Multi-modal signal processing and intelligent analysis motion load electrocardiogram test system

Abstract

The present application relates to the technical field of exercise load electrocardio test, and discloses a multi-modal signal processing and intelligent analysis exercise load electrocardio test system, which comprises: a multi-modal signal fusion module, which is used for constructing a joint data space and integrating multi-source physiological signals; a dynamic feature extraction module, which extracts features at global and local layers respectively; an adaptive decision optimization module, which constructs a knowledge rule base to generate a feature extraction strategy adapted to a new scene; and an abnormal state early warning module, which monitors and triggers an early warning strategy in real time. Through synchronous acquisition, multi-scale decomposition, hierarchical screening and other operations, the system realizes efficient fusion and feature extraction of multi-modal signals, can adaptively optimize the strategy according to different test scenes, monitors and warns in real time, improves the accuracy, anti-interference, adaptability and safety of the exercise load electrocardio test, and is suitable for the fields of sports medicine and health monitoring.

Description

Technical Field

[0001] This invention relates to the field of exercise stress electrocardiogram (ECG) testing technology, specifically an exercise stress ECG testing system that integrates multimodal signal processing and intelligent analysis. Background Technology

[0002] From a signal acquisition perspective, traditional systems often only acquire electrocardiogram (ECG) signals, making it difficult to simultaneously integrate motion parameters and other multi-source physiological signals, such as heart rate variability, respiratory rate, and electromyographic activity. Even some systems with multi-signal acquisition capabilities lack effective time alignment and calibration mechanisms, resulting in significant timestamp discrepancies between different signals. This makes it impossible to construct an accurate joint data space, hindering the effective analysis of temporal synchronization and frequency correlation between signals and severely impacting the accuracy of subsequent feature extraction and analysis.

[0003] In terms of feature extraction, traditional systems typically employ single-scale feature extraction methods, failing to achieve accurate feature extraction at both the global and local levels. For example, they can only extract simple global features such as signal mean and variance, while lacking the ability to extract more detailed local features that reflect physiological states, such as time-frequency energy, morphological inflection points, and feature peaks and valleys. This single feature extraction approach struggles to comprehensively capture the complex changes in human physiological signals during exercise, resulting in inaccurate and incomplete assessments of the human body's state under exercise load.

[0004] In terms of decision optimization mechanisms, traditional systems lack adaptive decision optimization capabilities. When faced with complex scenarios with different testing stages and varying exercise intensities, they cannot extract common rules from historical training and build a knowledge rule base. They also struggle to automatically generate adaptive feature extraction strategies based on new testing scenarios, resulting in poor system versatility and flexibility, and an inability to meet diverse testing needs.

[0005] In terms of abnormal state early warning, traditional systems have relatively simple and crude early warning mechanisms. They typically only set fixed early warning thresholds without considering the joint analysis of multi-source physiological signals or the possibility of stable operation but multiple triggering of abnormalities during actual testing. This results in insufficient accuracy and timeliness of early warnings, either leading to false alarms or failure to trigger in a timely manner when a warning is truly needed, thus failing to provide reliable protection for the safety of testing personnel.

[0006] Furthermore, traditional systems lack efficient sample screening and processing mechanisms when handling large amounts of signal data. They cannot properly classify and process samples of different values, leading to low data processing efficiency and impacting system real-time performance and response speed. Simultaneously, traditional systems lack effective methods for constructing and optimizing feature screening models, making it difficult to accurately assess feature importance and dynamically update them, further affecting system performance. In contrast, this device possesses powerful and precise detection technology, capable of accurately detecting and measuring various waveforms, amplitudes, and intervals in electrocardiograms. It also features high-speed, advanced filtering algorithms and digital signal processing algorithms, effectively filtering out various interferences and noise generated during high-intensity exercise, overcoming the shortcomings of traditional systems in feature extraction accuracy and anti-interference capabilities. Summary of the Invention

[0007] The purpose of this invention is to provide a multimodal signal processing and intelligent analysis exercise stress electrocardiogram testing system to solve the problems mentioned in the background art.

[0008] To achieve the above objectives, the present invention provides the following technical solution: a multimodal signal processing and intelligent analysis exercise stress electrocardiogram testing system, the system comprising:

[0009] The multimodal signal fusion module is used to construct a joint data space of electrocardiogram signals and motion parameters by using high-speed and high-precision synchronous signal acquisition, and to integrate multi-source physiological signals, including heart rate variability, respiratory rate and electromyographic activity. The joint data space includes a time-domain synchronization layer, a frequency-domain correlation layer and a time-frequency joint layer.

[0010] The dynamic feature extraction module executes global and local feature extraction strategies at the global and local levels, respectively, to accurately detect and measure various waveforms, amplitudes, and intervals in an electrocardiogram (ECG). The global feature extraction strategy uses a multi-scale decomposition method to extract a first feature set; the local feature extraction strategy extracts a second feature set by constructing a feature selection model. The first feature set includes signal mean, variance, and trend term; the second feature set includes time-frequency energy, morphological inflection points, and characteristic peak and valley values.

[0011] The adaptive decision optimization module is used to extract common rules from global and local feature extraction strategies trained under different testing stages and different exercise intensities, build a knowledge rule base, and generate feature extraction strategies adapted to new testing scenarios based on the knowledge rule base.

[0012] The abnormal state early warning module is used to monitor multi-source physiological signals in real time based on the joint data space, and trigger the first preset early warning strategy when an abnormal indicator is detected.

[0013] Preferably, a joint data space of electrocardiogram signals and motion parameters is constructed by employing high-speed, high-precision synchronous signal acquisition, and multi-source physiological signals are integrated. Specific steps include:

[0014] Use synchronous acquisition equipment to obtain timestamp data of ECG signals and motion parameters, and perform time alignment calibration;

[0015] By analyzing the correlation between signals, time-domain synchronization processing, frequency-domain correlation analysis, and time-frequency joint mapping are performed on the joint data space.

[0016] Real-time multi-source physiological signals from the actual test environment are acquired and input into the joint data space.

[0017] Preferably, the specific steps of the global feature extraction strategy include:

[0018] The signal data is segmented using the sliding window method. The data segment includes M samples, and each sample is represented by a first feature vector composed of a first feature set.

[0019] A multi-scale decomposition framework is constructed, which includes: a coarse-grained decomposition layer, a medium-grained decomposition layer, and a fine-grained decomposition layer.

[0020] A stratified screening method was used to process the samples;

[0021] The specific steps of the hierarchical screening method include:

[0022] S1, calculate the coarse-grained, medium-grained, and fine-grained features of the sample based on the multi-scale decomposition framework;

[0023] S2, based on feature importance, divides the samples into three groups: high-value samples, medium-value samples, and low-value samples;

[0024] S3 retains high-value samples, performs merging and recombination operations on medium-value samples, and performs high-probability screening operations on low-value samples.

[0025] S4. Based on the coarse-grained, medium-grained, and fine-grained features of the samples, hierarchical clustering is used to screen and select 3 typical samples. Among them, the samples with the most significant coarse-grained, medium-grained, and fine-grained features are selected as typical samples.

[0026] S5: When the processing length reaches the preset maximum processing length, output the current 3 typical samples and select the sample that best matches the current test scenario according to actual needs; otherwise, return to S2.

[0027] Preferably, the merging operation uses a time series splicing method to generate new samples. The operation process of the time series splicing method is as follows: the time series data of two parent samples are connected end to end in the order of timestamps to generate continuous child sample data segments, ensuring that the child samples retain complete ECG waveform temporal characteristics and improving the detection accuracy of waveforms, amplitudes and intervals.

[0028] The recombination operation uses the feature space transformation method to update the samples. The operation process of feature space transformation is as follows: the feature vector of the offspring sample is mapped to a high-dimensional space, the principal components are extracted through principal component analysis, and the dimensionality-reduced feature vector is generated.

[0029] Preferably, the global feature extraction strategy and the local feature extraction strategy are converted into global rule vectors and local rule vectors, respectively, and then concatenated to obtain a decision rule vector. For decision rule vectors of different test stages and different motion intensities, the repetition frequency of each rule in the decision rule vector is calculated in turn. When the repetition frequency is greater than a second preset threshold, it is determined that the rule has high universality across different stages and intensities and is regarded as a common rule. The extracted common rules are combined to obtain a knowledge rule base, which provides a general feature extraction framework for new test scenarios.

[0030] Preferably, the abnormal state early warning module monitors multi-source physiological signals in real time based on the joint data space. When the multi-source physiological signals exceed the preset safety range, it determines that the test system is operating abnormally and triggers the first preset early warning strategy. A statistical period is set to record the number of times the same abnormality is triggered within the period. If the same abnormality occurs multiple times within the statistical period and the test system is still running stably, the early warning threshold for the abnormality is adjusted.

[0031] Preferably, the specific steps of the local feature extraction strategy include:

[0032] Based on the operating status of the test system, a state matrix containing a second feature set is constructed. The expression of the state matrix is ​​a multidimensional array containing time steps, heart rate variability, respiratory rate, and electromyographic activity.

[0033] A lightweight computational model is used to construct a screening network. This screening network takes the state matrix as input and outputs a screening index for the second feature set. The screening index represents the importance score of the feature.

[0034] Construct a feature space, which represents the range of values ​​for the screening criteria of the second feature set;

[0035] Construct an evaluation function for the feature selection model;

[0036] The screening network is updated using an iterative optimization method.

[0037] Preferably, a synchronous acquisition device is used to acquire timestamp data of electrocardiogram signals and motion parameters, and time alignment calibration is performed. Specific steps include:

[0038] Separate timestamp recording units are configured for electrocardiogram signals and motion parameters;

[0039] All recording units are started synchronously by a hardware trigger signal to ensure that the initial timestamps are consistent.

[0040] The collected timestamp data is subjected to deviation detection, and the time offset is corrected by linear interpolation to achieve time alignment between signals.

[0041] Preferably, the first preset early warning strategy includes: an audio-visual alert strategy, a data storage strategy, and a test interruption strategy. The audio-visual alert strategy is used to send visual and auditory alarms to the test personnel, the data storage strategy is used to perform a complete backup of the signal data before and after the anomaly occurs, and the test interruption strategy is used to automatically stop the test process when the anomaly is severe.

[0042] Preferably, the knowledge rule base provides a general feature extraction framework for new testing scenarios. Specific application methods include: inputting the basic parameters of the new testing scenario into the knowledge rule base, matching the closest historical testing stage and exercise intensity, extracting the corresponding common rules, and combining the specific parameters of the new scenario to generate an adapted feature extraction strategy. The specific parameters include: the age, gender, and basic physical fitness indicators of the tester.

[0043] Compared with the prior art, the beneficial effects of the present invention are:

[0044] In terms of multimodal signal fusion, the system constructs a joint data space for ECG signals and motion parameters by employing high-speed, high-precision synchronous signal acquisition. It integrates multi-source physiological signals such as heart rate variability, respiratory rate, and electromyographic activity, including a time-domain synchronization layer, a frequency-domain correlation layer, and a time-frequency joint layer. During this process, timestamp data is acquired using synchronous acquisition equipment and time alignment calibration is performed. By analyzing the correlation between signals, time-domain synchronization processing, frequency-domain correlation analysis, and time-frequency joint mapping are conducted, enabling precise alignment and correlation of multi-source signals in a unified space. This lays a solid foundation for subsequent feature extraction and analysis, solves the problem of insufficient signal fusion in traditional systems, and improves the accuracy of signal processing.

[0045] The dynamic feature extraction module executes different feature extraction strategies at the global and local levels. The global feature extraction strategy uses a multi-scale decomposition method to extract the first set of features, such as the signal mean, variance, and trend term. The local feature extraction strategy extracts the second set of features, such as time-frequency energy, morphological inflection points, and feature peak and valley values, by constructing a feature selection model. In global feature extraction, a sliding window is used to segment the signal data. The multi-scale decomposition framework includes coarse, medium, and fine-grained decomposition layers. The hierarchical selection method processes the samples, including calculating features at each granularity, grouping by importance, and performing operations such as retention, merging and recombination, or high-probability selection on samples of different values. Then, hierarchical clustering is used to select typical samples. The merging operation uses time series splicing, and the recombination operation uses feature space transformation. This multi-scale, hierarchical processing method can comprehensively capture the global features of the signal while improving data processing efficiency. Local feature extraction constructs a state matrix based on the running state, uses a lightweight computational model to build a screening network, outputs feature importance scores, constructs a feature space and evaluation function, and uses an iterative optimization method to update the screening network, thus achieving accurate extraction of local subtle features. This comprehensively and meticulously reflects the changes in human physiological signals during movement, improving the comprehensiveness and accuracy of feature extraction.

[0046] The adaptive decision optimization module extracts common rules from global and local feature extraction strategies trained under different testing stages and exercise intensities, constructs a knowledge rule base, and generates feature extraction strategies adapted to new testing scenarios based on the knowledge rule base. Specifically, the global and local feature extraction strategies are transformed into rule vectors and concatenated to obtain decision rule vectors. The repetition frequency of each rule is calculated, and rules with a repetition frequency greater than a threshold are considered common rules. The knowledge rule base matches historical stages and intensities based on the basic parameters of the new scenario, extracts corresponding common rules, and combines them with specific parameters to generate adaptation strategies. This enables the system to automatically adjust and optimize feature extraction strategies according to different testing scenarios, greatly improving the system's versatility and flexibility, and meeting diverse testing needs.

[0047] The abnormal state early warning module monitors multi-source physiological signals in real time based on the joint data space. When an abnormal indicator is detected, it triggers a first preset early warning strategy that includes audible and visual alerts, data storage, and test interruption. Simultaneously, it records the number of times the same abnormality is triggered at a set statistical period. If the system operates stably, the early warning threshold is adjusted. This real-time monitoring and dynamic adjustment early warning mechanism improves the accuracy and timeliness of early warnings, better protects the safety of test personnel, and avoids false alarms and missed alarms. Attached Figure Description

[0048] Figure 1 This is a schematic diagram illustrating the working principle of the exercise stress electrocardiogram testing system for multimodal signal processing and intelligent analysis described in this invention.

[0049] Figure 2A flowchart for constructing a joint data space;

[0050] Figure 3 This is a flowchart of the sample merging and recombination operation;

[0051] Figure 4 This is a flowchart of a local feature extraction strategy. Detailed Implementation

[0052] 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.

[0053] Please see Figures 1-4 This invention relates to a multimodal signal processing and intelligent analysis exercise stress electrocardiogram testing system. The system includes: a multimodal signal fusion module, a dynamic feature extraction module, an adaptive decision optimization module, and an abnormal state early warning module. The specific implementation steps are as follows:

[0054] The multimodal signal fusion module is used to construct a joint data space of electrocardiogram signals and motion parameters by using high-speed and high-precision synchronous signal acquisition, and to integrate multi-source physiological signals, including heart rate variability, respiratory rate and electromyographic activity. The joint data space includes a time-domain synchronization layer, a frequency-domain correlation layer and a time-frequency joint layer.

[0055] The dynamic feature extraction module executes global and local feature extraction strategies at the global and local levels, respectively, to accurately detect and measure various waveforms, amplitudes, and intervals in an electrocardiogram (ECG). The global feature extraction strategy uses a multi-scale decomposition method to extract a first feature set; the local feature extraction strategy extracts a second feature set by constructing a feature selection model. The first feature set includes signal mean, variance, and trend term; the second feature set includes time-frequency energy, morphological inflection points, and characteristic peak and valley values.

[0056] The adaptive decision optimization module is used to extract common rules from global and local feature extraction strategies trained under different testing stages and different exercise intensities, build a knowledge rule base, and generate feature extraction strategies adapted to new testing scenarios based on the knowledge rule base.

[0057] The abnormal state early warning module is used to monitor multi-source physiological signals in real time based on the joint data space, and trigger the first preset early warning strategy when an abnormal indicator is detected. Example 1:

[0058] This embodiment details the specific implementation of the high-speed, high-precision signal synchronous acquisition and joint data space construction in the multimodal signal fusion module. The core function of this module is to construct a joint data space of electrocardiogram signals and motion parameters by using high-speed, high-precision signal synchronous acquisition, and to integrate multi-source physiological signals such as heart rate variability, respiratory rate, and electromyographic activity. The joint data space includes a time-domain synchronization layer, a frequency-domain correlation layer, and a time-frequency joint layer.

[0059] In high-speed, high-precision synchronous signal acquisition, the acquisition and calibration of timestamp data are fundamental steps. Specifically, independent timestamp recording units are configured for both ECG signals and motion parameters. These recording units utilize high-precision clock chips to ensure the accuracy and stability of the timestamps. For example, clock chips with nanosecond-level accuracy can be selected to ensure that time deviations are kept within a minimal range during long-term acquisition. All recording units are synchronously activated via a hardware trigger signal generated by a unified clock source, such as a crystal oscillator-based synchronous clock generator, ensuring that the initial timestamps of all recording units are consistent and avoiding initial deviations caused by asynchronous startup times.

[0060] After collecting timestamp data, deviation detection and correction are required. Timestamp data is monitored in real time, and when the deviation between adjacent timestamps exceeds a preset threshold, linear interpolation is used for correction. The time interval where the deviation occurs is determined, and the two valid timestamps before and after the deviation within this interval, along with their corresponding actual times, are obtained. The time offset and linear rate of change between these two timestamps are calculated. Interpolation is then performed on intermediate timestamps based on the linear rate of change to obtain the corrected timestamp values, thus achieving time alignment between signals. For example, if the offset of a timestamp relative to the preceding and following timestamps is 5 milliseconds, and the time interval between the preceding and following timestamps is 100 milliseconds, then linear interpolation is used to calculate the correction values ​​for each intermediate time point, ensuring the continuity and accuracy of the time series.

[0061] After time alignment, the joint data space needs to be processed in multiple dimensions. By analyzing the correlation between signals, correlation coefficient calculation methods are used to perform time-domain synchronization, frequency-domain correlation analysis, and time-frequency joint mapping on the joint data space. In the time-domain synchronization process, based on the time-aligned timestamps, the time series of ECG signals and motion parameters are precisely aligned to ensure that ECG signal data and motion parameter data at the same time point correspond. For example, each sampling point of the ECG signal is associated with a motion parameter sampling point at the same timestamp, forming a synchronized time series data pair.

[0062] In the frequency domain correlation analysis, Fourier transforms are performed on both the electrocardiogram (ECG) signal and motion parameters, converting the time-domain signals into frequency-domain features. Using the Fourier transform algorithm, the energy distribution of the signal at different frequency points is calculated, yielding their respective frequency-domain feature vectors. Then, the correlation coefficients between these two frequency-domain feature vectors at various frequency points are calculated to analyze their degree of frequency domain correlation. For example, the correlation coefficients between the ECG signal and motion parameters at different frequency points such as 1Hz and 2Hz are calculated. A high correlation coefficient at a certain frequency point indicates a strong correlation between the two at that frequency component.

[0063] Time-frequency joint mapping employs methods such as short-time Fourier transform (SFT) to convert ECG signals and motion parameters into the time-frequency domain. SFT, by selecting an appropriate window function, performs piecewise Fourier transforms on the signal, thereby obtaining the signal's distribution characteristics in both time and frequency dimensions. Constructing a time-frequency joint mapping map integrates the time-frequency characteristics of ECG signals and motion parameters, forming a three-dimensional mapping space containing information such as time, frequency, and energy, thus realizing the construction of a joint data space. For example, using time as the horizontal axis and frequency as the vertical axis, and representing energy intensity through color or grayscale, a time-frequency joint mapping map is constructed, visually displaying the signal's distribution in the time-frequency domain.

[0064] Real-time multi-source physiological signals from the actual testing environment are acquired and input into a joint data space. Physiological signals such as heart rate variability, respiratory rate, and electromyographic activity are acquired through various sensor devices. Among them, heart rate variability can be calculated from the RR interval of the electrocardiogram signal, respiratory rate can be analyzed from the signals collected by the respiratory sensor, and electromyographic activity is collected through electromyographic electrode pads. Specifically, when acquiring heart rate variability through electrocardiogram (ECG) signals, a peak detection algorithm is used to identify the peak value of the R wave in the ECG signal. The interval between adjacent R waves (RR interval) is calculated using timestamp data, and then statistical methods are used to analyze the RR interval sequence to obtain relevant parameters of heart rate variability. For respiratory rate, although it is mainly acquired through respiratory sensors, the respiratory sinus arrhythmia characteristics in the ECG signal can be used for auxiliary calculation. A spectral analysis algorithm is used to perform frequency domain transformation on the RR interval sequence to extract low-frequency components related to respiratory rate, thus enabling auxiliary verification of respiratory rate. For electromyographic (EMG) activity, in addition to direct acquisition, an independent component analysis algorithm can be used to separate ECG signals mixed with EMG interference, extract EMG components, and analyze their activity characteristics, further enriching the acquisition methods of multi-source physiological signals. Furthermore, the application of these algorithms all follow the time synchronization rules of the joint data space, ensuring the consistency of the acquired multi-source physiological signals in the time dimension and providing accurate data support for subsequent integration and processing. These multi-source physiological signals undergo preprocessing, such as filtering and amplification, and are then input into their corresponding positions in the constructed joint data space according to timestamp synchronization rules, thereby integrating the multi-source physiological signals. For example, the preprocessed heart rate variability data is mapped to the time-domain synchronization layer of the joint data space according to its timestamp, and then integrated with electrocardiogram signals and motion parameters to form a joint data space containing multi-source physiological signals, providing comprehensive data support for subsequent feature extraction and analysis.

[0065] The system employs high-speed, advanced filtering and digital signal processing algorithms to effectively filter out various interferences and noises generated during high-intensity exercise. The filtering operations encompass low-pass filtering, high-pass filtering, power frequency filtering, electromyography (EMG) filtering, and baseline filtering. Low-pass filtering removes high-frequency interference signals, high-pass filtering eliminates low-frequency baseline drift, power frequency filtering specifically suppresses 50Hz or 60Hz power frequency interference, EMG filtering weakens EMG noise generated by muscle activity, and baseline filtering further stabilizes the signal baseline. The digital signal processing algorithm dynamically sets different filtering parameters and algorithm combinations based on changes in exercise intensity, such as changes in signal characteristics from low-intensity to high-intensity exercise, and the characteristics of various multi-source biological signals, such as the proportion of low-frequency components in heart rate variability signals and the periodicity of respiratory rate signals. Simultaneously, it incorporates appropriate signal processing algorithms to ensure high-quality signal data acquisition even in complex exercise scenarios, providing a reliable foundation for the subsequent construction of a joint data space. Example 2:

[0066] This embodiment details the specific implementation of the global feature extraction strategy in the dynamic feature extraction module. The core of this strategy is to extract a first feature set containing the signal mean, variance, and trend term at the global layer using a multi-scale decomposition method. This is achieved through steps such as sliding window segmentation, multi-scale decomposition framework construction, and hierarchical filtering to complete feature extraction and sample selection of the signal data.

[0067] In the signal data segmentation stage, a sliding window method is used to process the original signal. The size of the sliding window is set to M samples, and the value of M needs to be flexibly adjusted according to the signal type and test requirements. For example, for ECG signals with high sampling frequency, considering their high-frequency characteristics and rich details, M can be set to 1000 to ensure that each window contains enough sample points to capture the waveform characteristics of the ECG signal; while for low-frequency motion parameters, such as signals reflecting slow changes in motion speed or acceleration, M can be set to 100, which can ensure the representativeness of the data within the window and reduce the amount of computation. Each sample consists of a first feature vector composed of a first feature set. The signal mean is calculated by summing all sample values ​​within the window and dividing by the number of samples, thus reflecting the average level of the signal within the window; the variance is calculated by first calculating the difference between each sample value and the mean, squaring it, and then averaging these squared values ​​to measure the degree of signal fluctuation; the trend term is obtained by fitting the sample values ​​within the window using a linear regression method, which can reflect the overall trend of the signal within the window.

[0068] Constructing a multi-scale decomposition framework is a crucial step in global feature extraction. This framework comprises three layers: a coarse-grained decomposition layer, a medium-grained decomposition layer, and a fine-grained decomposition layer. In the coarse-grained decomposition layer, a larger decomposition scale is used to process the signal, for example, by using larger filter coefficients or a lower decomposition frequency, thereby obtaining the overall contour and main features of the signal, i.e., coarse-grained features. These features reflect the macroscopic trends of the signal's changes. The medium-grained decomposition layer uses a medium decomposition scale to further decompose the signal based on the coarse-grained decomposition, capturing feature information between the macroscopic and microscopic levels, resulting in medium-grained features. These features reflect the signal's variation patterns at a medium scale. The fine-grained decomposition layer uses a smaller decomposition scale to finely decompose the signal, for example, by using high-resolution filters or high-frequency decomposition parameters, thereby extracting the signal's detailed features, resulting in fine-grained features. These features reflect subtle changes and local characteristics of the signal.

[0069] The hierarchical screening method involves several steps. First, coarse-grained, medium-grained, and fine-grained features of the samples are calculated based on a multi-scale decomposition framework. When calculating coarse-grained features, the samples are input into the coarse-grained decomposition layer, and parameters reflecting the overall signal characteristics are obtained through a corresponding decomposition algorithm. The calculation of medium-grained and fine-grained features is similar, obtained through processing by the medium-grained and fine-grained decomposition layers, respectively. Then, methods such as information gain are used to evaluate the importance of each feature to the signal description. Based on the sum of their importance, the samples are divided into high-value, medium-value, and low-value samples. High-value samples contain information crucial for the signal feature description, medium-value samples have features of intermediate importance, and low-value samples contain features that contribute little to the signal description.

[0070] Different processing methods are adopted for samples of different values. High-value samples are directly retained to ensure that key feature information is not lost. Medium-value samples undergo merging and recombination operations. The merging operation uses a time series splicing method, concatenating the time series data of two parent samples in time stamp order to generate continuous child sample data segments. This forms a longer-term time series sample while retaining the original information, capturing longer-term signal change trends. The recombination operation uses a feature space transformation method to map the feature vectors of the child samples to a high-dimensional space. Principal component analysis is used to extract the principal components. Specifically, the covariance matrix of the feature vectors is calculated, its eigenvalues ​​and eigenvectors are solved, and the top few eigenvectors with larger eigenvalues ​​are selected as principal components to generate dimensionality-reduced feature vectors. This reduces feature dimensionality while retaining key feature information, improving computational efficiency. Low-value samples undergo a high-probability screening operation, i.e., they are screened with a high probability to remove unimportant samples, thereby reducing the amount of data.

[0071] Based on the coarse-grained, medium-grained, and fine-grained features of the samples, hierarchical clustering is used for selection. Hierarchical clustering uses Euclidean distance as the distance metric between samples. By calculating the distances between samples, a clustering tree is constructed, and three representative samples are selected from the clustering results. The selection criteria are that the samples with the most significant coarse-grained, medium-grained, and fine-grained features are selected as representative samples. This ensures that the selected samples can comprehensively reflect the characteristics of the signal at different scales. When the processing length reaches the preset maximum processing length, the current three representative samples are output, and the sample that best fits the current scenario is selected based on the needs of the actual test scenario, such as the exercise intensity and physical condition of the testers. If the processing length does not reach the preset maximum processing length, the process returns to the step of dividing the sample groups according to feature importance and continues to process subsequent samples. Example 3:

[0072] This embodiment details the specific implementation of the adaptive decision optimization module. The core function of this module is to extract common rules from global and local feature extraction strategies trained under different testing phases and exercise intensities, construct a knowledge rule base, and generate feature extraction strategies adapted to new testing scenarios based on the knowledge rule base, thereby achieving adaptive optimization for different testing scenarios.

[0073] The global and local feature extraction strategies are transformed into global and local rule vectors, respectively. Elements in the global rule vector correspond to various rule parameters in the global feature extraction process, such as the scale parameters of each decomposition layer in the multi-scale decomposition framework and the feature importance threshold in the hierarchical selection method. Elements in the local rule vector correspond to various rule parameters in the local feature extraction strategy, such as the dimension setting of the state matrix and the structural parameters of the selection network. Then, the global and local rule vectors are concatenated to obtain the decision rule vector. The concatenation operation connects two vectors sequentially into a new vector. Let the global rule vector be... The local regular vector is Then the decision rule vector It can be represented as:

[0074] ;

[0075] in, Let be the dimension of the global rule vector. Let be the dimension of the local rule vector. Represents the first in the global rule vector One rule parameter, Represents the first in the local rule vector Each rule parameter.

[0076] For decision rule vectors at different testing phases and with different exercise intensities, the repetition frequency of each rule is calculated sequentially. The repetition frequency is calculated by counting the number of times each rule appears in all decision rule vectors to the total number of decision rule vectors. Assume there are a total of... A vector of decision rules, for a given rule parameter ( (where the position index is in the decision rule vector) If in If a decision rule appears in a vector, then its repetition frequency is... for:

[0077] ;

[0078] in, Let the total number of decision rule vectors be . For rule parameters Number of times it appears This indicates the frequency of repetition of the rule across all decision rule vectors.

[0079] When the repetition frequency exceeds a second preset threshold, the rule is considered to have high universality across different stages and intensities, and is thus regarded as a common rule. The second preset threshold can be set according to actual testing needs; for example, it can be set to 0.7, meaning that when the repetition frequency of a rule exceeds 70%, it is considered to have high universality. The extracted common rules are combined to obtain a knowledge rule base. The knowledge rule base is stored in the form of a database, containing all common rules and their related parameters, such as the rule parameter values ​​corresponding to each common rule, the applicable testing stage range, and the exercise intensity range.

[0080] The knowledge rule base provides a general feature extraction framework for new testing scenarios. Its specific application is as follows: The basic parameters of the new testing scenario are input into the knowledge rule base. These parameters include the testing phase (e.g., warm-up, high-intensity exercise, recovery), and exercise intensity (e.g., low, medium, high intensity). The knowledge rule base uses similarity calculation methods to match the closest historical testing phase and exercise intensity. Similarity calculation can employ methods such as cosine similarity for the basic parameter vector of the new testing scenario. and the basic parameter vector of historical test scenarios cosine similarity for:

[0081] ;

[0082] in, The dot product of two vectors. and Let be the magnitudes of the two vectors, respectively. The dimension of the basic parameter vector, Represents the first element in the basic parameter vector of the new test scenario. One parameter, The first element in the basic parameter vector of the historical test scenario represents the... One parameter.

[0083] By calculating the similarity between the new test scenario and various historical test scenarios, the historical test phases and exercise intensities with the highest similarity are identified, and corresponding common rules are extracted. Then, combined with the specific parameters of the new scenario, an adapted feature extraction strategy is generated. Specific parameters include the tester's age, gender, and basic physical fitness indicators. For example, the tester's age affects the normal range of heart rate variability, gender may be related to the characteristics of electromyographic activity, and basic physical fitness indicators (such as VO2 max and resting heart rate) affect physiological signal responses under exercise load. Based on these specific parameters, the extracted common rules are adjusted. For example, the decomposition scale in the multi-scale decomposition framework is adjusted to adapt to the electrocardiogram signal characteristics of testers of different ages, or the evaluation function parameters of the feature screening model are adjusted according to the basic physical fitness indicators, thereby generating a feature extraction strategy suitable for the new test scenario.

[0084] Throughout the implementation process, the knowledge rule base needs continuous updating and maintenance. When new test data is generated and analyzed, the new decision rule vectors are incorporated into the calculation, the repetition frequency of each rule is recalculated, and the set of common rules is updated. This ensures that the knowledge rule base can continuously optimize with the accumulation of test data, improving its adaptability to new test scenarios. Simultaneously, when generating feature extraction strategies adapted to new test scenarios, the rationality of the strategies needs to be verified. For example, the effectiveness of the strategies can be evaluated through pre-testing. If the strategy is found to be unsuitable for the current scenario, the weights of the specific parameters are readjusted, or a closer historical test scenario is selected, until a satisfactory feature extraction strategy is generated.

[0085] Example 4:

[0086] This embodiment details the specific implementation of the abnormal state early warning module. The core function of this module is to perform real-time monitoring of multi-source physiological signals based on a joint data space. When abnormal indicators are detected, a preset early warning strategy is triggered. Simultaneously, the early warning threshold is dynamically adjusted based on the number of abnormal triggers within a statistical period to ensure the safe operation of the testing system.

[0087] In the real-time monitoring phase, the abnormal state early warning module continuously acquires multi-source physiological signals from the joint data space, including heart rate variability, respiratory rate, and electromyographic activity. Each physiological signal corresponds to a preset safety range, which is determined based on medical standards and a large amount of historical test data. For example, the safety range for heart rate variability may be set to 50ms-150ms, which is based on the normal heart rate variability range of healthy adults at rest; the safety range for respiratory rate is usually 12-20 breaths / minute, which conforms to the normal physiological respiratory rate range; and the safety range for electromyographic activity is set according to the baseline electrical activity level of different muscle groups, such as the biceps brachii, whose electromyographic activity amplitude in a relaxed state is usually below 500μV.

[0088] When a physiological signal exceeds a preset safety range, the system will determine that the testing system is malfunctioning and trigger the first preset warning strategy. Taking heart rate variability as an example, if a test subject's heart rate variability suddenly drops to 30ms during exercise, below the safe lower limit of 50ms, the system will immediately identify this abnormal indicator. At this time, the audible and visual alert strategy will be activated, and the audible and visual alarm devices in the testing environment will sound an alarm. The visual alarm can be achieved through a flashing red icon on the test equipment display screen, while the warning light on the laboratory ceiling begins to flash at a high frequency; the auditory alarm will emit a continuous high-frequency sound through a buzzer, ensuring that test personnel and operators can detect the abnormality in a timely manner.

[0089] The data storage strategy is activated simultaneously, performing a complete backup of signal data before and after the anomaly occurs. Specifically, the system automatically records all signal data from 10 seconds before the anomaly to 20 seconds after it occurs, including data from every sampling point of the electrocardiogram (ECG) signal, acceleration and velocity change sequences in motion parameters, real-time calculated values ​​of heart rate variability, continuous monitoring data of respiratory rate fluctuation curves, and amplitude changes in electromyographic activity. This data is stored in a separate encrypted database for subsequent anomaly cause analysis and test process review. For example, if a test subject experiences an abnormally low heart rate variability during high-intensity exercise, the backup data can be consulted to analyze whether the anomaly is related to a sudden increase in exercise intensity, test subject fatigue, or poor equipment contact.

[0090] The system will correlate abnormal state warnings with abnormal electrocardiogram (ECG) events during exercise. Specifically, when abnormal physiological signals such as heart rate variability and respiratory rate are detected, real-time waveform data of ECG signals in the joint data space will be retrieved simultaneously. The system will focus on identifying typical ECG abnormal events such as abnormal QRS complex morphology, ST segment deviation, and arrhythmia. If abnormal physiological signals and abnormal ECG events occur simultaneously and are temporally correlated, it will be determined as a substantial cardiac function abnormality. At this time, the response level of the first preset warning strategy will be increased accordingly, such as shortening the interval between audible and visual prompts and speeding up the triggering speed of test interruption. If only abnormal physiological signals exist and no abnormal ECG events are detected, the system will further combine the trend of exercise intensity changes to determine whether it is a temporary physiological fluctuation, so as to reduce false warnings.

[0091] If the abnormal situation is severe, dynamic judgment based on exercise intensity is required: During non-maximum exercise, if heart rate variability remains below the safety lower limit (e.g., below 50ms) and accompanied by a sudden increase in respiratory rate to above 30 breaths / minute (exceeding the preset safety upper limit of 20 breaths / minute for this exercise intensity), the test interruption strategy will be triggered. The system will automatically send a stop command to the testing equipment, such as controlling the treadmill to gradually decelerate to a stop, stopping the resistance loading device of the power bicycle, and cutting off the power to the electromyography (EMG) electrodes. It should be noted that the safety range will be dynamically adapted according to exercise intensity: During maximal exercise, the human respiratory rate will physiologically increase, usually reaching 30-60 breaths / minute. At this time, the system will automatically increase the safety upper limit of respiratory rate for this stage (e.g., set to 60 breaths / minute). Only when the respiratory rate exceeds the dynamic safety range for the corresponding exercise intensity (e.g., exceeding 60 breaths / minute during maximal exercise) and is accompanied by other abnormal physiological signals (e.g., abnormal heart rate variability) will it be judged as abnormal and an interruption will be triggered, avoiding misjudging normal physiological responses under maximal exercise as abnormal.

[0092] At the same time, the system will generate an abnormal interruption report, recording the time of the abnormality, the specific value of the abnormal indicator, the triggered warning strategy, and the test status parameters at the time of interruption, such as the exercise intensity level and test duration, to provide a basis for subsequent safety assessment and test plan adjustment.

[0093] In addition, the module sets a statistical period to dynamically adjust the warning threshold. The statistical period can be set to 30 minutes, 1 hour, etc., according to testing needs. Taking a 1-hour statistical period as an example, within this period, the system will record the number of times the same anomaly is triggered in real time. Suppose that a test subject's body surface temperature reaches 37.8℃ multiple times within 1 hour during intermittent high-intensity training, exceeding the safety limit of 37.5℃, but after each anomaly, the test subject's other physiological indicators remain within the normal range, and the testing system operates stably. In this case, the system will adjust the warning threshold for this anomaly according to the preset adjustment logic. The adjustment method may be to temporarily raise the warning threshold for body surface temperature to 38℃ to avoid frequent warnings caused by individual differences or environmental factors, while continuing to closely monitor the changing trend of this indicator.

[0094] In practice, the abnormal state early warning module and the multimodal signal fusion module maintain real-time data interaction to ensure the time synchronization and accuracy of the acquired physiological signals. For example, when an abnormal respiratory rate occurs, the module simultaneously retrieves ECG and electromyography (EMG) data at the same timestamp, analyzes the correlation between the signals to determine whether the abnormality is caused by actual physiological changes or equipment interference. If the ECG signal shows baseline drift at the same time point, and the EMG activity also exhibits high-frequency noise, it may be determined that the abnormal respiratory rate is caused by equipment problems such as poor electrode contact, thus avoiding false triggering of the early warning strategy.

[0095] The warning module employs differentiated monitoring strategies for different testing phases. During the warm-up phase, when exercise intensity is low and physiological signals change relatively smoothly, the system sets stricter warning thresholds to promptly detect potential discomfort. However, during high-intensity exercise, as the range of physiological signal fluctuations increases, the system appropriately relaxes the warning thresholds for some indicators. For example, the lower limit for heart rate variability is temporarily adjusted to 40ms to avoid false warnings triggered by normal physiological fluctuations caused by exercise.

[0096] Example 5:

[0097] This embodiment details the specific implementation of the local feature extraction strategy in the dynamic feature extraction module. The core of this strategy is to extract a second feature set from multi-source physiological signals, including time-frequency energy, morphological inflection points, and feature peaks and valleys, by constructing a state matrix, filtering network, feature space, and evaluation function, in order to capture the local detailed features of the signal.

[0098] When constructing the state matrix, the system generates a multidimensional array based on the test execution status. Taking a test subject performing cycling exercise as an example, the time steps of the state matrix correspond to different moments in the exercise process, such as the 1st minute and the 5th minute after the start of exercise. At each time step, heart rate variability data is calculated from the RR interval of the electrocardiogram signal; for example, during high-intensity exercise at the 5th minute, the heart rate variability might be 60ms. Respiratory rate is obtained by analyzing signals collected by chest and abdominal respiratory sensors; for example, at the same time point, the respiratory rate is 22 breaths / minute. Electromyographic activity is collected by electrode pads attached to the quadriceps femoris muscle, and its amplitude might be 400μV. These data, along with the corresponding time-frequency energy, morphological inflection points, and characteristic peak-valley values, together constitute an element of the state matrix. Time-frequency energy can be calculated using short-time Fourier transform to determine the energy distribution of a specific frequency band, such as the time-frequency energy of an electrocardiogram signal in the 30-40Hz band; morphological inflection points refer to the key turning points of a signal waveform, such as the inflection point of a QRS complex; characteristic peak and valley values ​​are the peak and valley values ​​of a signal waveform, such as the peak and valley amplitudes of a respiratory signal.

[0099] When constructing a screening network using a lightweight computational model, taking a convolutional neural network as an example, its structure can include an input layer, several convolutional layers, pooling layers, and an output layer. The input layer receives multidimensional data from the state matrix; the convolutional layers extract local features from the data using convolutional kernels of different sizes; the pooling layers reduce the dimensionality of the features to reduce computational cost; and the output layer generates screening indices for the second set of features, i.e., the importance score of each feature. For example, when processing the time-frequency energy features of electrocardiogram (ECG) signals, the screening network may assign a higher score to the energy distribution features in the 30-40Hz frequency band because this band may be related to abnormal myocardial electrical activity under exercise load; while for the morphological inflection point features of respiratory rate, if it shows regular changes during exercise, the score may be lower, and conversely, if abnormal fluctuations occur, the score will increase.

[0100] When constructing the feature space, the value range of the screening index for the second feature set needs to be clearly defined. The screening index for time-frequency energy can be set to a range of 0-1, where 0 indicates that the feature is the least important and 1 indicates that it is the most important. The screening index for morphological inflection points can be set to a range of 0-10 based on the accuracy and importance of inflection point detection, with larger values ​​indicating that the inflection point is more critical to the description of the signal features. The screening index for feature peak and valley values ​​can be set in conjunction with the normal fluctuation range of signal amplitude. For example, for the peak and valley value feature of electromyography (EMG), the value range can be set to 0-500, corresponding to the μV unit of EMG amplitude.

[0101] The evaluation function of the feature selection model measures the degree of match between the output of the selection network and the actual importance of the feature. Taking the mean squared error function as an example, it calculates the average of the squared differences between the importance score output by the selection network and the manually labeled actual importance. For instance, when manually analyzing a segment of electrocardiogram (ECG) signal, if the time-frequency energy in the 30-40Hz band is found to be strongly correlated with ST segment changes under exercise load, then the actual importance of this feature is labeled as 0.8. If the score output by the selection network is 0.75, then the squared difference between the two is calculated as part of the evaluation function.

[0102] When updating the screening network using an iterative optimization method, the stochastic gradient descent algorithm is employed. In each iteration, a batch of state matrices is randomly selected from the training data as input. The screening network calculates the output screening index, then calculates the error based on the evaluation function, backpropagates the error, and updates the network parameters. For example, during training, if the screening index error for time-frequency energy features is found to be large, the algorithm adjusts the convolutional kernel parameters for the corresponding frequency band feature extraction in the convolutional layer, increasing the weight of that frequency band feature, thereby improving the screening network's ability to recognize that feature. After multiple iterations, the output of the screening network meets the preset requirements, such as the error of the evaluation function being reduced to below a certain threshold.

[0103] In actual testing, when the exercise intensity of the test subject switches from low to moderate, the local feature extraction strategy dynamically adjusts the focus of feature selection. For example, as exercise intensity increases, the time-frequency energy distribution of the electrocardiogram signal may shift from the low-frequency band to the high-frequency band. At this time, the selection network will automatically increase the importance score of the high-frequency time-frequency energy features to capture the impact of increased exercise load on myocardial electrical activity. At the same time, the morphological inflection point of respiratory rate may change due to deeper and faster breathing. The selection network will assign higher scores to newly emerging inflection point features to reflect changes in breathing patterns.

[0104] The local feature extraction strategy is adaptable to different testers. Taking young and old testers as examples, young testers usually have greater heart rate variability. Under the same exercise intensity, the peak and trough values ​​of their heart rate variability may differ from those of old testers. The screening network will automatically adjust the scoring criteria for the peak and trough values ​​of heart rate variability based on the age-specific parameters in the state matrix to ensure the accuracy of feature extraction.

[0105] It should be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, 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 process, method, article, or apparatus.

[0106] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A multimodal signal processing and intelligent analysis exercise stress electrocardiogram testing system, characterized in that, include: The multimodal signal fusion module is used to construct a joint data space of electrocardiogram signals and motion parameters by using high-speed and high-precision synchronous signal acquisition, and to integrate multi-source physiological signals, including heart rate variability, respiratory rate, and electromyographic activity. The joint data space includes a time-domain synchronization layer, a frequency-domain correlation layer, and a time-frequency joint layer. The dynamic feature extraction module is used to execute global and local feature extraction strategies at the global and local levels, respectively, to achieve accurate detection and measurement of various waveforms, amplitudes, and intervals in an electrocardiogram. The global feature extraction strategy uses a multi-scale decomposition method to extract a first feature set; the local feature extraction strategy extracts a second feature set by constructing a feature selection model. The first feature set includes the signal mean, variance, and trend term, while the second feature set includes time-frequency energy, morphological inflection points, and feature peak and valley values. The adaptive decision optimization module is used to extract common rules from global and local feature extraction strategies trained under different testing stages and different exercise intensities, build a knowledge rule base, and generate feature extraction strategies adapted to new testing scenarios based on the knowledge rule base. The abnormal state early warning module is used to monitor multi-source physiological signals in real time based on the joint data space, and trigger the first preset early warning strategy when an abnormal indicator is detected. The specific steps of the global feature extraction strategy include: The signal data is segmented using the sliding window method. The data segment includes M samples, and each sample is represented by a first feature vector composed of a first feature set. A multi-scale decomposition framework is constructed, which includes: a coarse-grained decomposition layer, a medium-grained decomposition layer, and a fine-grained decomposition layer. A stratified screening method was used to process the samples; The specific steps of the hierarchical screening method include: S1, calculate the coarse-grained, medium-grained, and fine-grained features of the sample based on the multi-scale decomposition framework; S2, based on feature importance, divides the samples into three groups: high-value samples, medium-value samples, and low-value samples; S3 retains high-value samples, performs merging and recombination operations on medium-value samples, and performs high-probability screening operations on low-value samples. S4. Based on the coarse-grained, medium-grained, and fine-grained features of the samples, hierarchical clustering is used to screen and select 3 typical samples. Among them, the samples with the most significant coarse-grained, medium-grained, and fine-grained features are selected as typical samples. S5: When the processing length reaches the preset maximum processing length, output the current 3 typical samples and select the sample that best matches the current test scenario according to actual needs; otherwise, return to S2. The merging operation uses a time series splicing method to generate new samples. The operation process of the time series splicing method is as follows: the time series data of two parent samples are connected end to end in the order of timestamps to generate continuous child sample data segments, ensuring that the child samples retain complete ECG waveform temporal characteristics and improving the detection accuracy of waveforms, amplitudes and intervals. The recombination operation uses the feature space transformation method to update the samples. The operation process of feature space transformation is as follows: the feature vector of the offspring sample is mapped to a high-dimensional space, the principal components are extracted through principal component analysis, and the dimensionality-reduced feature vector is generated. This enhances the ability of the feature vector to represent the details of the electrocardiogram waveform and helps to improve the accuracy of waveform parameter measurement.

2. The exercise stress electrocardiogram testing system for multimodal signal processing and intelligent analysis according to claim 1, characterized in that, By employing high-speed, high-precision synchronous signal acquisition to construct a joint data space for electrocardiogram signals and motion parameters, and integrating multi-source physiological signals, the specific steps include: It employs high-speed, high-precision bioinformatics acquisition technology to acquire multi-channel electrocardiogram signals with high fidelity and real-time. Use synchronous acquisition equipment to obtain timestamp data of ECG signals and motion parameters, and perform time alignment calibration; By analyzing the correlation between signals, time-domain synchronization processing, frequency-domain correlation analysis, and time-frequency joint mapping are performed on the joint data space. By collecting electrocardiogram signals, real-time multi-source physiological signals from the actual test environment are obtained, and these multi-source physiological signals are input into the joint data space.

3. The exercise stress electrocardiogram testing system for multimodal signal processing and intelligent analysis according to claim 1, characterized in that, The global feature extraction strategy and the local feature extraction strategy are transformed into global rule vectors and local rule vectors, respectively, and then concatenated to obtain decision rule vectors. For decision rule vectors of different test stages and different motion intensities, the repetition frequency of each rule in the decision rule vector is calculated in turn. When the repetition frequency is greater than a second preset threshold, the rule is determined to have high universality across different stages and intensities and is regarded as a common rule. The extracted common rules are combined to obtain a knowledge rule base, which provides a general feature extraction framework for new test scenarios.

4. The exercise stress electrocardiogram testing system for multimodal signal processing and intelligent analysis according to claim 1, characterized in that, The abnormal state early warning module monitors multi-source physiological signals in real time based on the joint data space. When the multi-source physiological signals exceed the preset safety range, it determines that the test system is operating abnormally and triggers the first preset early warning strategy. Set a statistical period and record the number of times the same anomaly is triggered within that period. If the same anomaly occurs multiple times within the statistical period and the test system is still running stably, then the warning threshold for that anomaly will be adjusted.

5. The exercise stress electrocardiogram testing system for multimodal signal processing and intelligent analysis according to claim 1, characterized in that, The specific steps of the local feature extraction strategy include: Based on the operating status of the test system, a state matrix containing a second feature set is constructed. The expression of the state matrix is ​​a multidimensional array containing time steps, heart rate variability, respiratory rate, and electromyographic activity. A lightweight computational model is used to construct a screening network. This screening network takes the state matrix as input and outputs a screening index for the second feature set. The screening index represents the importance score of the feature. Construct a feature space, which represents the range of values ​​for the screening criteria of the second feature set; Construct an evaluation function for the feature selection model; The screening network is updated using an iterative optimization method.

6. The exercise stress electrocardiogram testing system for multimodal signal processing and intelligent analysis according to claim 2, characterized in that, The time stamp data of ECG signals and motion parameters are acquired using a synchronous acquisition device, and time alignment calibration is performed. The specific steps include: Separate timestamp recording units are configured for electrocardiogram signals and motion parameters; All recording units are started synchronously by a hardware trigger signal to ensure that the initial timestamps are consistent. The collected timestamp data is subjected to deviation detection, and the time offset is corrected by linear interpolation to achieve time alignment between signals.

7. The exercise stress electrocardiogram testing system for multimodal signal processing and intelligent analysis according to claim 4, characterized in that, The first preset early warning strategy includes: an audio-visual alert strategy, a data storage strategy, and a test interruption strategy. The audio-visual alert strategy is used to send visual and auditory alarms to the testers. The data storage strategy is used to fully back up the signal data before and after the anomaly occurs. The test interruption strategy is used to automatically stop the test process when the anomaly is severe.

8. The exercise stress electrocardiogram testing system for multimodal signal processing and intelligent analysis according to claim 3, characterized in that, The knowledge rule base provides a general feature extraction framework for new testing scenarios. Specific application methods include: inputting the basic parameters of the new testing scenario into the knowledge rule base, matching the closest historical testing stage and exercise intensity, extracting the corresponding common rules, and combining the specific parameters of the new scenario to generate an adapted feature extraction strategy. The specific parameters include: the age, gender, and basic physical fitness indicators of the tester.

Images