Intelligent decision support system and method for elderly care robot based on knowledge graph

Abstract

The application discloses an old-age care robot intelligent decision support system and method based on a knowledge graph, relates to the field of intelligent robots, and dynamically constructs a dynamic situation graph capable of representing a current complex interaction relationship by fusing video, audio and physiological and other multi-modal sensor data and utilizing a static knowledge graph to perform semantic analysis on key entities such as people, objects and environments in a real-time situation. Based on this, a graph attention network is introduced to perform deep relationship reasoning and situation evolution perception on the dynamic situation graph, so as to quantify the correlation and conflict between different information sources, identify a user real intention with the highest credibility in multiple possibilities, and finally generate and execute to obtain robot actions. In this way, the problem of inaccurate intention inference caused by multi-modal information conflict can be solved, and the decision intelligence and service initiative of the old-age care robot in a complex dynamic environment are significantly improved.

Description

Technical Field

[0001] This application relates to the field of intelligent robots, and more specifically, to an intelligent decision support system and method for elderly care robots based on knowledge graphs. Background Technology

[0002] With the accelerating aging of the global population and changes in family structures, society's demand for efficient, precise, and humane elderly care services is becoming increasingly urgent. Traditional elderly care models are proving inadequate in the face of challenges such as staff shortages, high service costs, and the inability to provide 24 / 7 support. Against this backdrop, elderly care service robots integrating artificial intelligence serve as an important supplementary solution. Their core value lies in continuously sensing multimodal information such as the elderly's language, behavior, and physiological characteristics, and proactively providing support. To achieve this goal, the key is to endow the robots with the ability to deeply understand complex elderly care scenarios. Knowledge graph technology, through its powerful structured representation capabilities of entities, relationships, and attributes, provides an ideal framework for building the robot's scenario cognition and decision-making knowledge base. It can connect fragmented multimodal data into a meaningful whole, laying the foundation for achieving higher-level intelligent decision-making.

[0003] However, most existing solutions rely on pre-built static domain knowledge graphs. While these graphs contain rich, general knowledge about elderly care, they are inherently static and struggle to capture and model the rapidly changing dynamic contextual information in specific interaction scenarios in real time. This results in relatively rigid decision-making logic for robots, making them unable to flexibly adapt to the dynamic changes of people, objects, and the environment in real-world settings. This limitation is particularly pronounced when dealing with deep-level user intent inference problems, especially when faced with complex situations where user statements do not match their internal state. For example, when an elderly user verbally claims to be "in good health" to avoid troubling others, but their gait sensor data shows slow movement and physiological monitoring equipment detects abnormal heart rate fluctuations, existing systems often lack effective mechanisms to identify and quantify the inherent contradictions between these multimodal information. This insufficient ability to process weak signals and conflicting information prevents the system from penetrating surface information and effectively bridging the semantic gaps between different information sources, thus hindering accurate judgments about the user's true health status or potential needs.

[0004] Therefore, there is an urgent need for an optimized intelligent decision support system and method for elderly care robots based on knowledge graphs. Summary of the Invention

[0005] This application is made in order to solve the above-mentioned technical problems.

[0006] According to one aspect of this application, a knowledge graph-based intelligent decision support method for elderly care robots is provided, which includes: acquiring the raw sensor data stream of the target elderly object, wherein the raw sensor data stream includes video frame data, audio block data and physiological signal data.

[0007] Based on static knowledge graphs, the original sensor data stream is vectorized into contextual entity vectors to obtain a set of contextual entity vectors.

[0008] A dynamic context graph is constructed based on a set of context entity vectors.

[0009] Context awareness and node updates are performed on dynamic context graphs based on graph attention networks to obtain updated node embeddings.

[0010] The updated node embeddings are subjected to intent classification and credibility calculation to obtain a sorted intent list.

[0011] Decisions are generated and executed based on a sorted list of intentions to obtain robot actions.

[0012] According to another aspect of this application, a knowledge graph-based intelligent decision support system for elderly care robots is provided, which includes: a raw sensor data acquisition module for acquiring raw sensor data streams of the target elderly object, the raw sensor data streams including video frame data, audio block data and physiological signal data.

[0013] The context entity vectorization module is used to vectorize the raw sensor data stream into context entities based on a static knowledge graph to obtain a set of context entity vectors.

[0014] The Dynamic Context Graph Construction Module is used to construct dynamic context graphs based on a set of context entity vectors.

[0015] The context awareness and node update module is used to perform context awareness and node update on dynamic context graphs based on graph attention networks to obtain updated node embeddings.

[0016] The intent classification and credibility calculation module is used to classify intents and calculate credibility on the updated node embeddings to obtain a sorted list of intents.

[0017] The decision generation and execution module is used to generate and execute decisions based on a sorted list of intentions to obtain robot actions.

[0018] Compared with existing technologies, this application provides a knowledge graph-based intelligent decision support system and method for elderly care robots. It integrates multimodal sensor data (video, audio, and physiological data) and utilizes a static knowledge graph to perform semantic analysis on key entities such as people, objects, and the environment in real-time situations, thereby dynamically constructing a dynamic context graph that represents the current complex interactive relationships. Based on this, a graph attention network is introduced to perform deep relational reasoning and context evolution perception on the dynamic context graph, quantifying the associations and conflicts between different information sources. This identifies the user's true intent with the highest credibility among multiple possibilities and ultimately generates and executes the robot's actions. This solves the problem of inaccurate intent inference caused by multimodal information conflicts, thus significantly improving the decision-making intelligence and service initiative of elderly care robots in complex dynamic environments. Attached Figure Description

[0019] The above and other objects, features, and advantages of this application will become more apparent from the more detailed description of the embodiments of this application in conjunction with the accompanying drawings. The drawings are provided to further illustrate the embodiments of this application and form part of the specification. They are used together with the embodiments of this application to explain this application and do not constitute a limitation thereof. In the drawings, the same reference numerals generally represent the same components or steps.

[0020] Figure 1 This is a flowchart of a knowledge graph-based intelligent decision support method for elderly care robots according to an embodiment of this application.

[0021] Figure 2 This is a data flow diagram of the knowledge graph-based intelligent decision support method for elderly care robots according to an embodiment of this application.

[0022] Figure 3 This is a flowchart of sub-step S2 of the knowledge graph-based intelligent decision support method for elderly care robots according to an embodiment of this application.

[0023] Figure 4 This is a flowchart of sub-step S21 of the knowledge graph-based intelligent decision support method for elderly care robots according to an embodiment of this application.

[0024] Figure 5 This is a flowchart of sub-step S3 of the knowledge graph-based intelligent decision support method for elderly care robots according to an embodiment of this application.

[0025] Figure 6 This is a flowchart of sub-step S31 of the knowledge graph-based intelligent decision support method for elderly care robots according to an embodiment of this application.

[0026] Figure 7This is a flowchart of sub-step S5 of the knowledge graph-based intelligent decision support method for elderly care robots according to an embodiment of this application.

[0027] Figure 8 This is a block diagram of a knowledge graph-based intelligent decision support system for elderly care robots according to an embodiment of this application. Detailed Implementation

[0028] Embodiments of this disclosure will now be described in more detail with reference to the accompanying drawings. While some embodiments of this disclosure are shown in the drawings, it should be understood that this disclosure can be implemented in various forms and should not be construed as limited to the embodiments set forth herein. Rather, these embodiments are provided to provide a more thorough and complete understanding of this disclosure. It should be understood that the accompanying drawings and embodiments of this disclosure are for illustrative purposes only and are not intended to limit the scope of protection of this disclosure.

[0029] To address the problems mentioned above, this application proposes a knowledge graph-based intelligent decision support method for elderly care robots. Figure 1 This is a flowchart of a knowledge graph-based intelligent decision support method for elderly care robots according to an embodiment of this application. Figure 2 This is a data flow diagram for a knowledge graph-based intelligent decision support method for elderly care robots according to embodiments of this application. (See diagram below.) Figure 1 and Figure 2 As shown, the knowledge graph-based intelligent decision support method for elderly care robots includes the following steps: S1, acquiring the original sensor data stream of the target elderly object, the original sensor data stream including video frame data, audio block data, and physiological signal data; S2, based on a static knowledge graph, vectorizing the original sensor data stream into contextual entities to obtain a set of contextual entity vectors; S3, constructing a dynamic contextual graph based on the set of contextual entity vectors; S4, performing contextual perception and node updating on the dynamic contextual graph based on graph attention networks to obtain updated node embeddings; S5, performing intent classification and credibility calculation on the updated node embeddings to obtain a sorted intent list; S6, generating and executing decisions based on the sorted intent list to obtain robot actions.

[0030] In the aforementioned knowledge graph-based intelligent decision support method for elderly care robots, step S1 involves acquiring the raw sensor data stream of the target elderly individual. This raw sensor data stream includes video frame data, audio block data, and physiological signal data. It should be understood that because the state expressions of the elderly are multi-dimensionally correlated—body movements and facial expressions reflect external behavioral tendencies, voice content and tone convey linguistic intentions, and physiological indicators directly characterize internal health—relying on only a single type of data can lead to incomplete state perception. For example, when an elderly person verbally denies physical discomfort to avoid causing trouble, but simultaneously exhibits abnormal heart rate and painful expressions, a single data point can directly cause bias in subsequent intention inferences. Therefore, this application utilizes multimodal sensor deployment adapted to elderly care scenarios to simultaneously acquire video frame data, audio block data, and physiological signal data of the target elderly individual, thereby providing real-time raw data support for subsequent analysis. This effectively fills the gap in contextual information under a single data dimension, avoids misjudgment of context due to data bias, and ensures that the subsequent dynamic context map can accurately associate key entities and interaction relationships of people (elderly), objects (sensors), and environment (living space). At the same time, it provides original data basis for quantifying the correlation and conflict between different information sources, and ultimately lays a complete data foundation for accurately identifying the user's true intention and improving the decision-making intelligence of elderly care robots.

[0031] Specifically, in one possible embodiment, the implementation process of step S1 is as follows: Video frame data is collected by high-definition network cameras deployed on the ceiling or walls of each activity area. The camera resolution is set to 1920×1080 pixels, the frame rate is fixed at 25 frames / second, and the collection angle covers the complete field of vision of the elderly's main activity postures such as standing, sitting and walking. During the collection process, frame data is extracted according to a 3-second time window, 10 key images are extracted in each window, and 1 frame is extracted at an interval of 0.3 seconds. The data is stored in JPEG format to balance data accuracy and processing efficiency, ensuring that the elderly's limb movements (such as hand covering the chest, slow gait) and facial expressions (such as furrowed brows, downturned corners of the mouth) can be completely captured. Audio block data is acquired using an omnidirectional digital microphone deployed coaxially with the camera. The microphone sampling rate is set to 16kHz, and the sampling bit depth is 16 bits. The acquisition range covers an acoustic area with a radius of 5 meters, accurately capturing the elderly person's speech content (such as "I'm fine" or "Get me some water") and tone changes (such as weakness or rapid speech). The acquisition process is synchronized with the time window of the video frame data, generating one audio block data every 3 seconds and storing it in WAV format to preserve complete speech rhythm and emotional characteristics. Physiological signal data is acquired using a medical-grade smart bracelet worn by the target elderly person. The bracelet has a built-in photoelectric heart rate sensor, blood oxygen saturation sensor, and three-axis accelerometer, which can collect the elderly person's heart rate (acquisition frequency 1 time / second), blood oxygen saturation (acquisition frequency 1 time / 5 seconds), and basic activity level (acquisition frequency 1 time / second) in real time. The acquired data is synchronized in real time with the local data processing module of the elderly care robot via Bluetooth Low Energy technology. During the synchronization process, the physiological signal data is standardized and can be converted to JSON format to ensure that the data can be directly accessed in the subsequent multimodal signal feature extraction stage.

[0032] In the aforementioned knowledge graph-based intelligent decision support method for elderly care robots, step S2 involves vectorizing the raw sensor data stream into contextual entities based on a static knowledge graph to obtain a set of contextual entity vectors. It should be understood that since the video frame data, audio block data, and physiological signal data contained in the raw sensor data stream are unstructured or time-series raw information, they cannot be directly semantically associated with the structured entities and relationships in the static knowledge graph. Furthermore, the user's health background, medical history, and other information stored in the static knowledge graph are crucial for understanding the current context. Therefore, this application further performs contextual entity vectorization processing on the raw sensor data stream based on the static knowledge graph, thereby transforming the unstructured real-time sensor data into a set of contextual entity vectors that can be uniformly calculated with the entities in the static knowledge graph. This enables semantic alignment between multimodal real-time data and the static knowledge graph, eliminating the representation gap between different data types, providing a structured and computable entity foundation for the subsequent construction of dynamic context graphs, and avoiding contextual understanding biases caused by inconsistent data formats.

[0033] In particular, in one specific embodiment, Figure 3 This is a flowchart of sub-step S2 of the knowledge graph-based intelligent decision support method for elderly care robots according to an embodiment of this application. Figure 3 As shown, step S2 includes: S21, performing synchronous feature extraction and quantization of multimodal signals on video frame data, audio block data and physiological signal data in the original sensor data stream to obtain a quantized feature set; S22, converting each feature label in the quantized feature set into an observed entity; S23, using the user ID as the entry point, performing a graph query in the static knowledge graph to obtain the background entity; S24, inputting the observed entity and the background entity into the entity embedding model to obtain the set of context entity vectors.

[0034] Specifically, step S21 involves extracting and quantizing the synchronous features of multimodal signals from the video frame data, audio block data, and physiological signal data in the original sensor data stream to obtain a quantized feature set. It should be understood that, in elderly care scenarios, video frames, audio blocks, and physiological signals belong to different modalities, and their data structures and time dimensions differ. Processing them individually can easily lead to spatiotemporal asynchrony and fragmented information, failing to fully represent the real-time state of the elderly. For example, relying solely on video cannot determine whether abnormal movements are related to physiological discomfort. Therefore, this application achieves multimodal data synchronization through timestamp calibration, and then uses a dedicated mechanism to extract and quantify features for each modality's characteristics, thereby obtaining a unified format of feature information covering the behavior, emotions, and physiological conditions of the elderly. This avoids feature loss or conflict caused by data heterogeneity and asynchrony, providing accurate and complete basic data support for subsequent contextualization and semantic analysis, ensuring that subsequent decision-making can comprehensively reflect the current state of the elderly, and reducing decision-making bias caused by incomplete information.

[0035] In particular, in one specific embodiment, Figure 4 This is a flowchart of sub-step S21 of the knowledge graph-based intelligent decision support method for elderly care robots according to an embodiment of this application. Figure 4 As shown, step S21 includes: S211, inputting video frame data into the visual feature extraction branch to obtain action labels and their action confidence, as well as expression labels and their expression confidence; S212, inputting audio block data into the auditory feature extraction branch to obtain speech text and emotion labels and their emotion confidence; S213, inputting physiological signal data into the physiological feature extraction branch to obtain physiological indicators.

[0036] More specifically, in step S211, video frame data is input into the visual feature extraction branch to obtain action labels and their action confidence scores, as well as expression labels and their expression confidence scores. It should be understood that since video frame data is a continuous image sequence containing only pixel-level visual information, it cannot directly reflect the behavioral state and emotional tendencies of the target elderly person. If directly used for subsequent context analysis, it would fail to accurately capture the user's action intentions and emotional state due to its single information dimension and lack of semantic labeling. Therefore, this application further inputs the video frame data into a specially constructed visual feature extraction branch. Through targeted feature extraction and semantic mapping, action labels and expression labels with confidence scores are generated, thereby transforming the original visual data into structured information with semantic attributes and quantifiable confidence scores. This provides accurate visual dimension basis for subsequent dynamic context map construction, enabling the robot to judge user behavior risks through action labels and assist in identifying user emotional needs through expression labels. Simultaneously, high-reliability visual information is filtered using confidence scores to avoid interference from low-reliability data in subsequent intention inference, thus improving the accuracy of context perception.

[0037] Specifically, in one possible embodiment, step S211 is implemented as follows: The visual feature extraction branch includes an action recognition submodule and an expression recognition submodule. For obtaining action labels and action confidence scores, continuous video frames are input into a 3D-CNN network. This network identifies the action patterns of the target elderly person, such as getting up, walking, falling, and sitting, by capturing inter-frame spatiotemporal correlation features. Action labels are determined based on the category probability distribution output by the network, and the highest probability value is used as the action confidence score. For example, when recognizing a fall action, a fall label and a confidence score of 0.92 are output. For obtaining expression labels and expression confidence scores, a facial landmark detection algorithm is used to locate key areas such as the eyes and corners of the mouth from video frames, extract facial contours and muscle movement features, and input them into an expression classification model, such as a ResNet-based expression recognition network, to identify and output expression labels such as pain, calmness, and joy. Then, the probability of each category is calculated using the Softmax function, and the probability of the corresponding expression category is used as the expression confidence score. For example, when identifying a pain expression, the pain label and a confidence score of 0.88 are output. Finally, the action label, action confidence score, expression label, and expression confidence score are integrated to form a structured feature output in the visual dimension.

[0038] More specifically, in step S212, the audio block data is input into the auditory feature extraction branch to obtain speech text and emotion tags and their emotion confidence scores. It should be understood that since the audio block data is a continuous sound wave signal, containing only frequency and amplitude information, it cannot be directly converted into understandable language content and emotional tendencies. If directly used for subsequent decision-making, the lack of textual semantics and emotion labels would prevent the capture of user voice requests or emotional states, thus hindering accurate inference of the user's auditory intent. Therefore, this application further inputs the audio block data into a customized auditory feature extraction branch, using a dual-path processing of speech-to-text and emotion analysis to generate speech text and credible emotion tags. This transforms the original audio signal into structured information with both textual semantics and emotional attributes. This provides auditory dimension support for subsequent contextual analysis, enabling the robot to directly obtain the user's explicit requests through speech text and perceive the user's potential emotions through emotion tags, such as anxiety corresponding to possible health concerns. Simultaneously, emotion confidence scores are used to filter highly reliable emotional information, avoiding emotional misjudgments caused by environmental noise, ensuring the effectiveness of auditory data during multimodal information fusion, and improving the comprehensiveness of user intent inference.

[0039] Specifically, in one possible embodiment, step S212 is implemented as follows: The auditory feature extraction branch includes a speech-to-text submodule and an emotion recognition submodule. For speech-to-text acquisition, the audio block is first denoised using an adaptive filtering algorithm to remove environmental noise, and then input into an Automatic Speech Recognition (ASR) model. This model analyzes the spectral characteristics and prosody of the audio to convert the sound wave signal into a text sequence, such as converting the audio of "I'm a little dizzy" into corresponding text. For emotion tag and emotion confidence acquisition, Mel-frequency cepstral coefficients (MFCCs) are extracted from the denoised audio block as emotion features and input into an emotion classification model, such as an LSTM-based emotion recognition network, to identify and output emotion tags such as calm, anxiety, and help-seeking. Then, the probability value of the corresponding category in the model output layer is calculated as the emotion confidence. For example, when recognizing the emotion of help-seeking, the "help-seeking" tag and a confidence score of 0.91 are output. Finally, the speech-to-text, emotion tags, and emotion confidence scores are integrated to form a structured feature output in the auditory dimension.

[0040] More specifically, in step S213, physiological signal data is input into the physiological feature extraction branch to obtain physiological indicators. It should be understood that since physiological signal data is raw waveform data, containing only curve information of voltage or current changes over time, it cannot directly reflect the health status of the target elderly person. If directly used for subsequent decision-making, the lack of specific health indicator identifiers would make it impossible to determine whether the user's physiology is abnormal, such as excessively high heart rate or low blood oxygen, and thus unable to accurately identify the user's health-related needs, such as care needs when feeling unwell. Therefore, this application further inputs physiological signal data into a dedicated physiological feature extraction branch, generating specific physiological indicator values ​​through signal filtering and feature analysis, thereby transforming the raw physiological waveform into quantitative indicators that can be directly used for health status assessment. This provides core physiological dimensions for subsequent contextual reasoning, enabling the robot to judge the user's health risks through physiological indicators and verify the user's health-related intentions in conjunction with other modal information. For example, if the user claims to be healthy but their physiological indicators are abnormal, physiological data is given priority, ensuring accurate control of the user's health status during the decision-making process and improving the safety and accuracy of elderly care services.

[0041] Specifically, in one possible embodiment, step S213 is implemented as follows: First, the original physiological signal is filtered, and a wavelet transform algorithm is used to remove power frequency interference and motion artifacts, such as removing noise generated by body activity in the heart rate signal to ensure signal purity. For the heart rate signal, the number of heartbeats per unit time is calculated by analyzing the peak interval of the filtered signal, and the heart rate index is output. For the blood pressure signal, the amplitude change of the signal is analyzed using the oscilloscope method, and the systolic and diastolic blood pressure indices are calculated in combination with preset calibration parameters. For the blood oxygen saturation signal, the blood oxygen saturation value is derived by analyzing the absorption ratio of red light and infrared light. Finally, the specific values ​​of heart rate, systolic blood pressure, diastolic blood pressure, and blood oxygen saturation are integrated into a set of physiological indicators, which are output as structured features of the physiological dimension.

[0042] Specifically, step S22 transforms each feature label in the quantized feature set into observed entities. It should be understood that since the feature labels in the quantized feature set exist in text identifiers or numerical forms, they cannot be directly associated with entities in a static knowledge graph, making it difficult to integrate them into the semantic framework of the knowledge graph for subsequent contextual analysis and relational reasoning, resulting in a disconnect between real-time perceived data and the knowledge system. Therefore, this application further performs semantic mapping between each feature label in the quantized feature set and graph entities, thereby transforming low-level perceived features into structured entities with semantic attributes. The resulting observed entities can establish a connection between multimodal perceived data and the semantic system of the knowledge graph, enabling the real-time collected feature information to possess reasonable semantic attributes. This lays the semantic foundation for subsequent fusion of background knowledge to construct a dynamic context graph, ensuring that perceived data can participate in graph-level relational analysis and improving the depth of contextual understanding.

[0043] Specifically, in one possible embodiment, step S22 is implemented as follows: First, based on the three core categories and their subcategories defined in the static knowledge graph—behavioral entities, emotional entities, and physiological entities—a mapping rule base between feature labels and entities is established. For action labels extracted from video frames, slow gait corresponds to the gait abnormality sub-entity under the behavioral entity, and falling action corresponds to the dangerous behavior sub-entity under the behavioral entity. For facial expressions and audio emotion labels, painful expressions and anxious voices correspond to the negative emotion sub-entities under the emotional entity, and pleasant expressions and calm voices correspond to the positive emotion sub-entities under the emotional entity. For physiological indicators, heart rate exceeding the normal range corresponds to the heart rate abnormality sub-entity under the physiological entity, and blood oxygen saturation below the threshold corresponds to the blood oxygen abnormality sub-entity under the physiological entity. Subsequently, each feature label in the quantified feature set is converted into a corresponding observation entity according to the mapping rules, and each observation entity is assigned a unique identifier to ensure complete compatibility with the knowledge graph entity system.

[0044] Specifically, in step S23, the user ID is used as the entry point to perform a graph query in the static knowledge graph to obtain background entities. It should be understood that since observed entities only reflect the current real-time state of the elderly person, they lack background information such as their historical health records, long-term care needs, and past behavioral preferences. This information is crucial for accurately inferring true intentions, such as whether abnormal actions are related to underlying diseases. Relying solely on real-time information can easily lead to biased decision-making. Therefore, this application further uses the unique user ID of the target elderly person as the retrieval entry point to perform a targeted graph query in the pre-constructed static knowledge graph to obtain background information associated with that user and transform it into background entities. This supplements the historical context and personalized information beyond real-time observation. In this way, the fusion of real-time data and historical knowledge can be achieved, avoiding misjudgments of intention due to missing information, making subsequent decisions more tailored to the individual circumstances of the elderly person, and improving the personalization and accuracy of elderly care services.

[0045] Specifically, in one possible embodiment, step S23 is implemented as follows: The static knowledge graph includes user nodes, health record nodes, nursing need nodes, behavioral preference nodes, medication nodes, etc., and the relationships between nodes. During a query, starting from the user ID node, the user-health record association edges are traversed to obtain health record-related background entities such as history of hypertension and arthritis. The user-nursing need association edges are traversed to obtain nursing need-related background entities such as needing assistance to get up and avoiding strenuous nursing movements. The user-behavioral preference association edges are traversed to obtain behavioral preference-related background entities such as preferring morning care and being sensitive to sound and light stimulation. The user-medication association edges are traversed to obtain medication-related background entities such as taking antihypertensive medication at 9 am daily and having a history of drug allergies. After the query is completed, all background entities are aggregated to form a set, comprehensively covering the historical health and nursing-related information of the elderly.

[0046] Specifically, in step S24, the observed entities and background entities are input into the entity embedding model to obtain the set of context entity vectors. It should be understood that since both the observed entities and background entities are discrete semantic entities, they cannot be directly used for deep relational reasoning in graph attention networks. The discrete form lacks a numerical carrier to represent semantic associations, resulting in the inability to quantify the strength of associations and conflicts between entities. Therefore, this application further inputs the set of observed entities and background entities into a pre-trained entity embedding model, and transforms the entities into low-dimensional dense vectors, i.e., the set of context entity vectors, through model computation. The entity embedding model adopts the TransR model architecture, which preserves the semantic associations (such as the causal association between abnormal heart rate and a history of hypertension) and semantic differences (such as the behavioral differences between normal gait and abnormal gait) between entities through numerical representation. This provides computable node vectors for the construction of dynamic context graphs, enabling graph attention networks to quantify entity associations and conflicts based on vectors, improving the accuracy of context evolution perception and intent inference, and avoiding reasoning failures caused by the inability to compute discrete entities.

[0047] Specifically, in one possible embodiment, step S24 is implemented as follows: During the pre-training phase of the entity embedding model, entity triples that conform to the entity association logic of the elderly care scenario are first extracted from the static knowledge graph as training data. Positive sample triples must reflect effective associations between entities, such as a causal association between a history of hypertension and abnormal heart rate, a history of arthritis and abnormal gait, and a history of sleep disorders and anxious speech. Negative sample triples are combinations of entities without actual semantic association, such as no underlying medical history and no abnormal heart rate, or healthy lower limbs and unrelated fall movements. Simultaneously, effective entity pairs converted from previously collected multimodal data of the elderly care scenario are integrated to form a training dataset. Next, according to the entity category system defined in the static knowledge graph, such as physiological entities, behavioral entities, emotional entities, and health record entities, a unique identifier is assigned to each entity to ensure consistency between the entity and the semantic framework of the knowledge graph, and invalid triples whose relationship type is not defined in the static knowledge graph are filtered out. The model is then initialized, defining the vector dimensions for entities and relationships. The entity and relationship vector spaces are randomly initialized, and a projection matrix is ​​initialized separately for each preset relationship type (e.g., causal association, association, unrelated) for mapping entity vectors to the relationship space. During iterative training, a loss function adapted to triple training is used as the optimization objective, constrained by the principle that the vector distance after mapping positive samples is less than the corresponding distance for negative samples. The optimizer iteratively adjusts the model parameters. After each training round, a validation set is formed using triples extracted from the static knowledge graph that were not used in the training data to evaluate the model accuracy. Once the accuracy stabilizes after several rounds, the model is considered converged, and the trained model file is saved.

[0048] Then, in the context entity vector generation stage, the observed entities acquired in the current process (such as current gait abnormalities, real-time heart rate abnormalities) and background entities (such as the user's past history of hypertension, the user's need for assistance in getting up) are input into a pre-trained entity embedding model. Based on the pre-defined inter-entity relationships in the static knowledge graph, the entity embedding model maps each entity vector to the corresponding relationship space. After optimization by the pre-trained projection matrix, it outputs low-dimensional dense vectors, where semantically closely related entity vectors have higher similarity, and semantically unrelated or conflicting entity vectors have lower similarity. Finally, the low-dimensional dense vectors of all entities are collected to form the context entity vector set required for this application, used in the subsequent construction process of the dynamic context graph.

[0049] In the aforementioned knowledge graph-based intelligent decision support method for elderly care robots, step S3 involves constructing a dynamic context graph based on the set of context entity vectors. It should be understood that while the set of context entity vectors transforms observed and background entities into low-dimensional dense vectors, each context entity vector exists independently, failing to reflect the relationships between entities. However, the interaction relationships between entities in elderly care scenarios are the core of analyzing complex situations. Relying solely on independent vectors cannot capture these dynamic relationships, making it difficult for subsequent reasoning to reconstruct the complete context logic. Therefore, this application further constructs a dynamic context graph based on the set of context entity vectors, combining the structural information of the static knowledge graph with the semantic relationships of real-time entities. This transforms discrete entity vectors into a structured graph model capable of representing static relationships and dynamic interactions between entities, thereby significantly improving the robot's ability to analyze complex and dynamic elderly care scenarios.

[0050] In particular, in one specific embodiment, Figure 5 This is a flowchart of sub-step S3 of the knowledge graph-based intelligent decision support method for elderly care robots according to an embodiment of this application. Figure 5 As shown, step S3 includes: S31, instantiating graph nodes and initializing static structures based on the set of context entity vectors and the static knowledge graph to obtain a node set and a static relation matrix; S32, performing pairwise semantic relevance calculation on all nodes of type observation entity in the initial graph structure to obtain a dynamic relation matrix; S33, performing matrix fusion on the static relation matrix and the dynamic relation matrix to obtain a weighted adjacency matrix; S34, encapsulating the node set and the weighted adjacency matrix into graph objects to obtain the dynamic context graph.

[0051] Specifically, step S31 involves instantiating graph nodes and initializing the static structure based on the set of context entity vectors and the static knowledge graph to obtain a node set and a static relation matrix. It should be understood that the set of context entity vectors only contains numerical representations of entities, lacking a structured node form and predefined relationships between entities. The static knowledge graph, however, stores the inherent relational logic of entities in the elderly care scenario, such as the attribution relationship between users and health records, and the causal relationship between basic medical history and physiological abnormalities. Without constructing an initial graph structure using the static knowledge graph, the subsequent dynamic context graph would lack a stable semantic framework, making it impossible to reconstruct the basic relationships between entities. Therefore, this application further uses context entity vectors as the data foundation and the static knowledge graph as the semantic basis to instantiate graph nodes and initialize the static structure, thereby transforming discrete entity vectors into graph nodes and establishing inherent relationships between entities. In this way, a core framework can be built for the dynamic context graph, ensuring that the initial state of the graph structure has the semantic rationality of the elderly care scenario, avoiding the subsequent dynamic relationship calculation from deviating from the basic knowledge framework. At the same time, the identity attributes and static associations of the nodes are clearly defined, providing a stable basic structure for the subsequent dynamic relationship fusion and ensuring the semantic coherence of the context analysis.

[0052] In particular, in one specific embodiment, Figure 6 This is a flowchart of sub-step S31 of the knowledge graph-based intelligent decision support method for elderly care robots according to an embodiment of this application. Figure 6 As shown, step S31 includes: S311, extracting nodes of type user node and their indices from the node set; S312, traversing all other nodes and their indices in the node set, and setting the weight value of the edge between user node and other nodes to 1; S313, traversing all nodes of type background entity in the node set, determining whether there is a relationship between the nodes of two background entities based on the static knowledge graph, if there is a relationship, setting the weight value of the edge between the nodes of two background entities to 1, and if there is no relationship, setting the weight value of the edge between the nodes of two background entities to zero.

[0053] More specifically, step S311 involves extracting nodes of type user node and their indices from the node set. It should be understood that since the node set contains various types such as user nodes, observed entity nodes, and background entity nodes, with user nodes being the core associated subject in the elderly care scenario, and all observed and background entities revolving around a specific target elderly person, if user nodes and their indices are not explicitly extracted, the subsequent construction of relationships between entities will lead to confusion in the association logic due to the lack of a core subject, making it impossible to distinguish the target object to which an entity belongs. Therefore, this application further filters nodes of type user node from the node set and records their index positions in the set to establish the core associated object of the dynamic context graph. This provides a clear basis for subsequently constructing the attribution relationship between users and other entities, avoiding confusion in entity attribution in multi-user scenarios, ensuring that the dynamic context graph always revolves around the state and needs of the target elderly person, and guaranteeing the uniqueness and logical coherence of the subject in subsequent relationship reasoning.

[0054] Specifically, in one possible embodiment, step S311 is implemented as follows: First, each node in the node set is traversed. Based on the node type identifier in the node attributes, such as user node identifier as USER, observed entity node identifier as OBSERVE, and background entity node identifier as BACKGROUND, nodes of type user node are identified and filtered out. Next, the unique user identifier is extracted from the attribute information of the user node to confirm that the node is the user node corresponding to the current target elderly person. Finally, the index number of the user node in the node set is recorded. If the node set is arranged in order, and the user node is the second one, the index is 2. The user node and its index information are stored separately as the core reference for constructing the static relationship matrix later, ensuring that all subsequent associations between entities and users point to the target elderly person corresponding to the index.

[0055] More specifically, in step S312, all other nodes and their indices in the node set are traversed, and the weight value of the edge between the user node and other nodes is set to 1. Specifically, this application establishes the basic attribution relationship between the user node and all associated entities by traversing all other nodes in the node set except the user node and, based on the attribution relationship rules of the static knowledge graph, uniformly setting the weight value of the edge between the user node and these nodes to 1. This ensures that all associations in the dynamic context graph revolve around the target elderly person, avoiding cross-contamination of entity information from different users, and providing a basic framework for adding other relationships between entities subsequently, thus guaranteeing the main relevance and semantic rationality of the graph structure.

[0056] Specifically, in one possible embodiment, step S312 is implemented as follows: First, the previously extracted user nodes and their indices are retrieved, such as user node index 2. Then, all other nodes in the node set whose indices are not equal to 2 are traversed, including observed entity nodes, such as current gait abnormality and real-time heart rate abnormality, with indices 4 and 6 respectively, and background entity nodes, such as hypertension history and daily medication records, with indices 8 and 10 respectively. Subsequently, based on the preset rule in the static knowledge graph that "users have affiliation relationships with all associated entities", the positions corresponding to user node index (2) and other node indices (4, 6, 8, 10) are found in the static relation matrix, and the edge weight values ​​of these positions are uniformly set to 1. At the same time, the association type of the edges is marked as affiliation in the matrix attributes to ensure that the semantics of the weight assignment are clear. Finally, the construction of the basic association between user nodes and all other nodes is completed, forming the core association part about user affiliation in the static relation matrix.

[0057] More specifically, in step S313, all nodes of type background entity in the node set are traversed. Based on the static knowledge graph, it is determined whether there is a relationship between two background entity nodes. If there is, the weight value of the edge between the two background entity nodes is set to 1; if not, the weight value of the edge between the two background entity nodes is set to zero. Specifically, this application further traverses all background entity nodes and, based on the entity relationship data stored in the static knowledge graph, determines whether there is an association between each pair of background entity nodes, and sets the edge weight to 1 or 0 accordingly, thereby restoring the inherent semantic association between background entities. This allows the static relationship matrix to fully cover the two core static associations: user-entity affiliation relationships and semantic associations between background entities, providing more comprehensive basic knowledge support for the dynamic context graph, avoiding semantic gaps in subsequent reasoning due to missing background entity associations, and improving the logical integrity and knowledge coherence of context analysis.

[0058] Specifically, in one possible embodiment, the implementation process of step S313 is as follows: First, all nodes of type background entity are selected from the node set, such as history of hypertension, need for assistance to get up, daily medication record, and no underlying medical history, with indices 8, 10, 12, and 14 respectively. Then, these background entity nodes are traversed in pairs to form all possible node pairs, such as 8 and 10, 8 and 12, 8 and 14, 10 and 12, etc. Subsequently, for each pair of nodes, the entity relationship database stored in the static knowledge graph is queried to determine whether there is a preset relationship between the two: if there is a disease-nursing need association between history of hypertension (8) and need for assistance to get up (10), and a disease-medication association between history of hypertension (8) and daily medication record (12), then the position weight of the static relationship matrix corresponding to these two pairs of nodes is set to 1; if there is no association between history of hypertension (8) and no underlying medical history (14), and no association between need for assistance to get up (10) and no underlying medical history (14), then the position weight of the matrix corresponding to these node pairs is set to 0. Finally, the weights of all background entity node pairs are set, and the static relation matrix is ​​updated to include the inherent semantic relationships between background entities, laying the foundation for subsequent matrix fusion.

[0059] Specifically, step S32 involves calculating pairwise semantic relevance for all nodes in the initial graph structure that are observation entities to obtain a dynamic relation matrix. In a specific example of this application, step S32 includes: calculating pairwise semantic relevance for all nodes in the initial graph structure that are observation entities using the following formula:

[0060]

[0061] in, Given the context entity vectors of the nodes of two observed entities, This indicates the calculation of the Euclidean norm. The cosine similarity between the two. For the collaborative threshold, The conflict threshold, This represents the weight of the edge between nodes of two observed entities. In other words, if the cosine similarity of two vectors is higher than the collaboration threshold... For example, if the weight is 0.8, it is determined that there is a strong cooperative relationship between the two, and the edge weight is... A value set to 1 corresponds to multiple signals pointing to the same risk, such as gait abnormalities and heart rate abnormalities. If the cosine similarity is below the conflict threshold... For example, if the value is -0.5, then it is determined that there is a clear conflict between the two, and the edge weight is adjusted accordingly. A value set to -1 corresponds to contradictory signals, such as the verbal claim "I'm fine" versus the physiological tachycardia. If the cosine similarity falls between the conflict threshold and the cooperability threshold, the two signals are considered to have no significant strong correlation or strong conflict, and the edge weights are adjusted accordingly. It is set to 0. In this way, this application explicitly transforms the semantic similarity at the vector level into edge weights with clear meaning in the graph structure, thereby constructing a dynamic relation matrix that can accurately reflect the multimodal information interaction state in the current context, providing key dynamic information for subsequent fusion of static knowledge and comprehensive contextual reasoning.

[0062] In particular, in another possible preferred embodiment, step S32 further includes: calculating the mean vector of context entity vectors corresponding to all background entities to generate a user background vector representing the current context center; relativizing the context entity vectors of each observed entity relative to the user background vector to generate relativized context entity vectors of the nodes of the observed entities; performing confidence amplification processing on the relativized context entity vectors of the nodes of the observed entities to generate optimized context entity vectors of the nodes of the observed entities, wherein the confidence amplification processing is used to enhance the significance of strong positive or strong negative semantic shifts; and calculating pairwise semantic correlations between the nodes of the observed entities based on the optimized context entity vectors of the nodes of the observed entities.

[0063] Here, when calculating pairwise semantic relevance based on cosine similarity, what is essentially measuring is the angle between two vectors in the global absolute semantic space. However, the true meaning of the relationship between observed events is not absolute, but highly relative to the observed subject. For example, in a specific scenario, the two observed entities of a user's dizziness and high sugar intake have a very strong synergistic relationship for a known diabetic patient, but for a user without such a medical history, this correlation is much weaker.

[0064] At the same time, since the pre-trained word vectors are based on general knowledge to locate entities, the vector similarity between, for example, pain and nothing, may be negative, but the specific value is arbitrary and depends on the training corpus. Moreover, this calculation method cannot perceive that the two concepts were asserted by the same person at the same time. This fact makes the opposition between them particularly important and significant.

[0065] Therefore, semantic relevance calculation needs to be context-centric. That is, instead of directly comparing the absolute positions of observed entity semantic vectors, we first relativize them relative to the context vector representation, and then measure the consistency between the relativized vectors. In other words, an observed event (such as pain) represents the user's relative relationship to their context state. Therefore, the context vector corresponding to the user's context entity can be seen as the origin or context center of all current events. Thus, the entity vector will represent the semantic direction and magnitude of the user's state shifting from their context to their actual state.

[0066] Specifically, firstly, the mean vector of the context entity vectors corresponding to all background entities is calculated. This process generates a user background vector representing the current context center, and then relativizes the context entity vectors of each observed entity relative to the user background vector to generate relativized context entity vectors for the nodes of the observed entities.

[0067] in, This represents the context entity vector of the node representing the i-th observed entity. This represents vector subtraction. This represents element-wise multiplication, which in the above formula involves multiplying each element of the vector positionally. express The reciprocal vector, that is, the vector of Take the reciprocal of each element. The context entity vector represents the node of the i-th relativized observation entity, which characterizes the direction and magnitude of semantic offset. In other words, it represents the trajectory of user state changes in a high-dimensional semantic space. This trajectory can anchor all observation events to a common and meaningful origin, making subsequent comparisons context-dependent.

[0068] This approach provides a reference framework tailored to individual user characteristics for subsequent semantic analysis of observed entities, avoiding semantic misjudgments caused by using general background standards. It ensures that the interpretation of observed entities always revolves around the user's baseline state, making the semantics of observed entities more accurately aligned with the individual user's situation. For example, the user's current heart rate observation vector, after being relativized with a background vector containing a history of hypertension, can clearly reflect the degree of deviation of the current heart rate from its baseline heart rate level. A significant deviation indicates a health risk, without relying on a general normal heart rate range. This avoids semantic misjudgments caused by individual differences in baseline conditions, laying the foundation for accurate identification of the user's state.

[0069] At the same time, targeting The semantic shift direction and magnitude are mixed and represented, and confidence amplification based on fine-grained collaboration of feature values ​​is performed. That is, the context entity vector of the relativized observation entity's node is subjected to confidence amplification processing to generate an optimized context entity vector of the observation entity's node. The confidence amplification processing is used to enhance the significance of strongly positive or strongly negative correlated semantic shifts.

[0070] in, Indicates to Perform hyperbolic tangent function operation on each element. The L2 norm of a vector. The first norm of a vector is represented by its 1-norm. This represents an element-wise multiplication operation, where the ratio of the 2-norm to the 1-norm is added to each element of the vector in the above formula. This represents the context entity vector of the node of the i-th optimized observed entity. The hyperbolic tangent function is used to... The values ​​are normalized, and the confidence level is amplified using the ratio of the second norm to the first norm. That is, scores close to zero (indicating a lack of correlation in both direction and magnitude of semantic shift) remain close to zero after transformation, while scores with strong positive or negative correlations are further pushed towards their numerical representations, making the distinction between weakly and strongly correlated entities clearer. This makes the semantic features of key observed entities more prominent. For example, the synergistic association between a user's simultaneously present gait anomaly and heart rate anomaly vectors is significantly enhanced after confidence amplification, clearly pointing to a health risk context. Similarly, the strong negative correlation shift between the user's voice "I'm fine" vector and the physiological anomaly vector is amplified, clearly identifying information conflicts, avoiding interference from weakly related information in key decisions, and improving the accuracy of subsequent semantic relevance calculations.

[0071] Then, based on the context entity vectors of the optimized observed entity nodes, the pairwise semantic correlations between the nodes of the observed entity are calculated. Specifically, firstly, all possible vector pairs are extracted from the optimized set of observed entity vectors. Secondly, for each vector pair, the semantic correlation between the two is calculated using a cosine similarity algorithm, obtaining the similarity value by dividing the vector dot product by the product of the magnitudes of the two vectors. Then, the similarity value is judged according to a preset collaboration threshold (e.g., 0.8) and conflict threshold (e.g., -0.5). If the similarity is higher than the collaboration threshold, it is determined to be a collaboration relationship; if it is lower than the conflict threshold, it is determined to be a conflict relationship; and if it is between the two, it is determined to be an unrelated relationship. Finally, the correlation value and association type of each vector pair are recorded to form the pairwise semantic correlation results between observed entities. This result can be directly used to construct the subsequent dynamic relationship matrix, supporting the accurate generation of dynamic context graphs. This allows for improved calculation accuracy by enhancing context sensitivity; that is, the calculation is no longer based on an absolutely universal understanding of concepts, but rather on the context of the current state of a specific user.

[0072] Specifically, step S33 involves matrix fusion of the static relation matrix and the dynamic relation matrix to obtain a weighted adjacency matrix. In a specific example of this application, step S33 includes: performing matrix fusion of the static relation matrix and the dynamic relation matrix using the following formula:

[0073] in, It is a static relation matrix. It is a dynamic relation matrix. This is a weighted adjacency matrix. Specifically, this application further integrates the static and dynamic relational matrices to combine the inherent logical relationships between entities with real-time interactive relationships, forming a unified relational matrix that comprehensively reflects the current context. This allows the final weighted adjacency matrix to retain the core semantic foundations of the elderly care scenario, such as the user's unique attribution to entities and the knowledge relationships between background entities, while also including the real-time interactive states of observed entities, such as the collaborative or conflicting relationships among multiple abnormal observed entities in the current scenario. Simultaneously, the unified matrix format can directly adapt to the input requirements of subsequent graph attention networks, avoiding the reduction in reasoning efficiency caused by fragmented data structures. This enables the graph attention network to quantify the strength and conflict of relationships between entities based on complete relational information, accurately perceiving the evolutionary trend of the context. For example, the overlapping associations of a user's gait abnormalities, heart rate abnormalities, and history of hypertension can point to higher-priority health risk scenarios, ultimately improving the depth of contextual understanding and the reliability of intent inference for elderly care robots in complex and dynamic elderly care scenarios.

[0074] Specifically, in step S34, the node set and weighted adjacency matrix are encapsulated into graph objects to obtain the dynamic context graph. It should be understood that since the node set and weighted adjacency matrix are independent data structures, without encapsulation, the dispersed data structures cannot meet the input format requirements of graph attention networks, preventing the model from performing collaborative reasoning based on nodes and edges. Therefore, this application further integrates the node set and weighted adjacency matrix according to graph data specifications, encapsulating them into graph objects to construct a structurally complete, uniformly formatted dynamic context graph containing node features and edge relationships. This provides a standardized input carrier for subsequent graph attention networks, enabling the model to directly perform attention weight calculations, information transmission, and node embedding updates based on the node vector features and edge weights in the graph object, accurately capturing the complex relationships between entities in elderly care scenarios. Meanwhile, the encapsulated graph object can completely retain all semantic and relational information in the current context, avoiding information loss during data transmission or processing, ensuring that the graph attention network can perform reasoning based on complete contextual data, and ultimately improving the accuracy of the elderly care robot's contextual understanding and intent inference in the current dynamic elderly care scenario, laying the foundation for generating decision-making actions that meet the user's real needs.

[0075] In the aforementioned knowledge graph-based intelligent decision support method for elderly care robots, step S4 involves performing context perception and node updating on the dynamic context graph based on a graph attention network to obtain updated node embeddings. It should be understood that although the dynamic context graph has integrated the node set and weighted adjacency matrix, its node embeddings are still initial vectors output by the previous entity embedding model, failing to fully integrate the differences in the importance of relationships between nodes in the current context. If directly used for subsequent intent classification, this would cause the model to fail to focus on key entity relationships, reducing the accuracy of intent inference. Therefore, this application further utilizes a graph attention network to perform context perception and node updating on the dynamic context graph. By quantifying the importance of relationships between nodes through an attention mechanism, key information from neighboring nodes is aggregated to the target node, generating node embeddings that integrate contextual relationship weights. This ensures that the updated node embeddings not only retain the semantic features of the entity itself but also include key influence information from neighboring nodes in the current context. Meanwhile, the context awareness capability of graph attention networks can capture the underlying logic of context evolution, providing higher quality input data for subsequent intent classification and credibility calculation, significantly improving the deep understanding of complex situations and the accuracy of intent inference of elderly care robots, and avoiding decision bias caused by ignoring the differences in the importance of node associations.

[0076] Specifically, in one possible embodiment, step S4 is implemented as follows: First, the graph attention network is initialized, and pre-trained network parameters are loaded. These parameters are optimized through training on historical contextual graph data of elderly care scenarios. The number of attention heads and the dimension of hidden layers are set to ensure that the network input dimension is consistent with the node feature dimension of the dynamic contextual graph. Simultaneously, the node feature matrix and weighted adjacency matrix of the dynamic contextual graph are used as network input. Next, attention weight calculation is performed. For each node, such as a user node, a node with a history of hypertension, or a node observing gait abnormalities, the network calculates the attention weight of that node to each neighboring node based on its edge weights and node feature similarity with its neighbors using an attention scoring function. For example, among the neighboring nodes of a user node, gait abnormalities and a history of hypertension are closely related to the user's health status, with high edge weights and high feature similarity, resulting in significantly higher attention weights than environmental nodes such as those with normal room temperature. Subsequently, node information aggregation is performed. The network, based on calculated attention weights, weights and sums the features of each node's neighboring nodes. This weighted neighbor feature is then fused with the node's initial features. Specifically, this is achieved through linear transformations and activation functions to update node information. For example, the initial embedding of a user node is fused with weighted features of gait abnormalities and a history of hypertension to form a new embedding that reflects current health risk associations. Finally, after iterative processing via the network's multi-layered attention mechanism, the updated embeddings of all nodes are output. These embeddings fully cover the key association information between nodes in the current elderly care scenario and can be directly used in subsequent intent classification and credibility calculation stages.

[0077] In the aforementioned knowledge graph-based intelligent decision support method for elderly care robots, step S5 involves classifying and calculating the credibility of the updated node embeddings to obtain a sorted list of intentions. It should be understood that while the updated node embeddings have integrated key relationships between nodes in the current context, they still exist in a scattered manner as individual nodes, not integrated into a unified feature representing the global context, and lack a mapping relationship with specific intention categories in the elderly care scenario. Directly inferring intentions based on these scattered node embeddings would prevent a global assessment of the user's overall needs and would make it impossible to quantify the credibility of different intentions, thus making it difficult for the robot to select the decision direction that best matches the user's actual needs from among multiple potential intentions. Therefore, this application further performs intention classification and credibility calculation on the updated node embeddings. By integrating global node features and mapping them to preset intention categories, the potential intentions of the user in the current context are determined and sorted according to credibility. This enables the elderly care robot to clearly grasp the priority of various potential user needs. Meanwhile, credibility calculation can filter out low-credibility intentions, such as misjudged intentions due to abnormal behavior caused by environmental interference, ensuring that robot decisions are based on highly reliable contextual inferences, and significantly improving the accuracy and safety of elderly care services.

[0078] In particular, in one specific embodiment, Figure 7 This is a flowchart of sub-step S5 of the knowledge graph-based intelligent decision support method for elderly care robots according to an embodiment of this application. Figure 7 As shown, step S5 includes: S51, performing average pooling on the updated node embeddings to obtain graph vectors; S52, inputting the graph vectors into an intent classifier containing fully connected layers and a Softmax activation function to obtain the ordered intent list.

[0079] Specifically, in step S51, average pooling is performed on the updated node embeddings to obtain graph vectors. It should be understood that the updated node embeddings are independent vector representations of each node in the dynamic context graph, and these scattered node embeddings cannot be directly input into the intent classifier. Therefore, this application further performs average pooling on all updated node embeddings to aggregate the scattered node-level features into a single graph-level vector, preserving key association information of the global context. This provides a standardized input carrier for the subsequent intent classifier, ensuring that the classifier can make inferences based on global context information, laying the foundation for accurate intent classification.

[0080] Specifically, in step S52, the graph vector is input into an intent classifier containing a fully connected layer and a Softmax activation function to obtain the ordered intent list. It should be understood that although the graph vector integrates global features of the dynamic context graph, it is only a low-dimensional vector with numerical values, which cannot be directly mapped to specific intent categories in the elderly care scenario, nor can the credibility of each intent be quantified. Therefore, this application further inputs the graph vector into a preset intent classifier. This classifier processes the graph vector features through a fully connected layer and outputs the credibility of each intent using a Softmax activation function, thereby transforming the global context features into specific intent categories and corresponding credibility, and sorting them according to their credibility. This provides clear intent priority guidance for the elderly care robot, avoiding decision-making hesitation caused by ambiguity in intent credibility, and significantly improving the robot's initiative and accuracy in decision-making.

[0081] Specifically, in one possible embodiment, step S52 is implemented as follows: First, the intent classifier is initialized, loading model parameters pre-trained using historical contextual data of elderly care scenarios. The classifier includes two fully connected layers and a subsequent Softmax activation function. Preset intent categories include health risk warning, daily care assistance, emotional comfort needs, and emergency help response. Next, the validated graph vectors are input into the classifier. The first fully connected layer performs linear transformations and non-linear activations on the graph vectors, such as the ReLU function, to extract deep contextual features. The second fully connected layer maps the deep features to the feature dimensions corresponding to the four intent categories. Subsequently, the Softmax activation function converts the feature values ​​of each intent category into confidence levels (probability values) between 0 and 1, for example, health risk warning confidence level 0.92, emotional comfort needs 0.05, daily care assistance 0.03, and emergency help response 0.00. Finally, the intent categories are sorted from high to low confidence levels to form an intent list of "health risk warning > emotional comfort needs > daily care assistance > emergency help response," which is then output to the subsequent decision generation and execution stages.

[0082] In the aforementioned knowledge graph-based intelligent decision support method for elderly care robots, step S6 involves generating and executing decisions based on a sorted intent list to obtain robot actions. Specifically, this application further generates and executes corresponding robot action sequences based on the sorted intent list through preset decision rules and action mapping mechanisms. This transforms abstract user intents into concrete, actionable operations for the robot, while adhering to intent priority to ensure core needs are addressed first. This allows the robot to overcome the predicament of knowing intents but lacking execution paths, achieving a closed loop from intent recognition to service delivery. This significantly improves the initiative and accuracy of elderly care robot services, avoids decision-making stagnation due to a lack of action mapping, and ensures orderly service execution, preventing user experience degradation or safety risks caused by chaotic actions.

[0083] Specifically, in one possible embodiment, step S6 is implemented as follows: First, the decision rule base is initialized, and the mapping relationship and execution flow of core intentions and actions in the elderly care scenario are preset in the robot control system. For example, the intention of health risk warning corresponds to the action sequence of "first performing secondary monitoring of physiological indicators, then performing voice inquiry about the user's physical condition, then performing sending health reminders to preset caregivers, and finally performing continuous monitoring and recording of data". The intention of daily item retrieval corresponds to the action sequence of "first performing voice confirmation of the item retrieval target and location, then performing planning of the movement path, then performing starting the robotic arm to retrieve the item, and finally performing navigation to the user's location for delivery". The intention of emotional comforting corresponds to the action sequence of "first performing playing soothing music, then performing initiating gentle voice interaction, and finally performing adjusting the body lights to a warm tone". Next, intention-action mapping matching is performed to obtain a sorted list of intentions. If the health risk warning has the highest credibility, the action sequence corresponding to this intention is retrieved from the decision rule base to determine the action steps and their order to be executed. Then, action feasibility verification is performed. The robot self-checks the current hardware module status and environmental adaptability. After confirming that all conditions are met, the final executable action instruction is generated. Finally, the robot executes the action sequence and provides status feedback: First, it activates the physiological monitoring sensors to collect data on indicators such as heart rate and blood oxygen. Simultaneously, it asks the user via a voice module, "Are you feeling unwell?" At the same time, based on a preset communication protocol, it sends a reminder message containing the user ID, current physiological data, and risk level to the caregiver's mobile terminal. During the action execution, the robot provides real-time feedback to the system on the completion status of each step, such as completion of physiological monitoring and successful transmission of reminder messages, ensuring that the action is executed in a closed loop according to the process, ultimately forming a complete robot service action that meets the user's current core needs.

[0084] In summary, the knowledge graph-based intelligent decision support method for elderly care robots, based on the embodiments of this application, is elucidated. It integrates multimodal sensor data (video, audio, and physiological data) and utilizes a static knowledge graph to perform semantic analysis on key entities such as people, objects, and the environment in a real-time context, thereby dynamically constructing a dynamic context graph that represents the current complex interactive relationships. Based on this, a graph attention network is introduced to perform deep relational reasoning and context evolution perception on the dynamic context graph, quantifying the associations and conflicts between different information sources. This identifies the user's true intent with the highest credibility among multiple possibilities and ultimately generates and executes the robot's actions. This solves the problem of inaccurate intent inference caused by multimodal information conflicts, thereby significantly improving the decision-making intelligence and service initiative of elderly care robots in complex dynamic environments.

[0085] Figure 8 This is a block diagram of a knowledge graph-based intelligent decision support system for elderly care robots according to an embodiment of this application. Figure 8As shown, the knowledge graph-based intelligent decision support system 100 for elderly care robots according to an embodiment of this application includes: a raw sensor data acquisition module 110, used to acquire raw sensor data streams of the target elderly object, the raw sensor data streams including video frame data, audio block data, and physiological signal data; a context entity vectorization module 120, used to perform context entity vectorization on the raw sensor data streams based on a static knowledge graph to obtain a set of context entity vectors; a dynamic context graph construction module 130, used to construct a dynamic context graph based on the set of context entity vectors; a context perception and node update module 140, used to perform context perception and node update on the dynamic context graph based on a graph attention network to obtain updated node embeddings; an intent classification and credibility calculation module 150, used to perform intent classification and credibility calculation on the updated node embeddings to obtain a sorted intent list; and a decision generation and execution module 160, used to generate and execute decisions based on the sorted intent list to obtain robot actions.

[0086] As described above, the knowledge graph-based intelligent decision support system 100 for elderly care robots according to embodiments of this application can be implemented in various wireless terminals, such as servers with knowledge graph-based intelligent decision support algorithms for elderly care robots. In one possible implementation, the knowledge graph-based intelligent decision support system 100 for elderly care robots according to embodiments of this application can be integrated into the wireless terminal as a software module and / or a hardware module. For example, the knowledge graph-based intelligent decision support system 100 for elderly care robots can be a software module in the operating system of the wireless terminal, or it can be an application developed for the wireless terminal; of course, the knowledge graph-based intelligent decision support system 100 for elderly care robots can also be one of many hardware modules of the wireless terminal.

[0087] Alternatively, in another example, the knowledge graph-based intelligent decision support system 100 for elderly care robots and the wireless terminal can also be separate devices, and the knowledge graph-based intelligent decision support system 100 for elderly care robots can be connected to the wireless terminal via wired and / or wireless networks, and transmit interactive information in accordance with an agreed data format.

[0088] Here, those skilled in the art will understand that the specific operations of each step in the knowledge graph-based intelligent decision support system for elderly care robots described above have been referenced. Figures 1 to 7 The method for intelligent decision support for elderly care robots based on knowledge graphs has been described in detail, and therefore, its repeated description will be omitted.

Claims

1. A knowledge graph-based intelligent decision support method for elderly care robots, characterized in that, include: Acquire raw sensor data streams from the target elderly subjects. The raw sensor data streams include video frame data, audio block data, and physiological signal data. Based on a static knowledge graph, the original sensor data stream is vectorized into contextual entity vectors to obtain a set of contextual entity vectors. This includes: extracting and quantizing the synchronous features of multimodal signals from video frame data, audio block data, and physiological signal data in the original sensor data stream to obtain a quantized feature set; converting each feature label in the quantized feature set into an observed entity; performing a graph query in the static knowledge graph using the user ID as the entry point to obtain the background entity; and inputting the observed entity and the background entity into an entity embedding model to obtain the set of contextual entity vectors. A dynamic context graph is constructed based on a set of context entity vectors, including: instantiating graph nodes and initializing static structures based on the set of context entity vectors and a static knowledge graph to obtain a node set and a static relation matrix; calculating pairwise semantic relevance for all nodes of type observation entity in the initial graph structure to obtain a dynamic relation matrix; performing matrix fusion on the static relation matrix and the dynamic relation matrix to obtain a weighted adjacency matrix; and encapsulating the node set and the weighted adjacency matrix into a graph object to obtain the dynamic context graph. The process involves performing context awareness and node updating based on a graph attention network to obtain updated node embeddings for a dynamic context graph. This includes: first, initializing the graph attention network by loading pre-trained network parameters and using the node feature matrix and weighted adjacency matrix of the dynamic context graph as network input; next, calculating attention weights for each node by using an attention scoring function based on the edge weights of its neighbors and the similarity of its node features; then, aggregating node information by weighting the features of each node's neighbors according to the calculated attention weights, and fusing the weighted neighbor features with the node's initial features to update the node information; finally, iteratively processing through a multi-layer attention mechanism to output the updated embeddings of all nodes. The updated node embeddings are subjected to intent classification and credibility calculation to obtain a sorted intent list; Decisions are generated and executed based on a sorted list of intentions to obtain robot actions; Specifically, graph node instantiation and static structure initialization are performed based on the set of context entity vectors and the static knowledge graph to obtain the node set and static relation matrix, including: Extract the nodes of type User Node and their indices from the node set; Iterate through all other nodes and their indices in the node set, and set the weight of the edge between the user node and other nodes to 1; Iterate through all nodes of type background entity in the node set, and determine whether there is a relationship between two background entity nodes based on the static knowledge graph. If there is a relationship, the weight of the edge between the two background entity nodes is set to 1; if there is no relationship, the weight of the edge between the two background entity nodes is set to zero. Specifically, pairwise semantic relevance calculations are performed on all nodes of type observed entity in the initial graph structure to obtain a dynamic relation matrix, including: The pairwise semantic relevance of all nodes of type observed entity in the initial graph structure is calculated using the following formula: ； in, Given the context entity vectors of the nodes of two observed entities, Calculate the Euclidean norm. The cosine similarity between the two. For the collaborative threshold, The conflict threshold, This represents the weight of the edge between nodes of two observed entities.

2. The knowledge graph-based intelligent decision support method for elderly care robots according to claim 1, characterized in that, Synchronous feature extraction and quantization of multimodal signals are performed on video frame data, audio block data, and physiological signal data in the raw sensor data stream to obtain a quantized feature set, including: Video frame data is input into the visual feature extraction branch to obtain action labels and their action confidence, as well as expression labels and their expression confidence; The audio block data is input into the auditory feature extraction branch to obtain the speech text and sentiment labels and their sentiment confidence; Physiological signal data is input into the physiological feature extraction branch to obtain physiological indicators.

3. The knowledge graph-based intelligent decision support method for elderly care robots according to claim 1, characterized in that, The matrix fusion of the static and dynamic relation matrices to obtain a weighted adjacency matrix includes: fusing the static and dynamic relation matrices using the following formula: ; in, It is a static relation matrix. It is a dynamic relation matrix. It is a weighted adjacency matrix.

4. The intelligent decision support method for elderly care robots based on knowledge graphs according to claim 1, characterized in that, The updated node embeddings are then subjected to intent classification and credibility calculation to obtain a sorted list of intents, including: The updated node embeddings are then subjected to average pooling to obtain graph vectors. The graph vector is input into an intent classifier containing a fully connected layer and a Softmax activation function to obtain the ordered intent list.

5. A knowledge graph-based intelligent decision support system for elderly care robots, characterized in that, include: The raw sensor data acquisition module is used to acquire the raw sensor data stream of the target elderly object. The raw sensor data stream includes video frame data, audio block data and physiological signal data. The context entity vectorization module is used to vectorize the original sensor data stream based on a static knowledge graph to obtain a set of context entity vectors. This includes: extracting and quantizing the synchronous features of multimodal signals from video frame data, audio block data, and physiological signal data in the original sensor data stream to obtain a quantized feature set; converting each feature label in the quantized feature set into an observed entity; performing a graph query in the static knowledge graph using the user ID as the entry point to obtain a background entity; and inputting the observed entities and background entities into an entity embedding model to obtain the set of context entity vectors. A dynamic context graph construction module is used to construct a dynamic context graph based on a set of context entity vectors. This includes: instantiating graph nodes and initializing the static structure based on the set of context entity vectors and a static knowledge graph to obtain a node set and a static relation matrix; calculating pairwise semantic relevance for all nodes of type observation entity in the initial graph structure to obtain a dynamic relation matrix; performing matrix fusion on the static and dynamic relation matrices to obtain a weighted adjacency matrix; and encapsulating the node set and the weighted adjacency matrix into a graph object to obtain the dynamic context graph. The context awareness and node update module is used to perform context awareness and node update on a dynamic context graph based on a graph attention network to obtain updated node embeddings. This includes: first, initializing the graph attention network by loading pre-trained network parameters and using the node feature matrix and weighted adjacency matrix of the dynamic context graph as network input; next, calculating attention weights for each node by using an attention scoring function based on the edge weights of its neighbors and the similarity of its node features; then, aggregating node information by weighting the features of each node's neighbors according to the calculated attention weights, and fusing the weighted neighbor features with the node's initial features to update the node information; finally, through iterative processing using a multi-layer attention mechanism, outputting the updated embeddings of all nodes. The intent classification and credibility calculation module is used to classify intents and calculate credibility on the updated node embeddings to obtain a sorted list of intents; The decision generation and execution module is used to generate and execute decisions based on a sorted list of intentions to obtain robot actions. Specifically, graph node instantiation and static structure initialization are performed based on the set of context entity vectors and the static knowledge graph to obtain the node set and static relation matrix, including: Extract the nodes of type User Node and their indices from the node set; Iterate through all other nodes and their indices in the node set, and set the weight of the edge between the user node and other nodes to 1; Iterate through all nodes of type background entity in the node set, and determine whether there is a relationship between two background entity nodes based on the static knowledge graph. If there is a relationship, the weight of the edge between the two background entity nodes is set to 1; if there is no relationship, the weight of the edge between the two background entity nodes is set to zero. Specifically, pairwise semantic relevance calculations are performed on all nodes of type observed entity in the initial graph structure to obtain a dynamic relation matrix, including: The pairwise semantic relevance of all nodes of type observed entity in the initial graph structure is calculated using the following formula: ； in, Given the context entity vectors of the nodes of two observed entities, To calculate the Euclidean norm, The cosine similarity between the two. For the collaborative threshold, The conflict threshold, This represents the weight of the edge between nodes of two observed entities.

Images