A wireless sensing platform with edge computing function

Abstract

This invention proposes a wireless sensing platform with edge computing capabilities, relating to the field of wireless sensor technology. Addressing the problems of high data redundancy, high transmission latency, and weak ability to identify and respond to sudden events in complex scenarios in existing wireless sensor technologies, this invention employs a joint processing mechanism combining a causal event identification component, a causal data fusion component, an event separation component, and a data allocation component. By detecting causal change events in the environment, it dynamically extracts and fuses sensor data with strong causal relationships, reducing redundant data transmission caused by sudden events, lowering transmission latency, and improving the robustness of causal data fusion in concurrent event scenarios. This achieves efficient and accurate perception and response to sudden events in complex environments.

Description

Technical Field

[0001] This invention relates to the field of wireless sensor technology, and in particular to a wireless sensing platform with edge computing capabilities. Background Technology

[0002] As a core technology of the Internet of Things (IoT), wireless sensor networks (WSNs) have demonstrated tremendous application potential in fields such as environmental monitoring, smart homes, industrial automation, and smart cities due to their advantages of flexible deployment, wide coverage, and real-time sensing. With the rise of edge computing technology, which pushes computing power down to the network edge, wireless sensing platforms are endowed with real-time data processing and local decision-making capabilities, improving system response speed and reliability. This is driving sensing computing towards a new stage of distributed and intelligent development.

[0003] However, existing wireless sensing platforms still face significant challenges in multimodal sensing tasks in complex environments. To comprehensively capture environmental information, it is typically necessary to deploy multiple heterogeneous sensors, such as temperature, humidity, sound, and image sensors. The data collected independently by these sensors contains a large amount of redundant and complementary information in both spatiotemporal dimensions. Traditional processing methods either directly upload massive amounts of raw data to the cloud, resulting in a huge network transmission load, or employ simple, fixed data fusion strategies at the edge. As the accuracy of wireless sensors improves and the amount of data collected increases, transmission latency also increases. In response to sudden events such as impacts and explosions, the ability of existing wireless sensors to identify and output information in a timely and accurate manner remains significantly inadequate.

[0004] Therefore, there is an urgent need for a wireless sensing platform with edge computing capabilities that can identify the inherent correlations in data and adapt to causal changes in the environment. Summary of the Invention

[0005] To address this, the present invention provides a wireless sensing platform with edge computing capabilities to overcome the problems of large amounts of redundant data, high transmission latency, and weak ability to timely and accurately identify and output data in response to emergencies in the prior art.

[0006] To achieve the above objectives, the present invention provides a wireless sensing platform with edge computing capabilities, comprising:

[0007] The data acquisition unit is used to collect raw environmental data from the multimodal sensor and transmit it to the edge processing unit.

[0008] The edge processing unit, connected to the data acquisition unit, is used to perform data fusion on the raw environmental data input from the data acquisition unit and output fused features, specifically including:

[0009] The edge processing unit includes:

[0010] The feature encoding module is used to encode the features of the data transmitted by each sensor.

[0011] The dynamic relationship perception module, connected to the feature encoding module, is used to construct a hierarchical sensor relationship model, including instantaneous dynamic relationships, semantic functional relationships, and statistical dependencies.

[0012] The causal fusion module, connected to the dynamic relationship perception module, is used to identify, extract, and fuse data with the same representational meaning. It includes a causal event identification component for detecting causal change events occurring in the environment and identifying the affected sensor data types; a causal data fusion component for extracting and fusing multi-sensor data based on causal events; an event separation component for identifying and separating the effects of concurrent causal events; and a data allocation component for optimizing the allocation of shared sensor resources.

[0013] Furthermore, the data acquisition unit includes:

[0014] The hardware trigger synchronization submodule is used to generate a global trigger signal to enable all sensors to collect data synchronously.

[0015] The software timestamp alignment submodule is connected to the hardware trigger synchronization submodule and is used to align multi-sensor data through timestamp interpolation.

[0016] The data quality assessment submodule, connected to the software timestamp alignment submodule, is used to assess the integrity, noise level, and dynamic range of data from each sensor.

[0017] Furthermore, the dynamic relationship awareness module includes:

[0018] The multi-granularity relationship modeling submodule is used for parallel computation of instantaneous dynamic relationships and statistical dependencies;

[0019] The physical constraint correction submodule, connected to the multi-granularity relation modeling submodule, is used to correct the physical feasibility of relation weights based on sensor spatial location, field of view, and detection range.

[0020] The dynamic adjacency matrix construction submodule, connected to the physical constraint correction submodule, is used to construct an adaptively adjustable relation matrix through cosine similarity and temporal correlation calculations.

[0021] Furthermore, the causal fusion module includes:

[0022] The counterfactual fusion submodule is used to simulate sensor failure scenarios and fuse causal data;

[0023] The relation memory submodule, connected to the counterfactual fusion submodule, is used to store historical relation patterns and provide references through similarity retrieval;

[0024] The stability assessment submodule, connected to the counterfactual fusion submodule, is used to calculate the stability of relationships based on temporal consistency and causal effects.

[0025] Furthermore, the counterfactual fusion submodule includes:

[0026] Counterfactual simulation components are used to construct causal effect trees that transcend physical connections by simulating causal change events.

[0027] The causal event identification component, connected to the counterfactual simulation component, is used to detect causal change events occurring in the environment and identify the types of sensor data affected.

[0028] The causal data fusion component, connected to the causal event recognition component, is used to extract and fuse multi-sensor data based on causal events;

[0029] The selective transmission component, connected to the causal data fusion component, is used to control the content of data transmission and reduce redundant information transmission.

[0030] Furthermore, the steps of the causal event recognition component in detecting causal change events occurring in the environment and identifying the affected sensor data types include:

[0031] Detecting causal change events in the environment using pattern recognition algorithms;

[0032] Based on the causal impact tree, we analyze the expected types of sensor data and the extent of impact that each causal event will affect;

[0033] Generate an impact range vector for each causal event, where the vector elements represent the expected change patterns of each sensor data type;

[0034] Based on the influence range vector, the influence score of each sensor data type is calculated, and data types with influence scores exceeding a preset threshold are selected to generate a list of associated data types.

[0035] Based on the list of associated data types, determine the associated data types that need to participate in the fusion process.

[0036] Furthermore, based on the list of associated data types, the steps to determine the associated data types that need to participate in the fusion process include:

[0037] The severity of a causal event is determined based on its scope and extent of impact.

[0038] Based on the severity of the causal event, the associated sensors of the causal event are selected according to a preset boundary range;

[0039] Extract and determine the association type data of the preset type from the associated sensors.

[0040] Furthermore, the preset boundary range and preset type are positively correlated.

[0041] Furthermore, the steps for the event separation component to identify and separate the effects of concurrent causal events include:

[0042] Based on the relational memory submodule and the causal influence tree, data features that do not belong to a single causal event are identified.

[0043] Determine the quantity and type of causal events based on data characteristics;

[0044] Independent component analysis was used to separate different causal events and their corresponding data.

[0045] Furthermore, the steps by which the data allocation component optimizes the allocation of shared sensor resources include:

[0046] Identify shared sensors that are affected by a combination of different causal events;

[0047] Calculate the contribution weights of shared sensors to each causal event;

[0048] Sensor data fusion directions are assigned based on contribution weights and the severity of causal events.

[0049] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0050] Firstly, this invention uses a pattern recognition algorithm to monitor multi-sensor data streams in real time to detect causal change events occurring in the environment. Subsequently, based on a pre-constructed causal influence tree, it analyzes the sensor data types that the event is expected to affect and the degree of that impact, generating a quantified influence range vector for each causal event. The impact score for each sensor data type is calculated, and associated data types with influence scores exceeding a preset threshold are selected, generating a list of associated data types. Finally, strongly correlated data with the same representational meaning are extracted and merged. This solves the problems of large data redundancy and high transmission latency in traditional methods, reducing data redundancy caused by sudden events that have causal relationships.

[0051] Secondly, when multiple causal events are detected concurrently, the event separation component uses blind source separation techniques such as independent component analysis to decompose the mixed data stream into independent data streams corresponding to different causal events. Simultaneously, the data allocation component identifies shared sensors affected by multiple events and calculates their contribution weights to different causal events, thereby dynamically allocating their data processing resources. This solves the problems of mutual interference and resource competition when multiple events are concurrent, achieving parallel and accurate perception and fusion of multiple causal events in complex scenarios, and improving the robustness of causal data fusion.

[0052] Third, this invention utilizes a causal fusion module to dynamically adjust the preset boundary range and preset type of associated sensors based on the severity of causal events. When the severity of a causal event is detected to be low, the system automatically narrows the preset boundary range and reduces the preset data types, thus solving the problem of wasted data processing and transmission resources under normal conditions. When the severity of an event exceeds the allowable range, the system expands the preset boundary range and increases the preset data types, thus solving the problem of bandwidth occupation by normal or low-priority events on major or high-priority events, further reducing the transmission delay of major or high-priority events, and further improving robustness in complex scenarios. Attached Figure Description

[0053] Figure 1 This is a structural block diagram of a wireless sensing platform with edge computing capabilities provided in an embodiment of the present invention.

[0054] Figure 2 This is a structural block diagram of a causal fusion module of a wireless sensing platform with edge computing capabilities provided in an embodiment of the present invention.

[0055] Figure 3 A flowchart illustrating the identification of causal change events and the data types involved in the events in a wireless sensing platform with edge computing capabilities, provided as an embodiment of the present invention;

[0056] Figure 4 This is a flowchart illustrating how a wireless sensing platform with edge computing capabilities determines the associated data types that need to participate in fusion processing, as provided in an embodiment of the present invention. Detailed Implementation

[0057] To make the objectives and advantages of the present invention clearer, the present invention will be further described below with reference to embodiments; it should be understood that the specific embodiments described herein are merely for explaining the present invention and are not intended to limit the present invention.

[0058] Preferred embodiments of the present invention will now be described with reference to the accompanying drawings. Those skilled in the art should understand that these embodiments are merely illustrative of the technical principles of the present invention and are not intended to limit the scope of protection of the present invention.

[0059] It should be noted that in the description of this invention, the terms "upper", "lower", "left", "right", "inner", "outer", etc., which indicate the direction or positional relationship, are based on the direction or positional relationship shown in the drawings. This is only for the convenience of description and is not intended to indicate or imply that the device or element must have a specific orientation, or be constructed and operated in a specific orientation. Therefore, it should not be construed as a limitation of this invention.

[0060] Furthermore, it should be noted that, in the description of this invention, unless otherwise explicitly specified and limited, the terms "installation," "connection," and "linking" should be interpreted broadly. For example, they can refer to a fixed connection, a detachable connection, or an integral connection; they can refer to a mechanical connection or an electrical connection; they can refer to a direct connection or an indirect connection through an intermediate medium; and they can refer to the internal connection of two components. Those skilled in the art can understand the specific meaning of the above terms in this invention according to the specific circumstances.

[0061] This invention proposes a wireless sensing platform with edge computing capabilities, relating to the field of wireless sensor technology. Addressing the problems of high data redundancy and high transmission latency in existing wireless sensor technologies, this invention employs a causal event identification component for detecting causal change events in the environment and identifying the affected sensor data types; a causal data fusion component for extracting and fusing multi-sensor data based on causal events; an event separation component for identifying and separating the effects of concurrent causal events; and a data allocation component for optimizing the allocation of shared sensor resources. This invention reduces data redundancy with causal relationships caused by sudden events, reduces data transmission latency, and improves the robustness of causal data fusion in complex scenarios.

[0062] Example 1

[0063] like Figure 1 As shown, an embodiment of the present invention provides a wireless sensing platform with edge computing capabilities, comprising:

[0064] The data acquisition unit is used to collect raw environmental data from the multimodal sensor and transmit it to the edge processing unit.

[0065] The edge processing unit, connected to the data acquisition unit, is used to perform data fusion on the raw environmental data input from the data acquisition unit and output fused features, specifically including:

[0066] The edge processing unit includes:

[0067] The feature encoding module is used to encode the features of the data transmitted by each sensor.

[0068] The dynamic relationship perception module, connected to the feature encoding module, is used to construct a hierarchical sensor relationship model, including instantaneous dynamic relationships, semantic functional relationships, and statistical dependencies.

[0069] The causal fusion module, connected to the dynamic relationship perception module, is used to identify, extract, and fuse data with the same representational meaning. It includes a causal event identification component for detecting causal change events occurring in the environment and identifying the affected sensor data types; a causal data fusion component for extracting and fusing multi-sensor data based on causal events; an event separation component for identifying and separating the effects of concurrent causal events; and a data allocation component for optimizing the allocation of shared sensor resources.

[0070] For example, multimodal sensors include: temperature sensors, humidity sensors, audio sensors, vision sensors, force sensors, and smoke sensors;

[0071] The raw environmental data includes: temperature, humidity, audio information, visual information, stress information, and smoke information.

[0072] The feature encoding module performs standardized feature extraction on heterogeneous raw data;

[0073] Specifically, for visual information, a lightweight, pre-trained convolutional neural network model is used to extract the feature map before the last fully connected layer and globally average pool it into a 512-dimensional floating-point vector.

[0074] For audio information, the audio information is first converted into a Mel spectrogram, and then a convolutional neural network model is used to extract it into a 128-dimensional vector;

[0075] For scalar data such as temperature and humidity, they are first normalized, and then projected into a 16-dimensional vector through a fully connected neural network layer.

[0076] Ultimately, all sensor data is uniformly encoded into a fixed-dimensional feature vector.

[0077] The data acquisition unit includes:

[0078] The hardware trigger synchronization submodule is used to generate a global trigger signal to enable all sensors to collect data synchronously.

[0079] The software timestamp alignment submodule is connected to the hardware trigger synchronization submodule and is used to align multi-sensor data through timestamp interpolation.

[0080] The data quality assessment submodule, connected to the software timestamp alignment submodule, is used to assess the integrity, noise level, and dynamic range of data from each sensor.

[0081] For example, the hardware foundation of the data acquisition unit adopts a central acquisition gateway, which is equipped with a field-programmable gate array, a high-precision real-time clock, and multiple sensor interfaces.

[0082] Specifically, the clock circuit inside the field-programmable gate array (FPGA) generates a global trigger signal, which the FPGA then sends to all deployed sensors simultaneously. Upon receiving the global trigger signal, all sensors synchronously begin acquiring a frame of data.

[0083] For example, the software timestamp alignment submodule adds a high-precision kernel timestamp based on a precision clock protocol to each sensor data packet by running a driver integrated on the acquisition gateway's operating system.

[0084] For example, the data quality assessment submodule performs real-time checks on the input data stream to ensure data quality, specifically including integrity assessment, noise level assessment, and dynamic range assessment.

[0085] Specifically, the integrity, noise level, and dynamic range are evaluated by calculating the packet loss rate, signal-to-noise ratio, and duration above 90% of the range. If the evaluated data meets one of the following conditions, the packet loss rate is greater than 5%, the signal-to-noise ratio is less than 20dB, or the duration above 90% of the range is more than 30 seconds, the data is marked and discarded.

[0086] The dynamic relationship awareness module includes:

[0087] The multi-granularity relationship modeling submodule is used for parallel computation of instantaneous dynamic relationships and statistical dependencies;

[0088] The physical constraint correction submodule, connected to the multi-granularity relation modeling submodule, is used to correct the physical feasibility of relation weights based on sensor spatial location, field of view, and detection range.

[0089] The dynamic adjacency matrix construction submodule, connected to the physical constraint correction submodule, is used to construct an adaptively adjustable relation matrix through cosine similarity and temporal correlation calculations.

[0090] For example, the multi-granularity relationship modeling submodule uses the Pearson correlation coefficient to calculate short-term relationships and mutual information to calculate statistical dependencies.

[0091] Specifically, calculating instantaneous dynamic relationships involves: defining a sliding time window T = 5 seconds; extracting the principal components of the feature vectors of each sensor within the window, for example, using the first element of each vector as a representative value to form time series X1, X2, X3, X4; calculating the Pearson correlation coefficient between each pair of sensor time series; and normalizing the absolute value of the Pearson correlation coefficient to the [0,1] interval, using it as the instantaneous relationship weight. For example, within window T, if the feature value fluctuations detected by the visual sensor and the feature value fluctuations detected by the audio sensor are highly synchronized, then the instantaneous relationship weight between the visual sensor and the audio sensor is 0.9. Conversely, if the temperature changes slowly, then the instantaneous relationship weight between the temperature sensor and the visual sensor is 0.1.

[0092] Specifically, calculating statistical dependencies involves defining a long-term statistical window Ts = 24 hours, using histogram methods or kernel density estimation to calculate the mutual information between pairs of sensor data for the historical feature vector data within this window, and normalizing the mutual information to the [0,1] interval to obtain the statistical relationship weights. For example, historical data shows that whenever factory machines are started, the ambient temperature rises slowly, and even if the instantaneous correlation is not high, there is a significant statistical dependency between them.

[0093] For example, the physical constraint correction submodule performs physical feasibility correction on the relation weights based on the sensor's physical profile and spatial geometry calculations;

[0094] Specifically, for vision and audio, the overlap of their detection areas in the field of view or the overlap of their sound field coverage are calculated; for temperature and vibration, their Euclidean distance D_ij is calculated, and a distance attenuation function is applied. If the two sensors are physically impossible to perceive each other, the result is 0. For example, if the vision sensor and the temperature sensor are located in two isolated areas within a factory and are physically unrelated, even if their statistical data happen to depend on each other, this module will sever this unreasonable connection.

[0095] For example, the dynamic adjacency matrix construction submodule constructs a final adaptively adjustable dynamic relational adjacency matrix by fusing multiple relations. This matrix serves as a key input to the causal fusion module, guiding which closely related sensor data should be prioritized and fused under specific causal events.

[0096] This invention addresses the limitations of traditional methods that only consider single-type associations by constructing a hierarchical sensor relationship model that includes instantaneous dynamic relationships, semantic functional relationships, and statistical dependencies, thus enabling multi-dimensional in-depth mining of the collaborative sensing capabilities of sensor networks. By introducing a physical constraint correction mechanism based on sensor spatial location, field of view, and detection range, it solves the problem of physically infeasible connections that may occur in traditional relationship models, ensuring the physical rationality of relationship weight calculation. Furthermore, by constructing an adaptively adjustable dynamic adjacency matrix, it solves the problem that fixed relationship models cannot adapt to dynamic environmental changes, enabling real-time optimization and updating of the sensor relationship network and ensuring that the system always maintains optimal sensing performance.

[0097] like Figure 2 As shown, the causal fusion module includes:

[0098] The counterfactual fusion submodule is used to simulate sensor failure scenarios and fuse causal data;

[0099] The relation memory submodule, connected to the counterfactual fusion submodule, is used to store historical relation patterns and provide references through similarity retrieval;

[0100] The stability assessment submodule, connected to the counterfactual fusion submodule, is used to calculate the stability of relationships based on temporal consistency and causal effects.

[0101] For example, the counterfactual fusion submodule verifies and sanitizes causal relationships by constructing counterfactual assumptions, ensuring that the fusion result is based only on genuine causal connections.

[0102] Specifically, a graph neural network and attention mechanism are used for counterfactual simulation and data fusion. This involves: constructing a graph structure by using the current dynamic adjacency matrix as the connection relationships and the feature vectors of each sensor as the node features; inputting this graph structure into a graph attention network; and having the graph attention network calculate attention weights between nodes, aggregate information from neighboring nodes, and generate a context-aware fusion feature for each sensor node, thus obtaining a factual fusion result based on all current data relationships. Then, features of a key sensor node are randomly masked or discarded from the graph structure to create a counterfactual graph. This counterfactual graph is input into the same graph attention network to obtain the fusion result under the counterfactual scenario.

[0103] The difference between the factual and counterfactual results is calculated. If the difference is large, it indicates that the masked sensor has a significant causal impact on the final result, and its data is preserved and enhanced. If the difference is small, it indicates that the influence of the masked sensor may be redundant or non-causal, and its weight is reduced in the final causal fusion.

[0104] This invention effectively solves the problem of causal confusion in traditional fusion methods by simulating sensor failure scenarios and performing counterfactual reasoning. It can verify the true causal relationships between sensor data, eliminate spurious correlation interference, improve the accuracy of causal reasoning, and significantly enhance the reliability of the fusion results.

[0105] For example, the relational memory submodule is implemented using a vector database, specifically including memory storage, similarity retrieval, and reference provision.

[0106] Specifically, memory storage includes: after processing a typical causal event, such as a fire or an impact, storing the current situation as a memory fragment in a vector database. The key is the vectorized representation of the dynamic adjacency matrix under that event, representing the sensor relationship pattern unique to the event. The values ​​are the causal influence tree related to the event, successful fusion features, and the final system decision result.

[0107] Similarity retrieval includes: when a new causal event occurs, this submodule vectorizes the current dynamic adjacency matrix as a query vector and uses an approximate nearest neighbor search algorithm to quickly find the K most similar historical memory fragments to the current relational pattern in the vector database. For example, K=3.

[0108] The retrieval process includes providing retrieved historical memory fragments to the counterfactual fusion submodule and the stability assessment submodule as prior knowledge to accelerate the causal analysis process of the current event and improve its accuracy.

[0109] This invention establishes a vectorized storage and retrieval mechanism for historical relationship patterns, enabling the intelligent utilization of past experience. Its similarity retrieval function provides a reliable reference for current causal analysis, improves the speed of time-based identification, and reduces the risk of misjudgment through historical pattern comparison.

[0110] For example, the stability assessment submodule calculates a comprehensive stability score by combining time consistency assessment and causal effect stability assessment, and finally identifies which sensor relationships are long-term stable and causally strong, and which are transient, fluctuating or pseudo-correlated.

[0111] The counterfactual fusion submodule includes:

[0112] Counterfactual simulation components are used to construct causal effect trees that transcend physical connections by simulating causal change events.

[0113] The causal event identification component, connected to the counterfactual simulation component, is used to detect causal change events occurring in the environment and identify the types of sensor data affected.

[0114] The causal data fusion component, connected to the causal event recognition component, is used to extract and fuse multi-sensor data based on causal events;

[0115] The selective transmission component, connected to the causal data fusion component, is used to control the content of data transmission and reduce redundant information transmission.

[0116] For example, the counterfactual simulation component employs a causal discovery algorithm, combined with domain knowledge, to construct and update a directed acyclic graph, specifically including:

[0117] Preload a basic causal effect tree into the counterfactual simulation component; for example, a fire event will simultaneously affect the visual sensor, temperature sensor, humidity sensor, and smoke sensor.

[0118] Collect a large amount of historical data from multiple sensors under normal and abnormal conditions. Use PC algorithms to analyze the data through conditional independence tests, such as the G-squared test, to analyze the potential causal relationships between variables and generate an initial causal influence tree sketch.

[0119] Continuously conduct counterfactual simulations, for example: if the smoke sensor or vision sensor is damaged, can the data collected by the temperature sensor and humidity sensor accurately indicate the occurrence of a fire? By comparing the simulation results with real data, verify and quantify the strength of the edges in the causal influence tree.

[0120] Output a dynamically evolving causal influence tree represented by a weighted directed acyclic graph. The root node in the tree represents the root cause, such as fire or impact, the child nodes represent the observed phenomena affected by it, and the edge weights represent the strength of the causal influence.

[0121] The counterfactual simulation component in this invention achieves deep modeling and verification of complex causal relationships by constructing a causal influence tree that transcends physical connections. This component can simulate various hypothetical scenarios, effectively distinguishing between true causal relationships and superficial associations, significantly improving the system's understanding of causal mechanisms. Through counterfactual reasoning, the component enhances the accuracy and reliability of causal judgments, providing a more solid theoretical basis for system decision-making. This innovation endows the system with stronger causal reasoning capabilities, providing crucial technical support for emergency warning and intelligent decision-making, and overall improving the system's intelligence level and environmental adaptability.

[0122] like Figure 3 As shown, the steps of the causal event recognition component in detecting causal change events occurring in the environment and identifying the affected sensor data types include:

[0123] Detecting causal change events in the environment using pattern recognition algorithms;

[0124] Based on the causal impact tree, we analyze the expected types of sensor data and the extent of impact that each causal event will affect;

[0125] Generate an impact range vector for each causal event, where the vector elements represent the expected change patterns of each sensor data type;

[0126] Based on the influence range vector, the influence score of each sensor data type is calculated, and data types with influence scores exceeding a preset threshold are selected to generate a list of associated data types.

[0127] Based on the list of associated data types, determine the associated data types that need to participate in the fusion process.

[0128] For example, a strategy combining feature space anomaly detection and key signal mutation verification is used to detect causal change events in the environment. Specifically, an isolated forest model is established based on the data distribution under normal operating conditions, and the continuously flowing, standardized multi-sensor fusion feature vectors are input into the isolated forest model. When a new feature vector is identified as an anomaly by the model, it is initially marked as a potential causal change event. The CUSUM algorithm is run in parallel on the raw readings or feature values ​​of key sensors to monitor small but continuous shifts in their mean values. When the isolated forest and at least one CUSUM detector alarm simultaneously, the system confirms that a high-confidence causal change event has occurred and outputs the triggered causal event alarm.

[0129] For example, based on the established causal influence tree, i.e., a directed acyclic graph, the detected abnormal patterns are compared with the leaf node patterns under each root node in the causal influence tree using cosine similarity calculation. Finally, the root cause with the highest phase velocity is matched, such as a fire event or an impact event. Based on the connection of the fire event in the causal influence tree, the types of sensors that it is expected to directly affect are determined, such as temperature sensors, smoke sensors, and humidity sensors. The expected degree of influence is read from the edge weights of the causal influence tree.

[0130] For example, a fixed-length vector is defined that corresponds one-to-one with all sensor data types. Based on the analysis results of the causal influence tree, a value is assigned to each element of the vector. This value represents the expected change pattern and intensity of the data type. For example, the value is +0.95 for a temperature sensor, +0.90 for a smoke sensor, 0.0 for a vibration sensor, and -0.90 for a humidity sensor. Positive values ​​indicate an expected increase, negative values ​​indicate an expected decrease, and the magnitude of the absolute value indicates the intensity.

[0131] For example, in this embodiment, the element values ​​in the influence range vector are themselves influence scores. The preset influence score threshold is a fixed value pre-set based on the target requirements, which is used to distinguish strong correlation signals, weak correlation signals and noise. All vector elements are traversed, and vector elements that are strongly correlated with causal events are combined to generate a list of associated data types.

[0132] Furthermore, based on the list of associated data types, the causal event identification component extracts the corresponding sensor feature vector data within the time window of the causal event occurrence from the feature encoding module, packages these selected highly correlated data, and sends them to the causal data fusion component for the next step of fusion processing.

[0133] This invention addresses the high latency and high false negative rate of traditional single-threshold methods or fixed-rule event detection by employing a hybrid pattern recognition strategy that combines isolated forest unsupervised anomaly detection with the CUSUM mutation detection algorithm. This enables early, high-confidence detection of potential causal change events in the environment. By performing graph matching and similarity calculation between detected anomaly patterns and pre-generated causal influence trees, it overcomes the limitation of existing technologies that can only identify anomalies without understanding their specific types and root causes. This achieves accurate identification and characterization of causal event types, upgrading the system from perceiving anomalies to understanding events. Furthermore, by calculating influence scores based on influence range vectors and filtering sensor types exceeding preset thresholds to generate a list of associated data types, it solves the problem of massive data redundancy and wasted computational resources in multimodal sensing systems. This enables intelligent and dynamic selection of the data range participating in fusion processing, ensuring that subsequent processing modules only focus on key data subsets strongly correlated with the current core causal event.

[0134] like Figure 4 As shown, the steps for determining the subset of data to be included in the fusion process based on the list of associated data types are as follows:

[0135] The severity of a causal event is determined based on its scope and extent of impact.

[0136] Based on the severity of the causal event, the associated sensors of the causal event are selected according to a preset boundary range;

[0137] Extract and determine the association type data of the preset type from the associated sensors.

[0138] Preset boundary range and preset type are positively correlated

[0139] For example, the Euclidean norm is used to weight all elements in the influence range vector to obtain a severity score. Based on the calculation result, the causal event is divided into different degrees of severity. For example, the influence range vector of a fire event is [temperature sensor 0.95, delay sensor 0.90, brightness sensor 0.70], and the calculated severity score is 1.52.

[0140] Specifically, the severity levels are divided into low, medium, and high;

[0141] The severity score range corresponding to the lowest severity level is (0, 1);

[0142] The severity score range corresponding to intermediate severity is [1, 2];

[0143] The severity score range corresponding to the highest severity level is (2, +∞);

[0144] For causal events of low severity, only sensors with a direct causal link to the event are selected, i.e., the initial list of related sensors;

[0145] For causal events of moderate severity, the association list is expanded to include sensors physically adjacent to or within the same network subnet as the core sensor. For example, temperature and humidity sensors located in adjacent areas are included in the association scope.

[0146] For causal events of high severity, the scope is further expanded to include all sensors within the entire sensing network that have a potential semantic association with the event. For example, all temperature sensors and power consumption sensors on the entire floor are included in the association.

[0147] Furthermore, for causal events of low severity, only the core compressed features of the sensor are extracted. For example, for a vision sensor, only its 128-dimensional backbone feature vector describing global brightness and color is extracted;

[0148] For causal events of intermediate severity, more detailed features are added on top of the core features. For example, for visual sensors, an additional 256-dimensional feature vector describing texture is extracted;

[0149] For causal events of high severity, extract all available feature dimensions, or prepare raw data, such as image frames, for in-depth analysis.

[0150] The causal data fusion component merges the received data packets into a signal characterizing the relevant data of this emergency, and sends it to the selective transmission component, specifically including:

[0151] The data fusion component compares the received data packets with the causal events stored in the relational memory submodule;

[0152] Determine the type of causal event and generate a highly condensed causal fusion feature signal that can characterize the nature of this sudden event, such as a small fire occurring on the 4th floor of a building at 10:00, with a severity score of 1.52.

[0153] The generated causal fusion feature signal is output to the selective transmission component.

[0154] The causal event recognition component provided by this invention achieves deep perception and accurate understanding of environmental events through a multi-level causal reasoning mechanism. First, the component employs a hybrid recognition strategy combining isolated forest unsupervised anomaly detection and the CUSUM mutation detection algorithm to effectively capture abnormal patterns in sensor data streams, significantly improving the timeliness and accuracy of event detection and solving the response lag problem of traditional threshold detection methods. Based on a pre-constructed causal influence tree, the component compares detected abnormal patterns with historical causal knowledge through graph matching and similarity calculation. This not only identifies the occurrence of events but also accurately determines the event type and its root cause, thereby enhancing the system's depth of understanding of the event's essence and achieving a leap from perceiving anomalies to understanding events. By generating a quantified influence range vector, the component precisely characterizes the expected influence pattern and intensity of each causal event on each sensor data dimension, improving the accuracy of defining the event's influence range. Based on this, the component intelligently identifies key data dimensions strongly correlated with the current event through a dynamic threshold filtering mechanism, generating a list of associated data types. This significantly improves the targeting and efficiency of data fusion and effectively reduces resource consumption caused by redundant data processing.

[0155] The causal data fusion component provided by this invention achieves significant data simplification and transmission optimization during the data fusion stage through an innovative causal-driven data processing mechanism. Based on the list of associated data types output by the causal identification component, the component first intelligently compares the received data packets with causal event patterns stored in the relational memory to accurately determine the type and characteristics of the current emergency. On this basis, the component uses feature compression and reconstruction techniques to fuse multi-source heterogeneous sensor data to generate a highly condensed causal fusion feature signal. This signal can fully characterize the core features of the emergency. The intelligent fusion mechanism based on causal relationships significantly reduces redundant transmission of data repeatedly representing the same time period, thereby reducing network transmission load and improving data transmission speed, while ensuring the integrity and accuracy of key event information.

[0156] The steps involved in the event separation component to identify and separate the effects of concurrent causal events include:

[0157] Based on the relational memory submodule and the causal influence tree, data features that do not belong to a single causal event are identified.

[0158] Determine the quantity and type of causal events based on data characteristics;

[0159] Independent component analysis was used to separate different causal events and their corresponding data.

[0160] Specifically, when causal events occur concurrently, the system retrieves the historical event pattern most similar to the current observation data from the relational memory, establishes a residual vector between the observation data and the best matching pattern, calculates the Euclidean norm of the residual vector, and determines that there are unexplained data features when the residual vector exceeds a threshold.

[0161] When unexplained data features are identified, the original observation data and residual features are combined to construct an enhanced feature matrix. The features are then standardized using Z-scores to eliminate the influence of dimensions. The Gap Statistic method is applied to determine the optimal number of clusters K. The high-dimensional features are reduced in dimensionality using Laplace eigenmaps. For each cluster center, the Hausdorff distance between it and various event patterns in the causal influence tree is calculated. An event type matching degree matrix is ​​established, and the optimal allocation is completed based on the Hungarian algorithm. The determined number and specific type labels of concurrent events are output.

[0162] Furthermore, the observation data will be modeled as Where S represents each concurrent event, A represents the mixing matrix, and ε represents the observation noise, the FastICA algorithm is used to maximize the non-Gaussianity of the output components, the mutual information between the separated independent components and the data from each sensor is calculated, the component-event mapping relationship is established based on the mutual information, the time-frequency characteristics of each event are recovered by wavelet reconstruction technology, and finally the concurrent causal events are separated into independent causal events.

[0163] The event separation component of this invention effectively solves the identification challenge in complex scenarios where multiple causal events are intertwined through an advanced concurrent event processing mechanism. This component can accurately identify and separate multiple causal events occurring simultaneously. By establishing clear event boundaries and independent signal channels, the component ensures that each causal event receives dedicated processing and analysis, avoiding mutual interference and feature confusion between events. This innovation significantly enhances the system's perception reliability and decision-making accuracy in real-world environments, providing solid technical support for intelligent responses in complex scenarios and significantly improving the robustness of causal data fusion.

[0164] The steps by which the data allocation component optimizes the allocation of shared sensor resources include:

[0165] Identify shared sensors that are affected by a combination of different causal events;

[0166] Calculate the contribution weights of shared sensors to each causal event;

[0167] The sensor data fusion direction is assigned based on the contribution weights and the severity of the causal events.

[0168] When concurrent causal events involve a unified sensor, the data allocation component connects each causal event to the sensor affected by it, records the influence strength value of each connection (derived from the influence range vector), counts the number of events associated with each sensor, and selects those sensors that are simultaneously associated with two or more events as shared sensors. For example, a camera that is needed by both fire and intrusion events is marked as a shared sensor.

[0169] For each shared sensor, the system evaluates its contribution to each event through multiple factors, including the strength of causal relationship, historical performance, and data quality. The contribution weight of the shared sensor to each event is calculated through weighted summation and normalization.

[0170] The severity score for each event is obtained. The contribution weight is multiplied by the event severity to obtain a priority score. The share of sensor resources that each event should receive is determined according to the priority score ratio. Multiple data output channels are established for shared sensors. The sensor data stream is distributed to different channels according to the calculated ratio. Each channel corresponds to a specific event's fusion processing flow. When the event severity changes significantly, the score is recalculated and allocated.

[0171] This invention's data allocation component effectively solves the resource contention problem in the event of multiple concurrent causal events by intelligently identifying shared sensors and establishing a multi-factor evaluation system. Through dynamic weight calculation and priority allocation mechanisms, the component ensures that high-severity events receive priority access to high-quality sensor resources while maintaining overall system efficiency. By establishing multiple data output channels and a real-time reallocation mechanism, it achieves efficient reuse and intelligent scheduling of sensor data streams, significantly improving resource utilization and event processing accuracy in complex scenarios, and guaranteeing the processing performance of critical tasks.

[0172] The technical solution of the present invention has been described above with reference to the preferred embodiments shown in the accompanying drawings. However, it will be readily understood by those skilled in the art that the scope of protection of the present invention is obviously not limited to these specific embodiments. Without departing from the principles of the present invention, those skilled in the art can make equivalent changes or substitutions to the relevant technical features, and the technical solutions after these changes or substitutions will all fall within the scope of protection of the present invention.

Claims

1. A wireless sensing platform with edge computing function, characterized in that, The application relates to a multi-modal sensor data fusion system, comprising: a data acquisition unit for acquiring raw environmental data of multi-modal sensors and delivering the raw environmental data to an edge processing unit; an edge processing unit connected with the data acquisition unit, for data fusion on the raw environmental data input from the data acquisition unit, and outputting fused features, specifically comprising: a feature encoding module for feature encoding on the data delivered by each sensor; a dynamic relationship perception module connected with the feature encoding module, comprising a multi-granularity relationship modeling submodule for parallel calculation of instantaneous dynamic relationships and statistical dependency relationships between sensor data; a physical constraint correction submodule connected with the multi-granularity relationship modeling submodule, for physical feasibility correction on relationship weights based on sensor spatial positions, field angles and detection ranges; a dynamic adjacency matrix construction submodule connected with the physical constraint correction submodule, for construction of a self-adaptable relationship matrix through cosine similarity and time correlation calculation, as input of a causal fusion module; the causal fusion module connected with the dynamic relationship perception module, for identification, extraction and fusion of data with the same representation meaning, comprising an anti-fact fusion submodule for simulation of sensor failure scenarios and fusion of causal data; a relationship memory bank submodule connected with the anti-fact fusion submodule, for storage of historical relationship patterns and provision of references through similarity retrieval; a stability evaluation submodule connected with the anti-fact fusion submodule, for relationship stability calculation based on time consistency and causal effect; an event separation component for identification and separation of influences of concurrent causal events; a data distribution component connected with the event separation component, for optimization of distribution of shared sensor resources; the anti-fact fusion submodule comprises: an anti-fact simulation component for construction of a causal influence tree beyond physical connection through simulation of causal change events; a causal event identification component connected with the anti-fact simulation component, for detection of causal change events occurring in the environment and identification of affected sensor data types; a causal data fusion component connected with the causal event identification component, for extraction and fusion of multi-sensor data based on causal events; a selective transmission component connected with the causal data fusion component, for control of data transmission content and reduction of redundant information transmission. 2.The wireless sensing platform with edge computing capability of claim 1, wherein, the data acquisition unit comprises: a hardware trigger synchronization submodule for generation of a global trigger signal to synchronize all sensors for acquisition; a software timestamp alignment submodule connected with the hardware trigger synchronization submodule, for realization of multi-sensor data alignment through timestamp interpolation; a data quality evaluation submodule connected with the software timestamp alignment submodule, for evaluation of the integrity, noise level and dynamic range of each sensor data. 3.The wireless sensing platform with edge computing capability of claim 1, wherein, the step of detecting causal change events occurring in the environment and identifying affected sensor data types by the causal event identification component comprises: detection of causal change events in the environment through a pattern recognition algorithm; analysis of sensor data types expected to be affected by each causal event and the degree of influence based on a causal influence tree. generating an influence range vector for each causal event, wherein vector elements represent expected change patterns of each sensor data type; calculating influence scores of each sensor data type based on the influence range vector, and screening out data types with influence scores exceeding a preset threshold to generate a list of associated data types; determining associated type data that needs to participate in fusion processing based on the list of associated data types. 4.The wireless sensing platform with edge computing capability of claim 3, wherein, The step of determining associated type data that needs to participate in fusion processing based on the list of associated data types comprises: determining severity of a causal event based on its influence range and influence degree; selecting associated sensors of the causal event according to a preset boundary range based on the severity of the causal event; extracting and determining associated type data of a preset type from the associated sensors. 5.The wireless sensing platform with edge computing capability of claim 4, wherein, The preset boundary range and the preset type are in a positive correlation relationship. 6.The wireless sensing platform with edge computing capability of claim 1, wherein, The step of the event separation component identifying and separating influences of concurrent causal events comprises: identifying data features not belonging to a single causal event based on a relationship memory bank sub-module and a causal influence tree; determining the number and type of causal events based on the data features; separating different causal events and corresponding data through independent component analysis. 7.The wireless sensing platform with edge computing capability of claim 1, wherein, The step of the data allocation component optimizing allocation of shared sensor resources comprises: identifying shared sensors commonly affected by different causal events; calculating contribution weights of the shared sensors to each causal event; allocating sensor data fusion directions based on the contribution weights and the severity of the causal events.

Images