Spine touch sensing system and method based on multi-mode sensor

Abstract

The invention discloses a spine tactile sensing system and method based on a multi-modal sensor, and the method comprises the steps: obtaining an original tactile signal, carrying out the preprocessing of the original tactile signal, carrying out the feature extraction based on the preprocessed original tactile signal, and obtaining a spatial distribution feature and a time dynamic feature; on the basis of tail end coordinates collected by forward kinematics of the tail end of the mechanical arm, space mapping of the tactile signals and the spine anatomical structure is constructed, and anatomical position features are obtained on the basis of the preprocessed original tactile signals and the space mapping; constructing task-oriented feature vectors based on the spatial distribution features, the anatomical position features and the time dynamic features, wherein the task-oriented feature vectors comprise a sliding state recognition feature vector and a spine ordinal number positioning feature vector; and constructing a multi-stage adaptive evolution hidden Markov model, and realizing spine ordinal number positioning in combination with the task-oriented feature vector. According to the method, through multi-dimensional feature fusion and dynamic model adjustment, anatomical prior knowledge and real-time tactile signals are combined, and spine segments are accurately recognized.

Description

Technical Field

[0001] The present invention belongs to the technical field of tactile perception, and in particular relates to a spinal tactile perception system and method based on a multimodal sensor. Background Art

[0002] Traditional methods often rely on single-modal tactile information for spinal localization, such as data from pressure or capacitance sensors. However, this single-feature perception method is highly susceptible to factors such as complex environmental noise and soft tissue interference, resulting in insufficient positioning accuracy. For example, given the individual differences between patients, a single feature cannot accurately distinguish bony landmarks from soft tissue, leading to misjudgment and severely impacting the accuracy and reliability of spinal ordinal localization.

[0003] Furthermore, most existing models are static Hidden Markov Models (HMMs). These models have significant limitations in practical applications. On the one hand, their observation probabilities and state transition matrices cannot be dynamically adjusted based on real-time signals, making them difficult to adapt to the dynamic changes in contact states during spinal tactile perception. On the other hand, the models lack adaptability and stability in the face of individual differences and complex and changing tactile contact scenarios, making it impossible to guarantee sustained and stable positioning performance.

[0004] To solve the above problems, the present invention proposes a spinal tactile perception system and method based on a multimodal sensor. Summary of the Invention

[0005] In order to solve the above technical problems, the present invention proposes a spinal tactile perception system and method based on a multimodal sensor to solve the problems existing in the above-mentioned prior art.

[0006] To achieve the above objectives, the present invention provides a spinal tactile perception method based on a multimodal sensor, comprising:

[0007] Obtaining and preprocessing the original tactile signal, performing feature extraction based on the preprocessed original tactile signal to obtain spatial distribution features and temporal dynamic features;

[0008] Based on the end coordinates collected by the forward kinematics of the robotic arm end, a spatial mapping between the tactile signal and the spinal anatomical structure is constructed, and the anatomical position features are obtained based on the preprocessed raw tactile signal and the spatial mapping;

[0009] Constructing a task-oriented feature vector based on the spatial distribution feature, the anatomical position feature, and the temporal dynamic feature, wherein the task-oriented feature vector includes a sliding state recognition feature vector and a vertebral ordinal location feature vector;

[0010] A multi-level adaptive evolutionary hidden Markov model is constructed and combined with the task-oriented feature vector to achieve spine ordinal positioning.

[0011] Optionally, the original tactile signal includes an original pressure signal and a capacitance sensor signal;

[0012] The preprocessing process includes: determining the filter window length according to the noise spectrum, using a moving average filter algorithm to reduce the noise of the original pressure signal; and using polynomial fitting to calibrate the capacitance sensor signal.

[0013] Optionally, the spatial distribution characteristics include pressure gradient characteristics and contact area gradient change rate characteristics, and the temporal dynamic characteristics include pressure change rate, contact area change rate and contact point lateral velocity.

[0014] Optionally, a sliding state identification feature vector is constructed based on the contact area gradient change rate feature, the contact area change rate, and the contact point lateral velocity.

[0015] Optionally, a vertebral ordinal positioning feature vector is constructed based on pressure gradient features, anatomical position features, and pressure change rate.

[0016] Optionally, the multi-level adaptive evolutionary hidden Markov model includes a sliding state monitoring layer and a segment positioning layer;

[0017] The sliding state monitoring layer obtains the contact state of the mechanical fingertip based on the sliding state identification feature vector;

[0018] The segment positioning layer obtains the spinal segment based on the contact state and the spinal ordinal positioning feature vector.

[0019] Optionally, the segment positioning layer determines whether the contact state is fixed contact or sliding exploration. If it is fixed contact, a high-precision observation model is used for segment positioning; if it is sliding exploration, an anti-interference observation model is used for segment positioning, wherein the high-precision observation model integrates pressure gradient characteristics and displacement differences, and the anti-interference observation model strengthens the dynamic weights of the lateral velocity of the contact point and the contact area change rate.

[0020] Optionally, the multi-level adaptive evolutionary hidden Markov model also includes an adaptive adjustment mechanism: during the segment positioning process, a confidence factor is introduced to measure the confidence level of the current spinal segment identification result; if the confidence factor is lower than a preset threshold, the parameters in the model are locally recalibrated according to a preset learning rate; if the confidence factors of multiple consecutive frames are lower than the preset threshold, the state transfer matrix is ​​reset to an anatomical prior matrix based on anatomical prior knowledge, where the transition probability between adjacent segments is set to be greater than a preset value.

[0021] The present invention also provides a spinal tactile perception system based on a multimodal sensor, comprising:

[0022] A data processing module, used to obtain raw tactile signals and perform preprocessing;

[0023] A feature extraction module is configured to extract features based on the preprocessed raw tactile signals to obtain spatial distribution features and temporal dynamic features; construct a spatial mapping between the tactile signals and the spinal anatomical structure based on the end coordinates collected by the forward kinematics of the end of the manipulator, and obtain anatomical position features based on the preprocessed raw tactile signals and the spatial mapping;

[0024] A task-oriented module, configured to construct a task-oriented feature vector based on the spatial distribution feature, the anatomical position feature, and the temporal dynamic feature, wherein the task-oriented feature vector includes a sliding state recognition feature vector and a vertebral ordinal location feature vector;

[0025] The perception module is used to construct a multi-level adaptive evolutionary hidden Markov model and realize the spine ordinal location in combination with the task-oriented feature vector.

[0026] Compared with the prior art, the present invention has the following advantages and technical effects:

[0027] The present invention achieves accurate recognition of 20 segments from C3 to L5 through the fusion of multi-dimensional tactile features and dynamic modeling of AE-HMM, solving the problem of misjudgment of bony landmarks caused by a single feature.

[0028] The present invention suppresses circuit noise, temperature drift and soft tissue interference through hierarchical signal processing (moving average filtering, polynomial calibration) and anti-interference observation model switching, solves the problem of insufficient stability of existing technologies in complex environments, and enhances anti-interference capability.

[0029] The present invention utilizes the adaptive evolution strategy of AE-HMM (confidence-driven parameter calibration, anatomical prior matrix reset) to solve the problem of poor adaptability of traditional static models to individual differences and dynamic contact states, and realizes real-time optimization of model parameters. BRIEF DESCRIPTION OF THE DRAWINGS

[0030] The accompanying drawings, which constitute part of this application, are intended to provide a further understanding of this application. The exemplary embodiments and descriptions of this application are intended to explain this application and do not constitute an improper limitation on this application. In the accompanying drawings:

[0031] Figure 1 This is a flow chart of an average filtering algorithm according to an embodiment of the present invention;

[0032] Figure 2 This is a flow chart of capacitance sensor signal calibration according to an embodiment of the present invention;

[0033] Figure 3 Schematic diagram of the AE-HMM process according to an embodiment of the present invention.

[0034] Figure 4 4 is a flow chart of the overall method of an embodiment of the present invention. DETAILED DESCRIPTION

[0035] It should be noted that, in the absence of conflict, the embodiments and features of the embodiments in this application can be combined with each other. The present application will be described in detail below with reference to the accompanying drawings and in combination with the embodiments.

[0036] It should be noted that the steps shown in the flowcharts of the accompanying drawings can be executed in a computer system such as a set of computer-executable instructions, and that, although a logical order is shown in the flowcharts, in some cases, the steps shown or described can be executed in an order different from that shown here.

[0037] Example 1

[0038] like Figure 4 As shown, this embodiment provides a spinal tactile perception method based on a multimodal sensor, including:

[0039] Step 1: Data processing stage;

[0040] During the data processing phase, the original signal needs to be denoised, calibrated, and compensated for interference to improve signal quality, as follows:

[0041] The original pressure signal P of the piezoresistive sensor signal processing is easily interfered by circuit noise. The moving average filter algorithm is used to reduce the noise. Let the filter window length be N. According to the noise spectrum characteristics, when the sampling frequency is 1kHz, take N = 21 (corresponding to a cutoff frequency of about 48Hz), which can effectively suppress the circuit power frequency interference. The processing process is as follows: Figure 1 As shown, the pressure signal after filtering is:

[0042]

[0043] The formula suppresses high-frequency noise and retains the true trend of the pressure signal.

[0044] Capacitive sensor signal calibration;

[0045] When the multimodal sensor integrated at the end of the robotic arm touches the spine, the change in contact area (ΔA) output by the capacitive sensor (referring to the contact area between the capacitive sensor on the robotic arm and the human body) is subject to temperature drift and nonlinear errors. Polynomial fitting is used for calibration. Assuming the calibration coefficients a, b, and c, the change in contact area after calibration is:

[0046] ΔA1=a·(ΔA) 2 +b·ΔA+c;

[0047] Through experimental calibration, coefficients a, b, and c are obtained to eliminate nonlinear errors and improve the contact area measurement accuracy. The capacitance sensor signal calibration process is as follows: Figure 2 shown.

[0048] Step 2: Feature extraction;

[0049] (1) Spatial distribution feature extraction;

[0050] Extract the pressure gradient feature from the filtered pressure signal P1:

[0051] Pressure gradient characteristics The pressure distribution gradient on the contact surface of the mechanical fingertip is calculated to distinguish the edge contour of the bony landmark from the smooth contact of the soft tissue as follows:

[0052]

[0053] High gradient values ​​(>15 Pa / mm) indicate sharp edges (such as spinous processes, pedicle boundaries), and low gradient values ​​(<5 Pa / mm) indicate flat or soft tissue contact.

[0054] Extract the contact area gradient change rate feature from the contact area change ΔA1 after calibration:

[0055] Contact area gradient change rate characteristics Characterize the rate of change of the contact area on the fingertip surface and identify the sliding contact and fixed contact states as follows:

[0056]

[0057] Where S is the length of the contact boundary curve. It shows periodic fluctuations and approaches 0 when the pedicle is positioned.

[0058] (2) anatomical location characteristics;

[0059] 3D contact point coordinates (x, y, z): These coordinates are directly derived from the 3D position of the fingertip contact point using the robot's forward kinematics. Here, the z-axis is defined as the longitudinal direction of the spine, with the head end direction as the positive direction. This coordinate system allows for the description and location of the fingertip contact point on the spinal surface in a manner consistent with anatomical and mechanical manipulation conventions.

[0060] The longitudinal displacement of adjacent points (Δz): by calculating the z coordinate value of the current contact point (z t ) and the z coordinate value of the last valid contact point (z t-1 ) gives the difference between:

[0061] Δz=zt -z t-1 ;

[0062] Where Δz reflects the continuity of longitudinal translation along the spine. When moving across consecutive vertebral surfaces, |Δz| should be close to the typical value of the distance between adjacent vertebrae (e.g., ≈25 mm for the lumbar spine). Abnormal Δz may indicate skipping of vertebrae or contact with non-target structures.

[0063] (3) Temporal dynamic characteristics;

[0064] Capturing the rate of change of the signal over time is crucial for identifying dynamic contact states (such as sliding) and sequence modeling.

[0065] Pressure change rate (dP / dt):

[0066]

[0067] The dP / dt value (Δt is the sampling interval) is approximated using first-order differences. This value characterizes the instantaneous rate of change of contact pressure. A rapid rise (dP / dt > 0) may indicate contact with a bony prominence, while a rapid fall (dP / dt < 0) may indicate loss of contact. Sliding on a bony surface may be accompanied by small fluctuations.

[0068] Contact area change rate (d(ΔA) / dt):

[0069]

[0070] The first-order difference approximation d(ΔA) / dt is used to directly quantify the dynamic change of the contact area. It usually shows obvious non-zero fluctuations in sliding contact and approaches zero in stable contact. The lateral velocity of the contact point (v_xy):

[0071]

[0072] Characterizes the speed at which the contact point moves in a plane perpendicular to the longitudinal axis of the spine (z-axis). A significant v_xy is a strong indicator of sliding contact.

[0073] Get a core set of low-level features: x, y, z, Δz, dP / dt, d(ΔA) / dt, (v_xy). These features provide the basis for the subsequent task-oriented feature vector construction.

[0074] Step 3: Task-oriented feature vector construction;

[0075] Based on the core underlying features extracted in step 2, this step combines and constructs a multi-dimensional feature vector with high discrimination according to the specific requirements of different perception tasks, which serves as the input of the subsequent pattern recognition algorithm (AE-HMM).

[0076] (1) Sliding state identification feature vector (F_slide);

[0077] Task objective: Determine whether the current contact state is fixed or sliding (such as exploring on the transverse process surface).

[0078] Eigenvectors:

[0079]

[0080] Key Features:

[0081] Reflects the movement of the contact boundary in space.

[0082] d(ΔA) / dt: directly quantifies the dynamic changes in contact area, which fluctuates significantly during sliding.

[0083] v_xy: Direct evidence of lateral movement velocity.

[0084] (2) spine ordinal location feature vector (F_seg);

[0085] Task objective: Identify the specific spinal segment (C3-L5) to which the current contact point belongs.

[0086] Key Features:

[0087] Core bony contact strength characteristics (especially sensitive to the spinous processes).

[0088] z: Absolute longitudinal position, which is the main basis for segment positioning.

[0089] Δz: key indicator of sequence continuity (matched a priori with anatomical vertebral spacing).

[0090] dP / dt: Enhanced modeling capabilities for contact dynamics and sequence changes.

[0091] Eigenvectors:

[0092]

[0093] Dimensionality: 4 dimensions (including dP / dt). dP / dt is often more relevant for bony localization than d(ΔA) / dt.

[0094] Step 4: Spine localization based on adaptive evolutionary hidden Markov model (AE-HMM) with multi-feature fusion.

[0095] In this stage, the task-oriented feature vectors (F_slide and F_seg) generated in step 3 are used as input to construct a multi-level adaptive evolutionary hidden Markov model (AE-HMM). Its main purpose is to achieve ordinal recognition of the 20 spinal segments from C3 to L5, that is, to accurately determine the specific spinal segment where the current contact point is located. The core structure of the model consists of a two-level processing chain:

[0096] (1) Sliding state detection layer: Based on F_slide, the contact state (fixed / sliding) is identified and the state label S_t∈{0,1} is output. The F_slide feature vector is mainly used here to identify the contact state and determine whether it is currently in a fixed state or a sliding state. Finally, a state label is output to inform the subsequent processing layer whether the current contact is a fixed contact at a certain position by the fingertip or a sliding exploration on the spinal surface.

[0097] (2) Segment positioning layer: Using S_t as a conditional switch, dynamically select the HMM observation probability model:

[0098] In this layer, the appropriate HMM observation probability model is dynamically selected based on the state label S_t output by the primary layer. Simply put, the observation model is adjusted based on whether the current contact is fixed or sliding, making it more consistent with the actual contact situation, thereby more accurately positioning the segment.

[0099] When S_t=0, that is, in a fixed contact state, a high-precision observation model is used. This model incorporates the pressure gradient The two features of the longitudinal displacement difference Δz are combined with prior knowledge of the anatomical structure (such as the typical longitudinal spacing of different vertebral segments) to accurately locate the vertebral segments.

[0100] When S_t = 1, indicating a sliding contact state, the anti-interference observation model is enabled. The weights of the two dynamic features, lateral velocity v_xy and contact area change rate d(ΔA) / dt, are increased. These dynamic features better reflect changes in the contact state during sliding, helping to mitigate interference caused by sliding and improve positioning accuracy.

[0101] like Figure 3 As shown, the AE-HMM algorithm process is as follows:

[0102] Assume that the hidden state Q_t represents the spinal segment number Q_t∈{C3,C4,...,L5} at time t (with a total of 20 possible segment numbers), and the observation sequence is the series of feature vectors obtained in the previous step. The entire model is then iterated and updated according to the following process to continuously optimize the recognition of spinal segments:

[0103] (1) Forward-backward probability calculation;

[0104] The modified Baum-Welch algorithm is used here to calculate the state transition probability. The so-called state transition probability is simply the probability of transitioning from one spinal segment to another, which reflects the correlation between the contact sequences of different spinal segments.

[0105] The improved Baum-Welch algorithm is used to calculate the state transition probability:

[0106]

[0107] Among them, α t (j) is the forward probability, which refers to the probability of being in state j at time step t and observing the first t observations; b j (O t ) is the observation probability, that is, observing O in state j t The probability of (t-th observation), O t =[F_seg(t);F_slide(t)];a ij is the transition probability from state i to j (constrained by Δz, the transition probability decays when |Δz|>25mm).

[0108] During the calculation process, the state transition probability is constrained by the longitudinal displacement difference Δz. If the absolute value of Δz exceeds 25mm (because the typical value of the distance between adjacent vertebrae is about 25mm), the state transition probability will decay:

[0109] The attenuation factor w(Δz) is defined as the adjustment coefficient of the state transition probability, and its expression is:

[0110]

[0111] Where k is the attenuation coefficient.

[0112] According to the attenuation factor w(Δz), the attenuation probability after attenuation can be calculated:

[0113] a ij1 =a ij0 w(Δz)

[0114] where a ij0 is the original transition probability.

[0115] The logic behind this is: under normal circumstances, when moving from one vertebra to the adjacent vertebra, the longitudinal displacement difference should be close to this typical value. If it exceeds too much, it may mean that the contact point has jumped abnormally rather than the normal adjacent segment transfer. Therefore, the probability of state transfer in this case should be reduced.

[0116] f j (Ot ): Observation probability density function, modeled by Gaussian mixture model (GMM):

[0117]

[0118] Among them, ω jk is the weight of the kth Gaussian component in the jth state, N(O t ;μ jk ;∑ jk ) is the probability density function of the kth Gaussian component in the jth state; μ jk is the mean vector of the kth Gaussian component in the jth state; ∑ jk is the covariance matrix of the kth Gaussian component in the jth state.

[0119] The observation probability density function is modeled by the Gaussian mixture model (GMM), which can describe the probability of observing the current feature vector in a given spinal segment. By constructing such a probability model, the model can better understand and match the relationship between different feature patterns and the corresponding spinal segments.

[0120] (2) Adaptive evolution strategy;

[0121] Introducing confidence factor:

[0122]

[0123] A confidence factor C is introduced t , which ranges from 0 to 1, is used to measure the confidence level of the current spinal segment recognition results. t When the value of is high, it means that the current recognition result is relatively reliable; otherwise, there may be greater uncertainty.

[0124] Drive parameter adjustment: If C t <τlow (τlow=0.7), start local recalibration:

[0125] μ jk ←μ jk +η·(O t -μ jk );

[0126] ∑ jk ←(1-η)∑ jk +η(O t -μ jk )(O t -μ jk ) T ;

[0127] Where η = 0.05 is the learning rate.

[0128] If C t If the value of is less than a lower threshold τlow (here, 0.7), local recalibration is initiated. Specifically, this involves adjusting relevant model parameters (such as the mean vector μ and the covariance matrix ∑) based on a pre-set learning rate η (here, 0.05). This adjustment can be understood as fine-tuning the model when the current recognition result is uncertain, allowing it to better adapt to the current input features, thereby improving the accuracy of subsequent recognition.

[0129] If the Ct value for five consecutive frames (i.e., five consecutive time steps) is less than τlow, further action is taken to directly reset the state transition matrix to the anatomical prior matrix. This anatomical prior matrix is ​​constructed based on prior knowledge of the spinal anatomy, in which the transition probability between adjacent segments is set to greater than 0.8. This fully utilizes the known anatomical information and redirects the model toward more reasonable state transitions for learning and recognition, preventing the model from falling into unreasonable state transition patterns.

[0130] (3) Decision output;

[0131] Finally, the optimal state sequence is solved using the Viterbi algorithm. Essentially a dynamic programming algorithm, the Viterbi algorithm finds the state sequence path most likely to produce the current observed feature sequence among many possible state sequences. In other words, it determines the spinal segment sequence that best matches the current continuous tactile signal.

[0132] First, the probability of the optimal path is solved by the Viterbi algorithm:

[0133] δ t (j) = max i [δ t-1 (i) a ij ]·f j (O t );

[0134] Among them, δt(j) is the optimal path probability observed in the previous t frames when the time step t is in state j, δ t-1 (i) is the optimal path probability at state i at time step (t-1), a ij is the state transition probability from i to j, f j (O t ) is the observation probability density function.

[0135] Then find the optimal state:

[0136] Q^t=argmax j δ t (j);

[0137] Among them, Q^t is the optimal state at time step t.

[0138] Specifically, at time step t, traverse all possible spinal segments j and find the one that makes δ t (j) The largest segment is the most likely contact segment in the current frame.

[0139] The final output is the spinal segment number Q^t and confidence C.

[0140] This embodiment also provides a spinal tactile perception system based on a multimodal sensor, comprising:

[0141] A data processing module, used to obtain raw tactile signals and perform preprocessing;

[0142] The feature extraction module is used to extract features based on the preprocessed raw tactile signals to obtain spatial distribution features and temporal dynamic features. Based on the end coordinates collected by the forward kinematics of the end of the manipulator, a spatial mapping between the tactile signals and the spinal anatomical structure is constructed, and anatomical position features are obtained based on the preprocessed raw tactile signals and the spatial mapping.

[0143] A task-oriented module is used to construct a task-oriented feature vector based on spatial distribution features, anatomical position features, and temporal dynamic features. The task-oriented feature vector includes a sliding state recognition feature vector and a vertebral ordinal positioning feature vector;

[0144] The perception module is used to build a multi-level adaptive evolutionary hidden Markov model and combine it with task-oriented feature vectors to achieve spine ordinal positioning.

[0145] The above are merely preferred embodiments of the present application, but the scope of protection of the present application is not limited thereto. Any changes or substitutions that can be easily conceived by a person skilled in the art within the technical scope disclosed in this application should be included in the scope of protection of the present application. Therefore, the scope of protection of the present application should be based on the scope of protection of the claims.

Claims

1. A spinal tactile perception method based on a multimodal sensor, characterized in that: The following steps are involved: Obtaining and preprocessing the original tactile signal, performing feature extraction based on the preprocessed original tactile signal to obtain spatial distribution features and temporal dynamic features; Based on the end coordinates collected by the forward kinematics of the robotic arm end, a spatial mapping between the tactile signal and the spinal anatomical structure is constructed, and the anatomical position features are obtained based on the preprocessed raw tactile signal and the spatial mapping; Constructing a task-oriented feature vector based on the spatial distribution feature, the anatomical position feature, and the temporal dynamic feature, wherein the task-oriented feature vector includes a sliding state recognition feature vector and a vertebral ordinal location feature vector; A multi-level adaptive evolutionary hidden Markov model is constructed and combined with the task-oriented feature vector to achieve spine ordinal positioning.

2. The spinal tactile perception method based on a multimodal sensor according to claim 1, characterized in that: The original tactile signal includes an original pressure signal and a capacitance sensor signal; The preprocessing process includes: determining the filter window length according to the noise spectrum, and using a moving average filter algorithm to reduce the noise of the original pressure signal; The capacitance sensor signal is calibrated using polynomial fitting.

3. The spinal tactile perception method based on a multimodal sensor according to claim 1, characterized in that: The spatial distribution characteristics include pressure gradient characteristics and contact area gradient change rate characteristics, and the temporal dynamic characteristics include pressure change rate, contact area change rate and contact point lateral velocity.

4. The spinal tactile perception method based on a multimodal sensor according to claim 3, characterized in that: The sliding state identification feature vector is constructed based on the contact area gradient change rate characteristics, contact area change rate and contact point lateral velocity.

5. The spinal tactile perception method based on a multimodal sensor according to claim 3, characterized in that: A vertebral ordinal positioning feature vector is constructed based on pressure gradient characteristics, anatomical position characteristics, and pressure change rate.

6. The spinal tactile perception method based on a multimodal sensor according to claim 1, characterized in that: The multi-level adaptive evolutionary hidden Markov model includes a sliding state monitoring layer and a segment positioning layer; The sliding state monitoring layer obtains the contact state of the mechanical fingertip based on the sliding state identification feature vector; The segment positioning layer obtains the spinal segment based on the contact state and the spinal ordinal positioning feature vector.

7. The spinal tactile perception method based on a multimodal sensor according to claim 6, characterized in that: The segment positioning layer determines whether the contact state is fixed contact or sliding exploration. If it is fixed contact, a high-precision observation model is used for segment positioning. If it is sliding exploration, an anti-interference observation model is used for segment positioning. The high-precision observation model integrates pressure gradient characteristics and displacement differences, and the anti-interference observation model strengthens the dynamic weights of the lateral velocity of the contact point and the contact area change rate.

8. The spinal tactile perception method based on a multimodal sensor according to claim 7, characterized in that: The multi-level adaptive evolutionary hidden Markov model also includes an adaptive adjustment mechanism: during the segment positioning process, a confidence factor is introduced to measure the confidence level of the current spinal segment identification result; if the confidence factor is lower than the preset threshold, the parameters in the model are locally recalibrated according to the preset learning rate; if the confidence factors of multiple consecutive frames are lower than the preset threshold, the state transfer matrix is ​​reset to an anatomical prior matrix based on anatomical prior knowledge, where the transition probability between adjacent segments is set to be greater than a preset value.

9. A spinal tactile perception system based on a multimodal sensor for executing the method according to any one of claims 1 to 8, characterized in that: include: A data processing module, used to obtain raw tactile signals and perform preprocessing; A feature extraction module is used to extract features based on the preprocessed original tactile signal to obtain spatial distribution features and temporal dynamic features; Based on the end coordinates collected by the forward kinematics of the end of the manipulator, a spatial mapping between the tactile signal and the spinal anatomical structure is constructed, and the anatomical position features are obtained based on the preprocessed original tactile signal and the spatial mapping; A task-oriented module, configured to construct a task-oriented feature vector based on the spatial distribution feature, the anatomical position feature, and the temporal dynamic feature, wherein the task-oriented feature vector includes a sliding state recognition feature vector and a vertebral ordinal location feature vector; The perception module is used to construct a multi-level adaptive evolutionary hidden Markov model and realize the spine ordinal location in combination with the task-oriented feature vector.

Images