Artificial intelligence-based non-contact detection method for motion fatigue state

Abstract

The application provides a kind of non-contact detection method of exercise fatigue state based on artificial intelligence, it is related to non-contact exercise fatigue detection field.The application realizes non-contact evaluation of exercise fatigue state by analyzing the visible light video of the face of user;innovatively propose knowledge-guided nonlinear factorization alignment and projection framework, first based on exercise fatigue knowledge graph, obtain the semantic concept normal vector of semantic direction representing core medical index, then use the alignment network corresponding to each core medical index to perform nonlinear alignment conversion on global visual features, project each alignment result with the corresponding semantic concept normal vector dot product, and splice each projection result, effectively map high-dimensional, semantic entangled depth visual features into low-dimensional, structured physiological field knowledge concepts;and the dynamic factorization hidden Markov model provided can deeply understand the cumulative effect of exercise fatigue and capture its dynamic evolution law.

Description

Technical Field

[0001] This invention relates to the field of non-contact motion fatigue detection, specifically to a non-contact method for detecting motion fatigue state based on artificial intelligence. Background Technology

[0002] Exercise fatigue is a physiological parameter and an important medical indicator used to reflect the functional state of the human body. How to conduct non-contact exercise fatigue detection has become a hot topic.

[0003] With the widespread application of deep learning in visual analytics, related technologies have enabled the extraction of high-dimensional and rich bio-visual features from visible light facial videos for detecting motion fatigue. Furthermore, to further improve the data efficiency, generalization ability, and interpretability of deep learning models, these technologies fully utilize various prior knowledge domains in the target field to compensate for the shortcomings of traditional deep learning. For example, the paper (Feng Wei, Jucheng Yang, Yuan Wang, Liang Lin, Haibin Zhang, Prior knowledge-guided multi-information graph convolutional network for driver drowsiness detection, Expert Systems with Applications, Volume 275, 2025, 127028, ISSN 0957-4174.) proposes a prior knowledge-guided multi-information graph convolutional network (MIGCN) to address the driver drowsiness detection problem. Compared to CNN-based driver drowsiness detection methods, the structured MIGCN can effectively learn spatial facial features, enhancing feature representation.

[0004] However, the above-mentioned technical solutions have not effectively connected the black-box, high-dimensional depth visual features with low-dimensional, structured domain knowledge concepts with clear physiological significance, which greatly limits the application potential of deep learning models in sports fatigue assessment. Summary of the Invention

[0005] (a) Technical problems to be solved

[0006] To address the shortcomings of existing technologies, this invention provides a non-contact detection method for motion fatigue state based on artificial intelligence, solving the technical problem of how to connect black-box, high-dimensional depth visual features with low-dimensional, structured domain knowledge concepts with clear physiological significance.

[0007] (II) Technical Solution

[0008] To achieve the above objectives, the present invention provides the following technical solution:

[0009] A non-contact detection method for exercise fatigue state based on artificial intelligence includes:

[0010] The system collects and slices visible light videos of the user's face in a resting state, embeds and processes each video slice into a corresponding visual representation, and aggregates them to obtain a single global visual representation.

[0011] Based on the constructed knowledge graph of exercise fatigue, several semantic concept normal vectors are obtained to represent the semantic direction of different core medical indicators used to assess exercise fatigue; wherein the core medical indicators include at least the following: increased resting heart rate, decreased heart rate variability, anxiety, depression, irritability, and drowsiness.

[0012] Based on the current global visual representation, a nonlinear alignment transformation is performed using an alignment network that corresponds one-to-one with each of the core medical indicators; each alignment result is projected onto the corresponding semantic concept normal vector using a dot product; and each projection result is stitched together to obtain the current interpretable medical indicator sequence.

[0013] Based on the current interpretable medical indicator sequence, a dynamic factorized hidden Markov model is used to assess the user's exercise fatigue status.

[0014] Preferably, the process of constructing the exercise fatigue knowledge graph includes:

[0015] The causes and symptoms of exercise fatigue were statistically analyzed; wherein the symptoms included at least the core medical indicators.

[0016] Using any of the aforementioned causes as the head node and the aforementioned exercise fatigue as the tail node, or using the aforementioned exercise fatigue as the head node and any of the aforementioned symptoms as the tail node, construct a first triplet (head node-relationship-tail node) to obtain an initial exercise fatigue knowledge graph;

[0017] Based on the initial knowledge graph of exercise fatigue, a second triplet (head node-relationship-tail node) is constructed to represent the pairwise relationships between the core medical indicators, thus obtaining the final knowledge graph of exercise fatigue.

[0018] Preferably, the RotatE knowledge graph embedding method is used to learn each of the core medical indicators in the exercise fatigue knowledge graph into corresponding semantic concept normal vectors.

[0019] Preferably, the dynamic factorized hidden Markov model includes a historical encoder, a transition network, a emission network, and a hidden Markov model; the step of assessing the user's exercise fatigue state using the dynamic factorized hidden Markov model based on the current interpretable medical indicator sequence includes:

[0020] Based on the historical information of the current interpretable medical indicator sequence, a historical context vector is generated using the historical encoder; wherein the historical information refers to a deep time series sequence composed of all historical interpretable medical indicator sequences up to the current time step.

[0021] Based on the historical context vector, the transfer network is used to map it into a dynamic transfer probability matrix;

[0022] Based on the historical context vector, the latent variable fatigue state is encoded, and using the emission network that corresponds one-to-one with each of the core medical indicators, the mean and standard deviation of the one-dimensional Gaussian distribution of each core medical indicator under different latent variable fatigue states are obtained; wherein the latent variable fatigue state corresponds one-to-one with the preset exercise fatigue type.

[0023] Based on the dynamic transition probability matrix, and combined with the mean and standard deviation of the one-dimensional Gaussian distribution of each core medical indicator under different latent variable fatigue states, the joint emission probability of the current interpretable medical indicator sequence observed under different latent variable fatigue states is calculated using the hidden Markov model.

[0024] Determine the maximum joint launch probability and its corresponding latent variable fatigue state, and use the motion fatigue type corresponding to the latent variable fatigue state as the user's motion fatigue state assessment result.

[0025] Preferably, a sequence-level conditional random field loss is used as the core training objective, and the conditional random field loss is expressed as:

[0026]

[0027] Where the subscript CRF indicates a conditional random field; log is the logarithmic function; Let F' represent the set of all possible latent variable fatigue state sequences; let exp be an exponential function; let score be the nonnormalized logarithmic score of a path (i.e., a hypothetical latent variable fatigue state sequence of the same length as a given sequence of interpretable medical indicators); M = [m1, ..., m T ] represents a given sequence of interpretable medical indicators, where T represents the sequence length; F true Represents the actual fatigue state; (F,M) represents any path; logP(f1|h1) represents the log probability of the initial latent variable fatigue state f1, depending on the initial historical context vector h1; logP(f t |f t-1 ,h t ) indicates that it depends on the historical context vector h t From the latent variable fatigue state ft-1 to f t Logarithmic transition probability; log P(m t |f t ,h t ) indicates that it depends on the historical context vector h t In the latent variable fatigue state f t The following emission revealed an interpretable medical indicator sequence m. t The logarithmic emission probability.

[0028] Preferably, the TimeSformer model is used to embed video slices into visual representations.

[0029] Preferably, a nonlinear multilayer perceptron is selected as the alignment network.

[0030] Preferably, a multi-layer recurrent neural network is selected as the history encoder.

[0031] Preferably, a nonlinear multilayer perceptron is selected as the transfer network.

[0032] Preferably, a multilayer sensor is selected as the transmission network.

[0033] A non-contact detection system for motion fatigue state based on artificial intelligence, comprising:

[0034] The data processing module is used to collect visible light video of the user's face in a resting state and slice it, embed and process each video slice into a corresponding visual representation, and aggregate to obtain a single global visual representation.

[0035] The graph embedding module is used to obtain several semantic concept normal vectors based on the constructed exercise fatigue knowledge graph to represent the semantic direction of different core medical indicators used to assess exercise fatigue; wherein the core medical indicators include at least increased resting heart rate, decreased heart rate variability, anxiety, depression, irritability and drowsiness;

[0036] The feature extraction module is used to perform non-linear alignment transformation based on the current global visual representation using an alignment network that corresponds one-to-one with each of the core medical indicators; perform dot product projection of each alignment result with the corresponding semantic concept normal vector; and stitch together each projection result to obtain the current interpretable medical indicator sequence.

[0037] The status assessment module is used to assess the user's exercise fatigue status based on the current interpretable medical indicator sequence using a dynamic factorized hidden Markov model.

[0038] A storage medium storing a computer program for non-contact detection of motion fatigue state based on artificial intelligence, wherein the computer program causes a computer to execute the non-contact detection method of motion fatigue state as described above.

[0039] An electronic device, comprising:

[0040] One or more processors; a memory; and one or more programs, wherein the one or more programs are stored in the memory and configured to be executed by the one or more processors, the programs including methods for performing non-contact detection of motion fatigue states as described above.

[0041] (III) Beneficial Effects

[0042] This invention provides a non-contact detection method for motion fatigue state based on artificial intelligence. Compared with existing technologies, it has the following advantages:

[0043] This invention first processes a user's resting facial visible light video to obtain a single global visual representation. Second, using a knowledge-guided nonlinear factorized alignment and projection framework, it first obtains semantic concept normal vectors representing the semantic directions of core medical indicators based on a motion fatigue knowledge graph. Then, it uses an alignment network corresponding one-to-one with each core medical indicator to perform nonlinear alignment transformation on the global visual features, projecting each alignment result with the corresponding semantic concept normal vector using a dot product, and then concatenating each projection result. Finally, a dynamic factorized hidden Markov model is used to process the concatenated interpretable medical indicator sequence to assess the user's motion fatigue state. This invention integrates the powerful capabilities of deep learning in complex visual representation with the unique advantages of motion fatigue knowledge in semantic interpretation, greatly improving the accuracy of assessing the user's motion fatigue state. Attached Figure Description

[0044] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0045] Figure 1 A block diagram illustrating a non-contact detection method for motion fatigue state based on artificial intelligence, provided in an embodiment of the present invention.

[0046] Figure 2 The flowchart illustrates a non-contact detection method for motion fatigue state based on artificial intelligence, as provided in an embodiment of the present invention. Detailed Implementation

[0047] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions in the embodiments of the present invention are described clearly and completely. Obviously, the described embodiments are only some embodiments of the present invention, 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.

[0048] This application provides a non-contact detection method for motion fatigue state based on artificial intelligence, which solves the technical problem of how to connect black-box, high-dimensional depth visual features with low-dimensional, structured domain knowledge concepts with clear physiological significance.

[0049] The technical solution in this application is to solve the above-mentioned technical problems, and the general idea is as follows:

[0050] 1. Non-invasive, convenient, real-time, and low-cost exercise fatigue assessment:

[0051] In this embodiment of the invention, the detection state is the user's resting state, focusing on chronic exercise fatigue caused by the cumulative effects of high-intensity training or long-term exercise. The solution only needs to analyze the user's facial visible light video, without any wearable devices or invasive detection, providing a technical foundation for more comprehensive and humanized exercise fatigue management.

[0052] 2. Knowledge-Guided Non-linear Factored Alignment and Projection (KG-NFAP) Framework:

[0053] The KG-NFAP framework designed in this invention is a key innovation connecting black-box deep visual features with interpretable domain knowledge. This framework precisely transforms the high-dimensional, entangled global visual representations extracted by deep learning models (such as the TimeSformer model) into six independent, non-linear alignment networks guided by domain knowledge (KGE concept normal vectors), resulting in six interpretable, physiologically significant core indicators of exercise fatigue. This bridges the semantic gap in deep features and constructs a clinically meaningful set of high signal-to-noise ratio core features for exercise fatigue.

[0054] 3. Interpretable dynamic temporal modeling and transition matrix:

[0055] This invention constructs a Dynamic Factorized Hidden Markov Model (DF-HMM) as the core time series evaluation model, which introduces the following key improvements on the basis of traditional HMM.

[0056] Factorized emission mechanism: For six decoupled physiological indicators, six independent, small emission networks are designed to accurately model the probability distribution of each indicator under different fatigue states, and reasonably assume that the indicators are conditionally independent under given states.

[0057] Dynamic transition mechanism: A historical encoder is introduced so that its state transition probability can be dynamically adjusted according to historical and current observations, capturing the complex laws of the evolution of motion fatigue state over time.

[0058] To better understand the above technical solutions, the following will provide a detailed explanation of the technical solutions in conjunction with the accompanying drawings and specific implementation methods.

[0059] Example 1:

[0060] like Figure 1 As shown, this embodiment of the invention provides a non-contact detection method for motion fatigue state based on artificial intelligence, including:

[0061] S1. Collect visible light video of the user's face in a resting state and slice it. Embed and process each video slice into a corresponding visual representation to aggregate and obtain a single global visual representation.

[0062] S2. Based on the constructed knowledge graph of exercise fatigue, obtain several semantic concept normal vectors to represent the semantic direction of different core medical indicators used to assess exercise fatigue; wherein the core medical indicators include at least increased resting heart rate, decreased heart rate variability, anxiety, depression, irritability and drowsiness;

[0063] S3. Based on the current global visual representation, a non-linear alignment transformation is performed using an alignment network that corresponds one-to-one with each of the core medical indicators; each alignment result is projected onto the corresponding semantic concept normal vector using a dot product; and each projection result is stitched together to obtain the current interpretable medical indicator sequence.

[0064] S4. Based on the current interpretable medical indicator sequence, use a dynamic factorized hidden Markov model to assess the user's exercise fatigue status.

[0065] The embodiments of the present invention combine the powerful capabilities of deep learning in complex visual representation with the unique advantages of knowledge in the field of sports fatigue in semantic interpretation, greatly improving the accuracy of the assessment of users' sports fatigue status.

[0066] like Figure 2 As shown, Figure 2 A flowchart of a non-contact detection method for motion fatigue state based on artificial intelligence is disclosed. The following will combine... Figure 2 The above technical solutions are described in detail below:

[0067] First, it should be clarified that the exercise fatigue that this embodiment of the invention focuses on refers to the fatigue caused by the accumulation of training load (including training volume and intensity) and possible non-training stress (such as psychological, social, nutritional, sleep and other factors) that exceeds the individual's ability to adapt and recover.

[0068] Accordingly, this invention proposes an innovative two-stage exercise fatigue assessment technology, aiming to achieve a non-invasive, convenient, real-time, and low-cost assessment of cumulative / chronic exercise fatigue states through deep analysis of users' facial visible light video data. This solution integrates the powerful capabilities of deep learning in complex visual representations with the unique advantages of exercise science knowledge in semantic interpretation. Based on interpretable features, it constructs a Dynamic Factorized Hidden Markov Model (DF-HMM) to model the dynamic transition patterns of exercise fatigue states.

[0069] like Figure 2 As shown, the first stage is knowledge-guided feature extraction based on visual features (video data). The first stage involves interpretable medical indicator sequences, corresponding to S1 to S3; the second stage involves dynamic time-series modeling and assessment of exercise fatigue state (interpretable medical indicator sequences). (Exercise fatigue state), corresponding to S4. See below for specific steps:

[0070] In step S1, a visible light video of the user's face in a resting state is acquired and sliced. Each video slice is embedded and processed into a corresponding visual representation, so as to aggregate and obtain a single global visual representation.

[0071] For example, this step first uses a camera to capture 2 minutes of visible light video of the user's face. As mentioned above, since this embodiment of the invention targets the assessment of cumulative / chronic exercise fatigue, the captured video is limited to the user's face in a resting state; for example, the detection time point can be set to the day after exercise. Furthermore, the data acquisition requirements must meet the following conditions: normal lighting conditions, the user maintaining a natural expression during video capture, avoiding significant head movements, and a distance of approximately 40-60cm between the face and the camera.

[0072] Next, for each visible light video of a face, this step slices it into consecutive video segments of 4 seconds each. The aforementioned segmentation strategy ensures that the length of each video segment meets the input requirements of the pre-trained deep model and effectively captures the local spatiotemporal information of the video.

[0073] Finally, this step uses the TimeSformer model to embed each video slice into a 768-dimensional high-dimensional visual representation vector. Let v represent the set of real numbers. Furthermore, to obtain comprehensive features representing a single 2-minute video from a given day, v represents the values ​​of all slices of that video.subclip Vectors are aggregated using average pooling to obtain a unified global visual representation for a single day.

[0074] In step S2, based on the constructed knowledge graph of exercise fatigue, several semantic concept normal vectors are obtained to represent the semantic direction of different core medical indicators used to assess exercise fatigue; wherein the core medical indicators include at least increased resting heart rate, decreased heart rate variability, anxiety, depression, irritability and drowsiness.

[0075] For example, the construction process of the exercise fatigue knowledge graph includes:

[0076] S10. Analyze the causes and symptoms of exercise fatigue; wherein the symptoms include at least the core medical indicators, namely at least the above-mentioned elevated resting heart rate, decreased heart rate variability, anxiety, depression, irritability and drowsiness.

[0077] S20. Using any of the stated causes as the head node and the stated exercise fatigue as the tail node, or using the stated exercise fatigue as the head node and any of the stated symptoms as the tail node, construct a first triplet (head node-relationship-tail node) to obtain an initial exercise fatigue knowledge graph.

[0078] S30. Based on the initial exercise fatigue knowledge graph, a second triplet (head node-relationship-tail node) is constructed to represent the pairwise relationships between the core medical indicators to obtain the final exercise fatigue knowledge graph.

[0079] Specifically:

[0080] This step first constructs a domain knowledge graph centered on "exercise fatigue," encompassing its causes and symptoms (i.e., the initial exercise fatigue knowledge graph). This domain knowledge graph is stored in the form of first triples (head node-relationship-tail node). For example, first triples for cause types include: (high-intensity exercise - leads to exercise fatigue), (insufficient recovery - leads to exercise fatigue), and (training stress - leads to exercise fatigue); similarly, first triples for symptom types include: (exercise fatigue - leads to increased resting heart rate), (exercise fatigue - leads to decreased heart rate variability), and (exercise fatigue - leads to drowsiness). Furthermore, in this domain knowledge graph, increased resting heart rate, decreased heart rate variability, anxiety, depression, irritability, and drowsiness are selected as the six core medical indicators i (i = 1, ..., M, where M = 6) for assessing exercise fatigue. The reason for choosing these six core medical indicators is that they are the core indicators for assessing exercise fatigue, and the facial visible light video obtained in step S1 also contains information on the six core medical indicators.

[0081] To further clarify the relative positions and relationships of the core medical indicators in the knowledge graph of the above-mentioned fields, second ternary pairs are added among the six core medical indicators to establish their relationships with each other. For example, definitions are made for (anxiety - a type of - mood disorder), (depression - a type of - mood disorder), (irritability - a type of - mood disorder), and (anxiety - comorbidity - depression), etc., to obtain the final knowledge graph of exercise fatigue.

[0082] Continuing with the example above, this step then uses the knowledge graph embedding method RotatE to learn each of the core medical indicators in the exercise fatigue knowledge graph into corresponding semantic concept normal vectors.

[0083] Specifically:

[0084] Based on the constructed knowledge graph of exercise fatigue, the knowledge graph embedding method RotatE is used to learn a 128-dimensional semantic concept normal vector n for each entity in the graph. i (i = 1, ..., 6). The core of RotatE lies in modeling entities and relations in a complex vector space. Unlike previous methods that treated relations as "translations" (such as the TransE model), RotatE models each relation as a rotation. RotatE was chosen because it can transform logical relations in knowledge graphs (such as "is a kind of," "leads to") into rotation operations in vector space. This allows for a deeper capture and modeling of the complex symmetric and hierarchical dependencies between core medical indicators, ultimately providing the most accurate semantic direction for these indicators. i It is a precise representation of core medical indicators in the semantic space of exercise fatigue, representing the direction and representation of the corresponding node in the semantic space of exercise fatigue. This step provides a reliable and highly discriminative "knowledge benchmark" for subsequent feature purification.

[0085] In step S3, based on the current global visual representation, a nonlinear alignment transformation is performed using an alignment network that corresponds one-to-one with each of the core medical indicators; each alignment result is projected onto the corresponding semantic concept normal vector using a dot product; and each projection result is stitched together to obtain the current interpretable medical indicator sequence.

[0086] Continuing with the example above, this step will be based on the single global visual representation extracted by TimeSformer. The data are fed into six independent feature extraction layers in parallel. Each feature extraction layer consists of two parts: nonlinear alignment transformation and knowledge projection.

[0087] Specifically:

[0088] ① Nonlinear alignment transformation: For each of the six core medical indicators i (i = 1, ..., 6), an independent nonlinear multilayer perceptron (MLP) is used as its alignment network. v global It is simultaneously fed into 6 independent MLPs. Each It can learn an optimal nonlinear mapping for the 768-dimensional visual vector v. global Align to fit n i The 128-dimensional semantic subspace for knowledge concept measurement yields... This alignment process is supervised by the downstream task loss and guided by the corresponding knowledge concept n. i Guided by.

[0089] ② Knowledge projection: In a space already aligned to a unified 128-dimensional semantic subspace Above, using the i-th knowledge normal vector Perform a dot product projection on it and calculate its projection length. This projection length It indicates the degree to which current visual evidence matches the corresponding medical concept.

[0090] ③ Construction of interpretable medical indicator sequences: Ultimately, these 6 independent projection lengths When pieced together, they form a 6-dimensional, interpretable medical indicator vector with clear physiological significance.

[0091] In step S4, the user's exercise fatigue status is assessed using a dynamic factorized hidden Markov model based on the current interpretable medical indicator sequence.

[0092] This step corresponds to Figure 2 The second stage, which aims to utilize (steps S1 to S3 corresponding to...) Figure 2 The low-dimensional, interpretable medical indicator sequence m extracted in the first stage t This stage involves a comprehensive assessment of exercise fatigue status using both historical and current information. The core of this phase is the use of a Dynamic Factorized Hidden Markov Model (DF-HMM) as the time-series assessment model, employing an end-to-end supervised training paradigm. The DF-HMM model can simultaneously learn the dynamic transition patterns of fatigue status and the distribution of factorized core medical indicators, thus achieving a balance between high performance and high interpretability.

[0093] like Figure 2 As shown, the dynamic factorized hidden Markov model includes a history encoder, a transition network, a launch network, and a hidden Markov model.

[0094] Based on this, for example, the assessment of a user's exercise fatigue state using a dynamic factorized hidden Markov model based on the current interpretable medical indicator sequence includes:

[0095] S100. Based on the historical information of the current interpretable medical indicator sequence, a historical context vector is generated using the historical encoder; wherein the historical information refers to a deep time series sequence composed of all historical interpretable medical indicator sequences up to the current time step.

[0096] The 6-dimensional interpretable medical indicator vector sequence output from the first stage As input, the sequence is fed into a history encoder, here a multi-layer recurrent neural network GRU is chosen to capture sequence m. t Complex temporal dependencies and cumulative effects.

[0097] At each time step t, GRU will adjust according to m t Historical information of the sequence (m1 to m) t-1 This generates a high-dimensional historical context vector. Represented as:

[0098] h t =GRU(m 1:t-1 )

[0099] Where, m 1:t-1 For the observation sequence from time step 1 to t-1, h t This is the historical context vector generated at time step t, which encodes all key timing information up to the current time step t.

[0100] S200. Based on the historical context vector, the transfer network is used to map it into a dynamic transfer probability matrix.

[0101] An MLP is selected as the transfer network, and the historical context vector h is used. t Mapped to a dynamic transition probability matrix Where N is the number of fatigue state types, which can be 3 here, meaning the preset fatigue state types include healthy, moderate fatigue, and severe fatigue.

[0102] A t =softmax(MLP) trans (h t ))

[0103] Among them, MLP trans The softmax() function is used to transfer the network function and normalize the output to ensure that the sum of the probabilities of each row is 1.

[0104] Understandably, unlike the fixed transition matrix A in a traditional Hidden Markov Model (HMM), A...t It can dynamically adjust according to historical sequences, thereby more accurately capturing the complex, context-dependent non-Markov dynamics in the evolution of motion fatigue states. This enables the model not only to determine "what state it is", but also to understand "under what historical conditions state transitions will occur".

[0105] S300. Based on the historical context vector, the latent variable fatigue state is encoded, and using the emission network that corresponds one-to-one with each of the core medical indicators, the mean and standard deviation of the one-dimensional Gaussian distribution of each core medical indicator under different latent variable fatigue states are obtained; wherein the latent variable fatigue state corresponds one-to-one with the preset exercise fatigue type.

[0106] Based on historical context vector h t Encoding the latent variable fatigue state f t ∈{1,…,N,N=3} (usually in one-hot encoded form) ) as input, where f t =1 indicates health, f t =2 indicates moderate fatigue, f t =3 indicates severe fatigue.

[0107] For each core medical indicator i (i.e. m) t The i-th dimension, Each of them designed an independent MLP as a transmission network. Each Receive status indication and historical context vector h t As input, it outputs the mean of a 1-dimensional Gaussian distribution of the index in the given state and historical context. and standard deviation Represented as:

[0108]

[0109] in, At time step t, the state is f t When, the mean and standard deviation of the Gaussian distribution of the i-th index are given, and [;] represents the vector concatenation operation.

[0110] Finally, for each fatigue state f t The model outputs the mean and standard deviation of 6 independent, 1-dimensional Gaussian distributions.

[0111] S400. Based on the dynamic transition probability matrix, and combining the mean and standard deviation of the one-dimensional Gaussian distribution of each core medical indicator under different latent variable fatigue states, the joint emission probability of the current interpretable medical indicator sequence observed under different latent variable fatigue states is calculated using the hidden Markov model.

[0112] m under different fatigue states t The joint launch probability P(m) t |f t ,h t The product of these independent Gaussian distributions is modeled as the factorization hypothesis:

[0113]

[0114] Wherein, P(m) t |f t ,h t ) is a given state f t and h t m was observed t The probability of. Representing variables It follows a mean of μ and a variance of σ. 2 The probability density function value of the Gaussian distribution.

[0115] It is important to note that traditional Hidden Markov Models (HMMs) often assume that all observed indicators follow a joint Gaussian distribution, which can easily lead to model mismatch and optimization difficulties when dealing with high-dimensional and complex features. This invention, through factorization assumptions, decomposes the joint emission probability modeling of high-dimensional features into M independent one-dimensional probability density estimation tasks, greatly simplifying the modeling process and enhancing model robustness and interpretability. The independent modeling of each indicator also better aligns with the relative independence of indicators in physiology.

[0116] S500: Determine the maximum joint launch probability and its corresponding latent variable fatigue state, and use the motion fatigue type corresponding to the latent variable fatigue state as the user's motion fatigue state assessment result.

[0117] In particular, such as Figure 2 As shown, this embodiment of the invention also employs an end-to-end supervised training paradigm, specifically:

[0118] Sequence-level Conditional Random Field (CRF) loss is used as the core training objective. This loss function aims to maximize the model's performance on a given sequence of interpretable medical indicators M = [m1, ..., m]. T Under the condition of ], predict the true fatigue state F true (F true The conditional probability P(F) for the labeltrue |M). Its logarithmic form is:

[0119]

[0120] Where log is the logarithmic function; Let F' represent the set of all possible latent variable fatigue state sequences, and let F′ represent any possible latent variable fatigue state sequence; exp is the exponential function; score represents the nonnormalized logarithmic fraction of a path; and T represents the sequence length.

[0121] The joint probability logP(F,M) of any path (F,M) is calculated as follows:

[0122]

[0123] Where log P(f1|h1) represents the log probability of the initial latent variable fatigue state f1 of the sequence, depending on the initial historical context vector h1; log P(f t |f t-1 ,h t ) indicates that it depends on the historical context vector h t From the latent variable fatigue state f t-1 to f t Logarithmic transition probability; log P(m t |f t ,h t ) indicates that it depends on the historical context vector h t In the latent variable fatigue state f t The following emission revealed an interpretable medical indicator sequence m. t The logarithmic emission probability.

[0124] After the model is trained, it can be used to assess the user's current exercise fatigue state in real time, for the input m t The sequence, based on the learned dynamic transition rules and factorized emission probabilities, assesses the state of motion fatigue.

[0125] Thus, this embodiment of the invention completes the entire process of a non-contact detection method for motion fatigue state based on artificial intelligence.

[0126] Example 2:

[0127] This invention provides a non-contact detection system for motion fatigue state based on artificial intelligence, comprising:

[0128] The data processing module is used to collect visible light video of the user's face in a resting state and slice it, embed and process each video slice into a corresponding visual representation, and aggregate to obtain a single global visual representation.

[0129] The graph embedding module is used to obtain several semantic concept normal vectors based on the constructed exercise fatigue knowledge graph to represent the semantic direction of different core medical indicators used to assess exercise fatigue; wherein the core medical indicators include at least increased resting heart rate, decreased heart rate variability, anxiety, depression, irritability and drowsiness;

[0130] The feature extraction module is used to perform non-linear alignment transformation based on the current global visual representation using an alignment network that corresponds one-to-one with each of the core medical indicators; perform dot product projection of each alignment result with the corresponding semantic concept normal vector; and stitch together each projection result to obtain the current interpretable medical indicator sequence.

[0131] The status assessment module is used to assess the user's exercise fatigue status based on the current interpretable medical indicator sequence using a dynamic factorized hidden Markov model.

[0132] Example 3:

[0133] This invention provides a storage medium storing a computer program for non-contact detection of motion fatigue state based on artificial intelligence, wherein the computer program causes a computer to execute the non-contact detection method of motion fatigue state as described in Embodiment 1.

[0134] Example 4:

[0135] This invention provides an electronic device, comprising:

[0136] One or more processors; a memory; and one or more programs, wherein the one or more programs are stored in the memory and configured to be executed by the one or more processors, the programs including methods for performing non-contact detection of motion fatigue states as described in Example 1.

[0137] It is understood that the non-contact detection system, storage medium and electronic device based on artificial intelligence for motion fatigue state provided in the embodiments of the present invention correspond to the non-contact detection method based on artificial intelligence for motion fatigue state provided in the embodiments of the present invention. The explanation, examples and beneficial effects of the relevant contents can be referred to the corresponding parts of the method, and will not be repeated here.

[0138] In summary, compared with existing technologies, it has the following beneficial effects:

[0139] 1. The embodiments of the present invention realize non-invasive, convenient, real-time and low-cost real-time continuous monitoring of exercise fatigue.

[0140] 2. The embodiments of the present invention provide interpretable core features with high signal-to-noise ratio and strong discriminative power.

[0141] 3. The embodiments of the present invention construct a time series evaluation model that is deeply interpretable and highly accurate.

[0142] 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 a process, method, article, or apparatus. Without further limitations, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes said element.

[0143] The above embodiments are only used to illustrate the technical solutions of the present invention, and are not intended to limit it. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A non-contact detection method for motion fatigue state based on artificial intelligence, characterized in that, include: The system collects and slices visible light videos of the user's face in a resting state, embeds and processes each video slice into a corresponding visual representation, and aggregates them to obtain a single global visual representation. Based on the constructed knowledge graph of exercise fatigue, semantic concept normal vectors are obtained to represent the semantic direction of different core medical indicators used to assess exercise fatigue; wherein the core medical indicators include at least increased resting heart rate, decreased heart rate variability, anxiety, depression, irritability and drowsiness; using the knowledge graph embedding method RotatE, each of the core medical indicators in the exercise fatigue knowledge graph is learned into a corresponding semantic concept normal vector. Based on the current global visual representation, a nonlinear alignment transformation is performed using an alignment network that corresponds one-to-one with each of the core medical indicators. Project each alignment result onto the corresponding semantic concept normal vector using a dot product. And by stitching together each projection result, the current interpretable medical indicator sequence is obtained; wherein a nonlinear multilayer perceptron is selected as the alignment network; Based on the current interpretable medical indicator sequence, a dynamic factorized hidden Markov model is used to assess the user's exercise fatigue status.

2. The non-contact detection method for exercise fatigue state as described in claim 1, characterized in that, The process of constructing the exercise fatigue knowledge graph includes: The causes and symptoms of exercise fatigue were statistically analyzed; wherein the symptoms included at least the core medical indicators. Using any of the aforementioned causes as the head node and the aforementioned exercise fatigue as the tail node, or using the aforementioned exercise fatigue as the head node and any of the aforementioned symptoms as the tail node, construct a first triplet (head node-relationship-tail node) to obtain an initial exercise fatigue knowledge graph; Based on the initial knowledge graph of exercise fatigue, a second triplet (head node-relationship-tail node) is constructed to represent the pairwise relationships between the core medical indicators to obtain the final knowledge graph of exercise fatigue.

3. The non-contact detection method for exercise fatigue state as described in claim 1, characterized in that, The dynamic factorized hidden Markov model includes a historical encoder, a transition network, a firing network, and a hidden Markov model; the assessment of a user's exercise fatigue state using the dynamic factorized hidden Markov model based on the current interpretable medical indicator sequence includes: Based on the historical information of the current interpretable medical indicator sequence, a historical context vector is generated using the historical encoder; wherein the historical information refers to a deep time series sequence composed of all historical interpretable medical indicator sequences up to the current time step. Based on the historical context vector, the transfer network is used to map it into a dynamic transfer probability matrix; Based on the historical context vector, the latent variable fatigue state is encoded, and using the emission network that corresponds one-to-one with each of the core medical indicators, the mean and standard deviation of the one-dimensional Gaussian distribution of each core medical indicator under different latent variable fatigue states are obtained; wherein the latent variable fatigue state corresponds one-to-one with the preset exercise fatigue type. Based on the dynamic transition probability matrix, and combined with the mean and standard deviation of the one-dimensional Gaussian distribution of each core medical indicator under different latent variable fatigue states, the joint emission probability of the current interpretable medical indicator sequence observed under different latent variable fatigue states is calculated using the hidden Markov model. Determine the maximum joint launch probability and its corresponding latent variable fatigue state, and use the motion fatigue type corresponding to the latent variable fatigue state as the user's motion fatigue state assessment result.

4. The non-contact detection method for exercise fatigue state as described in claim 3, characterized in that, The sequence-level conditional random field loss is used as the core training objective, and the conditional random field loss is expressed as: Among them, subscript Represents a conditional random field; It is a logarithmic function; Represents the set of all possible hidden variable fatigue state sequences. This represents any possible sequence of latent variable fatigue states; It is an exponential function; Represents the nonnormalized logarithmic fraction calculated for a path; This represents a given sequence of interpretable medical indicators. Indicates the sequence length; This indicates a true state of fatigue. Represents any path; The representation depends on the initial historical context vector. The initial latent variable fatigue state of the sequence The logarithmic probability; The representation depends on the historical context vector. From the latent variable fatigue state arrive The logarithmic transition probability; The representation depends on the historical context vector. In the latent variable fatigue state The following emission revealed an interpretable medical indicator sequence. The logarithmic emission probability.

5. The non-contact detection method for exercise fatigue state as described in any one of claims 1 to 4, characterized in that, The TimeSformer model is used to embed video slices into visual representations.

6. The non-contact detection method for exercise fatigue state as described in claim 3 or 4, characterized in that, A multi-layer recurrent neural network was selected as the history encoder. And / or a nonlinear multilayer perceptron is selected as the transfer network; And / or a multilayer sensor may be selected as the transmission network.

7. A non-contact detection system for motion fatigue state based on artificial intelligence, characterized in that, For performing the non-contact detection method for exercise fatigue state as described in claim 1, comprising: The data processing module is used to collect visible light video of the user's face in a resting state and slice it, embed and process each video slice into a corresponding visual representation, so as to aggregate and obtain a single global visual representation. The graph embedding module is used to obtain semantic concept normal vectors based on the constructed exercise fatigue knowledge graph, so as to represent the semantic direction of different core medical indicators used to assess exercise fatigue; wherein the core medical indicators include at least increased resting heart rate, decreased heart rate variability, anxiety, depression, irritability and drowsiness; The feature extraction module is used to perform non-linear alignment transformation based on the current global visual representation using an alignment network that corresponds one-to-one with each of the core medical indicators; perform dot product projection of each alignment result with the corresponding semantic concept normal vector; and stitch together each projection result to obtain the current interpretable medical indicator sequence. The status assessment module is used to assess the user's exercise fatigue status based on the current interpretable medical indicator sequence using a dynamic factorized hidden Markov model.

8. A storage medium, characterized in that, It stores a computer program for non-contact detection of motion fatigue state based on artificial intelligence, wherein the computer program causes the computer to execute the non-contact detection method of motion fatigue state as described in any one of claims 1 to 6.

9. An electronic device, characterized in that, include: One or more processors; Memory; And one or more programs, wherein the one or more programs are stored in the memory and configured to be executed by the one or more processors, the programs including methods for performing the non-contact detection method for motion fatigue state as described in any one of claims 1 to 6.

Images