally extended to multi-channel array signals through channel weighting. It belongs to a general signal preprocessing framework with weak prior, no training required, and oriented towards multiple sensors.

[0012] 4. The "period segmentation and nonlinear alignment" in patent CN120582167A is more of an auxiliary step in the entire prediction chain. Its final output focuses on the numerical prediction of future frequency modulation commands, and whether the period segment itself achieves high-precision phase alignment is not the main evaluation index. In contrast, this patent regards "period segmentation + phase alignment" itself as the core problem to be solved. By introducing interval constraints, length smoothing penalties and aggregation similarity based on normalized cross-correlation in the optimization process, and supplemented by subsequent baseline correction, abnormal segment removal and ROI secondary truncation, it specifically optimizes the morphological consistency and alignment accuracy of multi-period tactile segments. The output contains the minimum effective interval of multiple effective periods, thereby significantly reducing the burden on the subsequent recognition module. The two have significantly different functional positioning and output forms in the system architecture.

[0013] Comparison with the technology in patent EP3773173B1, "Up-sampling of signals by analytic phase projection".

[0014] 1. Patent EP3773173B1 addresses quasi-periodic time series or image sequences where reference signals are available but target signal sampling is insufficient. It proposes a "hypersampling" method based on Analytic Phase Projection (APP), significantly increasing the effective sampling rate to reconstruct a single typical periodic waveform. Applications include high temporal resolution reconstruction of physiological signals such as cardiovascular pulsation and cerebral blood flow in imaging systems like MRI / EEG. Its focus is on "upsampling and reconstructing undersampled signals given a known reference signal." This patent, however, targets high-sampling-rate one-dimensional periodic / quasi-periodic signals generated by robotic tactile sensors. It requires automatically segmenting multiple complete periods and achieving phase alignment for subsequent tactile recognition. It does not focus on upsampling or reconstruction but rather addresses the problem of "how to perform optimal period segmentation and phase alignment of existing sampled signals without a reference signal or template." The two patents differ significantly in their task objectives and typical application areas.

[0015] 2. The core mechanism of patent EP3773173B1 is as follows: First, a reference signal and the corresponding undersampled target time series are acquired from the sensor. The reference signal is filtered to obtain a single-component signal. Then, an analytical signal is constructed through Hilbert transform, and continuous analytical phases are extracted. Subsequently, the target sequence or image samples are mapped to a standardized phase interval according to the reference phase corresponding to the acquisition time. Samples within a large number of periods are "rearranged" and aggregated on a unified phase coordinate, thereby obtaining a typical periodic waveform with significantly improved time resolution. Its essence is a phase domain resampling and reconstruction method that relies on an external reference signal. In contrast, this patent does not require any reference signal or use analytical phases. Instead, it obtains the periodic prior through autocorrelation, constructs a directed acyclic graph of candidate segmentation points on the time axis, jointly models all segmentation starting points and segment lengths as variables to be optimized, and uses time-shifted maximization of normalized cross-correlation and structural regularization terms between multi-period segments to construct a global objective function. Then, the optimal segmentation path is obtained through shortest path search, realizing explicit segmentation and alignment of the original tactile signal directly in the time domain. The two are fundamentally different in terms of whether they rely on a reference channel and whether the alignment is completed in the phase domain or the time domain.

[0016] 3. Patent EP3773173B1 assumes that the target signal mainly faces the problem of "insufficient sampling rate". Its APP process mainly projects sparsely sampled data from multiple cycles onto a high-resolution phase grid and obtains a high temporal resolution single-cycle estimate through aggregation. It does not explicitly model the start and end boundaries of each cycle, nor does it output the segmentation cut point of a single cycle. In contrast, this patent addresses the problem of segmentation difficulties of tactile signals under conditions of cycle length fluctuation, baseline drift, and local morphological distortion. It not only explicitly solves for the start point and length of each cycle segment, but also ensures the stability of the cycle distribution globally through interval constraints, length smoothing penalties, and graph structure design. After obtaining (𝒔*,𝑻*), it continues to perform operations such as baseline correction, outlier removal, and window of interest truncation. The final output is a minimum effective interval containing multiple phase-aligned effective cycle segments. The two patents differ significantly in output format and the depth of modeling the "cycle structure".

[0017] 4. In terms of system architecture, the APP module of patent EP3773173B1 is usually integrated into medical imaging or complex measurement systems. As an upsampling and reconstruction method tightly coupled with signal acquisition equipment such as MRI / EEG, it is highly dependent on the sensor type and the reference signal acquisition method. However, the joint optimization segmentation and phase alignment method of this patent is designed as an independent software / algorithm module. It can run by inputting one-dimensional tactile time series data without relying on additional hardware or reference channels. Moreover, the claims give a clear module division (period prior acquisition module, joint optimization modeling module, constraint solving module, fragment extraction module), which has good versatility and portability. It can be extended to segmentation and preprocessing tasks of other one-dimensional periodic / quasi-periodic signals such as ECG, pulse wave, and gait vibration.

[0018] In summary, existing technologies lack a mechanism to comprehensively solve non-stationary quasi-periodic signals such as robotic tactile feedback without prior template dependence, by incorporating period length variations, segmentation point position drift, and fragment morphological consistency into a unified framework. Therefore, a method for achieving integrated global optimization of period segmentation and phase alignment is urgently needed. Summary of the Invention

[0019] The technical problem this invention aims to solve is the shortcomings of existing segmentation and phase alignment methods for one-dimensional periodic or quasi-periodic tactile signals, which suffer from poor robustness and inability to globally guarantee the morphological consistency of multiple periodic segments under conditions such as period length fluctuations, amplitude changes, and baseline drift. The technical solution of this invention is to provide a period-phase optimized method for robot tactile signal segmentation and alignment. Under periodic prior constraints, the starting point of each period segmentation and the period length are constructed as joint optimization variables, achieving integrated global optimization segmentation and alignment with the overall similarity between multiple periodic segments as the target.

[0020] This invention provides a period-phase optimized method for segmenting and aligning robot tactile signals, applicable to processing discrete tactile sensing signals. This includes the following steps performed by the processor:

[0021] S1. Acquire discrete tactile sensing signals And through autocorrelation analysis, an approximate period is estimated within a preset search window. ;

[0022] S2. Set the target number of segments , to the signal Each segmentation starting point and The length of each segment The unified model is used as a joint variable to be optimized;

[0023] S3. Construct a joint optimization objective function The At least one is included for maximizing the Aggregate similarity term of morphological similarity between segments , and a constraint for the and Regularization terms for structural rationality ,in For the first A segment;

[0024] S4. In the case described Under the determined structural constraints, the objective function is solved using a numerical optimization algorithm. In order to obtain the optimal set of starting points for segmentation. and the optimal set of fragment lengths ;

[0025] S5. According to the above and From the signal Extract A segment that has achieved phase alignment.

[0026] Furthermore, the signal acquired in S1 is a multi-channel signal. The objective function in S3 Constructed as a channel-weighted aggregation form:

[0027]

[0028] Among them, weight It can be determined based on the consistency between channels.

[0029] Furthermore, in S1 The estimation employs an autocorrelation peak search, where the search window is limited to a preset period range located in the positive half-axis delay region of the autocorrelation function, and the peak with the strongest correlation is selected within this range.

[0030] Furthermore, the segment length in S2 Modeled as:

[0031] (a) Equal length mode, i.e. ,in For the scalar to be optimized; or

[0032] (b) Variable length mode, i.e. ,in .

[0033] Furthermore, the aggregation operator in S3 Specifically limited to the arithmetic mean, used to calculate all segment pairs Morphological similarity between The average value.

[0034] Furthermore, in S3, the aggregation operator The morphological similarity measure used Specifically defined as normalized cross-correlation with time-shift maximization The calculation method is as follows:

[0035]

[0036] in, and Each represents a segment and At any moment The signal value, and These are the mean values ​​of the corresponding segments. For the allowed time offset, This represents the maximum offset range.

[0037] Furthermore, the structural regularization term in S3 Includes deviation penalties for the interval sequence:

[0038]

[0039] Or a length smoothing penalty under the variable length mode as described in claim 4(b):

[0040]

[0041] in For the expected period, .

[0042] Furthermore, the structural constraints in S4 include at least constraints on the spacing between adjacent starting points:

[0043]

[0044] in The minimum interval coefficient, This is the maximum interval coefficient.

[0045] Furthermore, the numerical optimization algorithm in S4 is specifically implemented as a nested optimization shortest path search method, including:

[0046] Sa1. Iterate through a set of candidate fragment lengths that satisfy claim 4. When the equal-length mode of claim 4(a) is used, the length is configured as a single scalar. When the variable-length mode of claim 4(b) is used, the length is configured as a vector. ;

[0047] Sa2. For each candidate segment length In the candidate state grid Construct a directed acyclic graph, where the nodes of the graph represent candidate starting points. ;

[0048] Sa3. Define the connection between adjacent nodes. and Boundary rights The edge weight Constructed to jointly optimize the global objective function with that in step S3 of claim 1. The corresponding local cost.

[0049] When using the equal-length mode, its form is:

[0050]

[0051] When using the variable-length mode, its form is:

[0052]

[0053] Among them, morphological similarity measurement and structural regularization terms Determined according to claims 6 and 7 respectively;

[0054] Sa4 uses the shortest path algorithm to find a cumulative cost path in the graph. The minimum optimal path is obtained. The corresponding optimal starting sequence ;

[0055] Sa5 selects the objective function jointly optimized in step S3 of claim 1. To reach the maximum The combination serves as the optimal starting point set for the joint optimization. and the set of optimal fragment lengths .

[0056] Furthermore, the candidate state grid in step Sa2 above Constructed in the following manner: A reference grid is formed on the time axis for the basic step size. Where r is the bias, And define a finite set of discrete perturbations for each grid point. ,in Based on this, the candidate starting point set is obtained. and as defined in claim 8 Interval constraint filtering Infeasible point.

[0057] Furthermore, following S5, the extraction process also includes... The steps for baseline correction of each segment: For each segment Before the starting point The mean of the data is used as the baseline. And deduct it.

[0058] Furthermore, after baseline correction, an outlier removal step is included: by calculating the root mean square value of each fragment or its aggregated similarity score with the remaining fragments, and based on... Outlier detection is performed on the scores to remove abnormal segments.

[0059] Furthermore, following S5, a secondary cropping step is included: in the aligned... Within the common morphological region of each segment, a window of interest is extracted for subsequent tactile recognition tasks of the same type.

[0060] On the other hand, the present invention also provides an apparatus for implementing a period-phase optimized robot tactile signal segmentation and alignment method, including a period prior acquisition module, a joint optimization modeling module, a constraint solving module, and a segment extraction module; wherein, the period prior acquisition module is used to acquire a one-dimensional periodic or quasi-periodic tactile signal and determine the period prior information of the signal within a preset search range based on autocorrelation or autopower spectrum analysis; the joint optimization modeling module is used to construct the segmentation start point and period length of multiple periodic segments as joint optimization variables under the period prior constraints, and construct a global optimization objective function based on the contour similarity and structural constraint terms between segments; the constraint solving module is used to solve the global optimization objective function under the period prior and structural constraint conditions to obtain the optimal segmentation start point and period length of each periodic segment; the segment extraction module is used to automatically extract the smallest effective segment containing multiple effective periods from the original tactile signal according to the optimal segmentation start point and period length, and output the phase-aligned periodic signal for subsequent processing.

[0061] In another aspect, the present invention also provides a computer device, including a processor and a memory, wherein the memory stores a computer program that can run on the processor, and when the computer program is executed, causes the processor to perform the steps of the method of the present invention to achieve joint optimized segmentation and phase alignment of one-dimensional periodic or quasi-periodic tactile signals.

[0062] In another aspect, the present invention also provides a computer-readable storage medium storing a computer program thereon, which, when executed by a processor, performs the steps and flow of the method of the present invention, thereby realizing the periodic prior acquisition, joint optimization modeling, constraint solving, and fragment extraction of one-dimensional periodic or quasi-periodic tactile signals.

[0063] The beneficial effects of this invention are as follows: It elevates the segmentation and phase alignment of one-dimensional periodic signals from a traditional problem of local feature point detection and template matching to a global optimization problem of solving the segmentation parameter set under period length constraints. Using the overall similarity between all periodic segments as the optimization objective, it eliminates the need for pre-specified reference templates and does not rely on prominent rising edges, peaks, or other local feature points. This allows for stable segmentation and alignment results even in tactile signal scenarios with complex internal morphologies and indistinct local features. By introducing the approximate period range obtained from the power spectrum as a soft constraint and constructing the objective function using indicators such as normalized cross-correlation between periodic segments, this invention can simultaneously correct for period length jitter and initial phase deviations, ensuring optimal alignment of all effective periods overall. This significantly improves the morphological consistency of multi-period segments and the reliability of subsequent feature statistics. Compared to the sequential method of segmenting first and then aligning, this invention employs a joint optimization strategy. During the optimization process, it automatically removes or compresses invalid segments at the beginning, end, and middle of the signal, outputting the minimum effective interval containing multiple effective periods. This reduces redundant data and alleviates the computational burden on subsequent recognition modules. Furthermore, the joint optimization segmentation and phase alignment method proposed in this invention has good versatility and scalability. In addition to being applicable to robot tactile signals, it can also be applied to the segmentation preprocessing of other one-dimensional periodic or quasi-periodic signals such as electrocardiogram, pulse wave, and gait vibration, providing a robust front-end processing means for various periodic signal analysis and pattern recognition tasks. Attached Figure Description

[0064] Figure 1 This is a flowchart of the feature extraction method described in this invention.

[0065] Figure 2 Schematic diagram of DAG model.

[0066] Figure 3 Comparison of static channel signal clipping and alignment before and after.

[0067] Figure 4 Comparison of dynamic channel signal clipping and alignment before and after. Detailed Implementation

[0068] The technical solution of this application will be described in detail below with reference to the accompanying drawings.

[0069] like Figure 1 As shown, this application provides a method for segmenting and aligning robot tactile signals. This method is applicable to various tactile sensors such as PVDF piezoelectric films and resistive strain gauges. By unifying the modeling of the period length and segmentation phase, and utilizing the shortest path search strategy in graph theory, high-precision signal extraction is achieved without prior templates. In this embodiment, the data acquisition frequency is set to 1000Hz to ensure the capture of subtle texture features during the robot's sliding contact process.

[0070] Step S1: Estimate the average period using autocorrelation.

[0071] First, the discrete tactile sensing signals acquired by the sensor... Preprocessing is performed. Due to fluctuations in the robotic arm's sliding speed, the signal often exhibits non-stationary quasi-periodic characteristics. This step first performs DC removal and smoothing on the signal, and then uses autocorrelation analysis to obtain the prior periodic information of the signal. Specifically, the autocorrelation function of the signal is calculated, and within a reasonable delay range corresponding to the physical motion frequency (e.g., the frequency band corresponding to the robotic arm's movement speed), the first significant peak of the autocorrelation function is searched. The delay time corresponding to this peak is determined as the approximate period. This parameter will serve as the search benchmark for subsequent joint optimization steps, providing physical constraints for constructing the search space.

[0072] Step S2: Unify modeling of the set of segmentation tangent points With the set of segment durations

[0073] To address the issue of segmentation error accumulation in traditional serial methods, this step involves dividing the starting sequence of all periods. and the corresponding segment length sequence The unified model is used for the joint variables to be optimized. In specific implementation, a modeling method based on directed acyclic graphs (DAGs) is adopted, such as... Figure 2 As shown. First, with Using the basic step size, N stages are divided on the time axis, and a search window is defined near the theoretical segmentation position of each stage; each discrete-time sampling point within the search window constitutes a node in the graph, representing the tangent point. The potential locations are then identified. Subsequently, connections between nodes are established based on physical rationality constraints. That is, a directed edge is established between two nodes only when the time difference between two adjacent stages falls within a preset periodic fluctuation range. This directed edge represents a candidate signal periodic segment. In this way, the high-dimensional combinatorial optimization problem is transformed into a path search problem in graph theory.

[0074] Step S3: Construct an optimization objective function by aggregating similarity and structural constraints.

[0075] To quantitatively evaluate the merits of segmentation paths in the aforementioned DAG model, a joint objective function incorporating morphological similarity and structural stability needs to be constructed. To facilitate solving this using shortest path algorithms, this embodiment will "maximize the global objective function." This is transformed into the form of "minimizing the cumulative weight of the path". Specifically, it defines the connection between adjacent nodes. and edge weight As shown in the formula below:

[0076]

[0077] in, For the similarity term, normalized cross-correlation with time shift maximization (NCC) is preferred to calculate the current candidate fragment. The degree of matching with the current reference waveform; the larger the value, the more significant the waveform characteristics. This is a structure regularization term used to calculate the length of the current segment. Target length of outer iteration The smaller the deviation between the values ​​(e.g., the square of the deviation), the more stable the period is; and The coefficients are used to balance the weights. By defining these weights, we find the path with the minimum cumulative weight in the graph, which is equivalent to finding a set of segmentation schemes with the most consistent shape and the most uniform periodic distribution.

[0078] Step S4: By maximizing Joint iterative solution for optimal and

[0079] This step employs a nested optimization strategy to solve the above model. First, in the outermost loop, in... Within the neighborhood (e.g., region) Set a series of candidate global cycles For each fixed Run a dynamic programming algorithm (such as the Viterbi algorithm or Dijkstra's algorithm) on the DAG graph constructed in step S2 to search for a cumulative weight from the starting layer to the ending layer. The shortest path. Finally, compare all candidates. The minimum cost of the corresponding path is used to select the set with the lowest global cost as the final solution. The sequence of nodes traversed by this optimal path is the optimal splitting starting point set. The time difference between nodes is the set of optimal segment lengths. .

[0080] Step S5: Extract the segmented and aligned fragments and perform subsequent processing

[0081] Based on the optimal parameters calculated in step S4, N consecutive valid signal segments are directly extracted from the original signal. For example... Figure 3 and Figure 4The figures shown are schematic diagrams illustrating the effects of the above processing on the PVDF and SG channels, respectively. The upper subplot shows the precisely locked segmentation positions in the original cluttered signal, while the lower subplot shows the waveforms after segmentation and overlapping. It can be seen that after joint optimization processing, all periodic segments achieve high phase alignment, and the waveform consistency is significantly improved.

[0082] Building upon this foundation, to meet the input specifications of subsequent recognition algorithms, the aligned segments undergo baseline correction and outlier removal. Specifically, the mean of the first segment of data for each segment is calculated as the baseline and subtracted to eliminate DC drift from the sensor. Simultaneously, the root mean square (RMS) or other statistical characteristics of all segments are calculated to construct a feature distribution. A Z-score algorithm (e.g., setting a threshold of 2.5) is used to automatically detect and remove outlier segments caused by accidental robotic arm jitter. Finally, based on specific task requirements, a fixed-length region of interest (ROI) is extracted from the aligned waveform, outputting the final standardized dataset.

Claims

1. A period-phase optimized method for segmenting and aligning robot tactile signals, applied to processing discrete tactile sensing signals. Its characteristics are, This includes the following steps performed by the processor: S1. Acquire the discrete tactile sensing signal And through autocorrelation analysis, an approximate period is estimated within a preset search window. ; S2. Set the target number of segments The signal Each segmentation starting point and The length of each segment The unified model is used as a joint variable to be optimized; S3. Construct a joint optimization objective function The At least one is included for maximizing the Aggregate similarity term of morphological similarity between segments , and a constraint for the and Regularization terms for structural rationality ,in For the first A segment; S4. In the case described Under the determined structural constraints, the objective function is solved using a numerical optimization algorithm. In order to obtain the optimal set of starting points for segmentation. and the optimal set of fragment lengths ; S5. According to the above and From the signal Extract A segment that has achieved phase alignment.

2. The period-phase optimized robot tactile signal segmentation and alignment method according to claim 1, characterized in that, The signal acquired in S1 is a multi-channel signal. The objective function in S3 Constructed as a channel-weighted aggregation form: ； Among them, weight It can be determined based on the consistency between channels.

3. The period-phase optimized robot tactile signal segmentation and alignment method according to claim 1, characterized in that, In S1 The estimation employs an autocorrelation peak search, with the search window limited to a preset period range located in the positive half-axis delay region of the autocorrelation function, and the peak with the strongest correlation selected within this range.

4. The period-phase optimized robot tactile signal segmentation and alignment method according to claim 1, characterized in that, The segment length in S2 Modeled as: (a) Equal length mode, i.e. ,in For the scalar to be optimized; or (b) Variable length mode, i.e. ,in .

5. The period-phase optimized robot tactile signal segmentation and alignment method according to claim 1, characterized in that, The aggregation operator in S3 Specifically limited to the arithmetic mean, used to calculate all segment pairs Morphological similarity between The average value.

6. The period-phase optimized robot tactile signal segmentation and alignment method according to claim 1, characterized in that, The aggregation operator in S3 The morphological similarity measure used Specifically defined as normalized cross-correlation with time-shift maximization The calculation method is as follows: ； in, and Each represents a segment and At any moment The signal value, and These are the mean values ​​of the corresponding segments. For the allowed time offset, This represents the maximum offset range.

7. The period-phase optimized robot tactile signal segmentation and alignment method according to claim 1, characterized in that, The structural regularization term in S3 Includes deviation penalties for the interval sequence: ； Or a length smoothing penalty under the variable length mode as described in claim 4(b): ； in For the expected period, .

8. The period-phase optimized robot tactile signal segmentation and alignment method according to claim 1, characterized in that, The structural constraints in S4 include at least constraints on the intervals between adjacent starting points:

9. Among them The minimum interval coefficient, This is the maximum interval coefficient.

10. The period-phase optimized robot tactile signal segmentation and alignment method according to claim 1, characterized in that, The numerical optimization algorithm in S4 is specifically implemented as a nested optimization shortest path search method, including: Sa1. Iterate through a set of candidate fragment lengths that satisfy claim 4. When the equal-length mode of claim 4(a) is used, the length is configured as a single scalar. When the variable-length mode of claim 4(b) is used, the length is configured as a vector. ; Sa2. For each candidate segment length In the candidate state grid Construct a directed acyclic graph, where the nodes of the graph represent candidate starting points. ; Sa3. Define the connection between adjacent nodes. and Boundary rights The edge weight Constructed to jointly optimize the objective function with that in step S3 of claim 1. Corresponding local costs; When using the equal-length mode, its form is: ； When using the variable-length mode, its form is: ； Among them, morphological similarity measurement and structural regularization terms Determined according to claims 6 and 7 respectively; Sa4 uses the shortest path algorithm to find a cumulative cost path in the graph. The minimum optimal path is obtained. The corresponding optimal starting sequence ; Sa5 selects the objective function jointly optimized in step S3 of claim 1. To reach the maximum The combination serves as the optimal starting point set for the optimization. and the set of optimal fragment lengths .

11. The period-phase optimized robot tactile signal segmentation and alignment method according to claim 9, characterized in that, Candidate state grid in step Sa2 Constructed in the following manner: A reference grid is formed on the time axis for the basic step size. Where r is the bias, And define a finite set of discrete perturbations for each grid point. ,in Based on this, the candidate starting point set is obtained. and as defined in claim 8 Interval constraint filtering Infeasible point.

12. The period-phase optimized robot tactile signal segmentation and alignment method according to claim 1, characterized in that, Following S5, the extraction is also included. The steps for baseline correction of each segment: For each segment Before the starting point The mean of the data is used as the baseline. And deduct it.

13. The period-phase optimized robot tactile signal segmentation and alignment method according to claim 11, characterized in that, Following baseline correction, an outlier removal step is included: this is achieved by calculating the root mean square value of each fragment or its aggregated similarity score with the remaining fragments, and based on... Outlier detection is performed on the scores to remove abnormal segments.

14. The period-phase optimized robot tactile signal segmentation and alignment method according to claim 1, characterized in that, Following S5, a secondary cropping step is also included: in the already aligned Within the common morphological region of each segment, a window of interest is extracted for subsequent tactile recognition tasks of the same type.

Images