Helmet collision detection accuracy improvement method and system based on multi-factor coordination

Abstract

The present application relates to the technical field of helmet safety detection, and discloses a helmet collision detection accuracy improvement method and system based on multi-factor coordination, which comprises the following steps: S1, obtaining multi-dimensional time sequence data generated in the collision detection process of the helmet and performing data preprocessing; S2, inputting the preprocessed multi-dimensional time sequence data into a pre-trained cross-scale guided recurrent network model to perform multi-scale feature extraction and fusion, and outputting a fusion feature vector comprehensively representing the helmet assembly, collision area, dynamic performance and shell state; S3, inputting the fusion feature vector into a multi-constraint optimization judgment model to perform multi-factor coordinated optimization judgment, balancing each constraint and maximizing the confidence of the judgment result by dynamically adjusting the weight factor associated with each constraint condition, and outputting a judgment result containing the detection result and core influencing factors. The present application can realize accurate multi-dimensional feature extraction, cross-scale coordinated fusion and efficient multi-constraint judgment.

Description

Technical Field

[0001] This invention relates to the field of helmet safety testing technology, specifically to a method and system for improving the accuracy of helmet collision detection based on multi-factor synergy. Background Technology

[0002] As a core protective equipment for motorcycle and electric bicycle riders, the helmet's collision safety performance directly affects the user's life safety. The GB 811-2022 standard clearly specifies the core indicators for helmet collision testing, including peak acceleration, the duration of acceleration exceeding a specific threshold, and shell integrity. Currently, helmet collision testing mainly relies on a combination of traditional physical testing and manual judgment, or simple data analysis models, which has the following prominent problems:

[0003] (1) Incomplete feature extraction: Existing detection methods often focus on single-dimensional data (such as only focusing on the peak acceleration), ignoring the synergistic effects of multiple factors such as assembly parameters, environmental conditions, and collision areas. Furthermore, they are difficult to capture the multi-scale features of collision data (such as high-frequency dynamic impact velocity and low-frequency stable assembly gap), resulting in one-sided feature representation and affecting the accuracy of judgment.

[0004] (2) Lack of cross-scale feature fusion: Collision detection data includes high-frequency instantaneous dynamic features (such as instantaneous acceleration and impact velocity fluctuations) and low-frequency stable features (such as the average assembly gap and constant ambient temperature). Existing models lack effective cross-scale guidance mechanisms and cannot utilize the global stability of coarse-scale features to optimize the local dynamic decoding of fine-scale features, which easily leads to feature redundancy or loss of key information.

[0005] (3) Insufficient accuracy of multi-constraint judgment: Helmet collision detection involves multiple constraints such as assembly compliance, area validity, performance compliance, and shell integrity. Traditional judgment methods often use threshold comparison or simple weighted fusion, which makes it difficult to balance the priority and correlation of each constraint. Moreover, the optimization process has problems such as slow convergence and low solution accuracy, which makes the detection results easily interfered with and the misjudgment and missed judgment rates are high.

[0006] (4) Poor interpretability of test results: Most existing intelligent detection models are “black box” structures, which cannot clearly mark the core factors that lead to unqualified or invalid tests, making it difficult to provide accurate guidance for subsequent test optimization and helmet design improvement, thus limiting the engineering application of the technology.

[0007] These problems result in insufficient accuracy and reliability of existing helmet collision detection methods, making it difficult to meet high standards of safety management requirements. There is an urgent need for a technical solution that can achieve accurate extraction of multi-dimensional features, cross-scale collaborative fusion, and efficient determination of multiple constraints. Summary of the Invention

[0008] To address the shortcomings of existing helmet collision detection technologies, the core objective of this invention is to provide a method and system for improving the accuracy of helmet collision detection based on multi-factor synergy. This method and system are particularly suitable for collision performance testing of motorcycle and electric bicycle helmets, and can accurately determine key indicators such as the helmet's energy absorption performance and shell integrity during a collision, providing technical support for helmet safety certification and quality control.

[0009] To achieve the above objectives, the following technical solution is adopted:

[0010] In a first aspect, embodiments of the present invention provide a method for improving the accuracy of helmet collision detection based on multi-factor synergy, comprising the following steps:

[0011] Step S1: Acquire multi-dimensional time-series data generated by the helmet during the collision detection process and perform data preprocessing. The multi-dimensional time-series data includes helmet assembly parameters, collision area information, environmental parameters, and collision dynamic parameters.

[0012] Step S2: Input the preprocessed multi-dimensional time-series data into a pre-trained cross-scale guided recurrent network model for multi-scale feature extraction and fusion, including: performing multi-scale decomposition on the input multi-dimensional time-series data to separate feature components representing characteristics of different time scales; encoding the feature components at each scale separately, extracting key features at each scale and generating corresponding context vectors; using a cross-scale guided mechanism, leveraging the stability of coarse-scale features to modulate the decoding process of fine-scale features, and finally outputting a fused feature vector that comprehensively represents helmet assembly, collision area, dynamic performance, and shell state.

[0013] Step S3: Input the fused feature vector into the multi-constraint optimization judgment model for multi-factor collaborative optimization judgment. By dynamically adjusting the weight factors associated with each constraint, the model balances the constraints and maximizes the confidence of the judgment result. Finally, the model outputs a judgment result that includes the detection result and the core influencing factors.

[0014] Secondly, embodiments of the present invention also provide a helmet collision detection accuracy improvement system based on multi-factor synergy, comprising:

[0015] Data acquisition and preprocessing module: used to acquire and preprocess multi-dimensional time-series data generated by the helmet during the collision detection process. The multi-dimensional time-series data includes helmet assembly parameters, collision area information, environmental parameters, and collision dynamic parameters.

[0016] Feature extraction and fusion module: This module is used to input the preprocessed multi-dimensional time-series data into a pre-trained cross-scale guided recurrent network model for multi-scale feature extraction and fusion. This includes: performing multi-scale decomposition on the input multi-dimensional time-series data to separate feature components representing characteristics at different time scales; encoding the feature components at each scale separately, extracting key features at each scale, and generating corresponding context vectors; and using a cross-scale guided mechanism to modulate the decoding process of fine-scale features by leveraging the stability of coarse-scale features, outputting a fused feature vector that comprehensively represents helmet assembly, collision area, dynamic performance, and shell state.

[0017] The detection result output module is used to input the fused feature vector into the multi-constraint optimization judgment model for multi-factor collaborative optimization judgment. By dynamically adjusting the weight factors associated with each constraint, it balances the various constraints and maximizes the confidence of the judgment result, and finally outputs the judgment result containing the detection result and the core influencing factors.

[0018] Compared with the prior art, the present invention achieves the following beneficial effects:

[0019] 1. More accurate and comprehensive feature extraction: The cross-scale guided recurrent network (MRN-CSG) achieves hierarchical extraction and cross-scale guided fusion of multi-scale features through a three-level structure of multi-scale decomposition, scale component encoding, and cross-scale guided temporal attention decoding. It can capture high-frequency dynamic information (such as instantaneous acceleration and impact velocity fluctuations) during the collision process, while retaining low-frequency stable information (such as assembly gaps and ambient temperature). This enables accurate extraction of multi-dimensional features such as the rationality of assembly parameters, the effectiveness of the collision area, acceleration change trends, and shell state. The representation capability of multi-dimensional features is significantly improved, solving the problem of one-sided feature representation in traditional models and providing more reliable basic data for subsequent judgment.

[0020] 2. Higher efficiency in cross-scale fusion: The cross-scale guided temporal attention decoder (GTA-Decoder) uses the global stability of coarse-scale features to guide the local dynamic decoding of fine-scale features, avoiding redundant computations in cross-scale feature fusion in traditional models. It fuses cross-scale information through element-wise multiplication, eliminating the need for explicit scale fusion steps. While reducing the model structure, it ensures the relevance and effectiveness of feature fusion, making the fused features more aligned with the core requirements of helmet collision detection, improving feature utilization efficiency, and solving the problem of insufficient cross-scale feature collaboration.

[0021] 3. Significantly Improved Accuracy and Stability of Multi-Constraint Judgment: The multi-constraint optimization fusion algorithm achieves efficient decomposition of multi-constraint problems through the Lagrange relaxation method. It integrates multiple constraints such as assembly compliance, regional validity, performance compliance, and shell integrity into the objective function. Combined with the Bat Algorithm to optimize multiplier updates, it dynamically optimizes the Lagrange multiplier update process, improving convergence speed and solution quality. It achieves accurate fusion and judgment of multi-dimensional features, solving the problems of slow convergence and easy getting trapped in local optima in traditional optimization methods. This allows the judgment results to accurately balance the requirements of various constraints. Experimental verification shows that the detection accuracy is improved by more than 15% compared with traditional methods, and the standard deviation of multiple detections is smaller, with better stability. It also solves the problems of weak constraint balancing ability and low accuracy of traditional judgment methods.

[0022] 4. High interpretability of test results: This invention can automatically associate and label the core factors of unqualified or invalid tests (such as excessive strap tension, non-weak areas in the collision zone, excessive peak acceleration, etc.), breaking through the limitations of the traditional "black box" model. It helps testers quickly locate the root cause of the problem and provides direct and accurate guidance for optimizing the test process (such as adjusting assembly parameters and optimizing the selection of the collision zone) and improving helmet products (such as strengthening the shell structure and optimizing the strap design), thereby enhancing the engineering application value of the technology.

[0023] 5. Wide adaptability and strong practicality: The method is compatible with various helmet types such as A1, A2, A3, B1, B2, and B3, and supports testing requirements under various pretreatment environments such as high temperature, low temperature, normal temperature, and water immersion. Moreover, the testing process is highly consistent with the GB 811-2022 standard and can be directly integrated into existing helmet testing equipment without large-scale modification, making it easy to promote and apply.

[0024] It should be understood that the description in the Summary of the Invention is not intended to limit the key or essential features of the embodiments of the present invention, nor is it intended to restrict the scope of the invention. Other features of the invention will become readily apparent from the following description. Attached Figure Description

[0025] The above and other features, advantages, and aspects of the various embodiments of the present invention will become more apparent from the accompanying drawings and the following detailed description. The drawings are provided for a better understanding of the invention and are not intended to limit the invention. In the drawings, the same or similar reference numerals denote the same or similar elements, wherein:

[0026] Figure 1 This is a flowchart illustrating the method for improving helmet collision detection accuracy based on multi-factor collaboration provided in an embodiment of the present invention.

[0027] Figure 2 This is a schematic diagram of the structure of the cross-scale guided recurrent network (MRN-CSG) model provided in an embodiment of the present invention;

[0028] Figure 3 This is an algorithm flowchart of the multi-constraint optimization decision model provided in the embodiments of the present invention;

[0029] Figure 4 This is a schematic diagram of a helmet collision detection accuracy improvement system based on multi-factor collaboration according to an embodiment of the present invention;

[0030] Figure 5 This is a curve comparing the accuracy performance of the algorithm of the present invention and the traditional model in helmet collision detection, as provided in the embodiments of the present invention. Detailed Implementation

[0031] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0032] Furthermore, the term "and / or" in this article is merely a description of the relationship between related objects, indicating that three relationships can exist. For example, A and / or B can represent: A existing alone, A and B existing simultaneously, or B existing alone. Additionally, the character " / " in this article generally indicates that the preceding and following related objects have an "or" relationship.

[0033] Figure 1 This is a flowchart illustrating a method for improving helmet collision detection accuracy based on multi-factor synergy, according to an embodiment of the present invention. Figure 1 As shown, the present invention provides a method for improving the accuracy of helmet collision detection based on multi-factor synergy, comprising the following steps:

[0034] Step S1, Data Acquisition and Preprocessing: Acquire multi-dimensional time-series data generated by the helmet during collision detection and perform data preprocessing. The multi-dimensional time-series data includes helmet assembly parameters, collision area information, environmental parameters, and collision dynamic parameters. Specifically, this includes:

[0035] Step S11: Basic Parameter Acquisition

[0036] (1) Head mold and helmet assembly parameter collection: Use a feeler gauge with an accuracy of 0.02mm to collect the gap data between the head mold and the inner wall of the helmet to ensure that the gap value is ≤1.5mm; collect the helmet strap tension data through a special push-pull force gauge to accurately control the tension within the standard range of 5±1N, and record the gap fluctuation value and tension stability value in real time during the assembly process.

[0037] (2) Collision area location data acquisition: Based on finite element analysis technology, the helmet is subjected to gridded stress scanning, and spatial coordinate data of multiple key weak areas (preferably 12, covering typical locations such as the edge of ventilation holes and the sudden change of shell curvature) are collected to establish a regional location database; the “independent annotation-cross-verification-trimming test” mechanism is adopted to record the annotation consistency data and verification deviation value of each weak area.

[0038] (3) Environmental and equipment parameter acquisition: Collect standardized pretreatment parameters to simulate the real use environment and calibration verification data to verify the accuracy of the testing equipment itself. Specifically: Collect helmet pretreatment environment data through high and low temperature test chamber, including temperature stability value and heat preservation time data under high temperature (50±2)℃, low temperature (-20±2)℃, and normal temperature (23±5)℃ conditions, as well as soaking temperature and draining time data during water immersion treatment; Collect pre-collision calibration data using helmet impact test machine, including head shape selection parameters and impact test platform drop height data, to ensure that the peak acceleration generated by the standard test block collision is in the range of 400g~440g, record the peak acceleration and deviation value of six consecutive collisions (the first three are preheated and the last three are effective), and at the same time collect the action time and deviation data when the acceleration is 200g (the deviation is required to be ≤0.1ms).

[0039] (4) Core data acquisition during the collision process: Using a triaxial accelerometer (arbitration method), impact velocity data is collected in the range of 10mm to 60mm before the collision (ensuring that it is not less than 95% of the theoretical velocity); real-time dynamic data such as the peak acceleration, the duration of acceleration exceeding 150g, and the duration of acceleration exceeding 200g (except when the peak acceleration does not exceed 300g) are collected; the shell state data after the helmet collision is collected through visual monitoring equipment, including whether there are obvious fragments falling off (whether the fragment length exceeds 10mm), the amount of shell deformation, etc.

[0040] Step S12: Data Preprocessing

[0041] (1) Data cleaning: Remove abnormal data caused by equipment abnormalities (such as acceleration peak deviation exceeding 20g) or operational errors, and retain valid data that meet the test conditions of GB811-2022 standard; perform outlier detection on the collected gap data, tension data and acceleration data, and use the mean replacement method to process isolated outliers.

[0042] (2) Data standardization: Standardize the detection data of different units, such as unifying the acceleration unit to g (1g=9.80665m / s²), the time unit to ms, and the length unit to mm, and establish a standardized dataset to ensure the comparability of the data.

[0043] (3) Data association and annotation: The assembly parameter data, regional positioning data, environmental equipment data and collision core data are associated and annotated according to sample number, detection time and collision point to form a "multi-dimensional-full process" data chain, providing complete data support for AI judgment.

[0044] Step S2, AI Model Construction and Training: The preprocessed multi-dimensional time-series data is input into a pre-trained cross-scale guided recurrent network model for multi-scale feature extraction and fusion, including: multi-scale decomposition of the input multi-dimensional time-series data to separate feature components representing characteristics of different time scales; encoding the feature components of each scale separately, extracting key features at each scale and generating corresponding context vectors; and using the stability of coarse-scale features through a cross-scale guided mechanism to modulate the decoding process of fine-scale features, ultimately outputting a fused feature vector that comprehensively represents helmet assembly, collision area, dynamic performance, and shell state.

[0045] (i) Construction of training dataset: Collect more than 27 sets of comparative test data, covering different helmet types (A1, A2, A3, B1, B2, B3), different preprocessing environments, and different collision points. 80% of the data is used as the training set and 20% of the data is used as the validation set. The dataset includes input features (assembly gap, strap tension, collision area coordinates, environmental parameters, impact velocity, acceleration data, etc.) and output labels (qualified / unqualified, detection data dispersion level).

[0046] (II) Feature Extraction Model Design:

[0047] This invention proposes a cross-scale guided recurrent network (MRN-CSG), which achieves accurate extraction of multi-dimensional collision detection features through a three-level structure of multi-scale decomposition, scale component encoding, and cross-scale guided temporal attention decoding. Figure 2 This is a schematic diagram of the structure of the cross-scale guided recurrent network (MRN-CSG) model provided in an embodiment of the present invention, as shown below. Figure 2 As shown, the details are as follows:

[0048] Step S21: Multiscale decomposition

[0049] In step S21, the signal decomposition method employs Stationary Wavelet Transform (SWT) to decompose the multi-dimensional time-series data into multiple scale components. Specifically, a preset high-pass filter is used to extract rapidly changing components from the time-series data, yielding fine-scale components containing high-frequency dynamic features such as impact velocity and instantaneous acceleration; simultaneously, a preset low-pass filter is used to extract slowly changing components from the time-series data, yielding coarse-scale components containing low-frequency stable features such as assembly gaps and ambient temperature. The parameters of the high-pass and low-pass filters can be optimized during model training. The specific execution process includes the following steps:

[0050] S211. Filter initialization and processing: Convolution operation is performed on the multi-dimensional time series data using a high-pass filter and a low-pass filter; wherein, both the high-pass filter and the low-pass filter are initialized and configured based on the db2 wavelet basis function;

[0051] S212, Scale decomposition: Through the convolution operation of the high-pass filter, fine-scale components that retain the high-frequency dynamic characteristics of impact velocity and instantaneous acceleration are extracted from the multi-dimensional time-series data; at the same time, through the convolution operation of the low-pass filter, coarse-scale components that retain the low-frequency stable characteristics of assembly gap and ambient temperature are extracted.

[0052] S213, Recursive Decomposition: The filtering and decomposition process of steps S211 and S212 is recursively repeated on the coarse-scale component to finally generate a set containing J subsequences of different scales, where J is an integer greater than 1.

[0053] S214. Parameter Adaptive Optimization: During the end-to-end training of the cross-scale guided recurrent network model, the filter coefficients of the high-pass filter and the low-pass filter are simultaneously fine-tuned to adapt the filter characteristics to the features of the helmet collision detection data.

[0054] The specific implementation process is as follows: The standardized multi-dimensional time-series data is processed hierarchically using Stationary Wavelet Transform (SWT), and the input data is filtered through a high-pass filter (HPF) and a low-pass filter (LPF). Decomposed into fine-scale components and coarse-scale components The decomposition formula is as follows:

[0055]

[0056] Initially, db2 wavelet filter parameters are used, and the filter coefficients are adaptively fine-tuned through end-to-end training to adapt to the collision detection data features; the decomposition process can be executed recursively, ultimately yielding J subsequences of different scales. The fine-scale components retain high-frequency dynamic characteristics such as impact velocity and instantaneous acceleration, while the coarse-scale components retain low-frequency stable characteristics such as assembly gap and ambient temperature.

[0057] : A standardized multi-dimensional time-series data matrix with dimensions of It includes time-series records covering features such as assembly gaps, belt tension, collision zone coordinates, environmental parameters, impact velocity, and acceleration. The number of feature dimensions for helmet collision detection, i.e., the total number of different types of features collected, such as basic parameters, environmental equipment parameters, and core collision data. The time step of time-series data refers to the total number of consecutive time nodes in the collision detection process, corresponding to the complete data acquisition cycle before, during, and after the collision. : The temporal sequence of the k-th collision detection feature, where k is the feature index ( This records the continuous values ​​of the feature throughout the entire detection period. g and l represent the filters used to extract high-frequency and low-frequency components, respectively. The conjugate filter of the high-pass filter (HPF) is used to extract fine-scale high-frequency features. It initially uses db2 wavelet filter parameters and adaptively fine-tunes the coefficients through end-to-end training to adapt to the extraction of dynamic features such as impact velocity and instantaneous acceleration. The conjugate filter of the low-pass filter (LPF) is used to extract coarse-scale low-frequency features. The initial parameters are db2 wavelet filter, and the coefficients are adaptively fine-tuned through end-to-end training to adapt to the extraction of stable features such as assembly gaps and ambient temperature. : Convolution operator, used for convolution operations between filters and time series data to achieve feature scaling. The fine-scale component dataset, obtained through HPF convolution, is designed to separate and preserve key dynamic features for evaluating the energy absorption performance of helmets during a collision, such as impact velocity fluctuations and instantaneous acceleration changes. : The fine-scale temporal sequence of the k-th feature, recording the dynamic changes of this feature in high-frequency dimensions. J: The number of scale layers in the multi-scale decomposition, i.e., the total number of different scale subsequences obtained through recursive decomposition, adaptively determined by the feature complexity of the collision detection data. The set of J subsequences at different scales obtained after recursive decomposition, where the preceding subsequences are fine-scale high-frequency features and the following subsequences are coarse-scale low-frequency features, comprehensively covering the multi-scale information of collision detection.

[0058] Step S22: Scale component encoding

[0059] In step S22, for each scale component sequence obtained from the decomposition, a recurrent neural network incorporating an attention mechanism is used for encoding to adaptively extract key temporal features at each scale and generate a context vector representing the global information at that scale. Specifically, scale component encoding is performed by inputting an attention encoder, which encodes the subsequence at each scale step-by-step. The specific process includes:

[0060] The specific implementation is as follows: for subsequences at each scale The input attention encoder (IA-Encoder) is used to adaptively select key features. The specific process is as follows:

[0061] : The set of helmet collision detection subsequences at scale j, containing the feature vectors of all time steps at that scale, i.e. This is used to focus on feature encoding at this scale. The collision detection feature vector at time step t and scale j, with dimension . It covers the characteristic values ​​of assembly parameters, environmental parameters, collision dynamic data, etc. at this time step and scale. Time step index of time series data ( ), which corresponds to the specific data acquisition time during the collision detection process. : Scale index for multiscale decomposition ( ), distinguishing feature subsequences of different frequency dimensions.

[0062] Step S221: Calculate the input attention weights

[0063] In step S221, for the current time step, based on the hidden state and memory unit state of the encoder in the previous time step, the input attention weights of each feature dimension in the current input feature vector are calculated, including: concatenating the hidden state and memory unit state of the encoder in the previous time step as the first intermediate representation. Use the values ​​of each feature dimension in the current input feature vector as a second intermediate representation. The first and second intermediate representations are linearly transformed and summed, then added to a learnable bias term. This summation is followed by processing with a non-linear activation function to obtain the attention scores for each feature dimension. A normalized exponential function is then used to process the attention scores to obtain the input attention weights for each feature dimension. The specific implementation is as follows:

[0064] Based on the encoder's hidden state at the previous moment and memory cell state Calculate the attention score for each feature dimension k. The input attention weights are obtained by normalization using the softmax function. :

[0065]

[0066] : Input attention encoder (IA-Encoder) hidden state at time step t-1 and scale j, with dimension . (m is the dimension of the hidden layer), storing the feature encoding information at this time and scale. The input attention encoder (IA-Encoder) stores the memory cell state at time step t-1 and scale j, which, together with the hidden state, stores historical feature information and supports long-term dependency modeling. : Attention score of the k-th feature at time step t and scale j, which quantifies the importance of the feature in the collision detection result. : The learnable weight vector for calculating attention score, whose dimensions match the dimensions of the hidden layer, used to adjust the sensitivity of feature attention calculation. : The learnable weight matrix for hidden states and memory unit states, used to... Perform a linear transformation. : The learnable weight matrix of the k-th feature time series at the j-th scale, used to perform linear transformation on the feature itself. : A learnable bias term used to calculate attention score, which is used to adjust the baseline value of the score. Vector concatenation operation, which merges the hidden states With memory cell state Concatenate them into a single vector to serve as the weight matrix. The input. tanh The hyperbolic tangent activation function is used to perform nonlinear mapping on the features after linear transformation, thereby enhancing the expressive power of the model. Its output range is [-1, 1]. Weight vector The transpose operation is used to calculate the inner product with the vector after nonlinear mapping, thus obtaining the scalar attention score. The input attention weights for the k-th feature at time step t and scale j are obtained by normalization using the softmax function, and their values ​​range from [value range missing]. A higher weight indicates that the feature is more important to the current encoding. : Normalization function, used to convert the attention scores of multiple features into a probability distribution form, ensuring that the sum of the weights is 1. exp Exponential functions are used to amplify differences in attention scores and strengthen the weighting of important features.

[0067] Step S222: Weighted fusion features

[0068] In step S222, the values ​​of each dimension of the current input feature vector are weighted according to the input attention weights to form an optimized input vector. The optimized input vector is then input into the Long Short-Term Memory (LSTM) unit to update the hidden state at the current time step. The specific implementation is as follows:

[0069] The optimized input is obtained by weighting each feature dimension according to its attention weight. And input the LSTM cell to update the hidden state. :

[0070]

[0071] The optimized input vector after attention-weighted fusion at time step t and scale j is generated by weighting each feature dimension according to its corresponding weight. The weighted average yields the contribution of key features. The specific value of the k-th feature at time step t and scale j is what constitutes the optimized input vector. The basic elements. The forget gate output of the LSTM unit is calculated using the sigmoid activation function, with a value range of [value range missing]. It is used to control whether historical information in memory units is forgotten. The input gate output of the LSTM unit is calculated using the sigmoid activation function, with a value range of [value range missing]. Used to control the current optimized input vector Information input intensity. The output gate of the LSTM unit is calculated using the sigmoid activation function, and its value range is [value range missing]. It is used to control the output strength of the memory cell state to the hidden state. The sigmoid activation function is used to map the result of a linear transformation to... The interval enables the switching control of the gate control unit. The learnable weight matrices of the forget gate, input gate, and output gate in the LSTM unit are used to... Perform a linear transformation. : Learnable bias terms for the forget gate, input gate, and output gate in the LSTM unit, used to adjust the reference value of the gated output. The memory cell state of the LSTM unit at time step t and scale j is obtained by fusing the historical state controlled by the forget gate and the current input state controlled by the input gate, and stores the core feature information at that moment. The learnable weight matrix for updating the state of memory cells in an LSTM unit, used to... Perform a linear transformation. : Learnable bias terms for updating the state of memory cells in an LSTM cell, used to adjust the baseline value for state updates. Element-wise multiplication (Hadamard product) is used to multiply corresponding elements of two vectors with the same dimension, thereby enabling gating signals to control the state element by element. The output hidden state of an LSTM unit at time step t and scale j is controlled by the output gate, which in turn controls the output of the memory unit's state. The result obtained by function mapping is the encoding result of the core features at that time step and scale.

[0072] Step S223: Calculate the context vector

[0073] In step S223, the hidden states at each time step are weighted and fused based on temporal attention weights to obtain the context vectors of each scale subsequence. The calculation method for the temporal attention weights is the same as that for the input attention weights. The specific implementation is as follows:

[0074] Based on temporal attention weights (The calculation method is the same as the input attention weights, focusing on key time steps), and the hidden states at each time step are weighted and fused to obtain the context vector. Quantify the core features at this scale:

[0075]

[0076] The temporal attention weight at time step t and at scale j is calculated in the same way as the input attention weight. It is used to focus on the historical time steps that are key to the current encoding, and its value range is... . The context vector at time step t and scale j, composed of the hidden states at each time step. Attention weight by time sequence The weighted fusion yields the global core features at this scale. The hidden state at time step i and scale j constitutes the context vector. The basic elements.

[0077] Step S23: Cross-scale guided temporal attention decoding

[0078] In step S23, during the decoding stage, feature information from a coarser scale is used as a guiding signal to influence the decoding process of the finer-scale feature sequence, achieving cross-scale feature synergy and fusion, and outputting the fused feature vector. The specific implementation is as follows:

[0079] A cross-scale guided temporal attention decoder (GTA-Decoder) is employed, leveraging the global stability of coarse-scale features to guide the local dynamics of fine-scale features during decoding. Specifically:

[0080] Step S231: Feature Fusion

[0081] S231, Feature Fusion: This decoder maps the hidden state of the coarse-scale decoder at the previous time step using a nonlinear function and performs element-wise multiplication with the hidden state of the fine-scale decoder at the previous time step. This modulates the local dynamics of the fine-scale features through the global stability of the coarse-scale features during the decoding process. Specifically:

[0082] Hidden state of the coarse-scale j+1 decoder with fine scale Hidden state in the previous moment Cross-scale guided features are obtained through element-wise multiplication fusion:

[0083]

[0084]

[0085] The cross-scale guided temporal attention decoder (GTA-Decoder) stores globally stable features (such as the average assembly gap and constant ambient temperature) at the coarse scale (t-1 time step and j+1 time step) in the hidden state of the coarse scale, providing guidance for fine-scale decoding. The decoder stores the hidden state at time step t-1, at fine scale j, containing local dynamic features at the fine scale (such as instantaneous acceleration and impact velocity fluctuations), which will be optimized based on coarse-scale features. tanh: Hyperbolic tangent activation function, used to optimize the coarse-scale hidden state. Perform a nonlinear mapping and adjust its numerical range to , adapting to the fusion of fine-scale hidden states. Element-wise multiplication (Hadamard product) is used to hide the previous time-series state at a finer scale. The elements are multiplied one-to-one with the coarse-scale hidden state after tanh mapping to achieve guided fusion of cross-scale features. The target sequence and context vector fusion feature at time step t-1 integrates historical target determination and scale core features to assist the decoding process. The collision detection at time step t-1 determines the target label (such as the quantized value of qualified / unqualified / invalid detection), providing target guidance for decoding. The context vector at time step (t-1) and scale (j) quantifies the global core features at that scale, serving as the fused features. Provides scale feature support. Vector concatenation operation, which concatenates the target sequence With context vector Concatenate them into a unified vector, which will serve as the weight matrix. Input. : The learnable weight matrix for fusion feature calculation, used to perform linear transformation on the concatenated vector and adjust the feature fusion ratio. Weight matrix The transpose operation is used to perform matrix multiplication with the concatenated vectors, resulting in a linear transformation. : Learnable bias terms calculated from fused features, used to adjust the baseline values ​​of fused features and optimize the decoding input. The decoding LSTM unit has the same internal update logic as the encoding LSTM in step S222 (including forget gate, input gate, output gate, etc.), with only the parameters initialized independently. It is used for cross-scale guided features and fused features. Decode the code and output the current hidden state. The decoder's output hidden state at time step t and scale j is determined by cross-scale guided features and fused features. The results, obtained through decoding LSTM units, integrate coarse-scale stability and fine-scale dynamics.

[0086] Step S232: Final Feature Output

[0087] S232. Final Feature Output: The decoded hidden state at the finest scale in the final time step is concatenated with the context vector at that scale and subjected to a linear transformation to output the final fused feature vector. The specific implementation is as follows:

[0088] The finest-scale decoding hidden state is obtained through linear mapping. With context vector The process involves fusing the features and outputting the final fused feature vector. It covers core information such as the rationality of assembly parameters, the effectiveness of the collision zone, the trend of acceleration changes, and the state characteristics of the shell:

[0089]

[0090] : finest scale ( In the last time step ( Decoding the hidden state of collision detection focuses on the high-frequency dynamic core features of collision detection (such as acceleration changes and peak impact velocity during the collision process). : finest scale ( In the last time step ( The context vector of ) quantifies the global core features at this scale, and They complement each other. The vector concatenation operation concatenates the finest-scale decoded hidden state with the context vector into a unified vector, which serves as the input for the final feature fusion. The learnable weight matrix of the final feature linear mapping is used to perform linear transformation on the concatenated vector and extract multi-dimensional fused features. : The learnable bias term of the final feature linear mapping, used to adjust the baseline value of the linear transformation result. The final feature output is a learnable weight vector, whose dimension matches that of the vector after linear transformation, and is used to perform weighted summation on the transformation result. Weight vector The transpose operation is used to calculate the inner product with the vector after linear transformation, resulting in a scalar feature combination. : The learnable bias term in the final feature output, used to fine-tune the value range of the final feature vector and optimize subsequent judgment performance. feat: The final output fused feature vector, covering core information such as assembly parameter rationality features, collision area validity features, acceleration change trend features, and shell state features, providing input for multi-factor collaborative judgment.

[0091] Step S3, Multi-factor Collaborative Judgment: The fused feature vector is input into the multi-constraint optimization judgment model for multi-factor collaborative optimization judgment. By dynamically adjusting the weight factors associated with each constraint, the various constraints are balanced and the confidence of the judgment result is maximized. Finally, the judgment result containing the detection result and the core influencing factors is output, and the detection data and judgment result are fed back to the model training library to optimize the model.

[0092] This invention proposes a multi-constraint optimization fusion algorithm, which achieves accurate fusion and determination of multi-dimensional features through the synergy of Lagrange relaxation and metaheuristic algorithms. Figure 3 This is a flowchart of the algorithm for the multi-constraint optimization decision model provided in this embodiment of the invention, as shown below. Figure 3 As shown, the details are as follows:

[0093] Step S31: Determine the target and constraint model

[0094] Using helmet collision test results as qualified, unqualified, or invalid, and maximizing the confidence level as the objective function, a multi-constraint optimization judgment model is established, consisting of multiple constraints directly related to helmet safety performance. These constraints include at least assembly compliance constraints, area validity constraints, performance compliance constraints, and shell integrity constraints. The specific implementation is as follows:

[0095] Using the helmet collision detection result (pass / fail / invalid detection) as the optimization objective, a multi-constraint optimization decision model is constructed, with the objective function being to maximize the decision confidence after feature fusion.

[0096]

[0097] The optimized objective function value for helmet collision detection, i.e., the decision confidence score after feature fusion, has a value range of [value missing]. The larger the value, the higher the reliability of the judgment result. The confidence function is used to calculate the credibility of the collision detection result based on the final feature vector feat. It is essentially a variation of the sigmoid function. The decision weight vector has the same dimension as the feature vector (feat) and is used to quantify the contribution of each core feature to the decision result. It is adaptively optimized through training. : Decision bias term, used to adjust the baseline value of the confidence function and optimize the bias of the decision result; it is a learnable parameter. exp: Exponential function, used to adjust the linear combination term. feat Performing an exponential transformation is the core component of the sigmoid function. feat: The inner product operation of the weight vector and the feature vector realizes the linear combination of features and quantifies the comprehensive influence of features on the judgment result.

[0098] Set multiple constraints, including:

[0099] (1) Assembly compliance constraints: gap tension .

[0100] The assembly compliance constraints are set according to the requirements for testing preparation in helmet safety standards (such as GB 811-2022). These constraints include: the assembly gap between the head mold and the inner wall of the helmet should not exceed a first preset threshold (e.g., 1.5 mm); and the tension of the helmet straps should be maintained within a preset allowable fluctuation range (e.g., ±1 N) near the standard value (e.g., 5 N). The gap constraint function in the assembly compliance constraint is where gap is the measured assembly gap between the head mold and the inner wall of the helmet (unit: mm), which is a key indicator of assembly compliance. The constraint gap value shall not exceed 1.5 mm. The tension constraint function in assembly compliance constraints, where `tension` is the measured tension value of the helmet strap (unit: N), is a core indicator of assembly compliance. The constraint tension is... Within the range. Absolute value calculation: Used to calculate the absolute value of the deviation between the strap tension and the standard value of 5N, ensuring that tension fluctuations do not exceed [the specified value]. scope.

[0101] (2) Effective constraints for the region: .

[0102] : Region effective constraint function, used to determine whether the collision point is located in a critical weak region. Spatial coordinates of the collision point The straight-line distance (in mm) to the nearest critical weak point is the criterion for determining the effectiveness of the area. thr: Distance threshold (in mm), determined by the distribution characteristics of the helmet's weak points. The distance between the impact point and the weak point must not exceed this threshold to be considered valid. .

[0103] (3) Performance compliance constraints: , , .

[0104] Peak acceleration constraint function in performance compliance constraints. The peak acceleration (in g) measured during the collision is a core indicator of the helmet's ability to absorb collision energy. The peak value is constrained to not exceed 300g (compliant with GB811-2022 standard) and must meet certain requirements. . The acceleration duration constraint function in the performance compliance constraints, time The duration of an acceleration exceeding 150g (in milliseconds) and the constraint duration not exceeding 4ms are key indicators of helmet safety performance and must meet these requirements. . The acceleration duration constraint function in the performance compliance constraints. For measured action time (in milliseconds) of acceleration exceeding 200g, the constraint duration must not exceed 2ms (except when the peak acceleration does not exceed 300g). This is an important indicator of helmet safety performance and must meet the following requirements. .

[0105] (4) Shell integrity constraints: deform-deform , .

[0106] The deformation constraint function in the shell integrity constraint is given by the following formula: "deform" refers to the measured deformation of the helmet shell after the collision (unit: mm). This represents the maximum permissible deformation of the shell. "Deformation" refers to the measured deformation of the helmet shell after a collision; it is a core indicator of shell integrity and must meet certain requirements. deform The maximum allowable deformation of the helmet shell (unit: mm) is determined by the helmet material and structural design, and is the threshold for judging the integrity of the shell. The fragment length constraint function in shell integrity constraints is given by the following formula: The maximum length (in mm) of fragments detached from the helmet shell after a collision, with a constraint that the fragment length should not exceed 10 mm. The maximum measured length of debris detached from the helmet after a collision is a key indicator of shell integrity and must meet certain requirements. .

[0107] Step S32: Lagrange relaxation decomposition

[0108] In step S32, non-negative Lagrange multipliers are introduced to incorporate multiple constraints into the objective function, resulting in an unconstrained Lagrange relaxation function. The relaxed problem is then decomposed into multiple single-constraint subproblems, each corresponding to the local optimization of a constraint. The specific implementation is as follows:

[0109] By introducing non-negative Lagrange multipliers By incorporating the constraints into the objective function, we obtain the unconstrained relaxation objective function:

[0110]

[0111] : A non-negative Lagrange multiplier vector containing the multipliers corresponding to the 8 constraints, i.e. This is used to incorporate constraints into the objective function to achieve unconstrained relaxation. : The Lagrange multiplier corresponding to the z-th constraint Values When the constraint is satisfied When the constraint is violated If the value is negative, a penalty is imposed on the objective function. feat, The unconstrained objective function after Lagrange relaxation transforms a multi-constrained optimization problem into an unconstrained optimization problem.

[0112] When constraints When satisfied When the constraint is violated A negative value penalizes the objective function. The relaxed problem is decomposed into eight single-constraint subproblems. Each of these corresponds to a local optimization for each constraint. The z-th single-constraint subproblem is obtained by decomposing the relaxed objective function. It corresponds to the local optimization of a single constraint. Global optimization is achieved by solving all subproblems.

[0113] Step S33: Metaheuristic Multiplier Update

[0114] In step S33, the Bat Algorithm is used to dynamically optimize and update the Lagrange multipliers. This includes initializing the multiplier population, iteratively updating the frequency, velocity, and position of the multipliers, and accepting new multipliers based on fitness evaluation until convergence yields the optimal multipliers. The specific implementation is as follows:

[0115] The Bat Algorithm (BA) is used to optimize the Lagrange multipliers. By simulating the echolocation behavior of bats, the multipliers are dynamically adjusted to improve constraint satisfaction and objective function value.

[0116] Step S331 Multiplier Initialization

[0117] Randomly generate the initial multiplier population Initial frequency ,speed Loudness Pulse rate .

[0118] The first Lagrange multiplier in the initial population Each multiplier individual, Population Index ( The solution is obtained through random generation and provides initial candidate solutions for multiplier updates. The population size of the bat algorithm, i.e. the total number of initial multiplier individuals, is adaptively set by the complexity of the collision detection optimization problem to ensure population diversity. The bat algorithm in the first The frequency parameters of each multiplier individual are used to regulate the step size strength of the multiplier update, and their values ​​are limited to a range of values ​​within a certain range. It adapts to the multiplier adjustment requirements of helmet collision detection with multi-constraint optimization. : No. The initial velocity parameter of each multiplier is set to 0, which means that the position of the multiplier has no tendency to move in the initial stage of the iteration, and the search for the optimal solution starts from the static state. : No. The initial loudness parameter of each multiplier is preset by the algorithm parameters and is used to control the acceptance probability of the new multiplier. The initial value is assigned as follows: This affects the algorithm's initial strategy for selecting candidate solutions. Initial loudness baseline value, representing the initial loudness of all multiplier individuals. The unified assignment, based on the pre-defined complexity of the helmet collision detection optimization problem, influences the acceptance strategy in the initial search phase of the algorithm, ensuring the diversity of the initial search. : No. The initial impulse rate parameter of each multiplier is used to trigger the local perturbation behavior of the multiplier, and is initially assigned a value of [value to be filled in]. This controls the triggering frequency of local searches in the initial stage. The initial baseline value for the pulse rate parameter is the initial pulse rate for all multiplier individuals. Unified assignment adapts to the search rhythm optimized by helmet collision detection constraints, balancing global exploration and local exploitation.

[0119] Step S332: Multiplier Update

[0120] Frequency update: (a uniform random vector);

[0121] Speed ​​updates: ( (the optimal multiplier in the previous iteration);

[0122] Location update: ,like Then, based on the local perturbation of the optimal multiplier: ( It is a random number. (mean loudness of the population).

[0123] Fitness evaluation: Calculate the objective function value for each multiplier. If rand and Then accept the new multiplier and update the loudness. (attenuation coefficient) and pulse rate ( (This refers to the growth coefficient).

[0124] : The minimum value of the frequency parameter (takes a value of 0), which limits the lower limit range of the frequency. : The maximum value of the frequency parameter (takes a value of 100), which limits the upper limit of the frequency range. : A uniform random vector, with values ​​ranging from It is used to randomly generate the frequency value of each multiplier individual, thereby enhancing population diversity. In the m-th iteration, the first... The velocity parameters of each multiplier individual are used to control the magnitude of the multiplier position update, with initial values... . : No. In the nth iteration The velocity parameters of each multiplier individual provide historical data for the current iteration velocity update. : No. In the nth iteration The value of each multiplier individual, i.e., the position of the historical multiplier. : No. The optimal multiplier in the next iteration, i.e., the one that makes the objective function... The multiplier that achieves the maximum value is used to guide the population towards the optimal direction for updates. In the m-th iteration, the first... The updated value (new multiplier position) of each multiplier individual is obtained by adding the historical position and the current velocity. rand: Random numbers within the interval are used to trigger local perturbations of the multiplier and the acceptance determination of the new multiplier, simulating the randomness of bat foraging. : No. In the nth iteration The impulse rate of each multiplier individual, with a value range of... Used to control the triggering probability of local disturbances, with an initial value of . Random numbers within the interval are used to generate the amplitude of local perturbations of the multiplier, thereby enhancing the randomness of the local search. : No. The average loudness of the multiplier population in each iteration is determined by the loudness of all individuals. The average value is obtained and used to adjust the intensity of local disturbances. : No. In the nth iteration The loudness of each multiplier individual, with values ​​ranging from... Used to control the acceptance probability of the new multiplier, with an initial value of . In the m-th iteration, the first... A new multiplier The corresponding Lagrange relaxation objective function value is used to evaluate the fitness of the new multiplier. : No. In the nth iteration The sun is a chariot. The corresponding objective function value is used as a benchmark for comparing the fitness of the new multipliers. Loudness attenuation coefficient, with a value range of [value missing]. Used to update loudness This causes the loudness to gradually decrease with iteration, enhancing the convergence stability of the algorithm in the later stages. In the m-th iteration, the first... The updated loudness of each multiplier individual is obtained by multiplying the historical loudness by the attenuation coefficient. : Pulse rate growth coefficient, with a value greater than 0, used to update the pulse rate. This allows the pulse rate to gradually increase with iteration, reducing the frequency of local disturbances in the later stages. In the m-th iteration, the first... The updated pulse rate of each multiplier individual is determined by the initial pulse rate and the growth coefficient. The exponential decay term approaches 0 as the number of iterations m increases, causing the pulse rate to... Gradually approaching . : Iteration count index, which records the current iteration step of the algorithm and is used to control the dynamic update of loudness and impulse rate.

[0125] Step S333: Convergence Determination

[0126] When the number of iterations reaches a preset threshold or the change in the multiplier is less than... When the time comes, stop updating and obtain the optimal multiplier. .

[0127] The threshold for determining multiplier convergence is the change in the value of all individual multipliers in two consecutive iterations. When all values ​​are less than the threshold, the algorithm is considered to have converged.

[0128] The optimal Lagrange multiplier vector obtained after the algorithm converges contains the optimal multipliers corresponding to the 8 constraints. It is used for the final constraint satisfaction check.

[0129] Step S34: Final Decision Logic

[0130] In step S34, the satisfaction of each constraint is checked based on the optimal multiplier. If all constraints are satisfied, the judgment result of "qualified" or "unqualified" is output based on the judgment confidence calculated by the optimal eigenvector. If a constraint is violated, it is judged as "invalid detection" and the specific constraint factor that was violated is marked. The specific implementation is as follows:

[0131] Step S341: Constraint Satisfaction Check

[0132] The optimal multiplier Substituting the relaxation function, if all constraints correspond to If the test is successful, the system will proceed to performance evaluation; otherwise, the test will be deemed invalid, and the core factors that violate the constraints will be marked (such as excessive tether tension, non-weak areas in the collision zone, etc.).

[0133] The multiplier corresponding to the z-th constraint in the optimal multiplier vector is used to check whether the z-th constraint is satisfied. Optimal multiplier With the z-th constraint function The product of the constraints is 0 when the constraints are satisfied and negative when the constraints are violated.

[0134] Step S342: Performance Compliance Determination

[0135] Based on optimal features Calculate the confidence level of the decision ,like If so, it is deemed qualified (meets GB811-2022 standard); if If it is, then it is judged as unqualified; if , it is marked as suspected unqualified and requires manual review.

[0136] : Based on the optimal features The determined confidence level calculated, with a value range of , which is used for the final performance determination of helmet collision detection. : The final feature vector optimized by the optimal multiplier is the calculation basis of the determined confidence level . 0.8: The qualified determination threshold. When , it is determined that the helmet collision detection result is qualified (meeting the GB811-2022 standard). 0.5: The unqualified determination threshold. When , it is determined that the helmet collision detection result is unqualified. conf : The suspected unqualified determination interval. The detection results within this interval need to be manually reviewed to ensure the determination accuracy.

[0137] The method of the present invention further includes the following steps:

[0138] Step S4, Result output and feedback

[0139] The AI model outputs the detection result (qualified / unqualified / detection invalid), the detection data dispersion analysis report, and the key influencing factor identification result; if the detection result is unqualified or detection invalid, the core factors causing this result are automatically associated and marked (such as the lacing tension exceeding the standard, the collision area not being the weak area, the acceleration peak value exceeding the standard, etc.), providing precise guidance for subsequent detection optimization; at the same time, the current detection data and the determination result are fed back to the model training library to achieve continuous iterative optimization of the model.

[0140] Figure 4 is the module schematic diagram of the helmet collision detection accuracy improvement system based on multi-factor collaboration in the embodiment of the present invention. As Figure 4 shown, a helmet collision detection accuracy improvement system 200 based on multi-factor collaboration in an embodiment of the present invention includes:

[0141] Data acquisition and preprocessing module 210: Obtain multi-dimensional time-series data generated during the helmet collision detection and perform data preprocessing. The multi-dimensional time-series data includes the assembly parameters of the helmet, the collision area information, the environmental parameters, and the collision dynamic parameters;

[0142] Feature extraction and fusion module 220: Inputs the preprocessed multi-dimensional time-series data into a pre-trained cross-scale guided recurrent network model for multi-scale feature extraction and fusion, including: performing multi-scale decomposition on the input multi-dimensional time-series data to separate feature components representing characteristics of different time scales; encoding the feature components at each scale separately, extracting key features at each scale and generating corresponding context vectors; and using the stability of coarse-scale features through a cross-scale guided mechanism to modulate the decoding process of fine-scale features, outputting a fused feature vector that comprehensively represents helmet assembly, collision area, dynamic performance, and shell state.

[0143] The detection result output module 230 inputs the fused feature vector into the multi-constraint optimization judgment model for multi-factor collaborative optimization judgment. By dynamically adjusting the weight factors associated with each constraint, it balances the various constraints and maximizes the confidence of the judgment result, and finally outputs the judgment result containing the detection result and the core influencing factors.

[0144] The helmet collision detection accuracy improvement system based on multi-factor collaboration provided in this embodiment of the invention can execute the helmet collision detection accuracy improvement method based on multi-factor collaboration provided in any of the above embodiments of the invention. It has the corresponding functions and beneficial effects of executing the helmet collision detection accuracy improvement method based on multi-factor collaboration. For detailed process, please refer to the relevant operations of the helmet collision detection accuracy improvement method based on multi-factor collaboration in the foregoing embodiments, which will not be repeated here.

[0145] The method and system for improving helmet collision detection accuracy based on multi-factor collaboration provided by embodiments of the present invention design a cross-scale guided recurrent network (MRN-CSG), utilizes stationary wavelet transform to separate high-frequency dynamic and low-frequency stable features, and achieves accurate fusion of multi-dimensional features through attention encoding and cross-scale guided decoding; constructs a multi-constraint optimization judgment model, employs Lagrange relaxation to integrate multiple safety constraints such as assembly compliance and performance standards, and introduces the bat algorithm to dynamically optimize weights, achieving efficient balance of complex constraints; establishes an interpretable decision-making process of "constraint check - performance judgment," automatically identifying and labeling the core factors leading to non-compliance. Ultimately, it achieves significant benefits such as more comprehensive feature extraction, higher judgment accuracy (experiments show an accuracy rate of 95.5%), and stronger interpretability of results.

[0146] like Figure 5As shown, a simulation experiment was constructed using Python to compare the accuracy performance of the algorithm of this invention (MRN-CSG + multi-constraint optimization fusion algorithm) with traditional RNN, CNN, and LSTM models in helmet collision detection. The experiment simulated 8-dimensional collision detection feature data (including key indicators such as assembly gap, strap tension, and peak acceleration) conforming to the GB811-2022 standard. The number of training iterations was set to 1-20. Differentiated icon labels (circle for RNN, rhombus for CNN, square for LSTM, and triangle for the algorithm of this invention) were used to draw performance comparison curves. The results show that as the number of training iterations increases, the accuracy of all four models gradually increases and tends to stabilize. Among them, the algorithm of this invention maintains the best performance throughout, with an accuracy of 95.5% after 20 training iterations, which is significantly higher than LSTM's 90.5%, CNN's 85.4%, and RNN's 79.5%. Moreover, it shows a fast convergence advantage in the early stage of training (the first 5 iterations), which reflects the ability of cross-scale guided recurrent networks to accurately extract multi-dimensional features and the core value of multi-constraint optimization fusion algorithm in balancing multiple detection indicators and improving judgment accuracy. This fully demonstrates that the algorithm of this invention has better accuracy and convergence efficiency than traditional models in helmet collision detection tasks.

[0147] It should be noted that the various embodiments in this specification are described in a progressive manner, with each embodiment focusing on the differences from other embodiments. Similar or identical parts between embodiments can be referred to interchangeably. For the apparatus disclosed in the embodiments, since it corresponds to the method disclosed in the embodiments, the description is relatively simple; relevant parts can be referred to in the method section.

[0148] It should also be noted that, in the embodiments of this application, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitations, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes said element.

[0149] The above description of the disclosed embodiments enables those skilled in the art to make or use this application. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the general principles defined in the embodiments of this application may be implemented in other embodiments without departing from the spirit or scope of this application. Therefore, this application is not to be limited to the embodiments shown in this application, but is to be accorded the widest scope consistent with the principles and novel features disclosed in the embodiments of this application.

Claims

1. A method for improving the accuracy of helmet impact detection based on multi-factor synergy, characterized in that, The method comprises the following steps: Step S1: obtaining multi-dimensional time series data generated by the helmet during the collision detection process and performing data preprocessing, wherein the multi-dimensional time series data comprises assembly parameters, collision region information, environmental parameters and collision dynamic parameters of the helmet; Step S2: inputting the preprocessed multi-dimensional time series data into a pre-trained cross-scale guided recurrent network model to perform multi-scale feature extraction and fusion, comprising: performing multi-scale decomposition on the input multi-dimensional time series data to separate feature components representing different time scale characteristics; encoding each scale feature component, extracting key features at each scale, and generating corresponding context vectors; through a cross-scale guiding mechanism, the stability of coarse time scale features is used to modulate the decoding process of fine time scale features, and finally a fusion feature vector is output, which comprehensively represents the assembly, collision region, dynamic performance and shell state of the helmet; The specific process of the cross-scale guided recurrent network model for extraction and fusion includes multi-scale decomposition, scale component coding and cross-scale guided decoding performed in sequence: S21: the multi-scale decomposition is used to decompose the multi-dimensional time series data into fine scale components containing high frequency dynamic features and coarse scale components containing low frequency stable features through a signal decomposition method; S22: the scale component coding is used to encode each scale component sequence obtained by decomposition using a recurrent neural network with an attention mechanism to adaptively extract key time series features at each scale and generate context vectors representing global information of the scale; S23: the cross-scale guided decoding is used to use the decoded feature information of the j+1 time scale as a guide signal to optimize the decoding process of the feature sequence of the current time scale j in the decoding stage, realize the cooperation and fusion of cross-scale features, and output the fusion feature vector; Step S3: inputting the fusion feature vector into a multi-constraint optimization judgment model to perform multi-factor collaborative optimization judgment, balancing each constraint and maximizing the confidence of the judgment result by dynamically adjusting the weight factor associated with each constraint condition, and finally outputting a judgment result containing a detection result and core influencing factors.

2. The multi-factor synergy based helmet collision detection accuracy enhancement method of claim 1, wherein, In step S1, the multi-dimensional time series data specifically comprises: assembly gap data between the helmet and the headform and tension data of the helmet harness obtained by physical measurement; spatial coordinate data of 12 key weak areas on the helmet shell determined based on finite element analysis; standardized preprocessing parameters for simulating real use environment and calibration verification data for verifying the accuracy of the detection device itself; impact speed data before the helmet collision is collected by the sensor in real time, acceleration time history data during the collision is collected, and shell shape data and fragment data after the helmet collision are collected by the visual monitoring device.

3. The multi-factor synergy based helmet collision detection accuracy enhancement method of claim 1, wherein, Wherein, In S21, the signal decomposition method uses stationary wavelet transform, and the following steps are specifically performed: S211, filter initialization and processing: using a high-pass filter and a low-pass filter to perform convolution operation on the multi-dimensional time series data; wherein the high-pass filter and the low-pass filter are both initialized and configured based on db2 wavelet basis function; S212, scale decomposition: through the convolution operation of the high-pass filter, the fine scale component retaining the impact speed and the high frequency dynamic characteristics of instantaneous acceleration is extracted from the multi-dimensional time series data; at the same time, through the convolution operation of the low-pass filter, the coarse scale component retaining the assembly gap and the low frequency stable characteristics of the environment temperature is extracted; S213, recursive decomposition: the filtering and decomposition process of S211 and S212 is repeatedly performed recursively on the coarse scale component, and finally a set containing J different scale sub-sequences is generated, wherein J is an integer greater than 1; S214, parameter adaptive optimization: in the end-to-end training process of the cross-scale guided recurrent network model, the filter coefficients of the high-pass filter and the low-pass filter are fine-tuned synchronously to adapt the filter characteristics to the characteristics of the helmet collision detection data.

4. The multi-factor synergy based helmet collision detection accuracy enhancement method of claim 1, wherein, Wherein, In step S22, the scale component coding is performed by inputting an attention encoder, which encodes each scale sub-sequence by time step, and the specific process includes: S221: for the current time step, based on the hidden state and memory unit state of the encoder at the previous time step, the input attention weight of each feature dimension in the current input feature vector is calculated; S222: according to the input attention weight, the dimension values of the current input feature vector are weighted to form an optimized input vector; the optimized input vector is input into the long short-term memory unit to update the hidden state of the current time step; S223: based on the time sequence attention weight, the hidden states of each time step are weighted and fused to obtain the context vector of each scale sub-sequence; wherein the calculation method of the time sequence attention weight is the same as that of the input attention weight.

5. The multi-factor synergy based helmet collision detection accuracy enhancement method of claim 4, wherein, The calculation method of the input attention weight of each feature dimension in the current input feature vector includes: The hidden state and memory unit state of the encoder at the previous time step are spliced as the first intermediate representation; The values of each feature dimension in the current input feature vector are taken as the second intermediate representation; The first intermediate representation and the second intermediate representation are linearly transformed and summed respectively, and then added with a learnable bias term, and then processed by a nonlinear activation function to obtain the attention score of each feature dimension; The attention score is processed using a normalization exponential function to obtain the input attention weight of each feature dimension.

6. The multi-factor synergy based helmet collision detection accuracy enhancement method of claim 1, wherein, Wherein, In S23, the cross-scale guided decoding is realized by a cross-scale guided time attention decoder, which specifically includes the following steps: S231, feature fusion: the decoder maps the hidden state of the coarse scale decoder at the previous time step through a nonlinear function, and performs element-level multiplication operation with the hidden state of the fine scale decoder at the previous time step, so that the global stability of the coarse scale feature modulates the local dynamic decoding process of the fine scale feature; S232, final output: concatenating and linearly transforming the decoded hidden state of the finest scale at the final time step with the context vector of the scale, outputting the final fusion feature vector.

7. The multi-factorial synergy based helmet collision detection accuracy enhancement method of claim 1, wherein, The step S3 specifically comprises: S31, judging target and constraint modeling: taking the helmet collision detection result as qualified, unqualified and invalid detection, maximizing the judgment confidence as the objective function, establishing a multi-constraint optimization judgment model composed of multiple constraint conditions directly related to the safety performance of the helmet, the multiple constraint conditions at least including a fitting standard constraint, a region validity constraint, a performance compliance constraint and a shell integrity constraint; Step S32, Lagrange relaxation decomposition: introducing a non-negative Lagrange multiplier, integrating the multiple constraint conditions into the objective function to obtain an unconstrained Lagrange relaxation function, and decomposing the relaxed problem into multiple single-constraint sub-problems corresponding to the local optimization of each constraint; Step S33, meta-heuristic multiplier updating: updating the Lagrange multiplier dynamically by using a bat algorithm, including initializing the multiplier population, iteratively updating the frequency, speed and position of the multiplier, and accepting new multipliers according to the fitness evaluation until the optimal multiplier is obtained; Step S34, final judgment: checking the satisfaction of each constraint based on the optimal multiplier, if the constraints are satisfied, outputting the judgment result of "qualified" or "unqualified" according to the judgment confidence calculated by the optimal feature vector; if the constraints are violated, judging as "invalid detection" and marking the specific constraint factor violated.

8. The multi-factorial synergy based helmet collision detection accuracy enhancement method of claim 7, wherein, Step S33, meta-heuristic multiplier updating, specifically comprising: S331, population initialization: randomly generating an initial population containing N Lagrange multiplier individuals; initializing a frequency parameter, a speed parameter, a loudness parameter and a pulse rate parameter for each multiplier individual; wherein the initial value of the frequency parameter is randomly generated within a predetermined fixed range, the initial value of the speed parameter is set to zero, and the loudness parameter and the pulse rate parameter are initialized to a predetermined reference value, respectively; S332, iterative updating and evaluation: in each iteration, the following operations are performed on each multiplier individual in the population: (a) updating its frequency parameter according to the predetermined frequency range and a random factor; (b) calculating its new speed according to the difference between its historical speed, its own position and the position of the optimal individual in the current population, and the updated frequency; (c) updating its position according to its historical position and new speed as a candidate new multiplier; (d) with a predetermined probability, applying a random disturbance to the candidate new multiplier centered on the position of the optimal individual in the current population and with an amplitude of the average loudness of the population to generate a disturbed candidate new multiplier; (e) calculating the Lagrange relaxation function value corresponding to the candidate new multiplier as its fitness; comparing its fitness with its fitness in the previous iteration, and accepting the better candidate new multiplier to update its position with a certain probability; (f) if a new multiplier is accepted, reducing its loudness parameter by a predetermined decay coefficient and increasing its pulse rate parameter by a predetermined growth mechanism; S333, convergence determination: when the number of iterations reaches the preset maximum number, or the position change amount of all multiplier individuals in the population in adjacent two iterations is less than the preset minimum threshold, stop iteration, and output the optimal multiplier individual in the current population as the optimal multiplier combination.

9. A multi-factor synergy based helmet collision detection accuracy enhancement system, comprising: The system for implementing the method for improving the accuracy of helmet collision detection based on multi-factor synergy according to any one of claims 1-8 comprises: a data acquisition and preprocessing module for acquiring multi-dimensional time series data generated by the helmet during the collision detection process and performing data preprocessing, the multi-dimensional time series data including assembly parameters, collision region information, environmental parameters, and collision dynamic parameters of the helmet; a feature extraction and fusion module for inputting the preprocessed multi-dimensional time series data into a pre-trained cross-scale guided recurrent network model for multi-scale feature extraction and fusion, including: multi-scale decomposition of the input multi-dimensional time series data to separate feature components representing different time scale characteristics; encoding each scale feature component, extracting key features at each scale, and generating corresponding context vectors; using the stability of coarse time scale features to modulate the decoding process of fine time scale features through a cross-scale guiding mechanism, and outputting a fusion feature vector representing the assembly, collision region, dynamic performance, and shell state of the helmet; a detection result output module for inputting the fusion feature vector into a multi-constraint optimization judgment model for multi-factor synergy optimization judgment, balancing each constraint and maximizing the confidence of the judgment result by dynamically adjusting the weight factor associated with each constraint condition, and finally outputting a judgment result containing the detection result and the core influencing factors.

Images