A high-precision digital acceleration and angular velocity data joint processing system

Abstract

The application relates to the technical field of sensor data processing, in particular to a high-precision digital acceleration and angular velocity data joint processing system. The system comprises the following modules: a feature acquisition and joint extraction module, which is used for acquiring acceleration data and angular velocity data; joint feature extraction is carried out according to the acceleration data and the angular velocity data, and joint feature data is obtained; a stability evaluation module, which is used for evaluating the stability of the joint feature data and obtaining feature stability data; a joint state feature fusion module, which is used for carrying out joint state feature processing on the joint feature data and the stability data, and obtaining joint state feature data; and a state reasoning module, which is used for carrying out shallow neural network reasoning on the joint state feature data, and obtaining state reasoning data. Through the construction of the coupling feature extraction and stability-driven state reasoning of the acceleration and angular velocity data, the state recognition accuracy and the model generalization ability in a dynamic disturbance environment are improved.

Description

Technical Field

[0001] This invention relates to the field of sensor data processing technology, and in particular to a high-precision digital acceleration and angular velocity data joint processing system. Background Technology

[0002] With the widespread application of inertial measurement units (IMUs) in intelligent equipment, intelligent transportation, wearable devices, and other fields, how to effectively extract stable and accurate features from high-frequency acceleration and angular velocity data to achieve accurate state identification has become a core issue in inertial data processing. Current technologies often use single-source acceleration or angular velocity data, combined with simple statistical features (such as mean and standard deviation) or manual feature engineering methods to construct state classification models. However, when faced with dynamic disturbances, multimodal motion, or multidimensional operational states, it is often difficult to guarantee both identification accuracy and model stability. Summary of the Invention

[0003] To address the aforementioned technical problems, this invention proposes a high-precision digital acceleration and angular velocity data joint processing system, thereby resolving at least one of the aforementioned technical problems.

[0004] This application provides a high-precision digital acceleration and angular velocity data joint processing system, including:

[0005] The feature acquisition and joint extraction module is used to acquire acceleration data and angular velocity data; and to perform joint feature extraction based on the acceleration data and angular velocity data to obtain joint feature data.

[0006] The stability assessment module is used to assess the stability of the joint feature data and obtain feature stability data.

[0007] The joint state feature fusion module is used to process joint feature data and stability data into joint state feature data.

[0008] The state reasoning module is used to perform shallow neural network reasoning on the joint state feature data to obtain state reasoning data.

[0009] This invention simultaneously acquires acceleration and angular velocity data and performs joint feature extraction, which avoids the information deficiency problem caused by a single sensing parameter and enables complementary modeling of multi-source inertial information. The stability assessment module performs multi-dimensional analysis of the joint feature data, including distribution entropy and covariance, to promptly eliminate unstable or noisy features and improve the reliability of feature input. The joint state feature fusion module performs gated redistribution and cross-fusion of main features and stability information, so that the state representation combines action ontology information and stability constraints, enhancing the distinguishability of different behavior patterns. The state reasoning module uses a shallow neural network for reasoning, ensuring computational efficiency while accurately mapping the fused feature vector and outputting state results with confidence.

[0010] Preferably, the joint feature extraction specifically includes:

[0011] Multi-branch variation calculations are performed based on acceleration and angular velocity data to obtain multi-branch variation data;

[0012] Mutual disturbance residuals are calculated based on acceleration and angular velocity data to obtain mutual disturbance residual data;

[0013] By performing feature vectorization on the multi-branch variation data and the mutually disturbing residual data, joint feature data is obtained.

[0014] In this invention, by setting up multi-branch change calculation in the joint feature extraction stage, multi-dimensional features can be constructed from the perspectives of the trend slope of the dominant angular velocity direction, the coupling relationship of acceleration projection, and momentum-energy residuals, making the sensitivity to changes in the target motion state higher and avoiding the bias under single feature expression. At the same time, by calculating the mutual disturbance residuals, the difference between the rate of change of acceleration and the rate of change of angular velocity is modeled as a residual, and indicators such as average residuals and local peaks are extracted, which can effectively characterize the incoordination and sudden disturbance behavior between modes and improve the ability to identify abnormal states. By unifying the feature vectorization of multi-branch change data and mutual disturbance residual data, not only is the scale consistency and structural integrity of features from different sources guaranteed, but a joint feature representation that can be efficiently used by subsequent models is also formed.

[0015] Preferably, the multi-branch change calculation specifically involves:

[0016] The direction of maximum magnitude is extracted from the angular velocity data to obtain the direction data of maximum magnitude;

[0017] The acceleration data is projected onto the data in the direction of the maximum magnitude to obtain the acceleration projection data;

[0018] Linear trend slope fitting is performed on the acceleration projection data to obtain angular coupling data;

[0019] Momentum-energy data is obtained by calculating momentum-energy based on acceleration and angular velocity data.

[0020] The momentum-energy data is averaged to obtain the momentum mean data;

[0021] The momentum residual amplitude is calculated based on the momentum mean data to obtain the momentum residual amplitude data;

[0022] By integrating the angular coupling data and momentum residual amplitude data, multi-branch variation data is obtained.

[0023] This invention extracts the direction of maximum modulus from angular velocity data, adaptively capturing the dominant rotational axis of the target motion, thus avoiding feature instability caused by rapid attitude changes or complex attitudes. Projecting acceleration data onto this direction and fitting the trend slope quantifies the dynamic coupling relationship between acceleration and angular velocity, more accurately reflecting the trend changes of the target motion. Momentum-energy calculations are performed on acceleration and angular velocity, and the mean momentum is extracted. Then, the residual amplitude is calculated, effectively characterizing abnormal fluctuations and behavioral mutations in energy output within a short time window. By integrating angular coupling data with momentum residual amplitude data, the resulting multi-branch variation data not only possesses descriptive capabilities for both motion trends and energy stability but also enhances the system's sensitivity to abnormal disturbances, rapid switching, and complex states.

[0024] Preferably, the mutual interference residual calculation is specifically as follows:

[0025] The rate of change of acceleration and angular velocity data is calculated to obtain the rate of change of acceleration and the rate of change of angular velocity data.

[0026] The difference residuals are calculated based on the acceleration rate of change data and the angular velocity rate of change data to obtain the difference residual data;

[0027] The average residual and local maximum are calculated based on the difference residual data to obtain the average residual data and local maximum data respectively.

[0028] The mutual disturbance response index is calculated based on the average residual data and local maximum data to obtain the mutual disturbance residual data.

[0029] This invention calculates the rate of change for both acceleration and angular velocity data, enabling sensitive capture of minute fluctuations in the temporal dimension of these two types of inertial signals. Constructing a residual sequence based on their differences not only reveals the degree of coupling deviation between acceleration and angular velocity during dynamic changes but also provides a foundation for anomaly identification. Extracting the average residual and local maxima from the residual sequence allows for simultaneous measurement of both overall modal differences and instantaneous sudden deviations, thus achieving a dual characterization of both stationary deviations and severe disturbances. By calculating the mutual disturbance response index and normalizing and coupling the average level with peak information, the proportional relationship between intermodal coordination and the intensity of sudden disturbances can be effectively quantified.

[0030] Preferably, the stability evaluation module includes:

[0031] The stability distribution entropy is calculated on the joint feature data to obtain the stability distribution entropy data;

[0032] The joint feature data is subjected to multidimensional covariance tensor determination to obtain structural determination data.

[0033] The stability distribution entropy data and structure determination data are subjected to multi-head mask self-attention calculation to obtain feature stability data.

[0034] In this invention, by calculating the distribution entropy of joint feature data, the concentration and dispersion of features in discrete space can be characterized, thereby reflecting the stability or volatility of behavioral states. By using multidimensional covariance tensor determination, eigenvalue decomposition is performed on the linear correlation structure of joint features to extract principal component energy ratio indicators, which helps to discover the strength of coupling between features and whether the overall structure is stable. The distribution entropy and structure determination results are fused into a multi-head mask self-attention mechanism, which can not only dynamically allocate the weights of different stability indicators in the overall evaluation, but also suppress the interference of unreliable features on the results through masking.

[0035] Preferably, the stability distribution entropy is calculated as follows:

[0036] Discrete space processing is performed on the joint feature data to obtain discrete feature data;

[0037] Probability calculations are performed on the discrete feature data to obtain feature probability data;

[0038] Entropy is calculated based on the feature probability data to obtain feature entropy data;

[0039] The temporal entropy change rate is calculated from the feature entropy data to obtain the temporal entropy change rate data.

[0040] Feature map data is obtained by constructing feature maps from time-series entropy change rate data, feature entropy data, and feature discrete data.

[0041] The feature map data is convolved to obtain the stability distribution entropy data.

[0042] This invention performs discrete spatial processing on joint feature data and transforms it into a probability distribution, enabling feature states to be measured in a statistical distribution form and avoiding the uncertainty caused by single-point fluctuations. Calculating the entropy value of the probability distribution not only represents the current concentration of features but also better characterizes the feature distribution. By calculating the temporal entropy change rate of the entropy data, the evolution trend of feature stability over time can be reflected, thus enabling sensitive detection of state abrupt changes or slow drifts. Constructing a feature map from the entropy value, entropy change rate, and discrete distribution data achieves a structured expression of multi-source stability indicators, avoiding the limitations of single numerical features. Extracting local patterns from the feature map through convolution processing not only highlights the differences between stable and fluctuating intervals but also enhances the ability to capture hidden patterns in both spatial and temporal dimensions.

[0043] Preferably, the determination of the multidimensional covariance tensor specifically involves:

[0044] The joint feature data is subjected to joint feature centralization processing to obtain joint feature-centered data;

[0045] Multidimensional covariance data is obtained by calculating the multidimensional covariance based on the joint feature-centered data.

[0046] Eigenvalue decomposition is performed on the multidimensional covariance data to obtain the principal structure data;

[0047] The structural concentration index is calculated based on the main structural data to obtain structural judgment data.

[0048] This invention eliminates baseline bias between different feature dimensions by centralizing the joint feature data, ensuring the accuracy of subsequent correlation calculations. A multidimensional covariance tensor is constructed based on the centralized data, which not only captures the linear correlation between different feature components but also reflects their overall distribution structure. Eigenvalue decomposition of the covariance tensor extracts principal structure data characterizing the main directions of change and energy distribution, providing a solid basis for depicting the synergistic relationships between features. Calculating the structure concentration index based on the principal structure data quantitatively measures whether current features are concentrated in a few dominant directions, thereby identifying whether the feature distribution is in a stable and coordinated state.

[0049] Preferably, the joint state feature fusion module includes:

[0050] Stability gating weights are generated based on stability data to obtain stability gating weight data.

[0051] Based on the stability gating weights, the joint feature data and the stability data are redistributed to obtain the main feature data;

[0052] The stability data is fused by cross-attention based on the main feature data to obtain joint state feature data.

[0053] This invention generates gating weights based on stability data, which can quantify the reliability of different features at the stability level. By redistributing joint feature data and stability data through gating weights, unstable or noisy features are dynamically downweighted, while more stable features are highlighted, thus effectively improving the robustness of the input features. Cross-attention fusion of stability data with principal feature data not only uncovers the dependencies between the two types of features but also adaptively strengthens the contribution of key features to the representation through an attention mechanism. Through the above processing, the resulting joint state feature data possesses both global trend information and stability constraints, reducing the interference of environmental noise on recognition and enhancing the ability to distinguish complex motion patterns.

[0054] Preferably, the state reasoning module includes:

[0055] The joint state feature data is mapped in a low dimension to obtain low-dimensional feature data;

[0056] Feature compression is performed on low-dimensional feature data to obtain feature-compressed data;

[0057] State data is obtained by generating states based on feature-compressed data and a pre-set tag state library;

[0058] State confidence is generated based on the state data to obtain state inference data.

[0059] This invention transforms high-dimensional features into a more compact low-dimensional representation by performing low-dimensional mapping on joint state feature data. This reduces computational overhead and avoids interference from redundant information. Feature compression of the low-dimensional features allows for the further extraction of key representation factors, making the feature space more discriminative. Matching the compressed features with a pre-defined label state library enables accurate generation of different motion patterns or behavioral states, ensuring high consistency between the recognition results and the actual states. By generating confidence scores for the state data, not only can the state category be output, but a confidence assessment can also be provided simultaneously, facilitating result selection and correction in scenarios with high uncertainty or noise interference.

[0060] The beneficial effects of this invention are as follows: The system of this invention simultaneously collects acceleration and angular velocity data through a feature acquisition and joint extraction module, and performs joint feature extraction. This not only avoids the information loss problem caused by a single inertial parameter, but also constructs multi-dimensional features at the levels of maximum modulus direction, projection coupling, and momentum residuals, providing a rich representational basis for recognition. The stability assessment module performs distribution entropy and multi-dimensional covariance tensor analysis on the joint feature data, and generates stability data by combining a multi-head mask self-attention mechanism, thereby quantitatively assessing the concentration and structural consistency of features and effectively eliminating noise and unstable components. The joint state feature fusion module dynamically redistributes the main features based on stability gating weights, and fuses stability information and motion ontology information through cross-attention, realizing the modeling of dependencies between features and enhancing the robustness and discriminativeness of feature representation. The state inference module performs low-dimensional mapping and feature compression based on the fused features, and completes state generation and confidence output by combining a preset label state library, which not only ensures computational efficiency, but also provides reliable state discrimination results. Attached Figure Description

[0061] Other features, objects, and advantages of this application will become more apparent from the following detailed description of the non-limiting embodiments, taken with reference to the accompanying drawings:

[0062] Figure 1 A flowchart of a high-precision digital acceleration and angular velocity data joint processing system according to one embodiment is shown;

[0063] Figure 2 A flowchart of a joint feature extraction submodule according to one embodiment is shown;

[0064] Figure 3 A flowchart of a stability evaluation module according to one embodiment is shown;

[0065] Figure 4 A flowchart of a joint state feature fusion module according to one embodiment is shown;

[0066] Figure 5 A flowchart of a state reasoning module according to one embodiment is shown. Detailed Implementation

[0067] The technical method of this invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of this invention. Based on the embodiments of this invention, all other embodiments obtained by those skilled in the art without inventive effort are within the scope of protection of this invention.

[0068] Furthermore, the accompanying drawings are merely illustrative of the invention and are not necessarily drawn to scale. The same reference numerals in the drawings denote the same or similar parts, and therefore repeated descriptions of them will be omitted. Some block diagrams shown in the drawings are functional entities and do not necessarily correspond to physically or logically independent entities. These functional entities can be implemented in software, in one or more hardware modules or integrated circuits, or in different network and / or processor methods and / or microcontroller methods.

[0069] It should be understood that although the terms "first," "second," etc., may be used herein to describe various units, these units should not be limited by these terms. These terms are used merely to distinguish one unit from another. For example, without departing from the scope of the exemplary embodiments, a first unit may be referred to as a second unit, and similarly, a second unit may be referred to as a first unit. The term "and / or" as used herein includes any and all combinations of one or more of the associated listed items.

[0070] Please see Figures 1 to 5 This application provides a high-precision digital acceleration and angular velocity data joint processing system 10, comprising:

[0071] The feature acquisition and joint extraction module 01 is used to acquire acceleration data and angular velocity data; and to perform joint feature extraction based on the acceleration data and angular velocity data to obtain joint feature data.

[0072] Specifically, triaxial acceleration and triaxial angular velocity data are collected at a frequency of 100 Hz, and the data is divided into 2-second segments to obtain single-window data. The mean of the single-window data is calculated to estimate the zero bias. Based on the zero bias, debiasing is performed, and the debiased data is subjected to a fourth-order Butterworth bandpass filter in the 0.1 Hz to 15 Hz frequency band. The angular velocity covariance matrix is ​​calculated from the filtered acceleration data to obtain the angular velocity covariance matrix. The eigenvalue decomposition of the aforementioned angular velocity covariance matrix is ​​performed, and the eigenvector corresponding to the largest eigenvalue is taken as the principal rotation axis direction (unit vector).

[0073] The filtered acceleration data is projected along the principal rotation axis to obtain a scalar sequence; the scalar sequence is then subjected to least-squares linear fitting, and slope trend data is extracted; the slope trend data is divided by the interquartile range of the scalar sequence to obtain the angular coupling quantity, which represents the degree of driving force and directionality of the change in angular momentum on acceleration.

[0074] Calculate the acceleration and angular velocity energy indices within the calculation window, i.e. ,in This is the acceleration energy index within the current time window, representing the cumulative sum of squares of the acceleration magnitudes. This is the index of the sampling time within the current window. Let be the debiased and filtered acceleration vector at time t. The angular velocity energy index within the current time window represents the cumulative sum of squares of the angular velocity magnitudes; the mean acceleration energy and mean angular velocity energy of the previous window are calculated; the momentum residual magnitude between the current window and the previous window is calculated, which is the mean acceleration energy minus the absolute value of the mean plus the angular velocity energy minus the absolute value of the mean angular velocity energy.

[0075] Perform first-order differences on acceleration and angular velocity to obtain a series of rates of change; divide the median of the acceleration rate of change series by the median of the angular velocity rate of change series to obtain the weighting coefficients; construct the difference residual series based on this. ,in Let be the difference residual value at time t. Let be the vector of the first-order rate of change of acceleration at time t. This is the scale alignment factor. Let be the vector of the first-order angular velocity change rate at time t. Calculate the mean and maximum values ​​of the difference residual sequence, and construct the mutual disturbance response exponent. ,in The mutual interference response index, This represents the maximum value of the difference residual sequence within the current time window. The average value of the difference residual sequence. To prevent the use of tiny positive constants with a denominator of zero, the principal rotation axis direction, angular coupling quantity, momentum residual magnitude, and mutual disturbance response index are concatenated / assembled into a joint feature vector.

[0076] Stability assessment module 02 is used to assess the stability of joint feature data and obtain feature stability data.

[0077] Specifically, the joint feature sequence is represented as a set of feature vectors within a time window. For each component in the feature vector, a histogram bin number is set. This bin number is determined by taking the square root of the sample number, rounding down, and limiting it to the range [8, 32], thus adaptively adjusting the dimensional discretization precision. All joint feature sequences are mapped onto a multi-dimensional histogram grid to construct a joint probability distribution. The entropy of the joint distribution is calculated based on the joint probability distribution. The obtained entropy is normalized to obtain the normalized entropy. The current normalized entropy is subtracted from the previous normalized entropy to obtain the stability distribution entropy.

[0078] The system centers all eigenvector data by first calculating the average vector of all eigenvectors; then subtracting this average vector from each eigenvector yields the centered eigenma matrix. The system calculates the covariance matrix; after obtaining the covariance matrix, it performs eigenvalue decomposition, resulting in an eigenvector matrix and an eigenvalue diagonal matrix. Each diagonal element in the eigenvalue diagonal matrix represents the intensity of change in the corresponding direction. The system calculates the dominant ratio index, which is the ratio of the largest eigenvalue to the sum of all eigenvalues. The system calculates the eigenvalue entropy index by normalizing each eigenvalue and combining its logarithm to obtain the overall entropy value, i.e., the eigenvalue entropy. The system calculates the covariance structure drift by comparing the difference between the covariance matrix in the current time window and the covariance matrix in the previous time window, and dividing by the overall intensity of the covariance matrix at the previous moment (the square root of the sum of the squares of all elements in the covariance matrix), to obtain the relative magnitude of the structural change.

[0079] The system constructs two sets of input vectors: a statistical index vector and a structural index vector. The statistical index vector has two dimensions: the normalized mean of the feature entropy and the magnitude of its change within the current time window (i.e., normalized entropy and stability distribution entropy). The structural index vector has three dimensions: the dominance ratio, eigenvalue entropy, and covariance structure drift. If the normalized mean of the feature entropy is greater than 0.7, or the absolute value of its change exceeds 0.15, the current statistical stability is considered poor. In this case, the system applies a masking penalty to the statistical index vector, i.e., setting a minimum value at the corresponding position in the attention mechanism, thereby effectively reducing the vector's participation in stability assessment. If the dominance ratio is less than 0.45, or the eigenvalue entropy is greater than 0.8, it indicates that the principal components in the system structure are weakly dominant or the spectral distribution is unbalanced. In this case, the system applies weight attenuation processing to each dimension of the structural index vector, for example, reducing its value to 50% of its original value. Based on the above masking rules, the system concatenates the two input vectors and sends them to the multi-head attention mechanism module. During attention calculation, the system assigns weights based on the similarity between vectors and dynamically controls the process by overlaying a mask matrix, thereby obtaining a stable weighted representation. After obtaining the weighted representation, the system uses residual connections and layer normalization to fuse the output results. The attention output vector is added element-wise to the original input concatenated vector, and then layer normalization is used to eliminate numerical bias.

[0080] The joint state feature fusion module 03 is used to process the joint feature data and stability data into joint state feature data to obtain joint state feature data.

[0081] Specifically, the system performs a linear mapping on the stability feature data, and then inputs it into a nonlinear activation function after combining scaling and bias coefficients to obtain the weight value for each channel dimension. The weight values ​​are limited to the range of 0 to 1. ,in For stability modulation weight vector, The activation function is preferably the Sigmoid function. This is the scaling factor. Here is the stability mapping matrix. For stability feature vectors, The bias coefficient, The total dimension of the joint state feature channels.

[0082] Based on the aforementioned gating weights, the system performs weighted fusion of principal feature data and stability feature data to construct redistributed fused feature data. The principal feature data is directly multiplied by the weights to form the first part. The stability feature data undergoes a linear transformation to ensure its dimension matches that of the principal feature data, and is then multiplied by a factor of one minus the weight value to form the second part. The two parts are added together to obtain a new fused feature, representing the principal feature data after stability adjustment.

[0083] The system uses the aforementioned fused features as the query vector in the attention mechanism and, based on stability feature data, generates key and value vectors through linear mapping. The system calculates the matching degree between the query and the key based on a cross-attention mechanism, and then performs a weighted summation of the value vectors to obtain the fused context representation vector. This fused representation vector is input into a feedforward network for nonlinear transformation to obtain the fused output. The system employs a residual connection approach, adding the initial fused features to the feedforward network output, and then outputs joint state feature data through layer normalization.

[0084] State reasoning module 04 is used to perform shallow neural network reasoning on the joint state feature data to obtain state reasoning data.

[0085] Specifically, the system first inputs the joint state feature data into a shallow fully connected network, and embeds the features through linear transformation and nonlinear activation functions. This transformation compresses the original high-dimensional state representation into a lower-dimensional space, extracting key discriminative features. The embedding dimension is 32. The embedded intermediate features are further compressed through a set of bottleneck structures, that is, the feature dimension is compressed to half of the original. This is achieved through nonlinear activation functions and random deactivation (deactivation ratio of 20%), resulting in a 16-dimensional compression. The system performs a linear transformation on the compressed features to generate an output vector with the same dimension as the label state library. This label state library can contain multiple predefined behavioral categories, such as diving, foraging, and rolling over. If the system enables the state transition prior mechanism, it can perform prior smoothing adjustment on the current state score based on the predicted state of the previous time step. Specifically, a smoothing term consisting of the transition probabilities between the previous and current states is added to the current state score, and the strength of this smoothing term is controlled by a weight parameter (suggested value is 0.3). The system transforms the scores of each category into a probability distribution through normalization. The system performs calibration and normalization using a temperature coefficient of 1.5. The system selects the maximum value from the normalized probability distribution as the confidence score for the current prediction result and outputs a prediction state label based on the category corresponding to the maximum probability. The system performs uncertainty assessment: when the maximum confidence score is below 0.6, or the prediction entropy of the probability distribution is greater than 1.2, the system considers the current prediction result to have significant uncertainty, automatically marks it as "uncertain," and submits this state data to the upper-level strategy module for processing, such as delayed decision-making or multi-window majority voting fusion. The system outputs state inference data including the prediction state label, the corresponding confidence value, and (when necessary) the uncertainty label.

[0086] Preferably, the joint feature extraction specifically includes:

[0087] 11. Perform multi-branch change calculations based on acceleration and angular velocity data to obtain multi-branch change data;

[0088] Specifically, the system calculates the covariance matrix of angular velocities within a set time window. Eigenvalue decomposition is performed on the covariance matrix, and the eigenvector corresponding to the largest eigenvalue is extracted. This eigenvector is used as the principal rotation axis direction for that time window, and is in unit vector form. The system calculates the proportion of the largest eigenvalue to the sum of all eigenvalues ​​as the direction confidence score. If this confidence score is below 0.4, it indicates that the current rotational structure distribution is relatively discrete or that multi-axis interference exists. In this case, the system will abandon the use of the eigenvalue decomposition direction and instead select the direction of the angular velocity with the largest magnitude within the current window as the alternative direction, substituting its unit vector form into subsequent calculations.

[0089] The system projects the acceleration vector at each moment onto the principal rotation axis, obtaining a one-dimensional projection sequence. Linear fitting is performed on this sequence to obtain the slope of the time series, representing the trend of acceleration along the rotation axis. The system preferably uses the Huber regression algorithm, setting the robustness factor to 1.5 times the interquartile range of the projection sequence. The linear slope is divided by the interquartile range of the projection sequence and normalized by adding a small smoothing factor to obtain the normalized coupling slope index, representing the angular coupling slope. A positive or negative sign can be added to the slope to indicate "coupling" or "inverse coupling" status.

[0090] The system calculates the total energy of acceleration and angular velocity within the time window, i.e., the sum of the squares of the magnitudes at all moments, as the instantaneous momentum-energy index. The system updates the historical energy mean using an exponential moving average method, with a smoothing factor of 0.9. The system calculates the residual between the current instantaneous energy and the moving average energy, and takes the sum of the residuals of acceleration and angular velocity as the energy residual amplitude. The system concatenates the angular coupling slope, energy residual amplitude, and direction confidence into a multi-branch variation feature vector.

[0091] 12. Calculate the mutual disturbance residuals based on the acceleration and angular velocity data to obtain the mutual disturbance residual data;

[0092] Specifically, the system calculates the rate of change for both acceleration and angular velocity data, obtaining corresponding rate of change sequences. A Savitzky-Golay filter is preferably used for smoothing and differentiation, with parameters including a window length of nine and second-order polynomial fitting. The system calculates the median magnitude of the magnitudes of the acceleration and angular velocity rate of change sequences separately, and uses their ratio as a scaling factor. After scaling, the system calculates the residual sequence between the acceleration rate of change and the scaled angular velocity rate of change. From this residual sequence, the system extracts the following statistical features: mean residual value; local maxima, extracted using a five-point neighborhood peak detection method; cross-perturbation response index, i.e., the relative increase ratio between the local maxima and the mean; and optional percentile features, such as the 95th percentile of the residual. The system combines these statistical features to construct a cross-perturbation residual feature vector, including the mean residual, maximum residual, cross-perturbation response index, and percentile residual values.

[0093] 13. Perform feature vectorization on the multi-branch variation data and the mutually disturbed residual data to obtain joint feature data.

[0094] Specifically, the system concatenates the obtained multi-branch change data and interfering residual data to obtain joint feature data.

[0095] Preferably, the multi-branch change calculation specifically involves:

[0096] The direction of maximum magnitude is extracted from the angular velocity data to obtain the direction data of maximum magnitude;

[0097] Specifically, the system first calculates the magnitude of the angular velocity data for each frame within the time window to obtain the rotation intensity index at each moment. This magnitude represents the overall amplitude of the rotation velocity at that moment. Based on the distribution of all magnitudes, the system selects a subset of representative high-amplitude samples. For example, the top 10% of sample points by magnitude can be selected to form an index set, with a minimum sample size of ten points. For the selected high-amplitude samples, the system normalizes their original angular velocity vectors, converting them into direction vectors. The system averages all direction vectors to obtain an average value. This average vector is then normalized again to obtain the dominant direction within the window, i.e., the direction vector with the maximum magnitude. The system calculates the magnitude of the average direction vector as a direction reliability index, i.e., direction confidence. If the direction reliability is lower than a preset threshold (e.g., 0.4), it indicates that the rotation directions within the current time window are relatively dispersed. In this case, the system does not use the direction averaging result but instead downgrades to the instantaneous maximum magnitude direction, i.e., selecting the moment with the largest magnitude among all samples and directly using the angular velocity direction vector corresponding to that moment as the dominant direction. The system outputs the direction data of the maximum modulus within the current time window, as well as the corresponding direction reliability index.

[0098] The acceleration data is projected onto the data in the direction of the maximum magnitude to obtain the acceleration projection data;

[0099] Specifically, the system extracts the direction vector of the maximum modulus from the angular velocity data, using it as the dominant rotation direction within the current time window. This direction vector has been normalized to represent the direction with the strongest rotational tendency. For each moment within the time window, the system calculates the projection component of the acceleration vector onto the dominant rotation direction. The system then performs a dot product operation between the acceleration vector of each frame and the dominant rotation direction vector to obtain a scalar value. This projection process yields a scalar sequence that evolves over time, called the acceleration projection sequence. The system outputs the acceleration projection sequence, which consists of the projection scalars from all moments within the current time window.

[0100] Linear trend slope fitting is performed on the acceleration projection data to obtain angular coupling data;

[0101] Specifically, the system treats the acceleration projection sequence as time-series data, with the time index ranging from the first frame to the Nth frame. Linear regression modeling is performed with time as the independent variable and the projection value as the dependent variable. The system preferably uses the Huber regression method for trend fitting. In the regression results, a positive slope indicates co-coupling, meaning acceleration increases along the rotation axis; a negative slope indicates inverse coupling, meaning acceleration decreases along the axis. The system normalizes the linear slope. The system calculates the interquartile range of the acceleration projection sequence as a scale reference value, divides the fitted slope by this scale value, and adds a small smoothing term to obtain the normalized slope index, i.e., the angular coupling characteristic value, which is dimensionless and retains the sign information of the slope. If the confidence of the principal rotation direction extracted in the previous stage is lower than 0.2, or the interquartile range of the current projection sequence is extremely small (e.g., less than 10 to the power of -4), it is considered that the projection change within the current window is insufficient or the direction identification is unreliable. The system marks this angular coupling slope as unusable and sets the mask flag to zero. The system uses the effective angular coupling eigenvalues ​​from the previous time window for gentle extrapolation filling. The system outputs normalized angular coupling eigenvalues, along with an optional quality mask flag.

[0102] Momentum-energy data is obtained by calculating momentum-energy based on acceleration and angular velocity data.

[0103] Specifically, within a set time window, the system performs cumulative calculations of the squared magnitudes of the acceleration and angular velocity signals to obtain the total momentum energy for that time period. The total energy of acceleration and angular velocity is then divided by the number of frames within the time window to obtain the average energy density per unit time. The system outputs momentum energy data including linear and angular energy, which can be presented as the original total energy or as energy density.

[0104] The momentum-energy data is averaged to obtain the momentum mean data;

[0105] Specifically, the system uses an exponential moving average method to dynamically update the mean values ​​of acceleration kinetic energy and angular velocity kinetic energy. The momentum energy value within the current time window is weighted and merged with the historical moving average value from the previous window to obtain the updated mean momentum energy value. During the moving average update process, the system sets a weighting coefficient to control the relative influence of the current value and historical values. The weighting coefficient ranges from 0.9 to 0.95. The system outputs the updated mean values ​​of acceleration momentum and angular velocity momentum for the current time window.

[0106] The momentum residual amplitude is calculated based on the momentum mean data to obtain the momentum residual amplitude data;

[0107] Specifically, the system calculates the absolute differences between the acceleration kinetic energy and the angular velocity kinetic energy relative to their corresponding sliding mean values, and sums them to form the momentum residual amplitude for the current window. The system then performs logarithmic compression on the residual amplitude index. The system outputs the momentum residual amplitude data for the current window, specifically including the logarithmically compressed residual amplitude index.

[0108] By integrating the angular coupling data and momentum residual amplitude data, multi-branch variation data is obtained.

[0109] Specifically, the system standardizes the angular coupling eigenvalues ​​and momentum residual amplitude eigenvalues ​​separately. Using the rolling Z-score method, the mean of each eigenvalue within its historical window is subtracted, divided by the historical standard deviation, and a smoothing factor is added to obtain the normalized result. The rolling statistical window size is no less than thirty time segments. The standardized angular coupling data and momentum residual amplitude data are then concatenated to obtain multi-branch variation data.

[0110] Preferably, the mutual interference residual calculation is specifically as follows:

[0111] The rate of change of acceleration and angular velocity data is calculated to obtain the rate of change of acceleration and the rate of change of angular velocity data.

[0112] Specifically, the system employs a smoothed derivative method based on the Savitzky-Golay filter for rate of change calculation. Within each time window, the three component axes (X, Y, Z) of acceleration and angular velocity are processed separately. The preferred sliding window length is 9 or 11 (an odd number); the polynomial order is second-order; and a symmetric structure is used for local fitting. After obtaining the rate of change of the three directional components, the system performs vector-level aggregation processing, i.e., calculates the magnitude of the acceleration rate of change vector and the angular velocity rate of change vector at each time step. If more than 10% of the data is missing within the current window, or if any component axis is saturated for an extended period (e.g., continuously at its maximum or minimum value), the derivative calculation for that window is considered unreliable. The system sets an unreliable derivative flag for this window and performs weight reduction or skipping of the features of that time slice in subsequent modules. The system outputs the acceleration rate of change sequence and the angular velocity rate of change sequence, along with an optional derivative reliability flag.

[0113] The difference residuals are calculated based on the acceleration rate of change data and the angular velocity rate of change data to obtain the difference residual data;

[0114] Specifically, the system calculates the median of the magnitude of the rate of change of acceleration and the median of the magnitude of the rate of change of angular velocity at all times within the current time window, and constructs a scale alignment coefficient by the ratio of the two. ,in This is the scale alignment factor. This is the median operator in the time dimension, which takes the median value of all samples within the time window corresponding to time t. Let be the vector of the rate of change of acceleration at time t. Let be the vector of the rate of change of angular velocity at time t. To prevent the use of tiny positive constants with a denominator of zero, this coefficient serves as a scaling factor to numerically scale the rate of change of angular velocity, ensuring it is on the same order of magnitude as the rate of change of acceleration. After scaling, the system calculates the difference between the acceleration rate of change vector and the scaled angular velocity rate of change vector at each time step, obtaining the difference residual vector for that time step. The system calculates the magnitude of the residual vector at each time step, using it as the scalar value of the difference residual, forming a complete difference residual sequence. The system calculates the median of the residual sequence, i.e., the value at the middle of all residual values; it then calculates the absolute value of the difference between each residual value and this median, taking the median of these absolute values ​​as the absolute median difference. An empirical coefficient is set, with a value between 1 and 3. This empirical coefficient is multiplied by the absolute median difference to obtain the soft threshold parameter. If the residual value at a certain time step is lower than this threshold, it is considered to belong to the normal fluctuation background, and the system suppresses the value to zero. The system outputs the processed difference residual sequence.

[0115] The average residual and local maximum are calculated based on the difference residual data to obtain the average residual data and local maximum data respectively.

[0116] Specifically, the system averages the difference residual sequence over the entire time window to obtain the average residual value. The system then performs local peak detection within the residual sequence. It filters out sample points that satisfy the local maximum condition, i.e., the residual value at that moment is greater than the residual values ​​of the two adjacent time points. The system calculates the peak height index of each candidate peak relative to the maximum value within its neighborhood (e.g., the two adjacent time points) and compares it with a threshold formed by adding twice the absolute median difference to the median of the residual sequence. Only significant peaks exceeding this threshold are retained. Among all significant peaks, the maximum value is extracted as the local maximum residual value for the current window. If no significant peak is detected in the current window, the system sets this local maximum value as the average residual value and sets the peak flag to zero.

[0117] The mutual disturbance response index is calculated based on the average residual data and local maximum data to obtain the mutual disturbance residual data.

[0118] Specifically, the system constructs a standard cross-interference response index based on the difference between the local maximum residual value and the average residual value, representing the increase in local peak amplitude relative to the global noise background. The system uses the 95th percentile of the residual sequence to replace the local maximum value, obtaining a robust version of the cross-interference response index. If the previous derivative calculation step was marked as unreliable (e.g., missing derivative or signal saturation), or if no valid peak is detected in the current residual sequence, the system marks the current index as low confidence and sets the reliability flag to zero. The system outputs the cross-interference residual data for the current time window, including the average residual value; the local maximum residual value; the standard or robust version of the cross-interference response index; and the index quality flag.

[0119] Specifically, mutual interference response index calculation ,in The mutual interference response index, The maximum value of the residual sequence. The average value of the residual sequence. To prevent the use of tiny positive constants with a denominator of zero, values ​​such as 10 to the power of -6 are used.

[0120] Preferably, the stability evaluation module includes:

[0121] 21. Calculate the stability distribution entropy of the joint feature data to obtain the stability distribution entropy data;

[0122] Specifically, for each feature dimension, its mean and standard deviation are calculated over the last thirty historical windows. The mean is subtracted from each sample in the current window, and the result is divided by the standard deviation plus a smoothing factor to construct a standardized feature matrix. The system adaptively discretizes each standardized feature dimension, estimating the optimal bin width based on the interquartile range and sample size. The system calculates the range between the maximum and minimum values ​​of the feature dimension, derives a reasonable number of bins based on the bin width, and limits it to between 8 and 32. The system maps the multidimensional discrete features of each sample to a multidimensional grid space, generating a unique cell number (i.e., grid ID) through mixed carry encoding. The frequency of each grid point is counted across all samples, and the probability distribution of each grid point is calculated using a Laplace smoothing strategy. After obtaining the probability distribution of all grid points, the system calculates the entropy within the current window. The system divides the entropy value by the logarithm of the corresponding number of grid points to obtain the normalized information entropy. The system calculates the distribution sparsity index, which is the ratio between the number of all grid points with probability values ​​greater than the average level and the total number of grid points. The system outputs stability distribution entropy data, including normalized information entropy values ​​and distribution sparsity.

[0123] 22. Perform multidimensional covariance tensor determination on the joint feature data to obtain structure determination data;

[0124] Specifically, the standardized joint feature matrix is ​​centered by calculating the sample mean vector, which involves averaging each feature dimension within the current window. For each sample, this mean vector is subtracted to obtain the centered data. The standard sample covariance matrix is ​​then calculated based on the centered data. Finally, an identity matrix and its corresponding scale factor are constructed based on the standard sample covariance matrix. ,in As the scale factor, Covariance matrix The trace is the sum of the variances of all eigenvalues ​​along the main diagonal of the matrix. The standard sample covariance matrix, The feature dimension is the total dimension of the joint feature data. A shrinkage covariance matrix is ​​constructed using a shrinkage coefficient (ranging from [0.05, 0.3], or a preset value). After obtaining the shrinkage covariance matrix, the system performs eigenvalue decomposition, yielding a set of orthogonal eigenvectors and corresponding eigenvalue sequences, arranged in descending order. The system calculates the ratio of the largest eigenvalue to the sum of all eigenvalues ​​as the structure dominance ratio. The system calculates the spectral entropy based on the eigenvalue proportions. The system calculates the ratio of the largest to the smallest eigenvalue as the condition number. The system compares the first-order principal direction in the current time window with the principal direction in the previous time window, calculating the principal direction drift. The system outputs structure determination indicators including the structure dominance ratio, spectral entropy, condition number, and principal direction drift, forming a structure determination data vector.

[0125] 23. Perform multi-head mask self-attention calculation on the stability distribution entropy data and structure determination data to obtain feature stability data.

[0126] Specifically, the system constructs two decision vectors: an entropy index vector, containing normalized distribution entropy and sparsity components; and a structure index vector, containing structure dominance ratio, eigenvalue entropy, condition number, and principal direction drift. When the normalized entropy value is too high (greater than 0.75) or the entropy quality indicator is invalid, attention flow towards the entropy index direction is blocked. When the structure dominance ratio is too low (less than 0.45), the structure spectrum entropy is too high (greater than 0.85), or the structure quality indicator is invalid, attention flow towards the structure index direction is blocked. In cases where soft gating is required, the system does not completely block attention but instead numerically suppresses the attention weights. The system calculates attention weights through the mapping relationship between queries, keys, and values, and incorporates masking adjustments during weight calculation. The weighted results are used to aggregate the sequence vectors to obtain a fused feature representation. This representation is then nonlinearly mapped via a feedforward network. The network structure consists of two fully connected layers with a hidden dimension of 64, and the activation function is GELU. The system obtains a normalized fused output sequence through residual connections and layer normalization. The fused sequence is then subjected to mean pooling to obtain a stability vector representation, which serves as the feature stability data for the current window. The system generates a scalar confidence value, ranging from 0 to 1, through an activation function. If both the entropy vector and structure vector of the decision input are masked, the system degenerates into directly concatenating the original decision vector as the stability representation and setting the confidence value to zero. The system output includes the stability vector and the stability confidence.

[0127] Preferably, the stability distribution entropy is calculated as follows:

[0128] Discrete space processing is performed on the joint feature data to obtain discrete feature data;

[0129] Specifically, after obtaining the joint feature data, the system performs discrete space processing on it to obtain discrete feature data. When the dimension of the joint feature is greater than eight, the system maps the original features to a lower-dimensional space using a fixed random orthogonal transformation matrix, for example, compressing the dimension to between four and six. If the feature dimension is no more than eight, the original features are used directly. The system performs binning processing independently for each feature dimension. Specifically, the bin width is determined based on the interquartile range of the sample distribution and the number of samples; that is, the bin width is equal to twice the interquartile range multiplied by the reciprocal of the cube root of the number of samples. The number of bins required for each dimension is calculated based on the bin width and limited to between 8 and 32. After this processing, each feature dimension is divided into several intervals, and each sample can be mapped to a multidimensional grid index. This index is obtained through multidimensional mixed-base encoding, and the product of the number of bins for all dimensions is the total number of grid points. When the total number of grid points exceeds 10 to the power of 5, the system makes adjustments, including increasing the dimension compression ratio or limiting the maximum number of bins for each dimension to within 16. The system output includes a multidimensional grid index sequence corresponding to the sample, which serves as discrete feature data; and the binning configuration used for each dimension.

[0130] Probability calculations are performed on the discrete feature data to obtain feature probability data;

[0131] Specifically, the system statistically analyzes the discretized grid indexes, calculating the number of times each grid point is included in the sample count. This number is recorded as the frequency of that grid point, denoted as the count value at that location. Simultaneously, the total number of samples is counted. To avoid zero probability cases, the system sets a smoothing factor based on the frequency statistics. This is achieved by adding a constant to the frequency of each grid point, dividing by the total number of samples, and then multiplying the constant by the total number of grid points, thus obtaining the smoothed probability value for each grid point. The system calculates a sparsity index, such as the number of grid points with a statistical probability value greater than the average probability level, and uses this proportion of the total number of grid points as the sparsity index. The system output includes feature probability data, i.e., the probability distribution results of all grid points, satisfying a probability sum of one; and the sparsity index.

[0132] Entropy is calculated based on the feature probability data to obtain feature entropy data;

[0133] Specifically, the system calculates the probability distribution of all grid points, multiplies the probability value of each grid point by its logarithm, takes the negative, and sums the results to obtain the entropy value of the overall distribution. The system then normalizes the entropy value by dividing it by the logarithm of the total number of grid points to obtain the feature entropy data.

[0134] The temporal entropy change rate is calculated from the feature entropy data to obtain the temporal entropy change rate data.

[0135] Specifically, the system calculates the difference between the normalized entropy value of the current window and the normalized entropy value of the previous window to obtain the first-order rate of change of entropy over time. The system uses an exponential moving average method to smooth and update the first-order rate of change. The smoothing coefficient is set to 0.7, meaning that the current result is obtained by weighting the smoothed value of the previous time step with the current rate of change. The system further performs a threshold judgment on the entropy rate of change. When the change in entropy exceeds 0.15, it is considered that a significant abrupt change in distribution has occurred, and a change flag of 1 is output; otherwise, a change flag of 0 is output. The system outputs time-series entropy rate of change data, including the first-order rate of change of entropy; the smoothed rate of change; and the change flag.

[0136] Feature map data is obtained by constructing feature maps from time-series entropy change rate data, feature entropy data, and feature discrete data.

[0137] Specifically, based on the binning results for each dimension, the system calculates the sample proportion within each bin interval and obtains a normalized histogram. This histogram represents the distribution of each feature dimension in different intervals. The histograms of all dimensions are stacked vertically to form a two-dimensional matrix; if there are insufficient parts, zeros are padded on the right. Based on a two-dimensional matrix, the system constructs three channels: the first channel is the occupied channel, which is directly composed of histogram matrices of each dimension; the second channel is the global entropy channel, which broadcasts the normalized entropy value within the matrix space (i.e., this entropy value is a scalar. This scalar is expanded according to the number of rows and columns of the histogram matrix, and element-wise copied and filled to the same two-dimensional matrix size as the occupied channel, thus obtaining an entropy feature map with the same shape as the histogram matrix), forming an entropy feature map with the same shape as the occupied channel; the third channel is the entropy change rate channel, which broadcasts the temporally smoothed entropy change rate within the matrix space (the system obtains the temporally smoothed entropy change rate value of the current window, which is a scalar. This scalar is expanded according to the number of rows and columns of the histogram matrix, i.e., the same entropy change rate value is filled at each position of the two-dimensional matrix, thus forming a change rate feature map with the same shape as the occupied channel), obtaining the corresponding change rate feature map. The system performs a square root transformation on the occupied channel; and performs linear scaling on the global entropy channel and the entropy change rate channel, limiting their values ​​to the range of 0 to 1. The system stacks the three channels in a fixed order to form a feature map tensor containing three layers of channels.

[0138] The feature map data is convolved to obtain the stability distribution entropy data.

[0139] Specifically, the system employs a two-layer lightweight convolutional network to process the feature map: the first convolutional layer uses a 3x3 kernel, with 3 input channels, 16 output channels, a stride of 1, and padding the boundaries with one pixel. After convolution, an activation function is applied, followed by downsampling through a 2x2 max pooling operation. The second convolutional layer also uses a 3x3 kernel, with 16 input channels, 32 output channels, a stride of 1, and padding the boundaries with one pixel. After convolution, an activation function is applied, followed by downsampling through a 2x2 average pooling operation. The convolutional result is flattened and input to a linear mapping layer, outputting a fixed-length feature vector with dimensions set to 8 or 16. A soft gating mechanism is introduced after the first convolutional layer output. When the entropy quality indicator is invalid, or a sudden change in the entropy rate is detected, the system applies a reduction coefficient to the output for weight reduction, thereby weakening the impact of unstable samples on the results. The system obtains a stability distribution entropy data vector; by normalizing and mapping this vector, the stability confidence is obtained. When the spatial resolution of the feature map is too low (e.g., the dimension product is less than 32), the system no longer uses a convolutional network, but instead uses a two-layer fully connected network.

[0140] Preferably, the determination of the multidimensional covariance tensor specifically involves:

[0141] The joint feature data is subjected to joint feature centralization processing to obtain joint feature-centered data;

[0142] Specifically, the system calculates the central value for each dimension of the joint feature data. The Huber mean estimation method is used, combined with a threshold set based on the interquartile range of that dimension, which is set to 1.5 times the interquartile range of that dimension. The mean estimation result is obtained through 2 to 3 iterations. The corresponding mean estimate is subtracted from the value of each sample in each dimension to obtain the centered joint feature data matrix, making the numerical distribution of each dimension relatively concentrated around 0. If the dimensional differences between dimensions are large, the system performs rolling standardization on the centered data. Specifically, it dynamically adjusts the standard deviation of the most recent thirty windows to normalize the values ​​of each dimension. During the centering process, if missing values ​​are detected in the input data, the system sets the centering quality flag to zero, indicating that the result is unreliable; otherwise, the flag is set to one, indicating that the result is valid. The system output includes the centered joint feature data matrix and the centering quality flag.

[0143] Multidimensional covariance data is obtained by calculating the multidimensional covariance based on the joint feature-centered data.

[0144] Specifically, based on the centralized joint feature data, the system obtains a multidimensional covariance matrix according to the sample covariance calculation method. The system then weights the original sample covariance matrix with the identity matrix to obtain a contracted covariance matrix. The weighting coefficients are set between 0.05 and 0.3. When the sample size is insufficient to support the calculation of the current dimension (sample size less than dimension plus 5), or the condition number of the sample covariance matrix is ​​greater than 10,000, or the smallest eigenvalue is too small (below 1 x 10^-8), the system increases the contraction coefficient according to preset parameters. If the covariance matrix is ​​still unstable after contraction, the system adds a small correction term on the diagonal, with a size of one ten-thousandth of the total variance. The system outputs a stable multidimensional covariance matrix and the corresponding contraction coefficient.

[0145] Eigenvalue decomposition is performed on the multidimensional covariance data to obtain the principal structure data;

[0146] Specifically, the system performs symmetric eigenvalue decomposition on the covariance matrix, obtaining a set of eigenvalues ​​and corresponding eigenvectors arranged in ascending order. All eigenvalues ​​less than zero are truncated to zero. The system calculates the proportion of each eigenvalue relative to the sum of all eigenvalues, as the explained variance ratio for that direction. The system selects the principal subspace based on the cumulative explained variance ratio, specifically selecting the set of eigenvectors with the smallest cumulative explained variance ratio reaching or exceeding 90%, thus obtaining the principal subspace representation. The system compares the first principal eigenvector of the current window with the corresponding vector of the previous window, calculating the degree of deviation between them. The closer the result is to zero, the more stable the principal direction; the closer the result is to 1, the more significant the rotation or drift of the principal direction. When the explained variance ratio of the first eigenvalue is less than the average distribution level plus 0.05, and the remaining eigenvalues ​​are relatively evenly distributed, the system determines that there is no obvious dominant structure. The system outputs principal structure data, including the eigenvalue set; the explained variance ratio corresponding to each eigenvalue; the first principal direction vector; the set of basis vectors of the principal subspace; the dimension of the principal subspace; and the principal direction drift.

[0147] The structural concentration index is calculated based on the main structural data to obtain structural judgment data.

[0148] Specifically, the structural concentration index is calculated as follows: ,in As the dominant ratio, The largest eigenvalue, For the eigenvalue order term, For feature dimension, For the first There are several eigenvalues. The closer the dominance ratio is to 1, the more significant the concentration direction of the current structural distribution; when the value is close to 1 / d, it indicates that the distribution in each direction is relatively balanced, and there is no obvious dominant structure. The system will... As an output of structural concentration index.

[0149] Preferably, the joint state feature fusion module includes:

[0150] 31. Generate stability gating weights based on stability data to obtain stability gating weight data;

[0151] Specifically, the system performs a linear transformation and normalization on the stability vector to obtain an intermediate embedding representation, with the embedding dimension set to 32. The embedding result is then mapped to the same dimensional space as the main features. The system calculates a Sigmoid activation function on the mapped result and adds temperature parameter adjustments to obtain continuous gating weights ranging from 0 to 1. The parameters are set to an amplification factor of 1.5 and a bias value of -0.5. The system employs an exponential smoothing strategy, proportionally weighting the current weights with the weights from the previous time step, with a recommended smoothing coefficient of 0.7. A lower limit of 0.1 is set on the generated weights. The system constructs two types of masking coefficients: an entropy index mask (suppressing attention in the entropy index direction when the normalized entropy value exceeds 0.75, the entropy change rate exceeds 0.2, or the entropy quality indicator is invalid); and a structure index mask (suppressing attention in the structure index direction when the structure dominance ratio is below 0.45, the normalized eigenvalue entropy exceeds 0.85, the covariance drift exceeds 0.3, or the structure quality indicator is invalid). The mask strength is adjusted using a soft coefficient of 0.8. The attention weights of relevant channels are softly reduced to approximately 0.2; for complete masking, a hard mask is used, directly invalidating the attention values ​​at the corresponding positions. The system output includes a continuous gating weight vector and the masking coefficients for the entropy and structure indices.

[0152] 32. Based on the stability gating weights, perform principal feature redistribution on the joint feature data and stability data to obtain the principal feature data;

[0153] Specifically, the system maps the stability vector to a dimensional space with the same dimensions as the main features through a linear transformation. Both the mapped stability features and the original main features are then subjected to layer normalization. The system performs weighted fusion of the main features and stability features in each channel based on gated weights. When the weight of a channel is close to 1, the main features are prioritized; when the weight is close to 0, the stability features are used as compensation. The system dynamically adjusts the gated weights based on the stability confidence score: when the stability confidence score is below 0.4, the overall weight of the main features is increased, such as by multiplying by 1.1 or 1.5; when the stability confidence score is above 0.8, the stability features are allowed to play a greater role, and the weight of the main features is correspondingly reduced, such as by multiplying by 0.5 or 0.8. If there are abnormal quality indicators (such as unreliable direction indicators or unreliable mutual interference indicators) in the previous steps, the weight of the main features in the corresponding channels is halved. The system adds a portion of the residual terms of the original main features to the weighted result and performs layer normalization again. The output is the main feature data, in the form of a redistribution vector with the same dimensions as the original features.

[0154] 33. Perform cross-attention fusion on the stability data based on the main feature data to obtain joint state feature data.

[0155] Specifically, the system first divides the main feature data into several equal-length segments (4 segments in total) and maps each segment to a unified modeling dimension space. Each segment serves as a query vector after mapping. The system constructs multiple index tokens from the stability data, including entropy index tokens composed of normalized entropy value and entropy change rate, and structure index tokens composed of structure dominance ratio, eigenvalue entropy, condition number, and principal direction drift. Simultaneously, the overall stability vector is mapped to the same dimension and combined with the aforementioned tokens to form key and value vectors. The number of key and value vectors is 3 to 5. The system combines the previously generated entropy masking coefficients and structure masking coefficients to adjust the attention weights: if a token is determined to be untrustworthy, its attention flow is directly blocked in hard masking mode, and a minimum value penalty is added at the corresponding position; in soft masking mode, tokens determined to be weakly reliable are numerically suppressed, only their weights are reduced (e.g., multiplied by 0.8 or 0.5) rather than completely blocked. The system employs a multi-head cross-attention mechanism (4 attention heads are recommended) to calculate the correlation between the main feature query and the stability key value, obtaining the fused feature representation. The fusion result is processed through two layers of feedforward network (using the GELU activation function) and combined with residual connections and layer normalization operations to form the fused output. The system performs aggregation processing on the fused feature sequence, using average pooling or taking the first vector to obtain the state vector representation. This state vector is concatenated with the original main feature data, and then processed through linear projection and layer normalization to obtain joint state feature data. If the system detects that all key vectors are masked, a backoff mechanism is activated to directly generate the state representation based on the main feature data.

[0156] Preferably, the state reasoning module includes:

[0157] 41. Perform low-dimensional mapping on the joint state feature data to obtain low-dimensional feature data;

[0158] Specifically, the input joint state feature data includes a temporal dimension and a feature dimension, where the temporal dimension corresponds to the length of the observation sequence, and the feature dimension corresponds to the number of original features at each time step. The system transforms the original high-dimensional features into a low-dimensional spatial representation through a linear mapping. This process utilizes a pre-defined linear transformation matrix and bias vector to perform matrix operations, achieving dimensionality reduction of the original features. The system employs two methods for selecting the transformation matrix: one is to initialize the transformation matrix based on principal component analysis or linear discriminant analysis; the other is to obtain the transformation parameters through training the pre-layer of a neural network. Through the above processing, the system outputs low-dimensional feature data, forming a compact vector sequence that retains the main state information.

[0159] 42. Perform feature compression on the low-dimensional feature data to obtain feature-compressed data;

[0160] Specifically, low-dimensional feature data is taken as input and convolutional operations are performed on it along the time dimension. The convolution operation uses a one-dimensional convolution method with a kernel size of 3 and a stride of 1. After convolution processing, the system compresses the time dimension using global average pooling to obtain a fixed-dimensional feature vector. The system outputs a compressed feature vector with dimensions set to 32, 64, or automatically determined based on the specific model structure.

[0161] 43. Generate state data by compressing the feature data and using the preset tag state library;

[0162] Specifically, the compressed feature vector is used as input, and a pre-defined state label library containing several candidate state categories is invoked. The system maps the input features through a fully connected layer and uses a normalized probabilistic classifier to match each state category, thereby obtaining a classification score vector for each category. The system selects the category with the highest score as the currently identified state label, while retaining the complete score vector. The output state data includes two parts: the currently identified state label and the corresponding classification score vector.

[0163] 44. Generate state confidence based on state data to obtain state inference data.

[0164] Specifically, after completing state label recognition, the system evaluates the reliability of the recognition results. The system uses the state classification score vector as input and selects the highest probability value as the confidence index for the current recognized state. The system outputs the recognized state label and its corresponding confidence value together to form state inference data. The confidence value ranges from 0 to 1. The output state inference data consists of two parts: the inferred state label and the confidence value corresponding to that label.

[0165] Therefore, the embodiments should be regarded as exemplary and non-limiting in all respects, and the scope of the invention is defined by the appended application documents rather than the foregoing description. Thus, it is intended that all variations falling within the meaning and scope of the equivalents of the application documents be incorporated into the invention.

[0166] The above description is merely a specific embodiment of the present invention, enabling those skilled in the art to understand or implement the invention. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the invention. Therefore, the present invention is not to be limited to the embodiments shown herein, but is to be accorded the widest scope consistent with the principles and novel features of the invention herein.

Claims

1. A high-precision digital acceleration and angular velocity data joint processing system, characterized in that, The application relates to a feature acquisition and joint extraction module for acquiring acceleration data and angular velocity data. According to the acceleration data and the angular velocity data, multi-branch change calculation is performed to obtain multi-branch change data; mutual interference residual calculation is performed according to the acceleration data and the angular velocity data to obtain mutual interference residual data; and the multi-branch change data and the mutual interference residual data are subjected to feature vectorization to obtain joint feature data. A stability evaluation module is configured to evaluate the stability of the joint feature data to obtain feature stability data. A joint state feature fusion module is configured to perform joint state feature processing on the joint feature data and the stability data to obtain joint state feature data. A state reasoning module is configured to perform shallow neural network reasoning on the joint state feature data to obtain state reasoning data. The multi-branch change calculation specifically comprises the following steps: The maximum modulus direction data is obtained by performing maximum modulus direction extraction on the angular velocity data, wherein the maximum modulus direction data is obtained by performing modulus calculation on each frame of angular velocity data in a time window, screening out a part of representative high-amplitude samples, performing unitization processing on the original angular velocity vectors of the high-amplitude samples, uniformly converting the high-amplitude samples into direction vectors, performing average operation on all direction vectors to obtain an average value of the direction vectors, and performing unitization processing on the average vector again to obtain a dominant direction in the window, i.e. the maximum modulus direction vector. The acceleration data is projected onto the maximum modulus direction data to obtain acceleration projection data. Linear trend slope fitting is performed on the acceleration projection data to obtain angular kinetic coupling data. Momentum energy calculation is performed according to the acceleration data and the angular velocity data to obtain momentum energy data. Mean value calculation is performed on the momentum energy data to obtain momentum mean value data. Momentum residual amplitude calculation is performed according to the momentum mean value data to obtain momentum residual amplitude data. The angular kinetic coupling data and the momentum residual amplitude data are integrated to obtain the multi-branch change data. The mutual interference residual calculation specifically comprises the following steps: Change rate calculation is performed on the acceleration data and the angular velocity data to obtain acceleration change rate data and angular velocity change rate data. Difference residual calculation is performed according to the acceleration change rate data and the angular velocity change rate data to obtain difference residual data. Average residual calculation and local maximum value calculation are performed according to the difference residual data to obtain average residual data and local maximum value data, respectively. Mutual interference response index calculation is performed according to the average residual data and the local maximum value data to obtain mutual interference residual data, wherein the mutual interference response index calculation is performed by the system according to the difference between the local maximum residual value and the average residual value to construct the mutual interference response index. The stability evaluation module comprises the following steps:

2. The system of claim 1, wherein, Stability distribution entropy calculation is performed on the joint feature data to obtain stability distribution entropy data. Multi-dimensional covariance tensor determination is performed on the joint feature data to obtain structure determination data. Multi-head mask self-attention calculation is performed on the stability distribution entropy data and the structure determination data to obtain feature stability data. The stability distribution entropy calculation specifically comprises the following steps:

3. The system of claim 2, wherein, Discrete space processing is performed on the joint feature data to obtain feature discrete data. ​ The feature probability data is obtained by performing probability calculation on the feature discrete data; The feature entropy data is obtained by performing entropy calculation on the feature probability data; The time series entropy change rate data is obtained by performing time series entropy change rate calculation on the feature entropy data; The feature map data is obtained by constructing a feature map on the feature entropy data and the feature discrete data using the time series entropy change rate data; The stability distribution entropy data is obtained by performing convolution processing on the feature map data.

4. The system of claim 2, wherein, The multi-dimensional covariance tensor determination is specifically as follows: The joint feature centering data is obtained by performing joint feature centering processing on the joint feature data; The multi-dimensional covariance data is obtained by performing multi-dimensional covariance calculation on the joint feature centering data; The principal structure data is obtained by performing eigenvalue decomposition on the multi-dimensional covariance data; The structure determination data is obtained by performing structure concentration index calculation on the principal structure data.

5. The system of claim 1, wherein, The joint state feature fusion module includes: The stability gating weight data is obtained by generating a stability gating weight based on the stability data; The main feature data is obtained by performing main feature reassignment on the joint feature data and the stability data based on the stability gating weight; The joint state feature data is obtained by performing cross-attention fusion on the stability data based on the main feature data.

6. The system of claim 1, wherein, The state reasoning module includes: The low-dimensional feature data is obtained by performing low-dimensional mapping on the joint state feature data; The feature compression data is obtained by performing feature compression on the low-dimensional feature data; The state data is obtained by generating a state based on the feature compression data and a preset label state library; The state reasoning data is obtained by generating a state confidence based on the state data.

Images