A signal processing-based traction stretch movement state monitoring method

Abstract

The application discloses a traction stretching motion state monitoring method based on signal processing and relates to the technical field of signal processing, which comprises the following steps: collecting traction force, electrocardio, displacement and conductivity signals, performing normalization and filtering processing, extracting phase and amplitude components, and constructing a signal atlas containing trajectory changes and timing characteristics; establishing a structural response mapping relationship to generate a self-evolution characteristic map, aligning the amplitude components to form a channel-level characteristic matrix and a multi-channel characteristic representation map; adopting structural path perception convolution mapping and introducing a path dissipation factor to extract deep features, combining an energy binding factor and a scheduling matrix to realize direction enhancement, and obtaining an energy regulation characteristic map; and finally, splicing the features according to time steps to form a high-resolution sequence, performing iterative training to realize signal optimization, and improving state recognition accuracy and stability.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of signal processing, and particularly relates to a traction stretching motion state monitoring method based on signal processing. BACKGROUND

[0002] At present, traction stretching training is widely used in neuromuscular function reconstruction, bone rehabilitation and posture correction fields, and its state monitoring is related to the rehabilitation effect. Four types of multi-channel signals, including traction force, electromyography, displacement and conductivity, are generated during traction, which have non-stationary and strong coupling characteristics, increasing the complexity of state recognition. The existing monitoring methods mostly rely on template matching and simple feature extraction, which cannot describe the dynamic evolution of signals and the cooperative relationship between channels, resulting in low recognition accuracy and easy to ignore abnormal response, which is difficult to meet the precise rehabilitation needs. Therefore, there is an urgent need for a traction stretching motion state monitoring method based on signal processing to realize high-precision modeling and optimized monitoring of complex dynamic signals.

[0003] Publication No. CN112843563A discloses a traction training evaluation system based on force sensing and angle sensing fusion, which constructs a training trajectory by collecting displacement and tension information, and combines sensing data for evaluation to realize trajectory reconstruction and state quantization analysis of the training process. Publication No. CN113989529A proposes an electromyography signal state recognition method for rehabilitation training, which constructs a feature set by collecting electromyography signals, and uses a classifier model to recognize training actions to realize classification and evaluation of action states in the rehabilitation training process.

[0004] However, the existing technology does not fully consider the phase amplitude cooperative change of traction stretching signals under continuous time segments and the evolution relationship between segments, ignores the dynamic coupling characteristics between multi-channel signals, the correlation mechanism between trajectory changes and energy distribution, and the cross-channel state adaptive adjustment requirements, resulting in the problems of insufficient state recognition accuracy, poor feature expression continuity and low stability of monitoring results of the proposed method in the scene of complex traction training state and significant individual differences. SUMMARY

[0005] The application provides a traction and stretching motion state monitoring method based on signal processing, aims to build a multi-channel signal feature extraction framework with time sequence modeling and structure perception capability, acquires multi-channel original signal data collected by a traction and stretching device under different operating states, extracts phase components and amplitude components through normalization and filtering processing, and builds a signal atlas data set containing trajectory change structure and time sequence discrete sampling features; according to the phase components in the signal atlas data set, a structure response mapping relationship is established under a time sequence index, and a self-evolution feature map reflecting the traction and stretching state change process is built; the self-evolution feature map and the amplitude components are subjected to time sequence alignment processing, a channel-level feature matrix group is constructed, a multi-channel feature representation map is generated, and the collaborative evolution law exhibited between channels in the traction and stretching process is captured; according to the multi-channel feature representation map, a convolution mapping strategy based on structure path perception is established, a path dissipation factor is introduced to adjust the response strength of high-density paths, difference propagation and response combination operations between channels are performed, and a multi-channel deep feature map representing deep structure changes is obtained; on this basis, the energy distribution of each channel in the deep feature map in the time dimension is calculated, an energy binding factor is introduced to compress and adjust the energy response strength, a channel scheduling correlation matrix is established, and directional reconstruction processing is performed on the original features, and a direction-enhanced energy-adjusted feature map is generated; the feature vectors of each channel in the direction-enhanced feature map at consecutive time steps are spliced in turn to form a feature sequence reflecting the dynamic changes of the traction and stretching state; by building a traction and stretching motion state monitoring signal processing model, inputting the signal atlas data set, sequentially performing self-evolution modeling, channel-level alignment, deep feature extraction and direction enhancement operations, and through iterative training until convergence, precise modeling and optimization processing of the traction and stretching motion state monitoring signal are realized.

[0006] In order to achieve the above purpose, the application provides the following technical scheme: a traction and stretching motion state monitoring method based on signal processing, the specific steps are as follows:

[0007] S1, multi-channel original signals of a traction and stretching device under different operating states are collected, normalized and filtered, phase components and amplitude components are extracted, and a signal atlas data set containing trajectory change structure and time sequence discrete features is built;

[0008] S2, according to the phase components in the signal atlas data set, a structure response mapping relationship under a time sequence index is established, and a self-evolution feature map reflecting the traction and stretching state change process is built;

[0009] S3, the self-evolution feature map and the amplitude components are subjected to time sequence alignment, a channel-level feature matrix group is constructed, and a multi-channel feature representation map is generated;

[0010] S4, according to the multi-channel feature representation map, a structural path-aware convolutional mapping strategy is adopted, a path dissipation factor is introduced, a channel difference propagation and response combination operation is performed, and a multi-channel deep feature map is obtained;

[0011] S5, the time energy distribution of each channel of the deep feature map is calculated, an energy binding factor is introduced, a channel scheduling correlation matrix is established, a directional reconstruction is performed on the original feature, and a direction-enhanced energy-adjusted feature map is generated.

[0012] S6, the feature vectors of each channel in the direction-enhanced feature map are sequentially spliced according to the time steps, and a feature sequence reflecting the change of the traction and stretching state is obtained.

[0013] S7, a traction and stretching motion state monitoring signal processing model is constructed, the input signal atlas data set is sequentially subjected to steps S2 to S6, iterative training is performed until convergence, and the optimization processing of the traction and stretching motion state monitoring signal is realized.

[0014] Preferably, in the S1 step, for the construction of the traction and stretching state monitoring signal atlas data set, first, multi-channel original signal data collected by the traction and stretching device under different operating conditions is obtained, the original signal data including traction force response signals, physiological parameter monitoring signals, electromyographic signals, acceleration signals and attitude angle sensing signals; in view of the differences in physical quantity units and sampling methods of different channel signals, a channel normalization strategy is adopted to perform amplitude standardization processing on the original data, and an anti-interference preprocessing is performed based on a combination of a sliding window and a filter to remove electromagnetic noise, body motion artifacts and baseline drift interference; subsequently, the processed signal sequence is subjected to time sequence alignment operation, so that different channels form synchronous sampling results on the same time axis; in the feature extraction stage, the time sequence amplitude feature and the phase change quantity are extracted from each channel signal respectively, and the basic feature vector containing instantaneous response and periodic disturbance information is constructed; further, the feature vector is grouped by time window, a time period response unit with stable structure is constructed, and a three-dimensional structure signal atlas data set is formed by channel dimension stacking, which is used to reflect the trajectory change structure and discrete sampling mode of the signal under traction state.

[0015] Preferably, the multi-channel signals collected during the traction training process exhibit nonlinear coupling and dynamic mutation characteristics at the structural level, and there are significant phase coordination and amplitude linkage phenomena between different channel signals; based on this feature, the present application proposes a modeling strategy that fuses phase components and amplitude features, and constructs multi-source signals into a analyzable multi-dimensional structure through unified sequence alignment and cooperative enhancement mechanism; this method enhances the structural expression ability while maintaining the physical characteristics of the signal, and provides a stable and reliable input basis for subsequent self-evolution feature extraction, multi-channel modeling and direction enhancement processing.

[0016] Further, in the S2 step, the specific process of constructing the self-evolution feature map includes:

[0017] S21, according to the phase component information of each channel in the signal atlas data set, extracting the channel-level phase component sequence under the time sequence index to form a phase component matrix ;

[0018] S22, based on the phase component matrix , establishing a measure relationship of structural response change between adjacent time points, constructing a structural response mapping vector , the mathematical model is:

[0019] ;

[0020] Wherein, represents the structural response intensity at time index t;

[0021] S23, combining the structural response mapping vectors under all time indexes in time sequence to form a self-evolution feature map , the mathematical model is:

[0022] ;

[0023] Wherein, is the self-evolution feature map finally formed.

[0024] Further, in the S2 step, for the change rule of the phase information in the multi-channel signal in the traction stretching state, the application is based on the evolution characteristics of the phase components of each channel in the signal atlas data set in the time dimension, constructs a self-evolution feature map for describing the change process of the signal structure response, and specifically, first extracts the phase distribution characteristics of each channel in the continuous time period, and constructs a time sequence index structure for recording the phase evolution track of different time points; then, by comparing the change trend of the phase of each channel between adjacent time points, the local response mutation and cross-channel disturbance characteristics appearing in the traction stretching process between channels are captured, so as to form a response mapping reflecting the dynamic evolution relationship of the structure; on this basis, considering the synchronization, phase drift amplitude and interference correlation strength between channels, a structure response expression framework that can change with time is established, and a self-evolution feature map for describing the change process of the traction stretching state is generated; the feature map can be indexed by time, dynamically reflecting the state transition, disturbance spread and structural cooperative change behavior of the signal structure appearing in the traction process, providing an expression basis with high dynamic sensitivity for subsequent signal modeling and state recognition.

[0025] Preferably, during the traction stretching training process, the multi-channel physiological and structural response signals present the characteristics of frequent short-time disturbance in the time sequence dimension, asymmetric cross-channel response structure and nonlinear state transition process, the coupling relationship between the signals is jointly driven by traction strength, motion amplitude, tissue stress feedback factors, and shows significant dynamic evolution characteristics; based on such unstable and highly coordinated signal behaviors, the present application proposes a disturbance-aware structural response mapping mechanism, which combines phase difference trends and inter-channel synchronization relationships to model the fine-grained structural changes in the traction stretching motion state change process, and constructs a self-evolution feature expression graph reflecting the signal collaborative evolution law, providing structural consistency support and response expression basis for subsequent feature coding, structure enhancement and state classification processing.

[0026] Further, in the S3 step, the specific method of constructing the multi-channel feature representation graph aligned with the channels includes:

[0027] S31, according to the self-evolution feature graph and the amplitude component extracted from the signal atlas data set , a time index mapping function is established , the amplitude component is aligned with the structural response time index in the time dimension, and an aligned amplitude matrix is constructed The mathematical model of

[0028] , ;

[0029] S32, according to the amplitude matrix , the self-evolution feature value and the aligned amplitude on each channel are combined in order in the time dimension to construct a channel-level feature matrix group The joint feature column vector of the feature matrix of each channel is represented as:

[0030] ;

[0031] S33, stack the channel-level feature matrix corresponding to all channels along the dimension of the channel set C to obtain a multi-channel feature representation graph The mathematical model is:

[0032] .

[0033] Further, in the S3 step, in order to construct the multi-channel feature representation with structural consistency and time sequence correlation in the process of traction stretching movement, first, the self-evolution feature map obtained in the previous stage is accurately aligned with the amplitude component of the original signal in the time dimension, to ensure the response synchronization of each channel at consecutive time points; second, based on the aligned data, the amplitude response trend and disturbance change pattern of each channel at different time points are extracted, and a channel-level feature response unit is constructed; then, taking the time index as the main axis, combined with the structural dependency and disturbance synchronization behavior among channels, a multi-channel feature matrix group is assembled and generated, to capture the collaborative evolution law among different channels during training; then, the feature matrix group is structurally rearranged and the response relationship is enhanced through matrix reconstruction operation, to improve the expression ability of the channel-level feature in the time sequence and structural level; finally, the processed feature matrix group is mapped to a multi-channel feature representation graph in a unified format, to provide basic data support for subsequent structural modeling and state recognition.

[0034] Preferably, in the process of traction training, the multi-channel signal has time response rate imbalance, disturbance position inconsistency and channel excitation intensity difference, resulting in non-uniform collaborative characteristics of different channels in the response structure, and traditional methods based on simple superposition or local convolution cannot effectively express the structural dependency relationship among channels. The present application can enhance the modeling ability of the system to the dynamic response collaboration among channels while maintaining the time sequence integrity of the original signal, by constructing a channel feature matrix group with time sequence alignment and response relationship as the core, to provide stable and reliable structural input for subsequent traction state perception.

[0035] Further, in the S4 step, the specific method of constructing the convolution mapping strategy with structural path perception and generating the multi-channel deep feature map includes:

[0036] S41, according to the multi-channel feature representation graph , a convolution mapping strategy with structural path perception is constructed, and a path mapping kernel tensor is defined, and the mathematical model is:

[0037] :

[0038] Wherein, represents the structural path perception convolution weight from the channel in the rth feature dimension and time position s to the channel;

[0039] S42, based on the exponential decay relationship of path information density and channel propagation intensity, a path dissipation factor is constructed, and the mathematical model is:

[0040] ;

[0041] Wherein, represents the path reachability strength between channels c1 and c2, and a>0 is a regulation factor;

[0042] S43, performing an inter-channel difference propagation operation to define a channel response difference amount at each time t, and constructing an inter-channel difference propagation regulation factor matrix , and the mathematical model is as follows:

[0043] ;

[0044] wherein, represents a response difference regulation factor of channels c1 and c2 at time t;

[0045] S44, based on the path mapping weight tensor , the path dissipation factor and the propagation regulation factor matrix , a multi-channel depth feature map is generated , and the mathematical model is as follows:

[0046] ;

[0047] wherein, represents a depth feature response value of the c2th channel at the time point t.

[0048] Further, in the S4 step, for the depth modeling method of the multi-channel feature, first, based on the multi-channel feature representation graph, the path connection relationship between channels is extracted, and a structure path-aware convolution mapping strategy is established to capture the channel interaction characteristics in the traction and stretching state; secondly, the path dissipation factor is introduced, the channel structure density and the propagation path strength are combined to construct a regulation function to weaken the dominant effect of redundant paths in feature propagation, and the information coupling degree between effective channels is improved; then, the difference propagation operation between channels is performed, the response deviation of each channel in the path mapping is calculated, and the local path features are integrated through a response combination function; then, the convolution mapping strategy is embedded into the feature propagation structure, and channel-level depth convolution processing is performed to extract response features with path-aware characteristics; finally, a multi-channel depth feature map is output, which provides a structure-aware enhanced high-order feature basis for subsequent energy regulation and direction modeling.

[0049] Preferably, in the process of traction stretching training, each sensing channel presents uneven path accessibility, dynamic fluctuation of response distribution and significant difference in coupling characteristics in terms of force path, structural impedance and feedback response; based on such structural unevenness and response heterogeneity, the application adopts a path-aware convolution mapping mechanism and introduces a path dissipation factor to weaken the dominant role of the area with excessively high path response density in the propagation process, and to improve the structural awareness proportion of the low response area, so as to realize the balanced expression of the differences between channels, and to provide structural enhancement support for stable modeling and downstream state discrimination of multi-channel deep feature maps.

[0050] Further, in the S5 step, the specific method for constructing the direction-enhanced energy regulation feature map includes:

[0051] S51, calculating the energy distribution of each channel in the time dimension of the multi-channel deep feature map , and constructing an energy mapping matrix , the mathematical model is:

[0052] ;

[0053] wherein, represents the value of channel c in the deep feature map at time point t, represents the energy response value at the corresponding time point;

[0054] S52, based on the energy mapping matrix , the energy cooperative variation intensity between channels in the time dimension is calculated, and a scheduling correlation matrix is constructed, the mathematical model is:

[0055] , ;

[0056] wherein, represents the energy cooperative factor of channel i and channel j;

[0057] S53, according to the scheduling correlation matrix , a directional regulation node coefficient set is constructed, the original deep feature is directionally reconstructed to generate a direction-enhanced feature map , the mathematical model is:

[0058] ;

[0059] S54, based on the inverse decay relationship between the time point energy term and the regulation sensitivity, an energy binding factor is constructed, the mathematical model is:

[0060] ;

[0061] wherein β>0 represents an energy regulation sensitivity coefficient, ∈(0,1] represents a compression degree of the direction-enhanced feature update at time point t;

[0062] S55, the direction-enhanced feature map is fused with the energy mapping matrix to generate a direction-enhanced energy regulation feature map , and a mathematical model thereof is:

[0063] ;

[0064] wherein, represents a direction-enhanced feature value at time point t after fusion in channel c.

[0065] Further, in the S5 step, for the construction method of the direction-enhanced energy regulation feature map, first, based on the multi-channel depth feature map, the energy change trajectory of each channel in the continuous time sequence is extracted, and the local energy distribution in the time dimension is calculated to describe the dynamic evolution strength of the response of different channels in the traction stretching training process; second, an energy binding factor is introduced to embed an energy amplitude regulation structure in the energy time sequence of each channel, to constrain and adjust the upper and lower limits of the energy distribution interval, reduce the interference of the local abnormal energy peak value on the overall scheduling weight distribution, and improve the stability of the energy feature expression; then, a channel scheduling correlation matrix is constructed to establish a directionality linkage structure reflecting the guidance and response relationship between channels by fusing the channel energy synchronization relationship and the time scheduling cooperativity; next, the original depth feature map is directionally reconstructed based on the scheduling correlation matrix to adjust the distribution structure of the feature in the time domain and the channel domain, and to strengthen the significant response path; finally, the direction-enhanced energy regulation feature map is output to provide a feature basis with directionality and energy integrated expression capability for subsequent structure convolution mapping and dynamic trend modeling, effectively improving the cross-channel structure modeling accuracy and response focusing performance in the traction stretching state change process.

[0066] Preferably, during the traction stretching training process, the physiological response parts corresponding to different signal channels have an asymmetric coupling relationship in the structural path, resulting in a multi-stage activation, unstable peak disturbance and inter-channel coordination degree fluctuation characteristics in the energy distribution of the signal in the time dimension; based on this dynamic response characteristic, the present application introduces an energy binding factor to construct a time-energy constraint mechanism, and combines an inter-channel scheduling correlation matrix to model the inter-channel directional relationship, thereby realizing directional reconstruction of the original features on the basis of maintaining the stability of the signal energy structure, effectively improving the expression ability of the cross-channel signal coupling relationship under the traction stretching state and the discrimination accuracy under directional scheduling, and providing more directional focusing and energy constraint ability for subsequent state change detection and trend prediction.

[0067] Further, in the S6 step, the state sequence construction method based on the directional enhancement energy adjustment mechanism comprises:

[0068] S61, the direction-enhanced energy adjustment feature map is combined with the channel response value group at each time step to construct a time feature vector The mathematical model is:

[0069] ;

[0070] Wherein, represents the vector form response composed of each channel feature at the current time step;

[0071] S62, the time feature vectors corresponding to each time step are sequentially spliced according to the time sequence to construct a feature sequence matrix of the traction stretching state evolution process The mathematical model is:

[0072] ;

[0073] Wherein, represents the ordered expression structure of the multi-channel dynamic feature response in the traction stretching process.

[0074] ​​Further, in order to improve the time continuity and evolution rule expression ability of the state change modeling in the traction stretching motion process, the application proposes a time sequence feature sequence construction mechanism based on a direction enhanced feature map. First, according to the direction enhanced energy adjustment feature map, the feature vectors of each channel in the continuous time window are extracted, and are structured and organized in time index order to form a time sequence feature unit at the channel level. Second, the feature vectors in the time step are spliced to maintain the structural consistency between channels while highlighting the state evolution trend and direction response change characteristics, and a multi-channel fused time feature trajectory is constructed. Finally, the feature change context in the time direction is retained in the splicing result to form a feature sequence that can reflect the continuous evolution process of the traction stretching state, providing a high time sequence resolution structure expression basis for subsequent state discrimination, trend modeling and dynamic recognition modules.

[0075] Preferably, in the S7 step, for the signal optimization processing of the traction stretching motion state monitoring, a traction stretching motion state monitoring signal processing model is constructed, and the structure response modeling, difference convolution mapping, energy direction enhancement and time sequence feature reconstruction stages are executed in turn. First, in the structure response modeling stage, the preprocessed signal atlas data set is input, the phase component is extracted, and the structure response mapping relationship is established based on the time sequence index to construct a self-evolution feature map reflecting the dynamic evolution process of the traction stretching state and capture the continuous change characteristics of the structure response in the traction process. Then, in the difference convolution mapping stage, the self-evolution feature map is aligned with the amplitude component to generate a multi-channel feature representation map, a convolution mapping strategy based on structure path perception is constructed, and a path dissipation factor is introduced to perform difference propagation and response combination operations between channels to generate a multi-channel deep feature map with deep discriminability. Next, in the energy direction enhancement stage, the time energy distribution characteristics of each channel in the deep feature map are calculated, a channel scheduling correlation matrix is constructed by combining an energy binding factor to perform directional reconstruction processing on the original features to generate a direction enhanced energy adjustment feature map for improving the response sensitivity in the traction path change direction. Then, in the time sequence feature reconstruction stage, the feature vectors of each channel in the enhanced feature map are spliced according to the time step to obtain a feature sequence reflecting the evolution law of the traction stretching state. By integrating the above stages end-to-end, combining the iterative training mechanism and the structure regulation optimization strategy, the model can realize the joint optimization of structure response modeling, energy direction enhancement and time sequence feature construction according to the dynamic characteristics of the state evolution in the traction stretching process, and finally output a high-quality signal expression result reflecting the state characteristics of the traction stretching process, effectively improving the state recognition accuracy and response stability in complex traction environments.

[0076] Compared with the prior art, the application focuses on the space-time evolution characteristics of multi-channel signals in the process of traction stretching training, constructs a feature modeling mechanism for trajectory changes and state responses, proposes a signal atlas construction strategy, combines phase component and amplitude component extraction methods, realizes the structural expression of traction stretching signals in the state change process, and effectively improves the feature map's ability to maintain continuous dynamic information in the traction process; in the structural response modeling stage, the self-evolution feature map is constructed to reflect the response mode under the time sequence index, and the channel level alignment mechanism is introduced to generate the feature representation map, which enhances the model's ability to describe the change coordination and phase synchronization between channels; in the depth mapping modeling stage, a convolution mapping strategy based on structural path perception is designed, a path dissipation factor is introduced to adjust the propagation strength of high-density paths, and the difference propagation and response combination operations between channels are combined to obtain a structural perception multi-channel depth feature map, which significantly improves the direction consistency and structural discriminability of feature mapping; in the direction enhancement stage, an energy binding factor is introduced, a scheduling correlation matrix is established to model the energy response relationship between channels, and the direction enhancement energy regulation map construction is completed, solving the problems of weak directionality and energy leakage in the original features; finally, by constructing a traction stretching motion state monitoring signal processing model, integrating self-evolution modeling, convolution mapping, energy regulation and sequence splicing strategies, and combining an iterative training mechanism for optimization processing, stable modeling and fine prediction of multi-channel signals in the traction stretching state are realized, effectively improving the accuracy and robustness of traction state recognition. BRIEF DESCRIPTION OF DRAWINGS

[0077] Figure 1 is a flowchart of a traction stretching motion state monitoring method based on signal processing provided by the application.

[0078] Figure 2 is a structural diagram of a self-evolution feature map generation process based on signal atlas coupling phase component construction provided by the application.

[0079] Figure 3 is a processing flowchart for generating a multi-channel feature representation map by time sequence alignment based on amplitude component and self-evolution feature map to generate a channel-level feature matrix group.

[0080] Figure 4 is a structural diagram of a multi-channel evolution feature map generated by combining a path dissipation factor, a convolution mapping strategy, and a difference transmission response provided by the application.

[0081] Figure 5 is a structural diagram of an energy regulation feature map generated by combining an energy binding factor, an amplitude relationship matrix, and a directionality construction provided by the application.

[0082] Figure 6 is a generation flowchart of a feature sequence formed after the energy regulation feature map is spliced by time steps provided by the application.

[0083] Figure 7 is a channel feature fold line chart contrast chart provided by the present application, showing the timing feature changes of the original signal and the enhanced signal on multiple channels.

[0084] Figure 8 is a feature gray scale chart contrast chart provided by the present application, wherein Figure 8 (a) of is the original feature gray scale chart, Figure 8 (b) of is the enhanced feature gray scale chart. DETAILED DESCRIPTION

[0085] The present application provides a traction stretching motion state monitoring method based on signal processing, which aims to realize the structural evolution modeling and state recognition construction of multi-channel physiological signals; the method acquires multi-channel original signal data of the traction stretching device under different running states; performs normalization and filtering preprocessing operation on the original signal, extracts phase component and amplitude component reflecting the internal modulation law of the signal, and constructs signal atlas data set with trajectory change structure characteristics and timing sampling properties; based on the atlas data, performs structure response modeling, feature map construction, structure convolution mapping, direction enhancement adjustment and timing sequence splicing processing, and finally outputs the feature sequence reflecting the traction state change process, supporting the subsequent state recognition and dynamic analysis tasks.

[0086] Please refer to Figure 1 , a traction stretching motion state monitoring method based on signal processing in the embodiment of the present application, the specific steps are as follows.

[0087] S1, acquiring multi-channel original signal data collected by the traction stretching device under different running states, performing normalization and filtering processing on the signal data, extracting phase component and amplitude component, and constructing signal atlas data set containing trajectory change structure and timing discrete sampling features.

[0088] Further, in the S1 step, a multi-channel physiological signal acquisition device with a model of BIOPAC MP160 is used to collect the traction and stable traction training of the subjects on the traction stretching training platform; the number of signal acquisition channels is set to 8, and four types of physical quantity data, i.e., electrocardiogram signal, traction force sensing signal, displacement sensing signal and skin conductance signal, are collected by each channel respectively; the sampling rate of each channel is set to 512 Hz, the single signal sampling duration is 5 s, the acquisition window step is 1 s, and the total sampling point number is 3000; the original signals collected are saved in floating-point encoding format after unified processing; amplitude normalization processing is performed before sampling to compress the amplitude of each channel to the interval [-1, 1]; a third-order band-pass filter is used for filtering operation, and the cutoff frequency is set to 0.5 Hz to 45 Hz to eliminate baseline drift and high-frequency interference; then, the phase component is extracted by time domain uniform sampling while the amplitude component is retained, so as to form a signal expression object with physical meaning; finally, the phase and amplitude components are spliced according to the time index to construct a standardized signal atlas dataset of trajectory change structure and time sequence discrete sampling characteristics, thereby providing a basic input for subsequent traction stretching motion state recognition.

[0089] S2, according to the phase component in the signal atlas dataset, a structure response mapping relationship under time sequence index is established to construct a self-evolution feature map reflecting the traction stretching state change process.

[0090] Further, in the S2 step, in order to depict the structural disturbance change characteristics of the signal in the time evolution dimension in the traction stretching process, a structure response mapping mechanism is constructed based on the phase components of each channel in the signal atlas dataset, and the flowchart is as shown in Figure 2 The specific steps of constructing the self-evolution feature map include.

[0091] S21, according to the structure feature extraction results of each channel in the signal atlas dataset, the phase component values of the time index The mathematical model is:

[0092] , ;

[0093] Wherein, is the phase value of the cth channel at time point t, the unit is radian, and the value range is [-π, π], T=200 in this embodiment;

[0094] In this embodiment, and are the real part and imaginary part of the complex signal of channel c at time t, and the calculation formula is:

[0095] ​​ , ;

[0096] wherein, is the sampling frequency of channel c, taking the value , N is the signal frame length, taking the value N = 3000; t is the time index, obtained by converting the sampling point n; ;

[0097] In this embodiment, is the amplitude value of the original discrete signal of channel c at the n th sampling point, and the calculation formula is:

[0098] ;

[0099] wherein, is the initial amplitude modulation coefficient of channel c, in this embodiment, the value of is the time decay rate of channel c, in this embodiment = 0.05, is the modulation frequency parameter of channel c, is the phase offset of channel c, in this embodiment, the value range is [-π / 2, π / 2].

[0100] S22, according to the phase component matrix , define the square sum of the phase difference of each channel between adjacent time t and t+1 as the measure index of the structure disturbance amplitude, and construct the structure response time quantity , the mathematical model is:

[0101] ;

[0102] wherein, indicates the structure response intensity at time index t, the value range is mainly concentrated in [0.01, 0.25], when state mutation occurs, the local value exceeds 0.4, and the phase difference After standardization processing, the value range is [-1, 1] to keep the comparability of time neighborhood change.

[0103] S23, map the structure response vector under all time indexes according to time sequence, and combine to form the self-evolution feature map , the mathematical model is:

[0104] ;

[0105] wherein, is the self-evolution feature map finally formed, and the dimension is set to 1 × T, so The specific dimension of the self-evolution feature map is 1*200, indicating a structural response intensity sequence of 200 continuous time points in a time sequence.

[0106] S3, time sequence alignment processing of the self-evolution feature map and the amplitude component is performed to construct a channel-level feature matrix group and generate a multi-channel feature representation map.

[0107] Further, in the S3 step, to realize joint modeling of structural evolution characteristics and amplitude change trends between different signal channels in the traction stretching motion state, time sequence alignment processing of the self-evolution feature map and the amplitude component of each channel is required, and a multi-channel feature representation map is constructed, the process of which is shown in Figure 3 The construction process includes the following steps.

[0108] S31, according to the self-evolution feature map and the amplitude component extracted from the signal atlas data set , a time index mapping function is established. The mathematical model of the amplitude matrix after alignment is:

[0109] , ;

[0110] Among them, represents the amplitude value corresponding to the structural response index t of channel c and is normalized to the interval [0, 1];

[0111] In this embodiment, The calculation formula of is:

[0112] ;

[0113] Among them, is the original time domain signal of channel c at sampling point n, is the number of sampling points of the calculation window, and in this embodiment, The calculation result is normalized, and in this embodiment, ;

[0114] In this embodiment, the calculation formula of the time index mapping function is:

[0115] ;

[0116] Among them, t∈{1, 2, …, T} represents the time sequence position under the structural response index, and in this embodiment, L=3000.

[0117] S32, according to the amplitude matrix The self-evolving eigenvalues ​​of each channel in the matrix With the corresponding amplitude By combining them sequentially along the time dimension, a channel-level feature matrix group is constructed. The feature matrix of each channel The joint feature column vector is represented as:

[0118] ;

[0119] in, This represents the feature matrix of the c-th channel in the time series dimension, composed of structural response and amplitude components, preserving dual feature information of structural driving and signal energy; self-evolutionary feature map. The elements can be represented as Used to describe the structural response intensity of channel c at time point t, self-evolutionary eigenvalues Amplitude Obtained through amplitude envelope.

[0120] S33, Channel-level feature matrices corresponding to all channels Stacking along the dimension of the channel set C yields a multi-channel feature representation map. The mathematical model is:

[0121] ;

[0122] in, The three-dimensional multi-channel feature representation structure of the traction and stretching process is given by G, which has dimensions of C×2×T.

[0123] S4. Based on the multi-channel feature representation map, establish a structure path-aware convolutional mapping strategy, introduce a path dissipation factor, and perform inter-channel difference propagation and response combination operations to obtain a multi-channel depth feature map.

[0124] Furthermore, in step S4, to enhance the ability to model the structural coupling between multi-dimensional channels under traction and tension motion, a structural path mapping perception strategy needs to be constructed based on the multi-channel feature representation map to extract a depth map reflecting the propagation characteristics of response differences between channels. The process is as follows: Figure 4 As shown, the construction process includes the following steps.

[0125] S41. Based on the multi-channel feature representation diagram We construct a path-aware convolutional mapping strategy, define the path mapping kernel tensor, and the mathematical model is as follows:

[0126] ;

[0127] in, Indicates the first The channel is transmitted to the first Channel propagation structure perception convolution weight;

[0128] In the embodiment, The calculation formula is:

[0129] ;

[0130] Wherein, S is the size of the perception window, take The initialization of the convolution kernel tensor Uniform distribution ;

[0131] In the embodiment, g(·) is the bottom layer similarity function, and the calculation formula is:

[0132] ;

[0133] Wherein, Indicates the sampling value of channel c at time point s+n;

[0134] In the embodiment, Indicates the local weight kernel on the feature dimension r, and the calculation formula is:

[0135] ;

[0136] Wherein, n is the index within the local window, The Gaussian smoothing amplitude adjustment coefficient of feature dimension r is in the embodiment =1.0, The new Gaussian smoothing kernel attenuation factor is in the embodiment =0.1, The amplitude adjustment coefficient of the cosine oscillation term is in the embodiment =0.3, The frequency parameter of the cosine modulation term is in the embodiment , r is the feature dimension index.

[0137] S42, based on the exponential decay relationship of path information density and channel propagation intensity, the path dissipation factor is constructed The mathematical model is:

[0138] ;

[0139] Wherein, α>0 is the control factor, and in the embodiment α=2.0, Indicates the path reach strength between channel c1 and c2;

[0140] In the embodiment, The calculation formula is:

[0141] ;

[0142] wherein, is the Euclidean distance, is the distance attenuation factor, set =0.5, S is the total number of feature dimensions, S=32;

[0143] In the embodiment, the embedding vector and respectively represent the representation of the channels C1 and C2 under the feature dimension s, and the calculation formula is:

[0144] ;

[0145] wherein, is the dimension scaling coefficient, in the embodiment =0.8, K is the length of the embedding vector, in the embodiment K=5;

[0146] In the embodiment, represents the kth local feature component of the channel c under the feature dimension s, and the calculation formula is:

[0147] ;

[0148] wherein, K is the local neighborhood index, is the sampling value of the channel c at the time point s+k, is the smoothing adjustment coefficient of the feature dimension s, in the embodiment =1.0, is the exponential attenuation factor, which controls the attenuation speed of the neighborhood weight as k increases, in the embodiment =0.1, is the oscillation modulation amplitude, in the embodiment =0.2, is the oscillation frequency, in the embodiment .

[0149] S43, performing an inter-channel difference propagation operation, defining a channel response difference amount at each time t, and constructing an inter-channel difference propagation factor matrix , the mathematical model is:

[0150] ;

[0151] wherein, is the convolution combination influence adjustment factor, in the embodiment =1.0.

[0152] S44, based on path mapping kernel with the propagation adjustment factor matrix , combined to generate a multi-channel depth feature map , the mathematical model is:

[0153] , ;

[0154] wherein, represents the depth feature response value of the c2 channel at time point t;

[0155] In this embodiment, The calculation formula is:

[0156] ;

[0157] wherein, represents the input signal value of channel C1 at time t-n, is the convolution kernel weight under the corresponding feature dimension r, the initial value is subject to uniform distribution U(0.01, 0.01), is the length of the convolution perception window, in this embodiment, .

[0158] S5, calculate the time energy distribution of each channel in the depth feature map, establish the scheduling correlation matrix between channels, and perform directional reconstruction processing on the original feature to generate a direction-enhanced energy adjustment feature map.

[0159] Further, in step S5, in order to improve the directional modeling ability of the multi-channel signal response structure in the process of traction stretching motion, a direction-enhanced energy adjustment feature map needs to be constructed to enhance the channel expression discrimination of the signal in the complex motion state, and the construction process is as shown in Figure 5 , specifically including the following steps.

[0160] S51, calculate the energy distribution of each channel in the multi-channel depth feature map in the time dimension, construct an energy mapping matrix , the mathematical model is:

[0161] , ;

[0162] wherein, represents the energy response value of the corresponding time point.

[0163] S52, based on the energy mapping matrix , calculate the energy cooperative variation intensity between channels in the time dimension, construct a scheduling correlation matrix , the mathematical model is:

[0164] , , ;

[0165] wherein, represents the energy synergy factor of channel i and channel j, reflecting the consistency of their response patterns in the traction action; to enhance the robustness of the calculation, a sliding window mechanism is adopted in this embodiment, with a window length of 200 and a step length of 50, and the energy synergy factor is calculated for each local window and averaged to reduce the interference of local abnormalities on the overall structure judgment.

[0166] S53, constructing a directional adjustment node coefficient set according to the scheduling association matrix ,S54, reconstructing the original depth feature in a directional manner to generate a directional enhanced feature map , the mathematical model is:

[0167] , ;

[0168] wherein, represents the reconstructed feature value of channel j at time point t after directional enhancement, represents the feature value of the original depth feature map at channel time point.

[0169] S54, constructing an energy binding factor based on the inverse decay relationship between the time point energy term and the adjustment sensitivity , the mathematical model is:

[0170] ;

[0171] wherein, β>0 represents the energy adjustment sensitivity coefficient, and in this embodiment, β=0.6;

[0172] In this embodiment, the calculation formula of is:

[0173] ;

[0174] wherein, H is the historical difference cumulative window length, and in this embodiment, H=12, is the directional enhanced feature value of channel c at time point t, is the historical feature value at time t-h, is the difference intensity adjustment factor, and in this embodiment =1.8, is the difference energy index, and in this embodiment =1.5, is the time decay factor, and in this embodiment = 0.25.

[0175] S55, the direction enhanced feature map with the energy mapping matrix channel dimension weighted fusion, generating a direction enhanced energy adjusted feature map , the mathematical model is:

[0176] , ;

[0177] wherein, indicates the direction enhanced feature value of the channel c at time point t after fusion;

[0178] S6, the feature vectors of each channel in the direction enhanced feature map in the continuous time sequence are sequentially spliced according to the time step, and a feature sequence reflecting the change of the traction and stretching state is obtained.

[0179] Further, in step S6, in order to realize the sequence modeling of the traction and stretching motion state, the channel level features of each time step in the direction enhanced energy adjusted feature map need to be structurally combined to construct a dynamic feature expression under a unified time sequence structure, and the process includes the following two steps, as shown in Figure 6 .

[0180] S61, the channel response value combination in the direction enhanced energy adjusted feature map at each time step is constructed to build a time feature vector , and the mathematical model is:

[0181] ;

[0182] wherein, indicates the vector form response composed of each channel feature at the current time step, indicates the direction enhanced feature value of the channel c∈{1,…,C} at the time point t.

[0183] S62, the time feature vectors corresponding to each time step are sequentially spliced according to the time sequence to construct a feature sequence matrix of the traction and stretching state evolution process, and the mathematical model is:

[0184] ;

[0185] wherein, indicates the ordered expression structure of the multi-channel dynamic feature response in the traction and stretching process.

[0186] S7, a traction stretching motion state monitoring signal processing model is constructed, input signal atlas data set, sequentially through steps S2 to S6, and through iterative training until convergence, the optimization processing of the traction stretching motion state monitoring signal is completed.

[0187] In the S7 step, for the traction stretching motion state monitoring signal processing model, first input the signal atlas data set containing the trajectory change structure and the time sequence discrete sampling characteristics, and sequentially execute each processing stage defined in steps S2 to S6; Specifically, in the self-evolution feature modeling stage, the structural response mapping relationship under the time sequence index is constructed based on the signal phase component, and the self-evolution feature map reflecting the dynamic change of the traction stretching state is obtained, which provides the time sequence structure prior for the subsequent multi-channel feature representation; In the multi-channel deep feature extraction stage, through the introduction of the convolution mapping strategy and the path dissipation factor, the difference propagation and response combination operation between channels are completed, and the expression ability of the feature under the cross-channel coupling condition is enhanced; In the energy regulation modeling stage, the energy binding factor is introduced through the calculation of the time energy distribution, the inter-channel scheduling correlation matrix is established, the directional reconstruction of the original feature is performed, and the directional enhanced energy regulation feature map is generated; Then, in the feature sequence construction stage, the directional enhanced feature map is spliced in time steps, and the feature sequence which can completely reflect the evolution process of the traction stretching state is obtained; Through the end-to-end integration of the above stages, combined with the iterative training mechanism and the structured optimization strategy, the model can adaptively adjust the feature modeling and energy regulation strategy for different traction stretching states, and finally output the optimized traction stretching motion state monitoring signal, which significantly improves the recognition accuracy and stability under complex motion states.

[0188] Further, in the S7 step, the traction stretching motion state monitoring signal processing model constructed is written based on Python language, implemented using PyTorch deep learning framework, the training process adopts Adam optimizer, the initial learning rate is set to 0.0003, the batch size is 64, the total training round number is set to 200, and the cross-entropy loss function is used as the target optimization index; During the training process, the parameters of the structure convolution module and the direction modulation module are shared and updated, and the linear annealing learning rate strategy is adopted, the overall loss function of the model tends to be stable at about the 160th round, indicating that the feature mapping and sequence fusion process has reached the convergence state, and the signal corresponding to the state evolution characteristics in the traction process can be accurately restored.

[0189] Further, in the S7 step, the traction stretching state signal data is input into the traction stretching state monitoring model constructed for processing, and the experimental results are shown in Figure 7 and Figure 8 . Figure 7This is a line graph comparing the channel characteristics, showing the temporal characteristics of the original and enhanced signals across multiple channels; the horizontal axis represents the sampling point number (sampling interval = 20), and the vertical axis represents the characteristic value. The dashed line represents the trend of the original signal, and the solid line represents the trend of the enhanced signal. Figure 7 It can be seen that the original signal fluctuates greatly between sampling points and has obvious noise interference, while the overall trend of the enhanced signal is clearer and more continuous, and the cross-channel features show higher consistency, indicating that the model effectively suppresses noise and enhances the key state evolution features. Figure 8 It consists of two grayscale heatmaps: the original feature grayscale map (a) and the enhanced feature grayscale map (b); the horizontal axis represents the time step, the vertical axis represents the channel index, and the grayscale blocks represent the feature intensity distribution at different time points; from Figure 8 As can be seen in (a), the grayscale distribution in the original feature grayscale image is rather chaotic, the temporal evolution pattern is blurred, and the boundaries of state changes are not obvious; while Figure 8 In (b), the enhanced feature grayscale image exhibits clear grayscale transitions, regular feature distribution, and a prominent black-and-white boundary in the state change region, with high consistency in grayscale patterns between channels; in summary, from Figure 7 and Figure 8 It can be verified that the traction and tension state monitoring model proposed in this invention can effectively reduce noise interference while retaining and strengthening the core characteristics of the traction state, and realize a clear presentation of the state evolution process, thus proving the effectiveness and robustness of this invention in the task of traction and tension signal monitoring.

[0190] The above description is only a preferred embodiment of the present invention. Those skilled in the art can make several substitutions, adjustments and combinations without departing from the concept of the present invention. These modifications and improvements should all be considered within the scope of protection of the present invention.

Claims

1. A signal processing based monitoring method of the state of traction stretch exercise, characterized by, Comprise the following steps: S1, collect the multi-channel original signal of the traction stretching device under different operating conditions, perform normalization and filtering processing, extract the phase component and amplitude component, and construct a signal atlas data set containing trajectory change structure and time sequence discrete features; S2, according to the phase component in the signal atlas data set, establish the structure response mapping relationship under the time sequence index, and construct a self-evolution feature map reflecting the traction stretching state change process; S21, according to the phase component information of each channel in the signal atlas data set, extract the channel-level phase component sequence under the time sequence index to form a phase component matrix ; S22、based on the phase component matrix, establishing a measure relationship of structural response change between adjacent time points, constructing a structural response mapping vector The mathematical model is: ； wherein, denotes the structural response intensity at time index t; S23, mapping the structural response vector of all time indexes Combining in time sequence, constituting self-evolution feature map The mathematical model is: ； Wherein, is the finally formed self-evolution feature map; S3, time sequence alignment of the self-evolution feature map and the amplitude component, construction of channel level feature matrix group, generation of multi-channel feature representation map; S31、according to the self-evolution characteristic map with the amplitude component extracted from the signal atlas data set , a time index mapping function is established , the amplitude component is aligned with the structural response time index in the time dimension, and an aligned amplitude matrix is constructed The mathematical model of the above formula is: ， ； S32、according to the amplitude matrix evolving eigenvalues on each channel with the aligned amplitude combining in sequence in the time dimension, constructing a channel-level feature matrix group , the channel-level feature matrix of each channel The joint feature column vector is represented as: ； S33, Channel-level feature matrices corresponding to all channels Stacking along the dimension of the channel set C yields a multi-channel feature representation map. The mathematical model is: ； S4, according to the multi-channel feature representation map, using the convolution mapping strategy of structure path perception, introducing the path dissipation factor, performing channel difference propagation and response combination operation, obtaining the multi-channel deep feature map; S5, calculate the time energy distribution of each channel of the deep feature map, introduce the energy binding factor, establish the inter-channel scheduling correlation matrix, and directionally reconstruct the original feature to generate a direction-enhanced energy regulation feature map; S6, the feature vectors of each channel in the direction-enhanced feature map are spliced in time steps to obtain a feature sequence reflecting the traction stretching state change; S7, construct a traction stretching motion state monitoring signal processing model, input the signal atlas data set, and sequentially pass through steps S2 to S6, and perform iterative training until convergence to realize the optimization processing of the traction stretching motion state monitoring signal.

2. A signal processing based traction stretch movement state monitoring method according to claim 1, characterized in that, The signal atlas data set comprises multi-channel original signal data collected by the traction stretching device under different operating conditions, including traction force response signal, physiological monitoring signal, electromyographic signal, motion acceleration signal and attitude angle sensing signal, time sequence response sequence under continuous time window is established through time sampling mode, the signal data is processed by normalization and filtering, amplitude feature and phase change are extracted, channel feature atlas reflecting energy distribution structure and dynamic change trend is constructed, and signal atlas data set containing trajectory change structure and time sequence discrete sampling features is formed by stacking in channel dimension.

3. A signal processing based traction stretch movement state monitoring method according to claim 2, characterized in that, The process of constructing a multi-channel deep feature map includes: S41、According to the multi-channel feature representation graph , a structure path-aware convolution mapping strategy is constructed, a path mapping kernel tensor is defined, and the mathematical model is: ； wherein, represents the rth channel in the rth channel to the rth S42, based on the path information density and channel propagation intensity exponential decay relationship, the path dissipation factor is constructed The mathematical model is: ； wherein, represents the path accessibility between channels c1 and c2, and a > 0 is a regulating factor; S43, performing inter-channel difference propagation operation, defining the inter-channel difference propagation factor matrix The mathematical model is: ； wherein, represents the response difference adjustment factor of the channel c1 and c2 at time t; S44, based on the path mapping weight tensor , path dissipation factor and the propagation regulation factor matrix , combined to generate a multi-channel depth feature map , mathematical model: ； wherein, denotes the depth feature response value of the c2th channel at time point t.

4. The signal processing-based traction stretch movement state monitoring method according to claim 3, characterized in that, The process of constructing a direction-enhanced energy regulation feature map includes: S51、calculating the multi-channel depth feature map The energy distribution of each channel in the time dimension is constructed into an energy mapping matrix The mathematical model is: ； wherein, denotes the value of channel c in the depth feature map at time point t, denotes the energy response value for the corresponding time point; S52、based on the energy mapping matrix , calculate the energy collaborative variation intensity between channels in the time dimension, and construct a scheduling correlation matrix , the mathematical model is: , ; wherein, represents the energy synergy factor of channel i with channel j; S53、according to the scheduling association matrix , construct the directional adjustment node coefficient set , directionally reconstruct the original depth feature to generate a direction-enhanced feature map The mathematical model is: 。 5. A signal processing based traction stretch movement state monitoring method according to claim 4, characterized in that, The process of constructing a direction-enhanced energy regulation feature map includes: S501 constructs the energy binding factor based on the inverse attenuation relationship between the time point energy item and the adjustment sensitivity The mathematical model is: ； wherein β > 0 represents an energy regulation sensitivity coefficient, ∈ (0, 1] represents the compression degree of the direction enhanced feature update at the time point t; S502, the direction enhancement feature map with the energy mapping matrix channel dimension weighted fusion is performed to generate a direction-enhanced energy adjustment feature map The mathematical model is: ； wherein, denotes the directional enhancement eigenvalue at channel c time point t after fusion.

6. A signal processing based traction stretch movement state monitoring method according to claim 5, characterized in that, The process of constructing a traction stretching state evolution feature sequence includes: S61、the direction-enhanced energy regulation feature map combining the channel response values at each time step to construct a time feature vector The mathematical model is: ； wherein, represents a vector-form response composed of the channel features at the current time step; S62、concatenate the time feature vectors of each time step in sequence to construct a feature sequence matrix of the traction and stretching state evolution process concatenate, in sequence, to construct a feature sequence matrix of the traction and stretching state evolution process The mathematical model is: ， wherein, represents an ordered representation structure of the multi-channel dynamic characteristic response during the traction stretching process.

Images