Electrocardiogram data enhancement-based pilot high-risk psychological state identification method
Through electrocardiogram data enhancement technology, high-risk and low-information segments are combined to form a heavy-balanced data set, solving the problem of scarcity of high-risk psychological state data for pilots, improving the recognition ability and accuracy of the model, and ensuring flight safety.
Patent Information
- Application Number
- CN202410944256.1
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2024-07-15
- Publication Date
- 2024-11-08
- Estimated Expiration
- 2044-07-15
AI Technical Summary
The pilot's psychological risk data set is due to the scarcity and insufficient proportion of high-risk state samples, which leads to insufficient generalization ability and low accuracy when identifying high-risk states.
Through the data enhancement method based on ECG data, high-risk factor samples are combined with low information segments to achieve heavy equalization of training samples. Low-information sample segments are extracted using residual network model and data enhancement is performed through peak detection algorithms to form heavy equalization The data set of the data set trains a model that can identify high-risk psychological states.
It realizes effective identification of high-risk psychological states in the case of sufficient low-risk data and insufficient high-risk data, and improves the generalization ability and actual application accuracy of the model.
Smart Images

Figure CN118902453A_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the technical field of physiological signal recognition, and in particular to a method for identifying a pilot's high-risk psychological state based on electrocardiogram data enhancement. Background Art
[0002] Safety is the lifeline of the civil aviation industry. With the increasing application of modern science and technology in the field of aviation safety, the main tasks of pilots in flight have changed greatly, from the previous two-stick and one-rudder flight operation to the collection, organization, decision-making and execution of overall flight-related information during the flight. Among the current major human factors that endanger flight safety for civil aviation pilots, psychological risk factors are the most hidden, but also the most dangerous.
[0003] The risky psychological state of civil aviation pilots during flight includes risk emotions, mental stress, cognitive fatigue, risk perception, alertness and other psychological factors that seriously threaten flight safety. These psychological risk factors affect the pilots' information receiving and problem-solving abilities during normal flights and unexpected special situations from multiple perspectives, such as motivation, skill level, and the ability to perceive danger. High-risk psychological states such as highly aroused panic and extremely high mental stress may even swallow up the pilots' rationality, resulting in the inability to effectively control safety risks, and ultimately causing huge disasters.
[0004] Therefore, when identifying the psychological risk factors of pilots, the most important thing to pay attention to is these high-risk psychological states. However, due to the professional adaptability of pilots themselves, when facing emergencies and unexpected special situations, the psychological fluctuations they produce are usually not too drastic. Even if some unexpected situations during flight training trigger high-intensity psychological risks, this high-risk throbbing is only short-lived and one-time. This leads to the fact that whether in flight training or in actual flight, the situation where pilots' psychological risk factors are in a high-risk state is relatively rare.
[0005] The consequence of this problem is that in the process of collecting psychological risk data, high-risk psychological states are very likely to fall into a situation of insufficient data due to their difficulty in stimulating and obtaining. Therefore, the psychological risk data set of pilots tends to have a long-tail distribution with sufficient low-risk data and insufficient high-risk data, that is, calm and relaxed states with low attention and low danger account for the majority of the total data, while high-risk states such as panic and severe mental stress that need the most attention account for a small proportion. If no targeted measures are taken for data showing a long-tail distribution, deep learning methods that are highly dependent on data volume will face two major challenges. First, due to the scarcity of data, the generalization ability of high-risk state samples is insufficient, which makes it difficult for the model to learn the identifiable features of high-risk states, so the recognition performance of high-risk states cannot meet the needs. Second, because the amount of data in the head class is much larger than that in the tail class, the training of the model will be dominated by the head class, resulting in the effective learning of the tail class being overwhelmed by the head class, and ultimately the accuracy in the test set and actual application scenarios is much lower than that in the training set. Therefore, it is necessary to study a method for identifying pilots' high-risk psychological states based on ECG data enhancement to overcome the two major challenges of scarcity and insufficient proportion of high-risk psychological state data. Summary of the invention
[0006] The purpose of the invention is to provide a method for identifying high-risk psychological states of pilots based on ECG data enhancement in response to the deficiencies in the prior art. The method divides ECG samples into high-information segments and low-information segments according to psychological risk relevance, and performs data enhancement by combining high-risk factor samples with different low-information segments to achieve rebalancing of training samples.
[0007] The technical solution adopted by the present invention is:
[0008] A method for identifying a pilot's high-risk psychological state based on electrocardiogram data enhancement includes the following steps:
[0009] Step S1: Train the low-information sample segment extraction model f on the long-tail distribution ECG dataset D h That is, based on the preset first deep network model based on the residual network, the model is trained on the long-tail distribution ECG dataset D to obtain the low-information sample segment extraction model f h ;
[0010] Among them, the head class of the long-tail distribution ECG dataset D is low-risk psychological state, and the tail class is high-risk psychological state;
[0011] Low-information sample segment extraction model f h A deep network model whose backbone network is a residual network (ResNet);
[0012] Step S2: For all head samples in the ECG dataset D, the model f is used through feature visualization technology h Extract each segment head class sample x h The sample low information segment is stored in the low information sample queue Q;
[0013] Step S3: In the part outside the peak protection segment of all tail class samples in the ECG dataset D, the low information sample segment in the low information sample queue Q is used to realize data enhancement of the tail class signal samples, and a re-balanced ECG dataset D' is obtained:
[0014] Step S4: training the pilot risk psychological state recognition model f on the re-equalized ECG dataset D', wherein the pilot risk psychological state recognition model f is a deep network model whose backbone network is ResNet; that is, based on the preset second residual network-based deep network model, the model is trained on the re-equalized ECG dataset D' to obtain the pilot risk psychological state recognition model f;
[0015] Step S5: input the pilot's real-time electrocardiogram data into the trained pilot risk psychological state recognition model f to obtain the pilot's risk psychological state recognition result.
[0016] Furthermore, in step S1, a low-information sample segment extraction model f is trained h When , the loss function used is the cross entropy loss function.
[0017] Furthermore, in step S2, the sample low information segment is a continuous segment with a sample point activation value of 0 obtained by feature visualization technology.
[0018] Furthermore, in step S3, the peak protection segment of the tail class samples is obtained by using a peak detection algorithm.
[0019] Furthermore, in step S3, the peak detection algorithm is used to obtain the peak protection segment of the tail class sample, specifically:
[0020] The peak detection algorithm is used to determine the position r of the R peak (i.e., the R wave peak) of the ECG signal in the sample, and then the preset threshold is used to (The preferred upper limit is 20, that is, 20 is taken as the reference value, and a certain upper and lower deviation is allowed) Determine the protection segment of the tail class sample:
[0021] Furthermore, in step S3, the data enhancement of the tail signal samples is realized by using the low information sample segments in the low information sample queue Q as follows:
[0022] Extract a low information sample from the low information sample queue Q and will Paste to the tail class sample xt Except for the protection segment I, the new signal sample after data enhancement is obtained Will The ECG data set D is added to obtain the re-balanced ECG data set D'.
[0023] Furthermore, in step S4, when the pilot risk psychological state recognition model f is obtained through training, the loss function adopted is the cross entropy loss function.
[0024] Furthermore, low-risk psychological states include: calm emotions, low stress levels and low fatigue levels; high-risk psychological states include: panic emotions, extreme resistance, high stress levels and high fatigue levels.
[0025] The technical solution provided by the present invention brings at least the following beneficial effects:
[0026] The present invention can obtain a rebalanced ECG data set by performing data enhancement on sample data when the psychological risk data is in an unbalanced sample condition where the low-risk psychological state data is sufficient and the high-risk psychological state data is insufficient, and further realize the effective identification of psychological risk states, especially high-risk psychological states. BRIEF DESCRIPTION OF THE DRAWINGS
[0027] In order to more clearly illustrate the technical solutions in the embodiments of the present invention, the drawings required for use in the description of the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative work.
[0028] Figure 1 A process diagram of a method for identifying a pilot's high-risk psychological state based on electrocardiogram data enhancement provided by an embodiment of the present invention. DETAILED DESCRIPTION
[0029] To make the purpose, technical solutions and advantages of the embodiments of the present invention clearer, the technical solutions in the embodiments of the present invention will be described in detail and completely in conjunction with the drawings in the implementation of the present invention. Obviously, the described embodiments are only a part of the embodiments of the present application, rather than all of the embodiments. Generally, the components of the embodiments of the present invention described and shown in the drawings can be arranged and designed using different configurations. Therefore, the following detailed description of the embodiments of the present invention provided in the drawings is not intended to limit the scope of the present application for protection, but merely represents selected embodiments of the present invention.
[0030] The embodiment of the present invention provides a method for identifying the high-risk psychological state of pilots based on ECG data enhancement. The method first trains a conventional recognition model on all data, and then uses feature visualization technology to extract low-information sample segments of head-type samples in the unbalanced data set through the model, and splices it with the effective ECG information of the tail-type signal to achieve data enhancement of high-risk psychological state samples, and then achieve rebalancing of psychological risk data. Finally, a psychological risk identification model is trained on the rebalanced data set to ensure that the psychological state of pilots in actual flight warnings and simulated flights for training can be detected, recorded, and analyzed. The method proposed in the embodiment of the present invention is targeted at the imbalanced scenario of psychological risk samples caused by insufficient data on the high-risk psychological state of pilots, so that high-risk psychological states can be accurately captured and effectively identified, ensuring the flight safety of pilots during flight.
[0031] As a possible implementation, see Figure 1 The specific implementation steps of the method for identifying a pilot's high-risk psychological state based on ECG data enhancement provided by the embodiment of the present invention include:
[0032] Step S1: Train the low-information sample segment extraction model f on the long-tail distribution ECG dataset D h :
[0033] Among them, the head class of the ECG dataset D with long-tail distribution has sufficient data volume, which is a low-risk psychological state, and the tail class with scarce data volume is a high-risk psychological state;
[0034] In this embodiment, the ECG data set D is an ECG data set of psychological risk factors of pilots collected based on a simulated flight training program for flight trainees, wherein a low-risk psychological state is calm, stress-free, and fatigue-free, and a high-risk psychological state is panic, extreme resistance, extremely high stress, and extremely high fatigue;
[0035] Low-information sample segment extraction model f h The backbone network is a deep neural network with ResNet, preferably, a low-information sample segment extraction model f h The specific structure is ResNet-18;
[0036] Specifically, the model f is trained by cross entropy loss h , where the cross entropy loss is: Where n represents the number of samples involved in training, I(i=y i ) represents an exponential function. If the predicted result is consistent with the true label, the function value is 1, otherwise it is 0; f i It is the signal feature of the sample extracted by the feature extraction network, and p() represents the classification probability of the current task output by the model.
[0037] Step S2: For all head class samples in the ECG dataset D, use feature visualization technology to use f h Extract each segment head class sample x h The sample low information segment is stored in the low information sample queue Q:
[0038] Among them, feature visualization technologies include Grad-CAM, Grad-CAM++, etc., and Grad-CAM++ is usually used; preferably, Grad-CAM++ is used as the feature visualization technology in this embodiment; the sample low information segment is a continuous segment with a sample point activation value of 0 obtained by the feature visualization technology;
[0039] Specifically, feature visualization technology can be used to h Get a head class sample x h The activation value of each sample point in x corresponds to the possibility of the sample point being used as a feature for identifying psychological risk status. h The activation value is divided into two parts. The activation value of one part of the sample points is 0, and the activation value of the other part of the sample points is a floating point value in the interval of (0,1]. The larger the value, the higher the possibility that the sample point can be used as an identifiable feature of the psychological risk factor. The sample segment with a continuous activation value of 0 is regarded as a low-information sample segment, that is, this sample segment contains no or only a very small amount of information related to the psychological risk factor category;
[0040] In this embodiment, only f is selected h This step is performed for head class samples with a confidence level higher than 0.6 to ensure the reliability of the extracted psychological risk state identification features;
[0041] The low information sample queue Q is a queue storage structure of the low information sample segment.
[0042] Step S3: In the part outside the peak protection segment of all tail class samples in the ECG data set D, the low information sample segment in the low information sample queue Q is used to realize data enhancement of the tail class signal samples, and a re-balanced ECG data set D` is obtained:
[0043] The peak protection segment of the tail class sample is obtained by the peak detection algorithm. First, the peak position r of the R peak in the sample is determined by peak detection, and then the threshold is used to Determine the protection segment of the tail class samples:
[0044] In this embodiment, the threshold 20 sample points are selected, that is, for each ECG signal peak, 41 sample points including the peak are used as peak protection segments;
[0045] Extract a low information sample from the low information sample queue Q and will Paste to the tail class sample x t Except for the protection segment I. In order to adapt to the characteristics of ECG signals, since the non-protection segment of the tail class sample contains several segments in one sample, the non-protection segment is filled by copying the low-information sample and adding other low-information samples to obtain a new signal sample after data enhancement. Will Add the ECG data set D to obtain the re-balanced ECG data set D`;
[0046] Specifically, the re-balanced ECG dataset D' consists of two parts, one part is all the samples in the ECG dataset D, and the other part is the new tail class samples obtained by data enhancement.
[0047] Step S4: Train the pilot risk psychological state recognition model f on the re-equalized ECG dataset D`:
[0048] The pilot risk psychological state recognition model f is a deep learning model with ResNet as the backbone network;
[0049] In this embodiment, the pilot risk psychological state recognition model f structure is specifically ResNet-18;
[0050] Specifically, the model f is trained by cross entropy loss h , where the cross entropy loss is:
[0051] Step S5: Input the pilot's real-time ECG data x into the trained pilot risk psychological state recognition model f to obtain the pilot's risk psychological state recognition result r(x); that is, when the preset training convergence conditions are met, such as the number of training times reaches the preset upper limit, or the loss of the set model converges.
[0052] Specifically, the risk psychological state identification result r(x)=f(x).
[0053] In this embodiment, during the training process of model f, the batch size is set to 64, the training epoch is set to 100, the ADAM optimizer is used, and the learning rates of all sub-modules in the module are set to 0.001.
[0054] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit it. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the aforementioned embodiments, or make equivalent replacements for some of the technical features therein. However, these modifications or replacements do not deviate the essence of the corresponding technical solutions from the spirit and scope of the technical solutions of the embodiments of the present invention.
[0055] The above are only some embodiments of the present invention. For those skilled in the art, several modifications and improvements can be made without departing from the creative concept of the present invention, which all belong to the protection scope of the present invention.
Claims
1. A method for identifying pilots' high-risk psychological states based on ECG data enhancement, characterized in that: The following steps are involved: Step S1: Train the low-information sample segment extraction model f on the long-tail distribution ECG dataset D h ; Among them, the head class of the long-tail distribution ECG dataset D is low-risk psychological state, and the tail class is high-risk psychological state; Low-information sample segment extraction model f h A deep network model whose backbone network is a residual network; Step S2: For all head samples in the ECG dataset D, the model f is used through feature visualization technology h Extract each segment head class sample x h The sample low information segment is stored in the low information sample queue Q; Step S3: In the part outside the peak protection segment of all tail class samples in the ECG dataset D, the low information sample segment in the low information sample queue Q is used to realize data enhancement of the tail class signal samples, and a re-balanced ECG dataset D' is obtained: Step S4: training a pilot risk psychological state recognition model f on the re-equalized ECG data set D', wherein the pilot risk psychological state recognition model f is a deep network model whose backbone network is a residual network; Step S5: input the pilot's real-time electrocardiogram data into the trained pilot risk psychological state recognition model f to obtain the pilot's risk psychological state recognition result.
2. The method according to claim 1, characterized in that In step S1, a low-information sample segment extraction model f is trained h When , the loss function used is the cross entropy loss function.
3. The method according to claim 1, characterized in that In step S2, the sample low information segment is a continuous segment with a sample point activation value of 0 obtained by feature visualization technology.
4. The method according to claim 1, characterized in that In step S3, the peak protection segment of the tail class samples is obtained by using a peak detection algorithm.
5. The method according to claim 4, characterized in that In step S3, the peak detection algorithm is used to obtain the peak protection segment of the tail class samples: The peak detection algorithm is used to determine the position r of the R wave peak of the ECG signal in the sample, and then the protection segment of the tail class sample is determined by the preset threshold ζ: I = [r-ζ, r+ζ].
6. The method according to claim 5, characterized in that The threshold ζ is set to 20±△, where △ represents the allowable deviation.
7. The method according to claim 5, characterized in that In step S3, the data enhancement of the tail signal samples is realized by using the low information sample segments in the low information sample queue Q as follows: Extract a low information sample from the low information sample queue Q and will Paste to the tail class sample x t Except for the protection segment I, the new signal sample after data enhancement is obtained Will The ECG data set D is added to obtain the re-balanced ECG data set D'.
8. The method according to claim 1, characterized in that In step S4, when training to obtain the pilot risk psychological state recognition model f, the loss function used is the cross entropy loss function.
9. The method according to claim 1, characterized in that Low-risk psychological states include: calm emotions, low stress levels and low fatigue levels; high-risk psychological states include: panic, extreme resistance, high stress levels and high fatigue levels.
10. The method according to claim 1, characterized in that Low-information sample segment extraction model f h The pilot risk psychological state recognition model f and the pilot risk psychological state recognition model f both use ResNet-18.
Citation Information
Patent Citations
Electrocardiogram data compression method and decoding method based on optimum time frequency space structure code
CN101099669A
Electrocardiosignal identity recognition method based on PCA and LDA analysis and electrocardiosignal identity recognition system based on PCA and LDA analysis
CN108537100A
Electrocardiosignal joint recognition module and method based on InResNet network
CN110688942A
Emotion recognition method based on expanded high-quality electroencephalogram sample
CN112084935A
Electrocardiogram data compression method and device
CN112863653A
