IoT-based intelligent temperature sensor data fusion method

Abstract

This invention relates to the field of Internet of Things (IoT) technology and discloses a data fusion method for intelligent temperature sensors based on IoT. The method includes: sensing nodes uploading state data such as real-time temperature values, temperature change rate, ambient humidity, and ambient light intensity, as well as preliminary discrimination data types such as temperature fluctuation range labels, sensor operating status labels, and whether the temperature is abnormal, to a cloud platform. After receiving the current sensor network topology data, the cloud platform extracts features from the sensing data to obtain state feature vectors and preliminary discrimination feature vectors. Based on the topology data, the state feature vectors are optimized, and the two types of feature vectors are fused to obtain a multi-source fusion feature representation. A fusion type label is determined, and a preset strategy is matched from a fusion strategy library and distributed to the sensing nodes. This method achieves efficient temperature data fusion and enhances the value of data utilization.

Description

Technical Field

[0001] This invention relates to the field of Internet of Things (IoT) technology, specifically to a data fusion method for intelligent temperature sensors based on IoT. Background Technology

[0002] With the rapid development of IoT technology, smart temperature sensors are widely used in numerous fields. As application scenarios continue to expand, the demands on temperature data processing are also increasing. Traditional temperature sensor data processing methods face many challenges.

[0003] In practical applications, sensing nodes collect a large amount of temperature-related data. This data includes not only real-time temperature values, but also various state data such as temperature change rate, ambient humidity, and ambient light intensity. It also includes preliminary data types for the sensing side, such as temperature fluctuation range labels, sensor operating status labels, and whether temperature anomalies have occurred. However, traditional methods often fail to effectively integrate and process this multi-dimensional data.

[0004] Existing data fusion methods suffer from insufficient and incomplete feature extraction when processing this data. For the temporal characteristics of sensing node state data, traditional methods fail to fully utilize techniques such as temporal sequence coding for effective feature mining, resulting in an inability to accurately capture the changing patterns of temperature data over time. Furthermore, when processing the initial discrimination data types at the sensing side, there is a lack of efficient encoding and clustering methods, making it difficult to obtain accurate initial discrimination feature vectors.

[0005] Furthermore, in considering the impact of sensor network topology on data fusion, traditional methods do not make good use of the current sensor network topology data and cannot optimize the spatial topology association of the state feature vectors of sensing nodes. This results in the neglect of the spatial topology relationship between sensor nodes during the data fusion process, which affects the effectiveness of data fusion.

[0006] In the process of data fusion, there are also shortcomings in the fusion processing of different types of feature vectors. Traditional methods cannot handle the heterogeneity between the preliminary data type discrimination feature vector and the optimized sensing node state feature vector well, and cannot achieve cross-dimensional alignment and effective splicing, thus making it difficult to obtain accurate multi-source fusion feature representation of temperature data.

[0007] Moreover, when determining the fusion type label for temperature data, traditional methods lack efficient data classification models and cannot accurately determine the fusion type label based on the multi-source fusion feature representation, which in turn affects the accuracy and efficiency of matching preset fusion strategies from the fusion strategy library. Summary of the Invention

[0008] The purpose of this invention is to provide a data fusion method for intelligent temperature sensors based on the Internet of Things, so as to solve the problems mentioned in the background art.

[0009] To achieve the above objectives, the present invention provides a data fusion method for intelligent temperature sensors based on the Internet of Things, the method comprising:

[0010] Sensing nodes upload sensing data to the cloud platform. This sensing data includes sensing node status data and preliminary data type discrimination from the sensing side. The cloud platform receives current sensor network topology data from the IoT positioning system. The cloud platform extracts features from the sensing node status data and the preliminary data type discrimination from the sensing side to obtain sensing node status feature vectors and preliminary data type discrimination feature vectors. Based on the current sensor network topology data, the cloud platform performs spatial topology correlation optimization on the sensing node status feature vectors to obtain optimized sensing node status feature vectors. The cloud platform fuses the preliminary data type discrimination feature vectors and the optimized sensing node status feature vectors to obtain a multi-source fusion feature representation of temperature data. Based on the multi-source fusion feature representation of temperature data, the cloud platform determines a temperature data fusion type label. Based on the temperature data fusion type label, the cloud platform matches a preset fusion strategy from the fusion strategy library and distributes the preset fusion strategy to the corresponding sensing nodes.

[0011] Preferably, the sensing node status data includes real-time temperature value, temperature change rate, ambient humidity, and ambient light intensity, and the sensing side initially determines the data types including temperature fluctuation range label, sensor working status label, and whether a temperature anomaly has occurred.

[0012] Preferably, the cloud platform performs feature extraction on the sensing node state data and the sensing side's preliminary data type discrimination to obtain sensing node state feature vectors and data type preliminary discrimination feature vectors, including: performing temporal sequence encoding on each of the sensing node state data to obtain a set of sensing node state temporal embedding encoding vectors; and performing hierarchical clustering on the set of sensing node state temporal embedding encoding vectors to obtain the sensing node state feature vectors.

[0013] Preferably, the cloud platform extracts features from the sensing node state data and the preliminary data type of the sensing side to obtain sensing node state feature vectors and preliminary data type discrimination feature vectors, and further includes: binary encoding each of the preliminary data types of the sensing side to obtain a set of binary encoded vectors of the preliminary data types of the sensing side; performing mean-shift clustering on the set of binary encoded vectors of the preliminary data types of the sensing side to obtain M local discrimination feature vectors of the data types; and calculating a weighted average of the M local discrimination feature vectors of the data types to obtain the preliminary data type discrimination feature vector.

[0014] Preferably, the cloud platform performs spatial topology association optimization on the sensing node state feature vector based on the current sensor network topology data to obtain an optimized sensing node state feature vector, including: performing a topology feature extractor based on an adjacency matrix on the current sensor network topology data to obtain a sensor network topology feature encoding matrix; and mapping the sensing node state feature vector to the topology space of the sensor network topology feature encoding matrix through vector concatenation to obtain the optimized sensing node state feature vector.

[0015] Preferably, the cloud platform integrates the preliminary data type discrimination feature vector and the optimized sensing node state feature vector to obtain a multi-source fusion feature representation of temperature data, including: concatenating the preliminary data type discrimination feature vector and the optimized sensing node state feature vector to obtain the multi-source fusion feature representation of temperature data.

[0016] Preferably, the cloud platform determines the temperature data fusion type label based on the multi-source fusion feature representation of the temperature data, including: inputting the multi-source fusion feature representation of the temperature data into a data classifier based on a decision tree model to obtain the temperature data fusion type label.

[0017] Preferably, concatenating the preliminary data type discrimination feature vector and the optimized sensing node state feature vector to obtain the multi-source fusion feature representation of the temperature data includes: utilizing the matrix norm consistency under a heterogeneous spatial structure to perform cross-dimensional alignment on the preliminary data type discrimination feature vector and the optimized sensing node state feature vector to obtain an optimized preliminary data type discrimination feature vector; and concatenating the optimized preliminary data type discrimination feature vector and the optimized sensing node state feature vector to obtain the multi-source fusion feature representation of the temperature data.

[0018] Preferably, leveraging the matrix norm consistency under a heterogeneous spatial structure, cross-dimensional alignment is performed on the preliminary data type discrimination feature vector and the optimized sensing node state feature vector to obtain an optimized preliminary data type discrimination feature vector. This includes: performing joint spatial standardization on the preliminary data type discrimination feature vector based on the optimized sensing node state feature vector to obtain a joint spatial standardization matrix; performing a full-dimensional heterogeneous mapping in Euclidean space on the preliminary data type discrimination feature vector and the optimized sensing node state feature vector to obtain a full-dimensional heterogeneous mapping vector; and performing a linear transformation homogenization based on the heterogeneous spatial structure on the preliminary data type discrimination feature vector using a normalization matrix based on the joint spatial standardization matrix and the full-dimensional heterogeneous mapping vector to obtain an optimized preliminary data type discrimination feature vector.

[0019] Preferably, the process of performing temporal sequence encoding on the state data of each of the sensing nodes to obtain a set of temporal embedding encoding vectors for the state of the sensing nodes includes: dividing the state data of the sensing nodes into time windows to obtain multiple time series segments; performing sliding window sampling on each time series segment to obtain a temporal feature sub-vector; and concatenating the temporal feature sub-vectors after dimensional alignment to obtain the set of temporal embedding encoding vectors for the state of the sensing nodes.

[0020] Compared with the prior art, the beneficial effects of the present invention are:

[0021] The IoT-based intelligent temperature sensor data fusion method proposed in this invention has several significant advantages. Sensing nodes upload sensing data, including real-time temperature values, temperature change rates, ambient humidity, ambient light intensity, and other status data, as well as preliminary discrimination data such as temperature fluctuation range labels, sensor operating status labels, and whether temperature anomalies have occurred, to a cloud platform. The cloud platform receives current sensor network topology data from the IoT positioning system. By extracting features from the sensing node status data and the preliminary discrimination data, sensing node status feature vectors and preliminary discrimination feature vectors are obtained. Specifically, the sensing node status data undergoes temporal sequence encoding, including time window division, sliding window sampling, and dimensional alignment and concatenation of temporal feature sub-vectors, resulting in a temporal embedding encoded vector set. Hierarchical clustering is then performed to comprehensively and deeply mine the temporal features of the data. The preliminary discrimination data is binary encoded to obtain a binary encoded vector set. Mean-shift clustering is used to obtain local discrimination feature vectors, and a weighted average is calculated to obtain the preliminary discrimination feature vector, effectively processing the preliminary discrimination data.

[0022] Based on the current sensor network topology data, a topology feature encoding matrix is ​​obtained through a topology feature extractor based on the adjacency matrix. The state feature vectors of sensing nodes are mapped to the topology space for spatial topology association optimization. This fully considers the spatial topology relationships between sensor nodes, making the state feature vectors more reflective of the impact of the actual network topology on the data.

[0023] When fusing two types of feature vectors, cross-dimensional alignment is achieved using matrix norm consistency under heterogeneous spatial structures. This includes joint spatial standardization, full-dimensional heterogeneous mapping, and linear transformation homogenization, which solves the heterogeneity problem of feature vectors and makes the multi-source fused feature representation of the spliced ​​temperature data more accurate. Inputting the multi-source fused feature representation into a data classifier based on a decision tree model accurately determines the temperature data fusion type label, and then matches a preset fusion strategy from the fusion strategy library and distributes it to the corresponding sensing nodes. Attached Figure Description

[0024] Figure 1This is a schematic diagram illustrating the working principle of the IoT-based intelligent temperature sensor data fusion method described in this invention.

[0025] Figure 2 A flowchart for extracting the state feature vector of a sensing node;

[0026] Figure 3 A flowchart for the initial feature vector extraction for data type identification;

[0027] Figure 4 A flowchart for topology optimization of the feature vector space of sensing node states;

[0028] Figure 5 This is a flowchart for optimizing feature vector concatenation in heterogeneous spatial structures. Detailed Implementation

[0029] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0030] Please see Figures 1-5 This invention provides a data fusion method for intelligent temperature sensors based on the Internet of Things, the specific steps of which are as follows:

[0031] The sensing nodes upload sensing data to the cloud platform, which includes sensing node status data and preliminary judgment data types from the sensing side.

[0032] The cloud platform receives current sensor network topology data from the IoT positioning system.

[0033] The cloud platform extracts features from the state data of the sensing nodes and the preliminary data type discrimination data of the sensing side to obtain the state feature vector of the sensing nodes and the preliminary data type discrimination feature vector.

[0034] Based on the current sensor network topology data, the cloud platform performs spatial topology correlation optimization on the state feature vectors of sensing nodes to obtain optimized state feature vectors of sensing nodes.

[0035] The cloud platform integrates the data type preliminary discrimination feature vector and the optimized sensing node state feature vector to obtain a multi-source fusion feature representation of temperature data.

[0036] The cloud platform determines the temperature data fusion type label based on the multi-source fusion feature representation of temperature data.

[0037] The cloud platform matches a preset fusion strategy from the fusion strategy library based on the temperature data fusion type label, and then distributes the preset fusion strategy to the corresponding sensing nodes.

[0038] Example 1: The sensing node status data includes real-time temperature values, temperature change rate, ambient humidity, and ambient light intensity. This data represents the specific information acquired by the sensing node when sensing its surrounding environment. The real-time temperature value reflects the current ambient temperature, the temperature change rate indicates how quickly the temperature changes over time, ambient humidity represents the water vapor content in the environment, and ambient light intensity represents the intensity of light. The sensing side initially identifies data types including temperature fluctuation range labels, sensor operating status labels, and whether a temperature anomaly has occurred. The temperature fluctuation range label identifies the range of temperature fluctuations within a certain time period, the sensor operating status label indicates whether the sensor is working normally, and the temperature anomaly label determines whether a temperature anomaly has occurred.

[0039] The cloud platform extracts features from the state data of sensing nodes and the data types initially identified by the sensing side. This is to extract vectors that reflect the essential characteristics of the data from these raw data for subsequent processing and analysis. During feature extraction from the state data of sensing nodes, time-series encoding is performed on the state data of each sensing node. The state data of sensing nodes is then divided into time windows, resulting in multiple time series segments. The division of time windows needs to consider the temporal characteristics of the data and processing requirements; for example, time windows can be divided according to certain time intervals, with each time window containing data within a specific time period.

[0040] Sliding window sampling is performed on each time series segment to obtain temporal feature sub-vectors. The size of the sliding window and the sliding step size affect the sampling results and need to be set appropriately according to the characteristics of the data and the purpose of the analysis. Sliding window sampling allows for the extraction of features from different time periods within the time series segments. The temporal feature sub-vectors are then concatenated after dimensional alignment to obtain a set of temporal embedding encoding vectors for the perception node state. Dimensional alignment ensures that the dimensions of each temporal feature sub-vector are consistent for the concatenation operation. The concatenated set of temporal embedding encoding vectors for the perception node state contains the feature information of the perception node state data under different time windows and sliding windows.

[0041] After obtaining the set of temporal embedding encoding vectors of the sensing node states, hierarchical clustering is performed on this set to obtain the feature vectors of the sensing node states. Hierarchical clustering is a commonly used clustering method that groups similar samples together by calculating the similarity between samples, forming different hierarchical structures. Through hierarchical clustering, similar vectors in the set of temporal embedding encoding vectors of the sensing node states can be clustered into one class, thereby extracting feature vectors that can represent the state of the sensing node.

[0042] In this process, each step needs to be executed precisely to ensure that the extracted sensor node state feature vectors accurately reflect the characteristics of the sensor node state data. For example, if the time window size is not set appropriately during time window partitioning, important temporal features of the data may not be captured; if the sliding window step size is too large during sliding window sampling, the sampled features may be incomplete; if the dimension alignment is inaccurate during dimension alignment, the concatenated vector may not correctly reflect the characteristics of the data; and if the clustering parameters are not set properly during hierarchical clustering, the clustering results may be inaccurate, thus affecting the quality of the sensor node state feature vectors.

[0043] During the process of sensing nodes uploading sensing data to the cloud platform, it is necessary to ensure the accuracy and stability of data transmission to guarantee that the cloud platform can receive the correct sensing data. The cloud platform receives current sensor network topology data from the IoT positioning system. This data describes the location and connection relationships of each sensor node in the sensor network, providing a foundation for subsequent spatial topology association optimization.

[0044] Example 2: When the cloud platform extracts features from the sensing node status data and the preliminary discrimination data type from the sensing side, in addition to performing time-series sequence encoding and hierarchical clustering on the sensing node status data to obtain the sensing node status feature vector, it also needs to process the preliminary discrimination data type from the sensing side. The preliminary discrimination data type from the sensing side includes temperature fluctuation range labels, sensor operating status labels, and whether a temperature anomaly has occurred. These labels represent data features in a discrete form and need to be converted into a numerical vector form suitable for computer processing.

[0045] Each initially identified data type from the sensing side is initially encoded using binary encoding. Binary encoding converts each label into a binary number string. For example, the temperature fluctuation range label might have categories such as "normal," "larger," and "abnormal," with each category corresponding to a unique binary code, such as "normal" encoded as 001, "larger" as 010, and "abnormal" as 100. In this way, each initially identified data type from the sensing side is converted into a binary encoding vector, resulting in a set of binary encoding vectors for each initially identified data type from the sensing side. Each vector in this set corresponds to the encoding result of one initially identified data type from the sensing side, containing the category information of that data type.

[0046] Mean-shift clustering is applied to the set of binary encoded vectors of the initially identified data types from the sensing side. Mean-shift clustering is a density-based clustering method. Its core idea is to iteratively calculate the shift mean of each sample point, moving the sample points towards higher density regions, ultimately dividing the sample points into different clusters. In this process, an appropriate bandwidth parameter needs to be set, as the bandwidth parameter determines the fineness of the clustering. Too large a bandwidth may cause samples of different categories to be clustered into one category; too small a bandwidth may cause samples of the same category to be divided into different clusters. Through mean-shift clustering, the set of binary encoded vectors of the initially identified data types from the sensing side is divided into M clusters, each cluster corresponding to a local discriminant feature vector of a data type.

[0047] A weighted average of M local discriminant feature vectors for different data types is calculated to obtain a preliminary data type discriminant feature vector. During weighted averaging, the weight of each local discriminant feature vector can be determined based on factors such as the number of samples in its cluster and the cluster density. For example, clusters with a larger number of samples correspond to larger weights, ensuring that the preliminary data type discriminant feature vector better reflects the characteristics of the majority of samples. Through weighted averaging, the multiple local discriminant feature vectors for different data types are merged into a comprehensive feature vector, which contains the overall characteristics of the data type initially discriminated by the perceptual side.

[0048] After extracting features from the state data of the sensing nodes and the initial data types identified by the sensing side, the cloud platform needs to perform spatial topology correlation optimization on the state feature vectors of the sensing nodes based on the current sensor network topology data. The current sensor network topology data is provided by the IoT positioning system and describes the positional relationships and connection methods of each sensing node in the sensor network, such as which sensing nodes are adjacent and whether the communication links between them are stable.

[0049] A topology feature extractor based on the adjacency matrix is ​​used to process the current sensor network topology data. The adjacency matrix represents the connectivity between nodes in a graph, with each element indicating whether an edge connects the nodes. By processing the adjacency matrix using the topology feature extractor, topological features of the sensor network can be extracted, such as node degree, clustering coefficients, and shortest paths, resulting in a sensor network topology feature encoding matrix. This matrix converts the sensor network's topology into a numerical form, facilitating subsequent processing.

[0050] The state feature vectors of sensing nodes are mapped to the topological space of the sensor network topological feature encoding matrix through vector concatenation. Vector concatenation involves combining the state feature vectors of sensing nodes with corresponding elements in the sensor network topological feature encoding matrix, allowing the state feature vectors of sensing nodes to be integrated into the topological space of the sensor network. Through this mapping, the state feature vectors of sensing nodes not only contain their own state characteristics but also their spatial topological relationship characteristics within the sensor network, thus obtaining optimized state feature vectors of sensing nodes.

[0051] In this process, the accuracy of each step is crucial. For example, if the encoding rules are set improperly in binary encoding, information loss may occur; if the bandwidth parameter is not selected properly in mean-shift clustering, the accuracy of the clustering results will be affected; if the weights are not determined scientifically in weighted averaging, the preliminary data type discrimination feature vector will not accurately reflect the data characteristics; if the features extracted by the topology feature extractor are not comprehensive when processing the adjacency matrix, the quality of the sensor network topology feature encoding matrix will be affected; if the mapping method is incorrect during vector concatenation mapping, the optimized sensing node state feature vector will not be able to correctly integrate spatial topological relationship features.

[0052] When receiving current sensor network topology data, the cloud platform needs to ensure the real-time nature and accuracy of the data, because the topology of the sensor network may change as sensing nodes are added, removed, or fail. Only by processing the latest and most accurate topology data can an effective and optimized sensing node state feature vector be obtained.

[0053] Example 3: When the cloud platform merges the initial discrimination feature vector of data types and the optimized sensing node state feature vector to obtain the multi-source fusion feature representation of temperature data, it is necessary to perform cross-dimensional alignment processing on these two feature vectors to solve the feature fusion problem in heterogeneous spaces. The specific process is as follows:

[0054] By leveraging the matrix norm consistency under heterogeneous spatial structures, cross-dimensional alignment is performed on the preliminary data type discrimination feature vector and the optimized sensing node state feature vector to obtain the optimized preliminary data type discrimination feature vector. Here, the optimized sensing node state feature vector is denoted as V.s It is obtained by optimizing the spatial topological association of the state feature vectors of sensing nodes, and includes the state features and spatial topological relationship features of the sensing nodes; the preliminary data type discrimination feature vector is denoted as V. d It is obtained by binary encoding, mean-shift clustering and weighted averaging of the data types initially identified on the sensing side, and reflects the overall characteristics of the data types on the sensing side.

[0055] Based on the optimized sensing node state feature vector V s Preliminary identification of feature vector V for data types d Joint spatial normalization is performed to obtain the joint spatial normalization matrix N. The purpose of joint spatial normalization is to map two eigenvectors to the same spatial scale for subsequent processing. Specifically, the joint spatial normalization matrix N is calculated as follows: For V... s and V d Normalization is performed to ensure that their norms are within the same range. For example, for a vector V, its L2 norm is defined as... Where v i Let be the i-th element of vector V. By analyzing V... s and V d The norms are calculated and normalized separately to obtain the standardized vectors, and then the joint space normalization matrix N is constructed.

[0056] Preliminary identification of feature vector V for data types d and optimized sensor node state feature vector V s A full-dimensional heterogeneous mapping in Euclidean space is performed to obtain a full-dimensional heterogeneous mapped vector H. Full-dimensional heterogeneous mapping in Euclidean space maps vectors from two different dimensions or feature spaces to a common high-dimensional space to preserve their feature information. In this process, the mapping method needs to consider the feature distribution and spatial structure of the two vectors to ensure that the mapped vector accurately reflects the features of the original vectors. For example, linear or nonlinear mapping can be used to map V... d and V s Mapping to the same high-dimensional space yields a full-dimensional heterogeneous mapping vector H.

[0057] Based on the joint spatial normalization matrix N and the full-dimensional heterogeneous mapping vector H, the data type preliminary discrimination feature vector V is obtained using the normalization matrix W. d A linear transformation based on heterogeneous spatial structures is performed to homogenize the data, resulting in an optimized data type preliminary discriminant feature vector V′. d Here, the normalization matrix W normalizes the mapped vector, ensuring it conforms to a specific format or requirement. The process of homogenization through linear transformation can be represented as:

[0058] V′ d =W×(N×V) d +H)

[0059] Where × represents matrix multiplication or vector operation. In this formula, W is the normalization matrix, used to normalize the transformed vector; N is the joint space normalization matrix, used to perform space normalization on the preliminary data type discrimination feature vectors; V d The initial data type discrimination feature vector is generated from the original vector; H is a full-dimensional heterogeneous mapping vector used to introduce mapping features from the heterogeneous space. Through this linear transformation, the initial data type discrimination feature vector V is generated. d Transformed into an optimized sensor node state feature vector V s Preliminary identification of feature vector V′ based on optimized data types under the same spatial structure d This enables cross-dimensional alignment.

[0060] After completing the cross-dimensional alignment, the optimized data type is initially used to determine the feature vector V′. d and optimized sensor node state feature vector V s The data is concatenated to obtain the multi-source fusion feature representation F of the temperature data. The concatenation method involves joining the two vectors sequentially to form a new vector. For example, if V′ d The dimension is m, V s If the dimension of the multi-source fusion feature representation of temperature data is n, then the dimension of the concatenated temperature data multi-source fusion feature representation F is m+n. By concatenating, the preliminary data type discrimination features and the optimized sensing node state features are fused together to form a feature representation containing multi-source information, which can more comprehensively reflect the characteristics of temperature data.

[0061] In this process, each step needs to be executed precisely to ensure that the fused feature representation is accurate and effective. For example, during joint space normalization, if the normalization method is incorrect, it may lead to inconsistent scales of feature vectors, affecting subsequent mapping and transformation; during full-dimensional heterogeneous mapping, if the mapping method cannot preserve the information of the original feature vectors well, it may lead to feature loss; during linear transformation homogenization, if the normalization matrix W is not chosen reasonably, the transformed feature vector may not accurately reflect the original features; during concatenation, if the concatenation order or method is incorrect, the fused feature representation may not be able to correctly integrate multi-source information.

[0062] It is important to note that during cross-dimensional alignment and stitching, the accuracy and consistency of the data must be ensured. For example, the optimized sensor node state feature vector V... s Preliminary identification of feature vector V based on data type dData must correspond to the same sensing node or the same time period; otherwise, the fused feature representation will lose its meaning. Simultaneously, the cloud platform needs sufficient computing and storage capabilities to ensure processing efficiency and data security when processing these feature vectors.

[0063] Example 4: The cloud platform determines the temperature data fusion type label based on the multi-source fusion feature representation of temperature data. Specifically, the multi-source fusion feature representation of temperature data is input into a data classifier based on a decision tree model to obtain the temperature data fusion type label. Here, a specific application scenario is used as an example for detailed description: Assume that multiple sensing nodes are deployed in a greenhouse environment. Each sensing node collects data such as temperature, humidity, and light intensity in real time and uploads it to the cloud platform.

[0064] The cloud platform has already obtained the multi-source fusion feature representation of temperature data through the preceding steps. Taking one sensing node as an example, the sensing node status data uploaded by this node includes a real-time temperature value of 25℃, a temperature change rate of 0.5℃ / h, an ambient humidity of 60%, and an ambient light intensity of 3000 lux. The preliminary data type classification by the sensing side includes a temperature fluctuation range label of "normal," a sensor working status label of "normal," and a label indicating whether a temperature anomaly has occurred of "no." The cloud platform extracts features from this data to obtain a sensing node status feature vector and a preliminary data type classification feature vector. After processing such as spatial topology association optimization and cross-dimensional alignment, the final multi-source fusion feature representation of temperature data is obtained. This feature representation is a vector containing multiple feature information, for example, it can be represented as [25,0.5,60,3000,0,1,0,0,1,0,...] (this is just an example; the actual dimensions may be more). The first few elements correspond to the features of the sensing node status data such as the real-time temperature value and the temperature change rate, while the later elements correspond to the features of the preliminary data type classification by the sensing side such as the temperature fluctuation range label, as well as the features after spatial topology association optimization.

[0065] The multi-source fusion feature representation of the temperature data is input into a data classifier based on a decision tree model. The decision tree model is a common machine learning model that makes classification decisions by constructing a tree structure. This model has already been trained using a large amount of historical data, which includes various multi-source fusion feature representations of temperature data and corresponding temperature data fusion type labels. For example, the training data might contain samples where: when the real-time temperature value in the multi-source fusion feature representation is between 20-30℃, the temperature change rate is small, the ambient humidity is moderate, the light intensity is normal, and the initial data types are all judged to be normal by the sensing side, the corresponding temperature data fusion type label is "normal fusion"; when the real-time temperature value exceeds 35℃, the temperature change rate is large, and the label for whether a temperature anomaly has occurred is "yes," the corresponding temperature data fusion type label is "abnormal emergency fusion," and so on.

[0066] After receiving the multi-source fusion feature representation of temperature data, the decision tree model starts from the root node and judges each feature value in the feature vector, traversing downwards along the corresponding branches until it reaches a leaf node. The category corresponding to the leaf node is the predicted temperature data fusion type label. For example, for the example feature vector mentioned above, the decision tree model first determines whether the real-time temperature value is within a certain range. Suppose it judges whether 25℃ is between 20-30℃, and the result is yes. Then it judges whether the temperature change rate is less than a certain threshold, 0.5℃ / h is less than a preset threshold. Next, it judges whether features such as ambient humidity and light intensity are within the normal range, and whether the data types initially judged by the sensing side are all normal. When all these conditions are met, the decision tree model will traverse to the corresponding leaf node, thus obtaining the temperature data fusion type label as "normal fusion".

[0067] After obtaining the temperature data fusion type label, the cloud platform needs to match a preset fusion strategy from the fusion strategy library based on this label and then distribute the preset fusion strategy to the corresponding sensing nodes. The fusion strategy library stores multiple preset fusion strategies, each corresponding to a different temperature data fusion type label. For example, the fusion strategy library may contain "normal fusion," "emergency fusion," and "sensor fault fusion" strategies. For the "normal fusion" strategy, the sensing nodes may be required to collect data at a normal sampling frequency, and the cloud platform may use a standard fusion algorithm to process the data. For the "emergency fusion" strategy, the sensing nodes may be required to increase their sampling frequency and upload data in real time, and the cloud platform may use a more complex and precise fusion algorithm, providing real-time feedback of the fusion results to relevant personnel.

[0068] Continuing with the previous example, when the temperature data fusion type label is "Regular Fusion," the cloud platform will find the corresponding "Regular Fusion" preset fusion strategy from the fusion strategy library. This strategy may include specific fusion algorithm parameters, data transmission frequency, and other information. The cloud platform then distributes this preset fusion strategy to the corresponding sensing nodes. After receiving the preset fusion strategy, the sensing nodes collect and process data according to the strategy's specifications, such as collecting temperature data at a normal sampling frequency, and uploading it to the cloud platform in a specified manner. The cloud platform then fuses the data according to the fusion algorithm in the strategy to obtain more accurate temperature data.

[0069] In this process, attention must be paid to the construction and training quality of the decision tree model. If the training data for the decision tree model is incomplete, or if the parameters are set improperly during training, it may lead to inaccurate classification, resulting in an inappropriate pre-defined fusion strategy. For example, if the training data does not include samples from certain special cases, the decision tree model may misclassify such cases, leading to the issuance of incorrect fusion strategies. Furthermore, the pre-defined fusion strategies in the fusion strategy library need to be designed and updated appropriately according to actual application scenarios to ensure they meet the data fusion needs under different conditions. For example, as the greenhouse environment changes or new requirements emerge, it may be necessary to add new fusion strategies or adjust existing ones.

[0070] When matching preset fusion strategies and distribution strategies, the cloud platform needs to ensure the stability and accuracy of communication to guarantee that the sensing nodes can correctly receive the preset fusion strategies. If data loss or errors occur during communication, the sensing nodes may be unable to operate according to the correct strategy, affecting the effectiveness of data fusion.

[0071] Example 5: The time-series encoding process of the sensing node status data is a key processing step. Taking the temperature monitoring scenario in an industrial production workshop as an example, the sensing nodes deployed in the workshop continuously collect data such as real-time temperature values, temperature change rate, ambient humidity, and ambient light intensity. These data change dynamically over time, and the time dimension features need to be extracted through time-series encoding.

[0072] First, time windows are divided. Assuming the sensing nodes collect data at 1-minute intervals, the raw data can be divided into 10-minute time windows, each containing 10 data records. The division of time windows must consider the temporal correlation of the data. If the window is too short, it may fail to capture the trend of temperature changes; if the window is too long, too many irrelevant temporal features may be mixed in. For example, in industrial production, the heating process of some equipment may be completed within tens of minutes, so a 10-minute window division can better reflect the temperature change characteristics in the early stages of equipment operation. After division, the original continuous time series data is split into multiple independent time series segments, each segment corresponding to data within a time window.

[0073] Next, a sliding window sampling is performed on each time series segment. The sliding window size is set to 3 minutes, with a step size of 1 minute. Taking the first time window (0-10 minutes) as an example, the first sliding window covers the data from 0-3 minutes, extracting features such as the average temperature and maximum humidity within this time period to form the first time series feature sub-vector. The second sliding window slides to 1-4 minutes with a step size of 1 minute, similarly extracting features to form the second sub-vector, and so on, until the entire 10-minute time window is covered. The size and step size of the sliding window need to be adjusted according to the frequency of data fluctuations. If temperature changes drastically, the sliding window size can be reduced to capture instantaneous changes; if temperature changes are gradual, the window size can be increased to extract overall trend features.

[0074] After sliding window sampling, multiple temporal feature sub-vectors are obtained. The dimensions of these sub-vectors may differ depending on the extracted features, requiring dimension alignment. For example, each sub-vector might originally contain four features: mean temperature, rate of change, extreme humidity, and mean light intensity. However, due to temporary malfunctions of light sensors within some sliding windows, light intensity data may be missing. In this case, interpolation is needed to fill in the missing values, or it can be uniformly stipulated that each sub-vector must contain the same feature dimension, with missing features represented by a specific label (e.g., -999), ensuring that all temporal feature sub-vectors have consistent dimensions. After dimension alignment, these sub-vectors are concatenated in chronological order to form a temporal embedding encoding vector for the state of the sensing node. For example, a 10-minute time window, sampled through a sliding window, yields eight sub-vectors with consistent dimensions, which are concatenated to form a temporal embedding encoding vector for the state of the sensing node containing 8 × 4 = 32 dimensions.

[0075] In practical applications, different sensing nodes may have different sampling frequencies, and some nodes may experience inconsistent data upload times due to communication delays. In such cases, it is necessary to calibrate the data with timestamps before dividing the time window to ensure that the data within the same time window belongs to the same time interval. For example, if a sensing node experiences a 2-minute delay in data upload due to network fluctuations, it is necessary to classify the data into the correct time window based on the timestamp during processing to avoid incorrect division of time series segments.

[0076] After constructing the temporal embedding encoding vector for a single time window, the encoding vectors from all time windows are aggregated. Assuming the sensing node operates continuously for one hour, a total of six time windows are generated (one every 10 minutes). Each window generates a 32-dimensional encoding vector, resulting in a set of six sensing node state temporal embedding encoding vectors. Each vector in this set represents a feature of the sensing data within a different time period, collectively reflecting the temporal variation pattern of the sensing node's state data.

[0077] Hierarchical clustering is performed based on the similarity of vectors within the set. Euclidean distance is used to calculate the distance between vectors; vectors closer in distance have higher similarity. During clustering, each vector is initially treated as an independent cluster, and then the closest clusters are gradually merged until a preset number of clusters is reached or a clustering termination condition is met. For example, the preset clustering of 6 vectors might be into two classes: one corresponding to the temperature characteristics during normal device operation, and the other corresponding to the temperature characteristics during the device startup phase. After clustering, the center vector of each cluster becomes the sensing node's state feature vector. This vector integrates the common features of all time-series embedded encoding vectors within the same class, representing the state characteristics of the sensing node within a specific time period.

[0078] In an industrial workshop setting, if equipment near a sensing node starts up, its temperature data will initially rise rapidly and then stabilize. By using temporal sequence encoding and hierarchical clustering, the time window vectors of the startup phase can be clustered into one class, and the vectors of the stable phase into another. The resulting sensing node state feature vectors accurately reflect the characteristics of different stages of equipment operation. This processing method effectively filters out random noise in the data, retains valuable temporal features, and provides high-quality feature input for subsequent spatial topology association optimization.

[0079] Throughout the process, the parameter settings for the time window and sliding window need to be adjusted according to the specific application scenario. For example, in a greenhouse environment, where temperature changes are relatively slow, the time window can be set to 30 minutes and the sliding window size to 5 minutes. In a data center, due to the precise control of the air conditioning system, temperature changes are small but sensitive to abnormal fluctuations; therefore, the time window can be set to 5 minutes and the sliding window size to 1 minute to capture subtle temperature change characteristics. Furthermore, the method for handling missing values ​​during dimension alignment needs to be selected based on the data characteristics. For non-critical features such as light intensity, mean interpolation can be used; for critical features such as temperature, if missing values ​​occur, the data in that sliding window can be discarded to avoid erroneous features affecting the overall coding quality.

[0080] It should be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, method, article, or apparatus.

[0081] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A data fusion method for intelligent temperature sensors based on the Internet of Things, characterized in that, include: The sensing nodes upload sensing data to the cloud platform. The sensing data includes sensing node status data and data types that are initially identified by the sensing side. The cloud platform receives current sensor network topology data from the IoT positioning system; The cloud platform extracts features from the sensing node state data and the preliminary data type discrimination data from the sensing side to obtain sensing node state feature vectors and preliminary data type discrimination feature vectors. Based on the current sensor network topology data, the cloud platform performs spatial topology association optimization on the sensing node state feature vectors to obtain optimized sensing node state feature vectors. This includes: using an adjacency matrix-based topology feature extractor on the current sensor network topology data to obtain a sensor network topology feature encoding matrix; mapping the sensing node state feature vectors to the topology space of the sensor network topology feature encoding matrix through vector concatenation to obtain the optimized sensing node state feature vectors; fusing the preliminary data type discrimination feature vectors and the optimized sensing node state feature vectors to obtain a multi-source fusion feature representation of temperature data; determining a temperature data fusion type label based on the temperature data fusion type label; and matching a preset fusion strategy from the fusion strategy library based on the temperature data fusion type label, and distributing the preset fusion strategy to the corresponding sensing nodes. The cloud platform integrates the preliminary data type discrimination feature vector and the optimized sensing node state feature vector to obtain a multi-source fusion feature representation of temperature data, including: concatenating the preliminary data type discrimination feature vector and the optimized sensing node state feature vector to obtain the multi-source fusion feature representation of temperature data; utilizing the matrix norm consistency under heterogeneous spatial structure, performing cross-dimensional alignment on the preliminary data type discrimination feature vector and the optimized sensing node state feature vector to obtain an optimized preliminary data type discrimination feature vector; and concatenating the optimized preliminary data type discrimination feature vector and the optimized sensing node state feature vector to obtain the multi-source fusion feature representation of temperature data.

2. The data fusion method for intelligent temperature sensors based on the Internet of Things according to claim 1, characterized in that, The sensing node status data includes real-time temperature value, temperature change rate, ambient humidity, and ambient light intensity. The sensing side initially determines the data types including temperature fluctuation range label, sensor working status label, and whether a temperature anomaly has occurred.

3. The data fusion method for intelligent temperature sensors based on the Internet of Things according to claim 2, characterized in that, The cloud platform extracts features from the state data of the sensing nodes and the preliminary data type discrimination of the sensing side to obtain the state feature vector of the sensing nodes and the preliminary data type discrimination feature vector of the data type. This includes: performing temporal sequence encoding on the state data of each sensing node to obtain a set of temporal embedding encoding vectors of the sensing node state; and performing hierarchical clustering on the set of temporal embedding encoding vectors of the sensing node state to obtain the state feature vector of the sensing nodes.

4. The IoT-based intelligent temperature sensor data fusion method according to claim 3, characterized in that, The cloud platform extracts features from the sensing node state data and the preliminary data type discrimination data on the sensing side to obtain sensing node state feature vectors and preliminary data type discrimination feature vectors. The process also includes: binary encoding each of the preliminary data types on the sensing side to obtain a set of binary encoded vectors for the preliminary data type discrimination data; performing mean-shift clustering on the set of binary encoded vectors for the preliminary data type discrimination data to obtain M local discrimination feature vectors for the data type; and calculating a weighted average of the M local discrimination feature vectors for the data type to obtain the preliminary data type discrimination feature vector.

5. The IoT-based intelligent temperature sensor data fusion method according to claim 4, characterized in that, The cloud platform determines the temperature data fusion type label based on the multi-source fusion feature representation of the temperature data, including: inputting the multi-source fusion feature representation of the temperature data into a data classifier based on a decision tree model to obtain the temperature data fusion type label.

6. The IoT-based intelligent temperature sensor data fusion method according to claim 5, characterized in that, By leveraging the matrix norm consistency under a heterogeneous spatial structure, cross-dimensional alignment is performed on the preliminary data type discrimination feature vector and the optimized sensing node state feature vector to obtain an optimized preliminary data type discrimination feature vector. This includes: performing joint spatial standardization on the preliminary data type discrimination feature vector based on the optimized sensing node state feature vector to obtain a joint spatial standardization matrix; performing a full-dimensional heterogeneous mapping in Euclidean space on the preliminary data type discrimination feature vector and the optimized sensing node state feature vector to obtain a full-dimensional heterogeneous mapping vector; and performing a linear transformation homogenization based on the heterogeneous spatial structure on the preliminary data type discrimination feature vector using a normalization matrix based on the joint spatial standardization matrix and the full-dimensional heterogeneous mapping vector to obtain an optimized preliminary data type discrimination feature vector.

7. The IoT-based intelligent temperature sensor data fusion method according to claim 3, characterized in that, The process of performing temporal sequence encoding on the state data of each of the sensing nodes to obtain a set of temporal embedding encoding vectors for the state of the sensing nodes includes: dividing the state data of the sensing nodes into time windows to obtain multiple time series segments; performing sliding window sampling on each time series segment to obtain a temporal feature sub-vector; and concatenating the temporal feature sub-vectors after dimensional alignment to obtain the set of temporal embedding encoding vectors for the state of the sensing nodes.

Images