Personalized workload intelligent evaluation method based on Bayesian multi-source fusion

Abstract

The invention provides a personalized workload intelligent evaluation method based on Bayesian multi-source fusion, and relates to the field of workload evaluation. According to the method, firstly, multi-source asynchronous data are collected and processed; secondly, reasonably initializing a learnable parameter set and a workload state under the condition of lack of individual historical data through a personalized meta-prior method of hysteresis distribution modeling incorporating biochemical response, and avoiding prior deviation; thirdly, acquiring an atlas aggregation item used for representing the causal lag influence to construct a state transition model, introducing a quality adaptive mechanism to construct an observation model, mapping low-quality observation into larger effective noise to automatically reduce the weight, and reducing the interference of asynchronization and noise on fusion; and finally, posterior value and uncertainty estimation is completed in the same framework, and the current working load state is intelligently evaluated, so that the method is superior to the prior art in the aspects of accuracy, interpretation and operability, and has a cross-scene migration capability.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of work load evaluation, in particular to a personalized work load intelligent evaluation method based on Bayesian multi-source fusion. BACKGROUND

[0002] The term "work load" can be derived from the English "Training Load" or "Workload". In the context of physical activity, it refers to the degree of load caused by a task on the body, including both physical expenditure and physiological response.

[0003] In related art, there are three main types of work load evaluation methods: first, empirical models based on physiological signals, such as heart rate, heart rate variability, skin temperature, etc. Non-invasive indicators are used as the core, combined with feature weights and clustering algorithms to divide the load level. Second, stage evaluation based on biochemical indicators, such as collecting lactic acid, creatine kinase, testosterone, urine protein, etc. Detection results are combined with work cycles and trends for analysis. Third, comprehensive methods based on multi-source data fusion, by collecting multi-modal data and performing feature fusion, to present the changes in work load in real time.

[0004] However, under the condition of multi-source asynchrony and individual differences, the above methods are prone to fusion distortion due to inconsistent sampling frequencies and response time lags of different modalities, and it is difficult to reasonably initialize the model due to individual baseline and response differences, resulting in unstable evaluation results and lack of personalization. SUMMARY

[0005] (I) Technical problems solved To overcome the shortcomings of the prior art, the present application provides a personalized work load intelligent evaluation method based on Bayesian multi-source fusion, which solves the technical problems of existing methods that are prone to fusion distortion and difficult to reasonably initialize the model under the condition of multi-source asynchrony and individual differences.

[0006] (II) Technical solutions To achieve the above purpose, the present application is implemented by the following technical solutions: A personalized work load intelligent evaluation method based on Bayesian multi-source fusion, comprising: Collecting and processing multi-source asynchronous data, constructing physiological features, facial behavior features, biochemical features, working condition vectors, work situation labels and individual baseline features; wherein the physiological features and facial behavior features are derived from facial visible light video, and the biochemical features are derived from biochemical data; modeling a personalized meta-prior of a lag distribution based on the job context label and the individual baseline feature to obtain a probability vector of each lag gear of the biochemical feature relative to the workload state and a lag time; and modeling a Bayesian multi-source fusion framework and a personalized meta-prior of the workload state to correspond to initialize a learnable parameter set and the workload state; constructing a time-series causal knowledge graph based on the workload state, the physiological feature, the facial behavior feature, the biochemical feature and the working condition vector, and combining the probability vector and the lag time to obtain a graph aggregation item for representing a causal lag effect; constructing a state transition model based on the learnable parameter set, the working condition vector and the graph aggregation item, and constructing a first observation model of a video modality and a second observation model of a biochemical data modality based on the learnable parameter set and combining a quality self-adaptive mechanism; solving the prior mean and the prior uncertainty at the current time through the state transition model, and updating the posterior mean and the posterior uncertainty at the current time through the first observation model and the second observation model by using Kalman filtering to intelligently evaluate the current workload state.

[0007] Preferably, the modeling of the personalized meta-prior of the lag distribution based on the job context label and the individual baseline feature to obtain the probability vector of each lag gear of the biochemical feature relative to the workload state and the lag time comprises: concatenating the job context label and the individual baseline feature, linearly transforming and normalizing to obtain the probability vector of each lag gear of the biochemical feature relative to the workload state; based on the probability vector, and selecting a classification distribution as the personalized meta-prior of the lag distribution, obtaining the lag time of each lag gear of the biochemical feature relative to the workload state.

[0008] Preferably, the normal distribution is selected as the personalized meta-prior of the Bayesian multi-source fusion framework and the workload state, and the learnable parameter set and the workload state are initialized correspondingly, wherein: for the personalized meta-prior of the Bayesian multi-source fusion framework: concatenating the job context label and the individual baseline feature, and fitting a linear mapping parameter based on the least square method to obtain a first prior mean and a first prior covariance; based on the first prior mean and the first prior covariance, the learnable parameter set is initialized to follow a normal distribution; wherein the learnable parameter set includes a state inertia coefficient, a working condition driving weight, an observation sensitivity, and observation parameters of the first observation model and the second observation model; for the personalized meta-prior of the workload state: concatenate the work context label and individual baseline features, and obtain linear mapping parameters based on centralization to obtain a second prior mean and a second prior covariance; initialize the workload state to follow a normal distribution based on the second prior mean and the second prior covariance.

[0009] Preferably, based on the workload state, physiological features, facial behavior features, biochemical features, and working condition vectors, a time-series causal knowledge graph is constructed, and a graph aggregation term for representing causal lag effects is obtained by combining a probability vector and a lag time, including: The physiological features, facial behavior features, and working condition vectors are taken as first-type parent nodes of the workload state, and the biochemical features are taken as second-type parent nodes of the workload state. Based on the causal influence of the first-type parent nodes on the workload state, a directed edge and its weight between the first-type parent nodes and the workload state are defined. Based on the probability vector and the lag time of each lag gear of the biochemical features relative to the workload state, a directed edge and its weight between the second-type parent nodes and the workload state are defined. Based on each directed edge and its weight, and combined with a transformation function, a graph aggregation term for representing causal lag effects is aggregated and obtained.

[0010] Preferably, the state transition model is represented as: wherein, , Xt-1 and Xt represent the workload states at t-1 and t time points, respectively. represents a normal distribution. represents a state inertia coefficient. represents a working condition driving weight, and a superscript T represents transposition. represents an observation sensitivity. Xt represents a working condition vector at t time point. is a graph aggregation term at t time point. represents an inertial effect of the workload state at t-1 time point on t time point. represents a working condition driving effect. represents a causal lag effect. >0 represents a process noise variance. represents a posterior mean at t-1 time point.

[0011] Preferably, the prior mean and the prior uncertainty at the current time point are solved by the state transition model, and are represented as: wherein, Let represent the prior mean at time t; Let represent the posterior mean at time t-1; This represents the prior uncertainty at time t; This represents the posterior uncertainty at time t-1.

[0012] Preferably, the construction of a first observation model for the video modality and a second observation model for the biochemical data modality based on a learnable parameter set and a quality adaptive mechanism includes: Based on the facial visible light video, the signal-to-noise ratio quality score, occlusion rate quality score, tracking confidence quality score, and jitter amplitude quality score are calculated respectively. After weighted fusion, the video modal quality score is obtained to obtain the first quality weight of each physiological feature or facial behavior feature. Based on the deviation between the relative sampling delay and the target delay, the repeatability consistency, and the device status in the biochemical data, a biochemical modality quality score is obtained to obtain a second quality weight for each biochemical feature. Based on the observation parameters, first quality weights, and corresponding covariate features of the first observation model, the first observation model is constructed; and based on the covariate features, the observation parameters and second quality weights of the second observation model, the second observation model is constructed.

[0013] Preferably, the first observation model and the second observation model are respectively expressed as: in, These represent the workload status at time t; Represents a normal distribution; express t Time of the first i Actual observations of physiological or facial behavioral features; the superscript vid indicates the video modality. Indicates the observation parameters of the first observation model. Video modal pairs Sensitivity, covariate feature weights of video modalities, zero-point differences under the same load in video modalities, with the superscript T indicating transpose; express t Covariate characteristics of video modalities at any given time; Represents the baseline observation noise variance for the video modality; express t Time of the first i The first quality weight of each physiological or facial behavioral feature; express Time of the first i The actual observations of a biochemical characteristic, with the superscript "bio" indicating a biochemical mode. Indicates the first i The j-th lag level of a biochemical characteristic is... The lag time; Indicates the observation parameters of the second observation model. Representing biochemical modal pairs Sensitivity, covariate feature weights of biochemical modes, and zero-point differences under the same load in biochemical modes; express t Covariate characteristics of biochemical modalities at time points; The baseline observation noise variance represents the biochemical modality; express Time of the first i The second quality weight of each biochemical characteristic.

[0014] Preferably, the step of updating the posterior mean and posterior uncertainty at the current time using Kalman filtering through the first observation model and the second observation model includes: Based on the first observation model and the second observation model, the corresponding Kalman gains are calculated and expressed as follows: in, Represents the time t. i Kalman gain corresponding to a physiological feature or facial behavioral feature; This represents the prior uncertainty at time t; express Time of the first i Kalman gain corresponding to each biochemical characteristic; Based on the first observation model and the second observation model, the corresponding predicted values ​​of the observations are obtained respectively, as follows: in, express t Time of the first i Observational predictions of physiological or facial behavioral features; express Time of the first i Observed and predicted values ​​of individual biochemical characteristics; Based on the Kalman gain, and combined with the difference between the actual observation and the predicted observation, the posterior mean is updated as follows: in, Let be the posterior mean at time t. Let represent the prior mean at time t; Based on the Kalman gain, the updated posterior uncertainty is expressed as: wherein, represents the posterior uncertainty at time t.

[0015] A personalized workload intelligent evaluation system based on Bayesian multi-source fusion, comprising: A collection and processing module for collecting and processing multi-source asynchronous data, constructing physiological features, facial behavior features, biochemical features, working condition vectors, job situation labels and individual baseline features; wherein the physiological features and facial behavior features are derived from facial visible light video, and the biochemical features are derived from biochemical data; A learning module for modeling personalized meta-prior of lag distribution based on job situation labels and individual baseline features to obtain probability vectors of each lag gear of biochemical features relative to workload state and lag time; and modeling Bayesian multi-source fusion framework and personalized meta-prior of workload state to initialize a set of learnable parameters and workload state corresponding; An aggregation module for constructing a time series causal knowledge graph based on workload state, physiological features, facial behavior features, biochemical features and working condition vectors, and combining probability vectors and lag time to obtain graph aggregation items for representing causal lag effects; A construction module for constructing a state transition model based on the set of learnable parameters, working condition vectors and graph aggregation items; and constructing a first observation model of video modal and a second observation model of biochemical data modal based on the set of learnable parameters and combining quality adaptive mechanism; An update module for solving the prior mean and prior uncertainty at the current time through the state transition model, and updating the posterior mean and posterior uncertainty at the current time through the first observation model and the second observation model by using Kalman filtering to intelligently evaluate the current workload state.

[0016] (Three) beneficial effects The present application provides a personalized workload intelligent evaluation method based on Bayesian multi-source fusion. Compared with the prior art, the following beneficial effects are achieved: In the present application, firstly, multi-source asynchronous data is collected and processed; secondly, through the personalized meta-prior method of incorporating the lag distribution modeling of biochemical response, the set of learnable parameters and the workload state are reasonably initialized in the absence of individual historical data, avoiding prior bias; thirdly, the graph aggregation term used to represent causal lag effects is obtained to construct a state transition model, and a quality adaptive mechanism is introduced to construct an observation model, which maps low-quality observations to greater effective noise to automatically reduce the weight and reduce the interference of asynchrony and noise on fusion; finally, the posterior prior and uncertainty estimation are completed in the same framework, the current workload state is intelligently evaluated, which is superior to the prior art in accuracy, interpretability and operability, and has cross-scene migration capability. BRIEF DESCRIPTION OF DRAWINGS

[0017] In order to more clearly illustrate the technical solutions in the embodiments of the present application or the prior art, the following will briefly introduce the drawings needed to be used in the embodiments or prior art description. Obviously, the drawings in the following description are only some embodiments of the present application, and other drawings can be obtained by those skilled in the art without creative labor.

[0018] Figure 1 A technical concept framework diagram of a personalized workload intelligent evaluation method based on Bayesian multi-source fusion provided by the embodiments of the present application; Figure 2 A block diagram of a personalized workload intelligent evaluation method based on Bayesian multi-source fusion provided by the embodiments of the present application; Figure 3 A flowchart of a personalized workload intelligent evaluation method based on Bayesian multi-source fusion provided by the embodiments of the present application. DETAILED DESCRIPTION

[0019] In order to make the objectives, technical solutions and advantages of the embodiments of the present application more clear, the technical solutions in the embodiments of the present application are described clearly and completely. Obviously, the described embodiments are some of the embodiments of the present application, not all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative labor are within the scope of protection of the present application.

[0020] The embodiments of the present application provide a personalized workload intelligent evaluation method based on Bayesian multi-source fusion, which solves the technical problems that the existing method is easy to fuse distortion and difficult to reasonably initialize the model under the condition of coexistence of multi-source asynchrony and individual difference.

[0021] The technical solutions in the embodiments of the present application are as follows to solve the above technical problems: AsFigure 1 As shown, the solution disclosed in this embodiment of the invention includes steps such as collecting and processing multi-source asynchronous data, modeling personalized meta-priors, constructing a (temporal causal) knowledge graph, quality assessment, Bayesian multi-source fusion, and workload assessment. The core improvements involved include at least the following: I. Generation of Personalized Meta-Priorities: In the absence of individual historical data in the system, this embodiment of the invention provides reasonable initialization for state transition parameters, observation sensitivity and workload status through meta-learning method, and incorporates hysteresis distribution modeling to enable the model to adapt to the physiological / biochemical response delays of different individuals and work situations, avoid early "prior drift" and improve the applicability across populations and scenarios.

[0022] II. Quality Adaptive Mechanism: This invention introduces quality weights into a unified Bayesian multi-source fusion framework, automatically amplifying the noise variance corresponding to low-quality observations to achieve adaptive weight reduction. This ensures that the fusion results rely more on high-quality information without relying on heuristic elimination, thereby improving the robustness of the evaluation.

[0023] III. A Bayesian inference framework integrating personalized lag priors and quality adaptation: This invention does not merely employ conventional Bayesian fusion; instead, it simultaneously introduces "personalized lag-distributed priors" and a "quality adaptive weighting mechanism" into the causal graphical model. This enables the system to maintain stable inference under conditions of strong asynchrony, multi-source heterogeneity, and quality fluctuations. Within this framework, it can output posterior estimates of load levels and uncertainty intervals, and can also quantitatively assess risk changes brought about by different operating conditions, thereby elevating traditional "load monitoring" to "decision support."

[0024] 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.

[0025] Example 1: like Figure 2 As shown, this embodiment of the invention provides a personalized intelligent workload assessment method based on Bayesian multi-source fusion, including: S1. Collect and process multi-source asynchronous data to construct physiological features, facial behavior features, biochemical features, work condition vectors, work context labels, and individual baseline features; wherein, the physiological features and facial behavior features are derived from visible light facial videos, and the biochemical features are derived from biochemical data; S2. Based on the job context label and individual baseline features, model the personalized meta-prior of the lag distribution to obtain the probability vector and lag time of each lag level of the biochemical features relative to the workload state; and model the personalized meta-prior of the Bayesian multi-source fusion framework and the workload state to initialize the set of learnable parameters and the workload state accordingly. S3. Based on workload status, physiological characteristics, facial behavior characteristics, biochemical characteristics and working condition vectors, construct a time-series causal knowledge graph, and combine probability vectors and lag time to obtain graph aggregation terms used to characterize the causal lag effect; S4. Based on the learnable parameter set, working condition vector, and graph aggregation term, construct a state transition model; and based on the learnable parameter set, combined with a quality adaptive mechanism, construct the first observation model for the video modality and the second observation model for the biochemical data modality, respectively. S5. Solve for the prior mean and prior uncertainty at the current moment through the state transition model, and update the posterior mean and posterior uncertainty at the current moment using Kalman filtering through the first observation model and the second observation model to intelligently assess the current workload status.

[0026] In this embodiment of the invention, on the one hand, a quality adaptive mechanism and lag distribution modeling are introduced to reduce the interference of asynchronous and low-quality observations between different modalities on inference; on the other hand, a personalized meta-prior generator is designed to reasonably initialize parameters and states even without individual prior data, thereby improving the model's adaptability to different individuals and work scenarios. Within the same framework, this method can output robust posterior predictions, providing interpretable and quantifiable decision support for sports training and a wider range of production operation scenarios, and can be extended to a wider range of production operations and manual labor scenarios.

[0027] like Figure 3 As shown, Figure 3 A flowchart of a personalized intelligent workload assessment method based on Bayesian multi-source fusion is disclosed.

[0028] Next, we will combine Figure 3 The steps of the above scheme are described in detail: To unify the time and symbol conventions for data and models, and to avoid confusion in subsequent inference, alignment, and tracing, this embodiment of the invention first explains the following core parameter symbols: This indicates the step size; t = 1, 2, ..., where t is the time index.

[0029] express t The workload status at any given time, i.e. the potential intensity of the workload, is a continuous latent variable estimated by fusing observations from multiple sources such as heart rate and biochemical parameters.

[0030] representing t the working condition vector (type, duration, intensity, load, rest, environment) at the time t.

[0031] representing the observation value of the modality m (heart rate HR / heart rate variability HRV / facial motion unit AU / head pose Pose / eye gaze direction Gaze / blood lactate La / creatine kinase CK / urine protein UP / testosterone T…), where m takes the video modality vid or the biochemical modality bio in the embodiments of the present application.

[0032] , representing the individual baseline characteristics and the job context label respectively, and can be selected from t the aforementioned two groups of feature vectors at the time t to form the covariate features (a group of factors that have an impact on the observation results, but do not directly represent the level of ‘workload’).

[0033] representing the quality score of the modality m.

[0034] In step S1, multi-source asynchronous data is collected and processed to construct physiological features, facial behavior features, biochemical features, working condition vectors, job context labels, and individual baseline characteristics.

[0035] It should be noted that “multi-source” in the embodiments of the present application refers to data from different sensors, measurement methods or recording channels, such as videos, biochemical detection, job recording, etc. “Asynchronous” refers to the fact that the sampling frequency, time resolution and lag of different data sources are not synchronized, and it is not possible to “collect the same batch of data at the same time”. Herein, the above-mentioned multi-source asynchronous data is specifically multi-modal data, job data and personal baseline data, wherein the multi-modal data includes facial visible light video and biochemical data.

[0036] For the multi-source asynchronous data, the following collection and processing processes are performed respectively: (1) Facial visible light video Exemplarily, a visible light camera with a frame rate of 30 fps is used, with a fixed focal length, and the facial ROI covers the cheeks and forehead. At the same time, factors affecting the quality of the video such as the signal-to-noise ratio , the occlusion rate , the tracking confidence , and the jitter amplitude at the time t are extracted to provide a data basis for subsequent solving of the video modality quality score.

[0037] After obtaining the visible light video through the visible light camera, the face video containing the face ROI is selected, the remote photoplethysmography (rPPG) signal is extracted from the face video using the heart rate estimation algorithm, and the processed rPPG signal is subjected to spectral analysis using the peak detection algorithm, and the heart rate HR feature is calculated by combining the sampling frequency of the signal. Eliminate noise and retain the frequency range related to heart rate. The filtered signal is processed by HeartPy (a Python heart rate analysis toolbox) to calculate the heart rate variability HRV feature. That is, at least physiological features such as heart rate HR and heart rate variability HRV are obtained through the face visible light video.

[0038] In addition, based on OpenFace (an open source face recognition library), the face is detected and 68 face key points are located from the collected face visible light video frames, and a plurality of face behavior features are obtained based on the located face key points, including at least head pose Pose feature, eye gaze direction Gaze feature and face movement unit AU feature.

[0039] (2) Biochemical data Collect blood lactic acid La, creatine kinase CK, urine protein UP, testosterone T, strictly record the sampling time stamp relative to the work time (such as '5min, 10min, 15min after work'), record the test reagent batch, equipment self-checking result and repeated measurement variance, and the quantitative detection results of each index, such as blood lactic acid concentration, creatine kinase activity level, urine protein quantitative value, serum testosterone concentration, are directly used as the corresponding biochemical features.

[0040] (3) Work data Record the working condition vector , used to represent the work type (such as carrying / welding / assembly / patrol, etc.), duration, rhythm, intensity (such as tempo, carrying weight, tempo interval, etc.), etc.

[0041] Work context label , used to represent the work context (such as high temperature / night shift / outdoor, etc.).

[0042] Among them, in the process of encoding work data, categorical variables are encoded using one-hot encoding to form a vector, and numerical variables (duration, weight, intensity, etc.) are directly standardized to form a vector.

[0043] (4) Individual baseline data Individual baseline data includes height, weight, age, gender, physical fitness assessment, health history, etc., and individual baseline features are obtained after encoding .

[0044] Among them, in the process of encoding individual baseline data, categorical variables (gender, health history, etc.) are encoded into vectors using one-hot encoding, and numerical variables (height, weight, age, physical fitness assessment) are directly standardized to form vectors.

[0045] In step S2, based on the job context label and individual baseline features, a personalized meta-prior of lag distribution is modeled to obtain the probability vector of each lag gear of the biochemical feature relative to the workload state and the lag time; and a Bayesian multi-source fusion framework and a personalized meta-prior of the workload state are modeled to correspond to initialize a set of learnable parameters and the workload state.

[0046] It should be pointed out that biochemical data features have been fully proven to have obvious lag characteristics (see Durand, R., Galli, M., Chenavard, M., Bandiera, D., Freund, H., & Messonnier, L. A. (2021). Modelling of Blood Lactate Time-Courses During Exercise and / or the Subsequent Recovery: Limitations and Few Perspectives . Frontiers in Physiology, 12, Article 702252. and the like), for example: Blood lactate (La) usually reaches a peak value 5-15 minutes after high-intensity work or exercise, and then slowly decreases, reflecting the delayed effect of metabolic processes; creatine kinase (CK) usually increases significantly after several hours, which is a typical delayed indicator of muscle damage; urinary protein (UP) also has a lag of several hours after heavy physical load, which is related to the cumulative effect of impaired glomerular filtration function. In addition, the levels of hormones such as testosterone and cortisol are not only affected by the circadian rhythm, but also have a significant time lag in response to workload.

[0047] Therefore, it is necessary to consider the lag distribution of biochemical features in individualized workload modeling, otherwise it is easy to misjudge the causal relationship or underestimate its actual impact.

[0048] As mentioned above, the lag time interval is affected by individual physical fitness / post differences and needs to be personalized. Here, the input is the spliced job context label and individual baseline features; the output is the prior distribution of the lag gear (the impact on workload is divided into several intervals according to the length of the delay time). Under the premise of maintaining stability, improve the lag adaptability of different people / post.

[0049] In an optional implementation, in this step, the personalized meta-prior of lag distribution is modeled based on the job context label and individual baseline features to obtain the probability vector of each lag gear of the biochemical feature relative to the workload state and the lag time, including: S10. Concatenate the work context label and individual baseline features, perform linear transformation and normalization to obtain the probability vectors of each lag level of the biochemical features relative to the workload state, expressed as: in, Indicates the first i The lag level of each biochemical characteristic compared to the workload state The probability vector, The first one is given i The effect delay of a biochemical trait may fall on the first j The probability at each lagging gear. This represents a parameter matrix (which can be viewed as the weights of a small MLP or linear transformation) used to linearly transform input features (individual baseline and operating condition features) to unnormalized probability scores logits. This represents the concatenated job context labels and individual baseline features. Softmax represents the function that converts logits into probabilities. This represents the number of lag levels for the i-th biochemical characteristic. Indicates the first i The first biochemical characteristic j The probability weight of each lagging gear.

[0050] S20. Based on the probability vector, and selecting the classification distribution as the personalized meta-prior of the lag distribution, obtain the lag time of each lag level of the biochemical feature relative to the workload state, expressed as: in, This represents the j-th lag level of the i-th biochemical characteristic relative to the workload state. The lag time; cat represents the classification distribution function; This indicates that the lag time follows a classification distribution, and the lag time is sampled from this distribution.

[0051] Furthermore, when the system initially lacks historical data for each individual, parameter estimation becomes unstable even with proper parameter and state initialization. Personalized meta-prior not only initializes the model parameters but also prepares the workload states. This provides a basis for the starting point. It significantly improves the problem of being unable to reasonably initialize parameters when there is no historical data of individuals at the beginning of the system. When only the mean and variance are known, the normal distribution is chosen as the prior because it satisfies the maximum entropy principle (when only the mean and variance are known, the normal distribution is the distribution with the largest entropy and the least biased), which is consistent with the experience that the residuals of linear regression approximate normality. At the same time, it is computationally friendly in Gaussian state-space models and can obtain closed solutions or efficient approximations. Therefore, the normal distribution is an assumption that is both reasonable and mathematically convenient.

[0052] Therefore, the step selects a normal distribution as the personalized meta-prior of the Bayesian multi-source fusion framework and the workload state. Correspondingly, the corresponding initialization of the learnable parameter set and the workload state is performed, in which: In an optional implementation, for the personalized meta-prior of the Bayesian multi-source fusion framework, the initialization of the learnable parameter set comprises: S10’, concatenating the job context label and the individual baseline feature, and fitting a linear mapping parameter based on a least square method to obtain a first prior mean and a first prior covariance.

[0053] The first prior mean is defined as , and the first prior covariance is . Wherein: The linear mapping parameter fitted based on the least square method. Specifically: a batch of subject data is collected offline, each subject offline data has and the estimated individual parameters, a multivariate linear regression model is used, and a least square method is used for fitting. The least square method can provide an analytical, computationally efficient, and Gaussian prior consistent solution, ensuring interpretability and stability. The first prior mean is represented as: Wherein, represents the linear mapping parameter fitted based on the least square method; specifically: A is a first weight matrix, used to map the concatenated job context label and individual baseline feature to the parameter space; b is a first bias vector, used to correct the translation item of the linear mapping.

[0054] Further, the prediction parameter vector is given as: Wherein, X represents the independent variable (feature) matrix, and the features of one subject offline data ; represents the dependent variable (target) matrix, and each row is the predicted parameter vector θ of the subject offline data.

[0055] The residual matrix is: Then the first prior covariance is : Wherein, N represents the number of subject offline data, and the superscript T represents transposition.

[0056] S20', initializing the set of learnable parameters to follow a normal distribution based on the first prior mean and the first prior covariance; wherein the set of learnable parameters comprises state inertia coefficient, working condition driving weight, observation sensitivity, observation parameters of the first observation model and the second observation model; denoted as: wherein θ represents the set of all learnable parameters, denotes that the set of learnable parameters θ follows a normal distribution.

[0057] is the state inertia coefficient, is the working condition driving weight vector, is the observation sensitivity (i.e. the channel coefficient of injecting the observation information aggregated by the knowledge graph into the hidden state).

[0058] denotes the observation parameter of the modality m, wherein: denotes the observation parameter of the first observation model, respectively represent the sensitivity of the video modality to the load, the covariate feature weight of the video modality, and the zero-point difference under the same load in the video modality.

[0059] denotes the observation parameter of the first observation model, respectively represent the sensitivity of the biochemical modality to the load, the covariate feature weight of the biochemical modality, and the zero-point difference under the same load in the biochemical modality. In an optional embodiment, for the personalized meta-prior of the working load state, the initialization of the working load state comprises:

[0060] S10'', concatenating the work context label and the individual baseline feature, and obtaining the linear mapping parameter based on the centering to obtain the second prior mean and the second prior covariance thereof. defining the first prior mean as

[0061] and the first prior covariance as . Wherein: obtaining the linear mapping parameter based on the centering. Specifically: offline collection of a batch of subject data, each subject offline data has and the estimated individual parameters thereof, using the centering to train the linear mapping parameter of the initial working load state . Centering is a common operation in regression analysis, which is convenient for interpretation and can ensure that variables of different scales are comparable, which will not be described here. The second prior mean is denoted as: ​​​ wherein, denote the linear mapping parameters obtained based on the least square fitting; specifically, is a second weight matrix, which is also used to map the spliced job situation label and individual baseline features to the parameter space; is a second bias vector, which is also used to correct the translation term of the linear mapping.

[0062] Referring to the first prior covariance of the above-mentioned first prior covariance and the first prior covariance of the solution process, the second prior covariance and the first prior covariance can be solved in the same way. Here, it is not repeated.

[0063] S20'', based on the second prior mean and the second prior covariance, initializing the workload state to follow a normal distribution; denoted as: wherein, denote the initial workload state following a normal distribution.

[0064] In step S3, based on the workload state, physiological features, facial behavior features, biochemical features and working condition vectors, a time series causal knowledge graph is constructed, and combined with the probability vector and the lag time, a graph aggregation term for representing causal lag effect is obtained.

[0065] Some biochemical features have lag effects (such as La and CK lagging rise), which may cause misjudgment of causality if ignored. In order to be compatible with asynchronous sampling and biochemical delay, improve explainability and prediction accuracy, the present embodiment first constructs a time series causal knowledge graph, and then uses the time series of each feature and the aforementioned probability vector and lag time to obtain a graph aggregation term for representing causal lag effect.

[0066] In an optional embodiment, the present step includes: S31, taking the physiological features, facial behavior features and working condition vectors as the first type of parent nodes of the workload state, and taking the biochemical features as the second type of parent nodes of the workload state.

[0067] S32, based on the causal influence of the first type of parent nodes on the workload state, defining the directed edges and their weights between the first type of parent nodes and the workload state.

[0068] S33. Based on the probability vector and lag time of each lag level relative to the workload state of the biochemical characteristics, define the directed edge and its weight between the second type of parent node and the workload state.

[0069] S34. Based on each of the directed edges and their weights, and in conjunction with the transformation function, aggregate to obtain a graph aggregation term used to characterize the causal lag effect.

[0070] Specifically: In a temporal causal knowledge graph, a node set is defined as follows: Among them, physiological features HR and HRV, facial behavioral features Pose, AU, and Gaze, and working condition vectors are included. The workload state at time t First type of parent node .

[0071] Biochemical characteristics La, CK, T, and UP are used as the workload status at time t. Second type of parent node .

[0072] Furthermore, for the edge set Define directed edges For the first type of parent node, the weight of the directed edge is defined as follows: To measure the impact of the i-th feature other than biochemical characteristics on... The influence strength (can be positive or negative); and for the second type of parent node, the weight of the directed edge is defined as... To measure the first i The j-th lag level of a biochemical characteristic The strength of the influence (can be positive or negative).

[0073] Both of the above edge weights can be obtained using external labels combined with the least squares method, only using For example ( The complexity is relatively low, but the overall solution approach remains the same. The specific solution is as follows: Where Z represents the design matrix, with dimensions of 1. Each row corresponds to a time point, and each column corresponds to a lag value. This represents the design matrix element corresponding to the j-th lag position of the i-th biochemical characteristic at time t, which is derived from the original observations at a delay. After a monotonic transformation Then, combined with weighting coefficients get; Indicates the first i Individual biochemical characteristics in observed value or its statistical quantity (such as mean, peak, window mean, etc.) at the moment; denotes a set of edge weight vectors corresponding to biochemical characteristics, each number in the vector corresponds to an edge weight ; Y denotes the experimental label value formed by expert evaluation results; Z denotes a design matrix; superscript T denotes transposition.

[0074] Through the above analysis, considering the causal influence of different parent nodes on the workload state, a unified driving quantity (graph aggregation term) can be finally formed, which ensures that the model can correctly capture asynchronous, multi-source and delay effects; the graph aggregation term is expressed as: In step S4, a state transition model is constructed based on the set of learnable parameters, the working condition vector and the graph aggregation term; and a first observation model of the video modality and a second observation model of the biochemical data modality are respectively constructed based on the set of learnable parameters combined with a quality adaptive mechanism.

[0075] It should be noted that the Bayesian multi-source fusion framework in the embodiment of the present application includes a state transition model and a first observation model of the video modality and a second observation model of the biochemical data modality.

[0076] This step first constructs a state transition model, which describes the evolution of the individual's workload state over time, which contains both "inertia effect" (load does not disappear instantaneously) and "working condition driving" (job type, duration, intensity, etc. directly affect the load), and is also affected by causal lag, its input is the state of the last step, the current working condition and the lagging biochemical causal influence, and the output is a "current load level distribution" with uncertainty. The state transition model is expressed as: wherein, , denote the workload state at t-1 and t moments respectively; denotes a normal distribution; denotes a state inertia coefficient; denotes a working condition driving weight, and superscript T denotes transposition; denotes an observation sensitivity; denotes a working condition vector at t moment; is a graph aggregation term at t moment; denotes the inertia effect of the workload state at t-1 moment on t moment; denotes the working condition driving effect; denotes the causal lag effect; >0 denotes the process noise variance; represents the posterior mean at time t-1.

[0077] Next, the first observation model and the second observation model are constructed, and in an optional embodiment, the construction process includes: S100, based on the face visible light video, respectively calculating a signal-to-noise ratio quality score, an occlusion rate quality score, a tracking confidence quality score, and a jitter amplitude quality score, and obtaining a video modal quality score by weighted fusion to obtain a first quality weight of each physiological feature or facial behavior feature.

[0078] Due to the different quality of data from different sources, simple alignment may cause deviation in the prediction result. Therefore, high-quality observation is first ensured before multi-source Bayesian fusion to avoid deviation caused by low-quality data. Here, the signal-to-noise ratio, the occlusion rate, the tracking confidence, and the device state are first standardized to [0, 1]; then, the unified quality score Q is obtained by linear weighting.

[0079] Specifically: As mentioned in step S1, the video quality factors obtained include the signal-to-noise ratio , the occlusion rate , the tracking confidence , and the jitter amplitude at time t. The quality score is obtained from the influence factor, and then the quality score of this modal is obtained by factor standardization and weighting. Among them: (1) Signal-to-noise ratio quality score: wherein, represents the signal-to-noise ratio quality score at time t, and the range is [0, 1]. The closer to 1, the better the video signal-to-noise ratio quality at this time. represents the reference value of the signal-to-noise ratio, which is used to set a reference point. When it is lower than it, the quality is low, and when it is higher than it, the quality is high. k represents the scale parameter, which controls the change speed between SNR and the quality score. is a Sigmoid activation function, which is used to compress the input value (which may be a positive or negative real number) to the interval [0, 1] as the final quality score.

[0080] (2) Occlusion rate quality score: wherein, represents the occlusion rate quality score at time t. represents the occlusion rate index factor, which can be set to an initial value according to the field experience to control the influence of the occlusion rate on the quality score.

[0081] (3) Tracking confidence quality score: wherein, This represents the tracking confidence quality score at time t. The function that takes the smaller value... This indicates that the quality score of the tracking confidence level will not exceed 1.

[0082] (4) Shaking amplitude quality score: in, The jitter amplitude quality score at time t represents the value of the jitter amplitude. This represents the baseline jitter threshold, and an initial value can be set based on domain experience. This represents the scale parameter, and an initial value can be set based on domain experience to control the rate of change between jitter amplitude and quality score.

[0083] Accordingly, the video modal quality score is represented as: in, This represents the standardized score of the i-th quality factor; The weight of the i-th quality factor can be set manually based on domain experience. For example, the occlusion rate has a greater impact on the rPPG signal, so a higher weight of 0.5 is assigned, while slight jitter will not have much impact on the result, so a lower weight of 0.1 is assigned. The weights of other quality factors are evenly distributed at 0.2.

[0084] All video features (i.e., physiological features or facial behavioral features) at the same time point are set to share a single quality score: The first quality weight for each of the aforementioned physiological features or facial behavioral features is expressed as follows: in, In video mode, time t represents the first... i The first quality weight of a physiological or facial behavioral feature. This represents the first shape parameter, used to control... The position and steepness of the activation function's curve, specifically: This represents the first position parameter, used for control. The horizontal offset of the activation function; This represents the first slope parameter, used to control... The steepness of the activation function. For example, here's an example: =0, =1.

[0085] S200. Based on the deviation between the relative sampling delay and the target delay, the repeatability consistency, and the device status in the biochemical data, obtain the biochemical modality quality score to obtain the second quality weight of each biochemical feature.

[0086] Unlike video modalities, biochemical modalities do not require factor weighting. Instead, each biochemical feature (La / CK / T / UP) uses the same formula to independently obtain its quality score at different times, and this quality score is directly used for the feature's own observation likelihood, rather than performing a weighted fusion within the "biochemical modality." This approach is consistent with the physical process of biochemical sampling and is also simpler and more stable. Specifically: Selecting factors affecting biochemical data quality relative to sampling delay in biochemical modalities Delay deviation from target, repeatability consistency Equipment status Solve for the biochemical modal mass components: in, This represents the quality score of the i-th biochemical feature in a biochemical modality. Represents the time t. i The actual delay of a biochemical feature sampling relative to the task can only be calculated by strictly recording the relative task time during sampling. Represents the time t. i Each biochemical characteristic has an ideal delay set based on the hysteresis effect, and the delay time varies for different characteristics. Represents the time t. i The scale of time delay bias of individual biochemical characteristics. Indicates the first i Each biochemical characteristic corresponds to a device status factor, with a value ranging from 0 to 1, which is used to indicate whether the testing device is in good condition. Represents the time t. i The repeatability index of a biochemical characteristic (the larger the index, the greater the fluctuation of repeated measurements / the worse the consistency, such as by using repeated measures variance or coefficient of variation). Indicates the first i A measure of consistency of biochemical characteristics.

[0087] If each biochemical characteristic is used to calculate its quality score independently, then the second quality weight of each biochemical characteristic is expressed as follows: in, This represents the second quality weight of the i-th biochemical feature at time t under the biochemical modality. This represents the second shape parameter, also used for control. The position and steepness of the activation function's curve, specifically: This represents the second position parameter, also used for control. The horizontal offset of the activation function; This represents the second slope parameter, also used to control... The steepness of the activation function. For example, here's an example: =0, =1.

[0088] S300. Based on the observation parameters, first quality weights, and corresponding covariate features of the first observation model, construct the first observation model; and based on the covariate features, the observation parameters, and second quality weights of the second observation model, construct the second observation model.

[0089] In state-space modeling and observation modeling, the embodiments of the present invention assume that the observation noise follows a normal distribution because the Gaussian noise assumption not only conforms to the central limit theorem but also brings computational convenience: under Gaussian conditions, the likelihood function and posterior distribution have analytical forms, which ensures the interpretability and stability of the model.

[0090] Accordingly, the first observation model and the second observation model are respectively expressed as: in, These represent the workload status at time t; Represents a normal distribution; express t Time of the first i Actual observations of physiological or facial behavioral features; the superscript vid indicates the video modality. Indicates the observation parameters of the first observation model. Video modal pairs Sensitivity, covariate feature weights of video modalities, zero-point differences under the same load in video modalities, with the superscript T indicating transpose; express t Covariate characteristics of video modalities at any given time; Represents the baseline observation noise variance for the video modality; express t Time of the first i The first quality weight of each physiological or facial behavioral feature; express Time of the first i The actual observations of a biochemical characteristic, with the superscript "bio" indicating a biochemical mode. Indicates the first i The j-th lag level of a biochemical characteristic is... The lag time; Indicates the observation parameters of the second observation model. Representing biochemical modal pairs sensitivity of the biochemical modality, covariate characteristic weight of the biochemical modality, zero-point difference under the same load in the biochemical modality; denotes t covariate characteristic of the biochemical modality at the time point; denotes reference observation noise variance of the biochemical modality; denotes denotes the second quality weight of the biochemical characteristic at the time point. i

[0091] It should be noted that the covariate characteristics , are selected from part or all of the characteristics in the work situation label and the individual baseline characteristic according to actual needs.

[0092] For example, for the physiological characteristic heart rate HR under the video modality, the covariate characteristics can include individual baseline factors (age, gender, physical fitness level), work situation factors (environmental temperature) and work condition related factors (work intensity, duration) and the like, to represent a group of variable information capable of adjusting the fluctuation of the heart rate but not belonging to the work load state itself.

[0093] In addition, it is necessary to point out that the embodiment of the present application introduces quality weight in the unified Bayesian framework, automatically amplifies the noise variance corresponding to the low-quality observation, realizes adaptive weight reduction, so as to ensure that the fusion result depends more on high-quality information without relying on heuristic elimination, and improves the robustness of the evaluation.

[0094] In step S5, the prior mean and the prior uncertainty at the current time point are solved through the state transition model, and the Kalman filter is used to update the posterior mean and the posterior uncertainty at the current time point through the first observation model and the second observation model, so as to intelligently evaluate the current work load state.

[0095] This step fuses multi-modal and different quality observations in real time, and gives the state posterior and uncertainty.

[0096] Specifically: First, prediction (prior propagation) is performed, that is, the prior mean and the prior uncertainty at the current time point are solved through the state transition model, and are represented as: Wherein, denotes the prior mean at the time point t; denotes the posterior mean at the time point t-1; denotes the prior uncertainty at the time point t; denotes the posterior uncertainty at the time point t-1.

[0097] ​Next, modal updates are performed. That is, using the first and second observation models, Kalman filtering is applied to update the posterior mean and posterior uncertainty at the current time step, including: Based on the first observation model and the second observation model, the corresponding Kalman gains are calculated and expressed as follows: in, Represents the time t. i Kalman gain corresponding to a physiological feature or facial behavioral feature; This represents the prior uncertainty at time t; express Time of the first i Kalman gain corresponding to each biochemical characteristic; Based on the first observation model and the second observation model, the corresponding predicted values ​​of the observations are obtained respectively, as follows: in, express t Time of the first i Observational predictions of physiological or facial behavioral features; express Time of the first i Observed and predicted values ​​of individual biochemical characteristics; Based on the Kalman gain, and combined with the difference between the actual observation and the predicted observation, the posterior mean is updated as follows: in, Let be the posterior mean at time t. Let represent the prior mean at time t; Based on the Kalman gain, the updated posterior uncertainty is expressed as: in, express t Posterior uncertainty at time.

[0098] It is not difficult to understand that, due to Therefore, this embodiment of the invention can intelligently assess the current workload status.

[0099] Furthermore, in obtaining the current workload status Then, given a future trajectory of T steps, the following conditions are considered. The embodiments of the present invention repeatedly execute the above scheme steps, which can be used to process the entire future hidden state path. The prediction is obtained through progressive forward propagation of the state transition model, and is expressed as follows: .

[0100] So far, the embodiment of the present application completes the whole process of the personalized work load intelligent evaluation method based on Bayesian multi-source fusion.

[0101] Embodiment 2: The embodiment of the present application provides a personalized work load intelligent evaluation system based on Bayesian multi-source fusion, comprising: The acquisition and processing module is used for acquiring and processing multi-source asynchronous data, constructing physiological features, facial behavior features, biochemical features, working condition vectors, job situation labels and individual baseline features; wherein the physiological features and the facial behavior features are derived from facial visible light videos, and the biochemical features are derived from biochemical data; The learning module is used for modeling the personalized meta-prior of the lag distribution based on the job situation labels and the individual baseline features, so as to obtain the probability vector of each lag gear of the biochemical features relative to the work load state and the lag time; and modeling the Bayesian multi-source fusion framework and the personalized meta-prior of the work load state, so as to correspond to initialize the learnable parameter set and the work load state; The aggregation module is used for constructing a time series causal knowledge graph based on the work load state, the physiological features, the facial behavior features, the biochemical features and the working condition vectors, and combining the probability vector and the lag time to obtain a graph aggregation item for representing the causal lag effect; The construction module is used for constructing a state transition model based on the learnable parameter set, the working condition vector and the graph aggregation item; and constructing a first observation model of the video mode and a second observation model of the biochemical data mode based on the learnable parameter set and combining a quality adaptive mechanism; The update module is used for solving the prior mean and the prior uncertainty of the current time through the state transition model, and updating the posterior mean and the posterior uncertainty of the current time through the first observation model and the second observation model by using Kalman filtering, so as to intelligently evaluate the current work load state.

[0102] It can be understood that the personalized work load intelligent evaluation system based on Bayesian multi-source fusion provided by the embodiment of the present application corresponds to the personalized work load intelligent evaluation method based on Bayesian multi-source fusion provided by the embodiment of the present application, and the explanation, examples and beneficial effects of the related content can refer to the corresponding parts in the method, which will not be repeated here.

[0103] In summary, compared with the prior art, the following beneficial effects are achieved: 1、The embodiment of the application introduces a quality self-adaptive mechanism and a lag distribution modeling on the one hand, reduces the interference of different modalities and low-quality observations on inference; on the other hand, a personalized meta-prior generator is designed, which can still reasonably initialize parameters and states when there is no individual prior data, thereby improving the adaptability of the model to different individuals and job situations. Within the same framework, the method can output robust posterior predictions, providing interpretable and quantifiable decision support for sports training and more extensive production scenarios, and can be extended to more extensive production and physical labor scenarios.

[0104] 2、The embodiment of the application introduces a quality weight in a unified Bayesian framework, automatically amplifies the noise variance corresponding to low-quality observations, realizes adaptive weight reduction, and thus ensures that the fusion result relies more on high-quality information without relying on heuristic rejection, thereby improving the robustness of evaluation.

[0105] 3、The embodiment of the application proposes a Bayesian inference framework that fuses personalized lag priors and quality adaptation, simultaneously introduces 'personalized lag distribution priors' and 'quality adaptive weighting mechanisms' in causal graph models, so that the system can maintain stable inference under strong asynchronous, multi-source heterogeneous and quality fluctuation conditions.

[0106] It should be noted that in this paper, 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. Moreover, the terms "include", "contain" or any other variant thereof are intended to cover non-exclusive inclusion, so that the process, method, article or device including a series of elements not only includes those elements, but also includes other elements not explicitly listed or inherent to such process, method, article or device. Without more limitations, the element defined by the statement "including a" does not exclude the presence of additional identical elements in the process, method, article or device including the element.

[0107] The above embodiments are only used to illustrate the technical solutions of the present application, but not to limit them; although the present application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that they can still modify the technical solutions recorded in the foregoing embodiments, or make equivalent replacement to part of the technical features; and these modifications or replacements do not make the essence of the corresponding technical solutions deviate from the spirit and scope of the technical solutions of the embodiments of the present application.

Claims

1. A personalized intelligent workload assessment method based on Bayesian multi-source fusion, characterized in that, include: Collect and process multi-source asynchronous data to construct physiological features, facial behavior features, biochemical features, work condition vectors, work context labels, and individual baseline features; wherein, the physiological features and facial behavior features are derived from visible light facial videos, and the biochemical features are derived from biochemical data; Based on job context labels and individual baseline features, we model personalized meta-prior of lag distribution to obtain the probability vector and lag time of each lag level of biochemical features relative to the workload state; and model personalized meta-prior of Bayesian multi-source fusion framework and workload state to initialize the set of learnable parameters and workload state accordingly. Based on workload status, physiological characteristics, facial behavior characteristics, biochemical characteristics, and working condition vectors, a time-series causal knowledge graph is constructed, and combined with probability vectors and lag time, graph aggregation terms are obtained to characterize the causal lag effects. A state transition model is constructed based on a set of learnable parameters, a condition vector, and a graph aggregation term; and a first observation model for video modality and a second observation model for biochemical data modality are constructed based on a set of learnable parameters and a quality adaptive mechanism, respectively. The prior mean and prior uncertainty at the current moment are solved by a state transition model, and the posterior mean and posterior uncertainty at the current moment are updated by Kalman filtering through a first observation model and a second observation model, so as to intelligently assess the current workload status.

2. The personalized workload intelligent assessment method as described in claim 1, characterized in that, The personalized meta-prior, based on job context labels and individual baseline features, models a lag distribution to obtain the probability vectors and lag times of each lag level of biochemical characteristics relative to the workload state, including: By splicing the work context label and individual baseline features, performing linear transformation and normalization, the probability vectors of each lag level of the biochemical features relative to the workload state are obtained. Based on the probability vector, and selecting the classification distribution as the personalized meta-prior of the lag distribution, the lag time of each lag level of the biochemical feature relative to the workload state is obtained.

3. The personalized workload intelligent assessment method as described in claim 1, characterized in that, The normal distribution is chosen as the personalized meta-prior for the Bayesian multi-source fusion framework and the workload state, corresponding to the initialization of the learnable parameter set and the workload state, wherein: Personalized meta-prior for the Bayesian multi-source fusion framework: The task context label and individual baseline features are spliced ​​together, and linear mapping parameters are obtained by fitting based on the least squares method to obtain the first prior mean and its first prior covariance; Based on the first prior mean and the first prior covariance, the set of learnable parameters is initialized to follow a normal distribution; wherein the set of learnable parameters includes state inertia coefficient, operating condition driving weight, observation sensitivity, and observation parameters of the first observation model and the second observation model; Personalized meta-prior for workload states: The task context label and individual baseline features are concatenated, and a linear mapping parameter is obtained based on centering to obtain the second prior mean and its second prior covariance; Based on the second prior mean and the second prior covariance, the workload state is initialized to follow a normal distribution.

4. The personalized workload intelligent assessment method as described in claim 1, characterized in that, The method constructs a time-series causal knowledge graph based on workload status, physiological characteristics, facial behavioral characteristics, biochemical characteristics, and work condition vectors. It then combines probability vectors and lag times to obtain graph aggregation terms used to characterize the causal lag effects, including: The physiological features, facial behavior features, and work condition vectors are used as the first type of parent nodes for workload status, and the biochemical features are used as the second type of parent nodes for workload status. Based on the causal influence of the first type of parent nodes on the workload state, the directed edges between the first type of parent nodes and the workload state and their weights are defined. Based on the probability vector and lag time of each lag level relative to the workload state according to the biochemical characteristics, the directed edge between the second type of parent node and the workload state and its weight are defined. Based on each directed edge and its weight, and combined with a transformation function, a graph aggregation term is obtained to characterize the causal lag effect.

5. The personalized workload intelligent assessment method as described in claim 1, characterized in that, The state transition model is expressed as follows: in, , These represent the workload status at times t-1 and t, respectively. Represents a normal distribution; Represents the state inertia coefficient; This indicates the weights driven by operating conditions; the superscript T indicates transpose. Indicates observation sensitivity; Represents the operating condition vector at time t; For the spectral aggregation term at time t; This represents the inertial effect of the workload state at time t-1 on the workload state at time t. Indicates the impact of operating conditions; Indicates the lagged effect of causation; >0 indicates process noise variance; Let represent the posterior mean at time t-1.

6. The personalized workload intelligent assessment method as described in claim 5, characterized in that, The method for solving the prior mean and prior uncertainty at the current time step using the state transition model is expressed as follows: in, Let represent the prior mean at time t; Let represent the posterior mean at time t-1; This represents the prior uncertainty at time t; This represents the posterior uncertainty at time t-1.

7. The personalized workload intelligent assessment method as described in claim 3, characterized in that, The method, based on a set of learnable parameters and combined with a quality adaptive mechanism, constructs a first observation model for the video modality and a second observation model for the biochemical data modality, including: Based on the facial visible light video, the signal-to-noise ratio quality score, occlusion rate quality score, tracking confidence quality score, and jitter amplitude quality score are calculated respectively. After weighted fusion, the video modal quality score is obtained to obtain the first quality weight of each physiological feature or facial behavior feature. Based on the deviation between the relative sampling delay and the target delay, the repeatability consistency, and the device status in the biochemical data, a biochemical modality quality score is obtained to obtain a second quality weight for each biochemical feature. Based on the observation parameters, first quality weights, and corresponding covariate features of the first observation model, the first observation model is constructed; and based on the covariate features, the observation parameters and second quality weights of the second observation model, the second observation model is constructed.

8. The personalized workload intelligent assessment method as described in claim 7, characterized in that, The first observation model and the second observation model are respectively expressed as: in, These represent the workload status at time t; Represents a normal distribution; express t Time of the first i Actual observations of physiological or facial behavioral features; the superscript vid indicates the video modality. Indicates the observation parameters of the first observation model. Video modal pairs Sensitivity, covariate feature weights of video modalities, zero-point differences under the same load in video modalities, with the superscript T indicating transpose; express t Covariate characteristics of video modalities at any given time; Represents the baseline observation noise variance for the video modality; express t Time of the first i The first quality weight of each physiological or facial behavioral feature; express Time of the first i The actual observations of a biochemical characteristic, with the superscript "bio" indicating a biochemical mode. Indicates the first i The j-th lag level of a biochemical characteristic is... The lag time; Indicates the observation parameters of the second observation model. Representing biochemical modal pairs Sensitivity, covariate feature weights of biochemical modes, and zero-point differences under the same load in biochemical modes; express t Covariate characteristics of biochemical modalities at time points; The baseline observation noise variance represents the biochemical modality; express Time of the first i The second quality weight of each biochemical characteristic.

9. The personalized workload intelligent assessment method as described in claim 8, characterized in that, The process of updating the posterior mean and posterior uncertainty at the current time using Kalman filtering through the first and second observation models includes: Based on the first observation model and the second observation model, the corresponding Kalman gains are calculated and expressed as follows: in, Represents the time t. i Kalman gain corresponding to a physiological feature or facial behavioral feature; This represents the prior uncertainty at time t; express Time of the first i Kalman gain corresponding to each biochemical feature; Based on the first observation model and the second observation model, the corresponding predicted values ​​of the observations are obtained respectively, as follows: in, express t Time of the first i Observational predictions of physiological or facial behavioral features; express Time of the first i Observed and predicted values ​​of individual biochemical characteristics; Based on the Kalman gain, and combined with the difference between the actual observation and the predicted observation, the posterior mean is updated as follows: in, Let be the posterior mean at time t. Let represent the prior mean at time t; Based on the Kalman gain, the updated posterior uncertainty is expressed as: in, express t Posterior uncertainty at time.

10. A personalized intelligent workload assessment system based on Bayesian multi-source fusion, characterized in that, include: The acquisition and processing module is used to acquire and process multi-source asynchronous data to construct physiological features, facial behavior features, biochemical features, work condition vectors, work context labels, and individual baseline features; wherein, the physiological features and facial behavior features are derived from visible light facial videos, and the biochemical features are derived from biochemical data; The learning module is used to model the personalized meta-prior of the lag distribution based on the job context label and individual baseline features, so as to obtain the probability vector and lag time of each lag level of biochemical features relative to the workload state; and to model the personalized meta-prior of the Bayesian multi-source fusion framework and the workload state, so as to initialize the set of learnable parameters and the workload state accordingly. The aggregation module is used to construct a time-series causal knowledge graph based on workload status, physiological characteristics, facial behavior characteristics, biochemical characteristics, and working condition vectors, and to obtain graph aggregation items to characterize the causal lagged effects by combining probability vectors and lag time. The module is used to construct a state transition model based on a set of learnable parameters, a condition vector, and a graph aggregation term; and to construct a first observation model for video modalities and a second observation model for biochemical data modalities based on a set of learnable parameters and a quality adaptive mechanism, respectively. The update module is used to solve the prior mean and prior uncertainty at the current time through the state transition model, and to update the posterior mean and posterior uncertainty at the current time through the first observation model and the second observation model using Kalman filtering, so as to intelligently evaluate the current workload status.

Images