Multi-source behavior and physiological feature joint early warning method based on time series deep learning

Abstract

The application discloses a multi-source behavior and physiological feature joint early warning method based on time sequence deep learning, which comprises the following steps: acquiring aligned multi-source behavior and physiological feature time sequence data; performing parallel processing on the multi-source time sequence data, extracting time sequence features, and learning causal structure information containing causal adjacency relationship and effect strength by using a differentiable causal discovery algorithm; generating a static causal mask and a dynamic modulation term based on the causal structure information, performing self-attention weighting of the time sequence features under the guidance of causality, generating deep fusion features, and outputting an early warning signal.When the early warning is triggered, the key causal path is automatically traced according to the causal structure information, and counterfactual simulation is performed.The application effectively removes the pseudo-correlation noise in the data, improves the robustness of the model early warning, breaks the network black box limit, and realizes the explainable intelligent early warning with structured attribution and quantitative intervention guidance.

Description

Technical Field

[0001] This invention relates to the field of intelligent data processing technology, and in particular to a joint early warning method based on time-series deep learning for multi-source behavior and physiological features. Background Technology

[0002] In recent years, with the popularization of the Internet of Things and wearable smart devices, it has become increasingly possible to continuously and dynamically collect multi-source behavioral trajectories and physiological data of individuals. Traditional smart device early warning systems often rely on single-dimensional static threshold judgments, which are prone to false alarms or missed alarms. Existing technologies typically utilize multimodal fusion networks to splice physiological states and behavioral patterns within a latent feature space or to dynamically weight them using conventional attention mechanisms. This allows for the discovery of nonlinear mapping relationships between multi-source time-series data, aiming to achieve joint judgment and early warning of potential health crises, abnormal behaviors, or extreme physiological states.

[0003] However, existing joint early warning schemes based on temporal deep learning still have some technical shortcomings. Conventional data-driven self-attention mechanisms can only capture statistical correlations (i.e., co-occurrence relationships or spurious correlations) between temporal features, making it difficult to effectively remove confounding factors from the data and uncover the true causal topological structure between multi-source features. This results in a lack of robustness of the model when facing distribution drift in complex dynamic environments, and feature fusion is susceptible to noise interference. Furthermore, existing early warning systems are mostly highly opaque black-box architectures, and their output early warning signals are often only single-dimensional probability values, lacking interpretability for decision-making. When the system triggers a high-risk warning, the business side cannot automatically trace the specific causal transmission chain that led to the abnormal state, and it is even more difficult to generate scientific quantitative guidance suggestions by evaluating the counterfactual state after variable intervention.

[0004] In summary, how to integrate causal inference mechanisms into the temporal neural network framework, break the black box of the model while realizing the correlation of causal representation features, and endow the early warning system with structured attribution explanation and forward-looking intervention simulation capabilities, has become a technical challenge that needs to be solved in this field. Summary of the Invention

[0005] The purpose of this section is to outline some aspects of embodiments of the present invention and to briefly describe some preferred embodiments. Simplifications or omissions may be made in this section, as well as in the abstract and title of this application, to avoid obscuring the purpose of these documents; however, such simplifications or omissions should not be construed as limiting the scope of the invention.

[0006] In view of the aforementioned existing problems, this invention is proposed. Therefore, this invention provides a multi-source behavioral and physiological feature joint early warning method based on temporal deep learning to solve the problems mentioned in the background art.

[0007] To address the aforementioned technical problems, this invention provides the following technical solution: a multi-source behavior and physiological feature joint early warning method based on temporal deep learning, comprising: Acquire aligned multi-source time-series data, which includes various behavioral and physiological characteristics; The multi-source time-series data are processed in parallel to extract time-series features and learn causal structure information, which includes causal adjacency relationships and causal effect strengths between features. Based on the causal structure information, the temporal features are subjected to causal-guided attention weighting to generate deep fusion features; Based on the deep fusion features, an early warning signal is generated.

[0008] Compared with the prior art, the beneficial effects of the present invention are as follows: 1. This invention introduces a differentiable causal discovery algorithm to learn the causal structure information between multi-source features and transforms it into a static causal mask and dynamic modulation terms, which are then injected into the underlying computation of the self-attention mechanism. This causal-guided weighted processing rigidly blocks the information interaction between non-causal feature pairs from the network's physical structure, effectively removing statistical pseudo-correlation and environmental noise that are easily learned in traditional data-driven models. This ensures that deeply fused features retain only those with genuine physical triggering logic, reducing the false alarm rate.

[0009] 2. This invention can automatically search for and lock the key causal path that leads to risk in a directed acyclic graph by directly calling the weighted causal adjacency matrix and combining it with the real-time abnormal deviation of nodes, thus revealing the transmission chain of abnormal states intuitively and accurately.

[0010] 3. Furthermore, after identifying key causal paths, this invention, combined with pre-defined modifiable characteristic variables and a counterfactual simulation engine, can virtually simulate the expected risk reduction after altering specific subjective behaviors along the forward propagation of a structural equation model. Based on this, it automatically generates scientific and quantitative intervention suggestions containing specific numerical values ​​and natural language descriptions, enhancing the practical value and security effectiveness of early warning systems in real-world scenarios such as medical monitoring and intelligent security. Attached Figure Description

[0011] To more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort. Wherein: Figure 1 This is a flowchart illustrating the overall process of a multi-source behavior and physiological feature joint early warning method based on temporal deep learning, as described in one embodiment of the present invention. Detailed Implementation

[0012] To make the above-mentioned objects, features, and advantages of the present invention more apparent and understandable, specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of the present invention, and not all of them. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort should fall within the protection scope of the present invention.

[0013] Many specific details are set forth in the following description in order to provide a full understanding of the invention. However, the invention may also be practiced in other ways different from those described herein, and those skilled in the art can make similar extensions without departing from the spirit of the invention. Therefore, the invention is not limited to the specific embodiments disclosed below.

[0014] Secondly, the term "one embodiment" or "embodiment" as used herein refers to a specific feature, structure, or characteristic that may be included in at least one implementation of the present invention. The phrase "in one embodiment" appearing in different places in this specification does not necessarily refer to the same embodiment, nor is it a single or selective embodiment that is mutually exclusive with other embodiments.

[0015] Furthermore, in the description of this invention, it should be noted that the terms "upper," "lower," "inner," and "outer," etc., indicate the orientation or positional relationship based on the orientation or positional relationship shown in the accompanying drawings. These terms are used solely for the convenience of describing the invention and for simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation. Therefore, they should not be construed as limitations on the invention. In addition, the terms "first," "second," or "third" are used for descriptive purposes only and should not be construed as indicating or implying relative importance.

[0016] Example 1

[0017] Reference Figure 1 This is the first embodiment of the present invention, which provides a multi-source behavior and physiological feature joint early warning method based on temporal deep learning, including: S1: Obtain aligned multi-source time series data, which contains various behavioral and physiological characteristics.

[0018] It should be noted that conventional sliding window truncation methods for time series often destroy the potential causal time differences between features and easily introduce a large amount of redundant and worthless background data. In order to accurately capture the causal triggering relationships between multimodal data and provide high-quality input for the model, data reconstruction is required.

[0019] Furthermore, multi-source heterogeneous raw data streams are collected and constructed in real time.

[0020] Specifically, in specific application scenarios (such as wearable health monitoring or smart security), raw user characteristic data is acquired through deployed sensor networks. In this embodiment, this raw characteristic data is divided into two categories: one is physiological characteristic sequences, which are continuous, high-frequency sampled data, such as continuous heart rate, blood oxygen saturation, or skin conductance response; the other is behavioral characteristic sequences, which are discrete or low-frequency sampled data, such as gait frequency changes, action categories (e.g., falls), and human-computer interaction frequency. It should be noted that due to the differences in the attributes of the acquisition devices and sensors, these two data streams exhibit significant timestamp misalignment and sampling rate differences in their initial state, representing typical multi-source heterogeneous data.

[0021] Furthermore, preset anchor events are identified in the multi-source time-series data.

[0022] Specifically, in order to determine causal occurrence intervals with analytical value, this invention will continuously monitor the aforementioned heterogeneous raw data stream in real time within a relatively wide background time sliding window to identify anchor events. It should be explained that an anchor event refers to a mutation point or key action indicator that occurs in the behavioral or physiological dimension and exceeds a preset statistical safety boundary.

[0023] Specifically, a standardized mutation score is derived by calculating in real time the degree to which the feature values ​​of each dimension deviate from their historical baseline mean, and combining this with the historical dynamic fluctuation amplitude (e.g., standard deviation) of the feature. In this process, the invention assigns different weights to features of different dimensions. When the sum of the weighted mutation scores of all features exceeds a preset empirical trigger threshold, an abnormal mutation is determined to have occurred at the current moment, and the timestamp of that moment is recorded as the anchor event time point.

[0024] It should be noted that by adaptively calculating the fluctuation amplitude, the present invention can automatically filter out stable, non-abnormal daily background data, so that causal analysis is only activated when there are potential risk signs (such as a sudden increase in heart rate or sudden collapse), thereby reducing the continuous computing power consumption at the edge device.

[0025] Furthermore, using the anchor event's time point as the center, an asymmetric cause-and-effect time window is employed to extract data from multiple sources of time series data.

[0026] Specifically, existing technologies often employ symmetrical time windows centered on the mutation point (e.g., 5 minutes before and after) for data extraction. However, this invention argues that causality inherently possesses asymmetry in its temporal evolution scale. That is, the antecedents leading to anomalies often require a long evolutionary period to lie dormant and accumulate, while the consequences of anomalies can usually be confirmed within a shorter period. Therefore, this invention uses a defined anchor event time point as the time base to construct an asymmetric time extraction window. The duration of this window extending in the historical direction (i.e., the antecedent tracing window length) is strictly greater than the duration extending in the future direction (i.e., the consequence observation window length). For example, a 15-minute forward extension is used to fully capture the slowly changing behavioral trajectory that triggers the anomaly, while a 3-minute delay is used to assess the acute stress response of physiological characteristics. This asymmetric window is then used to extract data from the original data stream.

[0027] It's important to explain that the asymmetric window design aligns with the objective laws governing the evolution of people's physiological, pathological, or behavioral psychology. More importantly, it provides ample data margin for dependent variables in causal discovery algorithms, effectively preventing tracing failures caused by premature truncation of the causal chain at the head.

[0028] Furthermore, the extracted heterogeneous time-series data are interpolated, aligned, and normalized to form standardized analysis samples.

[0029] Specifically, low-frequency discrete behavioral data and high-frequency continuous physiological data, extracted within an asymmetric window, are mapped to a global discrete-time grid with a predefined uniform sampling frequency. For continuous physiological characteristic data, smooth interpolation techniques (e.g., cubic spline interpolation) are used to supplement missing time points to maintain the smooth evolution trend of the physiological curve. For discrete behavioral event data, forward padding or zero-order preservation strategies are employed to ensure that the continuous state of discrete events is correctly extended within the grid.

[0030] Meanwhile, after aligning the time axis as described above, to address the significant dimensional differences between different sensor data (e.g., heart rate beats / minute versus acceleration G), extreme value mapping or standardization methods are used to normalize the values ​​of each dimension, so that they converge to the same dimensionless interval (i.e., between 0 and 1).

[0031] It should be noted that, through the above alignment and normalization, the originally messy heterogeneous data will be reconstructed into a well-structured, time-synchronized multidimensional feature matrix, eliminating the influence of feature range caused by different dimensions.

[0032] S2: Perform parallel processing on multi-source time series data, extract time series features and learn causal structure information respectively. Causal structure information includes causal adjacency relationships between features and the strength of causal effects.

[0033] It should be noted that traditional multimodal deep learning models typically combine feature extraction and the learning of relationships between features within the same feedforward network framework. This highly entangled approach can easily lead the network to learn spurious correlations between data points rather than genuine physical triggering logic. To remove confounding factors from the data, this invention constructs a dual-branch parallel architecture: the first branch focuses on global contextual semantic encoding in the time dimension (i.e., extracting temporal features); the second branch independently focuses on latent topology generation in the variable dimension (i.e., learning causal structures).

[0034] Furthermore, extracting temporal features aims to capture the nonlinear dependencies of multi-source sequences over long time spans using a self-attention mechanism. In this process, to address the computational complexity and local semantic fragmentation issues inherent in traditional point-by-point computation methods, this invention employs a patch-based temporal fragmentation coding strategy.

[0035] Furthermore, the multi-source time series data is divided into multiple time series patches along the time dimension.

[0036] Specifically, by using a sliding window with a set length and step size, the aligned multi-source feature matrix is ​​truncated on the time axis and divided into multiple continuous, non-overlapping or partially overlapping time-series data blocks, i.e., time-series patches.

[0037] It should be noted that in real-world scenarios, data from a single time step (e.g., the ECG voltage value at 1 millisecond) often lacks actual physiological or behavioral semantics, while a temporal patch (e.g., a waveform segment containing a complete cardiac cycle, or a complete arm-raising motion trajectory) contains a complete and independent local physical state. Packaging time steps into patches can reduce the dimensionality of the sequence in the time dimension, reduce the memory usage and computational complexity of self-attention processing, and also preserve the semantic integrity of the local feature flow.

[0038] Furthermore, the temporal patch is input into a temporal encoder model based on a self-attention mechanism, and the encoded temporal features are output by calculating the interdependencies between patches.

[0039] Specifically, the data flow of this temporal encoder model is as follows: First, a learnable linear projection layer maps each temporal patch in the original data space to a high-dimensional latent feature space. Simultaneously, to prevent disruption of the temporal order of the sequence, positional encoding vectors are superimposed on the mapped features to preserve the directional information of temporal evolution. Then, the feature sequence carrying positional information is fed into a temporal encoder composed of alternating stacks of multi-head self-attention layers and multilayer perceptrons.

[0040] Simultaneously, in the self-attention computation phase, the model generates query variables, key variables, and value variables by performing linear transformations on different patch features, and calculates the similarity score between any two patch features, using this score as weights to weight and aggregate global information. Then, after forward propagation through the deep network, a encoded temporal feature matrix with global temporal dependencies is finally output. At this point, each feature vector in the feature matrix not only represents the local state of the current time patch but also represents information from other historical patches highly correlated with it within the entire time window.

[0041] Furthermore, it learns causal structure information. This branch is executed in parallel with the extraction of temporal features mentioned above, aiming to overcome the black box problem of network interpretability from a purely data-driven perspective, thereby constructing a directed acyclic causal topology graph between various physiological and behavioral feature variables.

[0042] Furthermore, a differentiable causal discovery algorithm is employed to learn a weighted adjacency matrix representing the directed acyclic causal relationships between features from multi-source time-series data or their statistical summaries in a data-driven manner.

[0043] Specifically, traditional causal discovery algorithms are typically based on discrete conditional independence tests. As the feature dimension increases, their computational cost explodes, and they cannot be integrated into end-to-end gradient updates in deep learning. Therefore, this invention introduces a differentiable causal discovery model within a continuous optimization framework. This causal discovery model aims to construct a joint optimization objective function that includes a data reconstruction loss, a sparsity penalty term, and a strictly acyclic constraint term.

[0044] Specifically, in this objective function, the mean squared error (MSE) is used for data reconstruction loss. Its main function is to encourage the model to try to predict the current feature variable using combinations of other feature variables. The smaller the prediction error of the model, the stronger the correlation between the features. The sparsity penalty term is introduced by L1 regularization, which forces a large number of weak or redundant connection weights to zero. Its significance lies in the fact that an abnormal event in the objective world (e.g., a fall causing a sudden change in heart rate) is usually caused by only a few of the most direct antecedents. Therefore, the sparsity penalty term is mainly responsible for conforming to this objective law. The strict acyclic constraint term constructs a mathematical boundary through specific matrix trace exponent operations and forcibly blocks the occurrence of cyclic paths in the graph structure during the backpropagation iteration of the model (e.g., prohibiting the causal dead loop paradox of A leading to B, B leading to C, and C in turn leading to A).

[0045] Specifically, in this embodiment, the trace exponent operation of a specific matrix is ​​represented as follows: Where W is the weighted adjacency matrix; d represents the Hadamard product, which is the square of each element in the matrix; d is the total number of feature nodes. It is the trace of a matrix, which is the sum of all elements along the main diagonal (from the top left to the bottom right) of the matrix.

[0046] It should be noted that, based on the joint optimization of the objective function, through iterative updates using gradient descent and the Lagrange multiplier method, the model will eventually converge and output a stable weighted adjacency matrix.

[0047] Furthermore, the topological structure of this weighted adjacency matrix is ​​represented as a causal adjacency relationship, and the weight values ​​of this weighted adjacency matrix represent the causal effect strength.

[0048] Specifically, after the model training is complete and the weighted adjacency matrix is ​​output, the physical mapping relationship of the elements within the matrix is ​​as follows: The spatial distribution of non-zero elements in a matrix (i.e., topological structure) clearly indicates whether a certain behavior or physiological characteristic is the cause of a change in another characteristic, revealing the true causal adjacency relationship between multiple source variables; The absolute value of the non-zero elements in the matrix quantifies the degree and sensitivity of the unidirectional triggering effect, i.e., the strength of the causal effect.

[0049] S3: Based on causal structure information, attention-weighted processing guided by causal factors is applied to temporal features to generate deep fusion features.

[0050] It should be noted that traditional multimodal data fusion typically relies on standard data-driven self-attention mechanisms. However, this mechanism essentially performs fully connected statistical co-occurrence calculations, which easily assigns high attention weights to pseudo-correlated feature pairs that merely co-occur within a time window but lack actual physical triggering logic. This introduces a large amount of environmental noise, leading to a decrease in the model's robustness in complex scenarios. To address this issue, this invention uses the causal structure information extracted from S2 as a strong inductive bias, injecting it into and reshaping the underlying attention computation flow of the neural network to achieve feature desiccation and deep fusion.

[0051] Furthermore, a static causal mask is generated based on causal adjacency relationships.

[0052] Specifically, in the architecture of deep neural networks, in order to structurally block redundant information transmission between feature variables that have no real causal relationship, it is first necessary to parse the causal structure information of the output. For the extracted adjacency matrix indicating the directed acyclic topology between features, it is transformed into a structured boundary in the network feature space, i.e., a static causal mask is constructed. The mapping logic of this static causal mask is as follows: When a causal adjacency relationship indicates that there is no causal connection path between two physiological or behavioral feature nodes (e.g., a sudden drop in ambient temperature and abnormal gait), a very large negative penalty value (approaching negative infinity in network computation logic) is assigned to the corresponding position in the mask matrix; conversely, if a direct causal connection exists, a neutral value is assigned.

[0053] It should be noted that the advantage of this mapping is that when the attention network applies an exponentially normalized activation function (e.g., the Softmax function) for probability transformation, the negative penalty value in the mask will be mathematically mapped to zero weight. This means that the present invention can sever the information interaction path between non-causal feature pairs from the bottom layer of the network physical architecture, forcing the model to avoid allocating computing power to focus on pseudo-correlated noise that is merely statistically co-occurring, thereby improving the model's anti-interference ability when dealing with unknown, complex, and interference-filled dynamic environments.

[0054] Furthermore, this invention introduces a dynamic modulation term determined by the intensity of the causal effect when calculating attention scores.

[0055] Specifically, the causal effect values ​​representing physical intensity are directly extracted from S2, and combined with learnable scaling parameters and smoothing gain mechanisms in the network to construct a dynamic modulation term. This dynamic modulation term can nonlinearly map the absolute intensity of the causal effect to the numerical reward or bias score in the attention mechanism.

[0056] It's important to explain that the triggering levels of causal events in nature vary significantly (e.g., the causal effect of vigorous running on increased heart rate is necessarily much stronger than that of slow walking). This dynamic modulation term can transform this objectively existing physical triggering level into an attention bias during neural network computation. This allows the model to spontaneously and dynamically allocate more attention to feature association paths with stronger and more direct causal effects when processing real-time data streams. Simultaneously, the introduction of learnable scaling parameters ensures that the model can automatically find and optimize the optimal balance between pre-defined causal prior knowledge and the current real-time input feature state, avoiding overfitting caused by prior knowledge.

[0057] Furthermore, during the calculation of attention scores using the self-attention mechanism, the static causal mask and dynamic modulation terms are applied to constrain or enhance the attention weights between causally related feature pairs, generating deeply fused features.

[0058] Specifically, during the model's forward propagation, the temporal features extracted by S2 are first input into the feature mapping layer. Through linear transformations within the network, query feature matrices, key feature matrices, and value feature matrices containing actual physical state content are generated for similarity retrieval and matching. In the feature interaction calculation stage, the model first calculates the inner product between the query features and key features to obtain an initial statistical similarity score matrix entirely based on the current data. Then, before entering the normalization activation process, the model parallelly injects the generated static causal mask and the generated dynamic modulation term into the aforementioned similarity score using tensor addition.

[0059] Specifically, after the above processing, the model's original attention score at this processing node is essentially reshaped by a dual causal constraint: The first layer is the constraint effect: through the bottom-level injection of static causal mask, the scores of features that do not have causal logic support are instantly suppressed to negative infinity, and the weights are completely reduced to zero after normalization.

[0060] The second layer is the enhancement effect: through the numerical injection of dynamic modulation terms, features with strong causal support receive positive bias rewards for their scores, and their representation weight is further amplified after normalization.

[0061] Furthermore, the attention probability matrix, after being reshaped by the aforementioned dual causal model, is weighted and aggregated with the value feature matrix. Then, after passing through a feedforward neural network and hierarchical normalization operations, the final deep fusion feature matrix can be output.

[0062] S4: Generate early warning signals based on deep fusion characteristics.

[0063] It should be noted that most existing early warning systems are end-to-end black-box models, outputting only a single-dimensional risk probability value. While this approach can indicate risk to some extent, in high-value specific application scenarios (such as intelligent medical monitoring and high-risk operation safety), business users or end users not only need to know whether danger will occur, but also why danger will occur and how to intervene to avoid it. Therefore, this invention, while generating basic early warning probabilities, incorporates the topology structure in S2 to construct a complete early warning and decision-making mechanism that includes risk prediction, automatic source tracing, and counterfactual simulation.

[0064] Furthermore, the deep fusion features are pooled to obtain a global context representation, which is then input into the early warning classifier to output an early warning signal.

[0065] Specifically, after weighting in S3, the model outputs a feature sequence matrix containing multiple time dimensions. To transform this serialized data into fixed-length features suitable for classification tasks, a global average pooling operation is first performed on the time dimension of the feature sequence. This pooling operation condenses the features of multi-source heterogeneous data within the entire asymmetric time window. The single fixed-length vector output after pooling (i.e., the global context representation) is stripped of specific temporal length constraints but fully preserves the global spatial semantics and comprehensive situational information of the features evolving under causal constraints. This global context representation is then input into a pre-built early warning classifier. This classifier consists of a multi-layer feedforward neural network and a non-linear activation function (e.g., the sigmoid function). Inside the classifier, the feature vector is compressed into a numerical range of 0 to 1 after weight mapping and bias stacking in fully connected layers, generating the final early warning signal.

[0066] It should be noted that this warning signal is a continuous probability value predicting the occurrence of a specific high-risk event (such as sudden fall, fainting, or acute arrhythmia) in the target individual within the short term in the future (i.e., within the consequence observation time window set in S1).

[0067] Furthermore, after the warning signal triggers the warning conditions, based on the learned causal structure information and the values ​​of each feature in the current sample, the key causal path leading to the warning is automatically traced and identified, and a structured attribution explanation is generated.

[0068] Specifically, a risk and security threshold based on business experience is pre-set. When the probability value of the real-time calculated warning signal reaches or exceeds this threshold, the warning trigger condition is determined to be met. At this time, the warning system issues an alarm command and immediately activates an automatic tracing engine in the background. It is important to emphasize that, unlike the traditional model's analysis mechanism that relies on black-box gradients, the tracing engine of this invention directly calls the weighted causal adjacency matrix representing the objective physical topology extracted in S2 through the differentiable causal discovery algorithm.

[0069] Simultaneously, during the source tracing calculation, the real-time abnormal deviation of each physiological and behavioral feature in the current analysis sample (i.e., the absolute difference between the current feature value and its historical stable baseline value) is first extracted. Then, along the directed edges indicated by the causal adjacency matrix, the transmission effect of connected feature nodes in the network is comprehensively evaluated. The specific evaluation process is as follows: By performing a correlation product operation between the network topology weights learned over a long period of time (i.e., the intensity of causal effect noise) and the real-time difference (i.e., the degree of mutation) of the two connected features at the current moment when they undergo abrupt changes, a real-time path activation score for a directed connection is obtained.

[0070] Specifically, in the entire directed acyclic graph (DAG) containing multiple feature nodes, a graph search algorithm (e.g., longest path search) is then used to find one or more connected paths with the highest cumulative real-time activation scores and lock them as key causal paths.

[0071] Furthermore, after identifying the key causal path, one or more modifiable characteristic variables in the causal path are selected, counterfactual simulations are performed on the modifiable characteristic variables, the expected early warning risk after intervention is calculated, and quantitative intervention recommendations are generated.

[0072] Specifically, in real-world applications, risk factors can be categorized into uncontrollable variables (e.g., environmental temperature, individual age) and subjectively controllable variables (e.g., current cadence, posture, breathing rhythm). Based on this, this invention pre-configures a dictionary of modifiable feature variables in the background. After extracting the key causal path, this dictionary automatically traverses and filters upstream nodes of that path, extracting feature variables within its scope. Next, a counterfactual simulation engine is activated to answer the decision-making question: how much would the system risk be reduced if a certain behavior were changed immediately?

[0073] Furthermore, in the simulated data flow, this invention, based on the intervention logic in the field of causal inference, virtually severs the connection between the intrusive variable and its upstream historical cause nodes in the model, forcibly setting the value of the variable to a hypothetical safe target value (e.g., setting the current excessively high step frequency feature to be forcibly reduced by 30%). Subsequently, based on the causal structure information extracted in S2, a forward propagation operation flow based on structural equation model is constructed. Specifically, the fixed weighted causal adjacency matrix is ​​used as the information transmission coefficient matrix between network nodes. When the upstream intrusive variable is virtually changed to the target value, the target value is multiplied by its corresponding causal edge weight, and a preset baseline bias is added to update the values ​​of its directly adjacent downstream feature nodes in sequence; subsequently, the updated downstream nodes are again used as preconditions to continue triggering the numerical calculation of the next lower layer nodes. Through this layer-by-layer, downstream linear or nonlinear weighted summation operation along the topology graph, the new numerical state of each physiological feature (e.g., expected heart rate, expected blood pressure) at the end of the path is re-deduced under this assumption. Finally, this new feature matrix, modified with counterfactual correction and representing possible future states, is fed back into the early warning classifier to calculate the expected early warning risk probability after intervention. By comparing the difference in risk probabilities before and after intervention, the effectiveness of different intervention plans is automatically evaluated, and the plan that can reduce the early warning signal below the safety threshold is selected as a quantitative intervention recommendation.

[0074] Furthermore, the structured attribution explanations or quantitative intervention recommendations will be output in a format that includes the causal chain, specific values ​​of each feature node on the path, the degree of anomaly, and natural language text descriptions.

[0075] Specifically, in order to enable terminal business personnel without a computer professional background (such as on-site security personnel, clinical medical staff, or ordinary users of wearable devices) to intuitively and quickly understand the causal inference results of this invention, it is also necessary to format, convert, and package the generated matrix and score data for display.

[0076] Specifically, by invoking a preset Natural Language Generation (NLG) template, the data is mapped into a comprehensive, human-readable early warning report. The output data package and front-end display interface include: Graphical causal chain: A risk triggering graph that is visually drawn using nodes and directional arrows; Node status parameters: directly label the specific sensor readings of each feature node in the current causal chain, and use different color levels (e.g., red represents extreme anomaly, yellow represents slight deviation) to indicate the degree of anomaly; Natural Language Description: Integrates early warning, source tracing, and intervention logic into actionable guidance text. For example, the output text might be: High-risk warning triggered! Source tracing analysis shows that the target individual's recent excessively high pace has led to a sharp disruption in respiratory rate, increasing the probability of arrhythmia to 88%. Intervention recommendation: If the pace is immediately reduced by 30% and a deep breathing approach is adopted, the risk of arrhythmia is expected to drop significantly to the safe range of 32% within 3 minutes. Please take immediate deceleration measures! It should be noted that 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 preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention, and all such modifications or substitutions should be covered within the scope of the claims of the present invention.

Claims

1. A multi-source behavior and physiological feature joint early warning method based on temporal deep learning, characterized in that, include: Acquire aligned multi-source time-series data, which includes various behavioral and physiological characteristics; The multi-source time-series data are processed in parallel to extract time-series features and learn causal structure information, which includes causal adjacency relationships and causal effect strengths between features. Based on the causal structure information, the temporal features are subjected to causal-guided attention weighting to generate deep fusion features; Based on the deep fusion features, an early warning signal is generated.

2. The multi-source behavior and physiological feature joint early warning method based on temporal deep learning as described in claim 1, characterized in that, The acquisition of aligned multi-source time series data includes: Identify preset anchor events in the multi-source time-series data; Centered on the time point of the anchor event, an asymmetric cause-and-effect time window is used to extract and align the multi-source time series data to form an analysis sample.

3. The multi-source behavior and physiological feature joint early warning method based on temporal deep learning as described in claim 1, characterized in that, The extraction of temporal features includes: The multi-source time-series data is divided into multiple time-series patches along the time dimension; The time series patch is input into a time series encoder model based on a self-attention mechanism. By calculating the interdependencies between the patches, the encoded time series features are output.

4. The multi-source behavior and physiological feature joint early warning method based on temporal deep learning as described in claim 1, characterized in that, The learned causal structure information includes: A differentiable causal discovery algorithm is used to learn a weighted adjacency matrix representing the directed acyclic causal relationship between features from the multi-source time series data or its statistical summary in a data-driven manner. The topological structure of the adjacency matrix represents the causal adjacency relationship, and the weight values ​​of the adjacency matrix represent the strength of the causal effect.

5. The multi-source behavior and physiological feature joint early warning method based on temporal deep learning as described in claim 1, characterized in that, The temporal features are subjected to causal-guided attention-weighted processing, including: Based on the aforementioned causal adjacency relationship, a static causal mask is generated; In the process of calculating attention scores using the self-attention mechanism, the static causal mask is applied to constrain or enhance the attention weights between feature pairs that are causally related.

6. The multi-source behavior and physiological feature joint early warning method based on temporal deep learning as described in claim 5, characterized in that, Also includes: When calculating the attention score, a dynamic modulation term determined by the strength of the causal effect is introduced; The dynamic modulation term adjusts the attention score in real time based on the strength of the causal effect between features, so that feature pairs with stronger causal effects receive higher attention bias.

7. The multi-source behavior and physiological feature joint early warning method based on temporal deep learning as described in claim 1, characterized in that, The generation of the early warning signal includes: The deep fusion features are pooled to obtain a global context representation; The global context representation is input into an early warning classifier, and the output probability value representing future risk is used as the early warning signal.

8. The multi-source behavior and physiological feature joint early warning method based on temporal deep learning as described in claim 1, characterized in that, The method further includes: After the warning signal triggers the warning conditions, based on the learned causal structure information and the values ​​of each feature in the current sample, the key causal path that led to the warning is automatically traced and identified, and a structured attribution explanation is generated.

9. The multi-source behavior and physiological feature joint early warning method based on temporal deep learning as described in claim 8, characterized in that, Also includes: After identifying the key causal path, select one or more modifiable feature variables from the path; Counterfactual simulations are performed on the modifiable characteristic variables. The simulations calculate the expected early warning risks after intervention by setting target values ​​for the modifiable characteristic variables and calculating the impact of the target values ​​set by forward propagation of the causal structure information, thereby generating quantitative intervention recommendations.

10. The multi-source behavior and physiological feature joint early warning method based on temporal deep learning as described in claim 8 or 9, characterized in that, The structured attribution explanation or the quantitative intervention suggestion is output in a format that includes the causal chain, the specific values ​​of each feature node on the path, the degree of anomaly, and a natural language text description.

Images