Low-power data fusion method and system for multi-protocol heterogeneous internet of things terminal

Abstract

This invention discloses a low-power data fusion method and system for multi-protocol heterogeneous IoT terminals, relating to the technical field of data fusion processing. The method includes: identifying the communication protocol types in the multi-protocol heterogeneous IoT terminals through a fusion gateway, and collecting raw sensor data from each terminal; performing noise removal and outlier cleaning and converting the data into structured data; introducing an overcomplete dictionary for dimensionality reduction to obtain dimensionality-reduced data features, and constructing a correlation network reflecting the degree of correlation between trusted data sources; adjusting the weights of conflicting evidence, and weighted fusion outputting the comprehensive data fused from the multi-protocol heterogeneous IoT terminals. This invention solves the technical problems of low data fusion accuracy, high power consumption, and the tendency for conflicting evidence to dominate the results in existing multi-protocol heterogeneous IoT terminals at the edge, achieving the technical effect of improving the credibility and accuracy of edge-side fused data while reducing the power consumption of terminal-gateway communication and computation.

Description

Technical Field

[0001] This invention relates to the field of data fusion processing technology, specifically to a low-power data fusion method and system for multi-protocol heterogeneous IoT terminals. Background Technology

[0002] Smart cities, Industry 4.0, smart agriculture, and environmental monitoring exhibit multi-protocol heterogeneity. Terminal devices may transmit data based on different physical layer and data link layer protocols such as ZigBee, BLE, LoRa, Wi-Fi, 6LoWPAN, and even wired Modbus. This heterogeneity poses a severe challenge to data aggregation and fusion processing at the edge. Traditional IoT platforms typically deploy independent gateways or parsing services for each protocol, resulting in redundant hardware resources, complex management, and poor scalability at edge nodes. When a large number of heterogeneous terminals simultaneously aggregate raw sensor data to the edge, additional computing resources must be consumed for protocol identification, format conversion, and time synchronization. Wireless sensor networks are constrained by channel interference, multipath effects, and node hardware drift, inevitably resulting in noise impulses, outliers, and missing segments mixed into the collected raw data. If low-quality data is not effectively removed, it will severely distort the results of decision analysis. Furthermore, data cleaning based on thresholds or statistical distributions is insufficient to distinguish between "real physical correlations" and "false coincidences." This leads to the accidental deletion of valid information during the preprocessing stage. Most existing fusion strategies assume that each sensor source has equal credibility or preset fixed weight coefficients. However, in real-world scenarios, the remaining energy of nodes, channel quality, sensor aging, and attack / defense status of different protocol terminals change dynamically, resulting in significant time-varying differences in the credibility of each data source. Without the ability to dynamically perceive the credibility of evidence sources, the final output composite data will deviate significantly from the true physical state. In addition, the storage and transmission bandwidth of edge gateways are extremely limited. Uploading all high-dimensional structured data for fusion may drastically increase the uplink load and increase the communication power consumption between the terminal and the gateway.

[0003] Therefore, current technologies suffer from technical problems such as low data fusion accuracy, high power consumption, and conflicting evidence dominating the results at the edge of multi-protocol heterogeneous IoT terminals. Summary of the Invention

[0004] This application provides a low-power data fusion method and system for multi-protocol heterogeneous IoT terminals, which solves the technical problems of low data fusion accuracy, high power consumption, and conflicting evidence dominating the results in the prior art. It achieves the technical effect of improving the credibility and accuracy of edge-side fused data, while reducing the power consumption of terminal and gateway communication and computing.

[0005] This application provides a low-power data fusion method for multi-protocol heterogeneous IoT terminals. The method includes: identifying the communication protocol type used by each terminal in the multi-protocol heterogeneous IoT terminals through a fusion gateway deployed at the network edge, and calling the corresponding protocol parsing engine to collect the raw sensor data uploaded by each terminal; performing noise removal and outlier cleaning on the raw sensor data, and converting the cleaned data into structured data in a unified format; introducing an overcomplete dictionary to perform dimensionality reduction processing on the structured data to obtain dimensionality reduction data features corresponding to each terminal, and constructing an association network reflecting the degree of correlation between trusted data sources based on the dimensionality reduction data features; adjusting the weights of target conflict evidence with the association network as constraints, and performing weighted fusion based on the adjusted weights to output the comprehensive data fused from the multi-protocol heterogeneous IoT terminals.

[0006] In a possible implementation, noise removal and outlier cleaning are performed on the original sensor data, and the cleaned data is converted into structured data in a unified format. This includes: performing noise removal processing on the original sensor data and obtaining the data distribution of the noise-removed data; sequentially calculating the first quartile and the third quartile of the data distribution, using the difference between the first quartile and the third quartile as the interquartile range; using the first quartile minus a first multiple of the interquartile range as a lower threshold, and the third quartile plus a second multiple of the interquartile range as an upper threshold to form a filtering constraint; identifying data that does not meet the filtering constraint as outliers and removing them to obtain preprocessed data, and converting the preprocessed data into the structured data according to preset rules.

[0007] In a possible implementation, the preprocessed data is converted into the structured data according to a preset rule, including: obtaining the communication protocol type corresponding to the preprocessed data; determining whether the data structure complexity of the communication protocol type is higher than a preset complexity threshold; if it is higher, calling a first number of parallel computing nodes to perform format conversion on the data of the communication protocol type, and releasing more than a second number of parallel computing nodes after the conversion is completed to reduce energy consumption; if it is not higher, calling a single computing node to perform format conversion on the data of the communication protocol type.

[0008] In a possible implementation, an overcomplete dictionary is introduced to reduce the dimensionality of the structured data, obtaining the dimensionality-reduced data features corresponding to each terminal. This includes: constructing an overcomplete dictionary with the number of atoms being a multiple of the dimension of the structured data; initializing the dictionary atoms and sparse coefficient matrix of the overcomplete dictionary; in each iteration, fixing the current dictionary atoms and updating the sparse coefficient matrix, then fixing the current sparse coefficient matrix and updating the dictionary atoms, alternating iterations until the reconstruction error is lower than a preset tolerance or the number of iterations reaches the upper limit; and finally using the obtained sparse coefficient matrix as the dimensionality-reduced data features.

[0009] In a possible implementation, constructing an overcomplete dictionary with an atomic number that is a first multiple of the structured data dimension includes: obtaining the data dimension of the structured data and using the first multiple of the data dimension as the initial atomic number of the overcomplete dictionary; randomly initializing the initial atomic number of dictionary atoms and normalizing each dictionary atom to obtain an initial overcomplete dictionary; obtaining any data sample from the structured data and using the absolute value of the inner product of the arbitrary data sample and each dictionary atom in the initial overcomplete dictionary as a dictionary atom selection constraint; selecting a target dictionary atom set from the initial overcomplete dictionary with the goal of maximizing the dictionary atom selection constraint, and using the target dictionary atom set as the overcomplete dictionary.

[0010] In a possible implementation, constructing an association network reflecting the degree of correlation between trusted data sources based on the dimensionality-reduced data features includes: performing clustering processing on each of the dimensionality-reduced data features to obtain clustering results; identifying and removing untrusted nodes based on the clustering results, and filtering out target feature sets corresponding to each trusted node from each of the dimensionality-reduced data features; and constructing the association network based on the similarity between features in the target feature set and using the similarity as the association weight between each trusted node.

[0011] In a possible implementation, identifying and removing untrusted nodes based on the clustering results includes: extracting the cluster center of any cluster in the clustering results and calculating the inter-cluster distance between any cluster and its neighboring clusters; identifying neighboring clusters whose inter-cluster distance is less than a predetermined distance threshold as overlapping clusters; marking the overlapping data nodes corresponding to the overlapping clusters as candidate abnormal nodes; obtaining the historical data deviation rate of the candidate abnormal nodes, and using the historical data deviation rate being higher than a predetermined deviation rate as a removal constraint, removing the candidate abnormal nodes as untrusted nodes.

[0012] In a possible implementation, the target conflict evidence is weighted and corrected using the association network as a constraint, and then weighted and fused based on the corrected weights to output the comprehensive data fused from the multi-protocol heterogeneous IoT terminals. This includes: extracting first conflict data corresponding to the first conflict data source in the target conflict evidence; analyzing the first conflict data under the association network based on a predetermined feature dimension, and constructing a first data feature parameter set for the first conflict data; performing weighted analysis on the first data feature parameter set to obtain a first weight coefficient; weighting the first conflict data with the first weight coefficient to obtain the target valid data of the target conflict evidence; and performing data fusion of the multi-protocol heterogeneous IoT terminals based on the target valid data to obtain the comprehensive data.

[0013] In possible implementations, the predetermined feature dimensions include sample size, distribution characteristics, and data importance.

[0014] This application also provides a low-power data fusion system for multi-protocol heterogeneous IoT terminals. The system includes: a sensor data acquisition module, used to identify the communication protocol type used by each terminal in the multi-protocol heterogeneous IoT terminals through a fusion gateway deployed at the network edge, and call the corresponding protocol parsing engine to collect the raw sensor data uploaded by each terminal respectively; a structured data acquisition module, used to perform noise removal and outlier cleaning on the raw sensor data, and convert the cleaned data into structured data in a unified format; a data dimensionality reduction processing module, used to introduce an overcomplete dictionary to perform dimensionality reduction processing on the structured data, obtain the dimensionality reduction data features corresponding to each terminal, and construct an association network reflecting the degree of correlation between trusted data sources based on the dimensionality reduction data features; and a data fusion output module, used to adjust the weights of target conflict evidence with the association network as constraints, and perform weighted fusion based on the adjusted weights to output the comprehensive data fused from the multi-protocol heterogeneous IoT terminals.

[0015] This application proposes a low-power data fusion method and system for multi-protocol heterogeneous IoT terminals. The method identifies the communication protocol types of the multi-protocol heterogeneous IoT terminals through a fusion gateway and collects raw sensor data from each terminal. Noise removal and outlier cleaning are performed, and the data is converted into structured data. An overcomplete dictionary is introduced for dimensionality reduction to obtain dimensionality-reduced data features, and a correlation network reflecting the degree of correlation between trusted data sources is constructed. Target conflict evidence is weighted and adjusted, and a weighted fusion output is obtained, resulting in comprehensive data fused from the multi-protocol heterogeneous IoT terminals. This solves the technical problems of low data fusion accuracy, high power consumption, and the tendency for conflict evidence to dominate the results in existing multi-protocol heterogeneous IoT terminals at the edge. It achieves the technical effect of improving the credibility and accuracy of edge-side fused data while reducing the power consumption of terminal-gateway communication and computation. Attached Figure Description

[0016] To more clearly illustrate the technical solutions of the embodiments of this disclosure, the accompanying drawings of the embodiments of this disclosure will be briefly described below. Flowcharts are used in this application to illustrate the operations performed by the system according to the embodiments of this application. It should be understood that the preceding or following operations are not necessarily performed precisely in sequence. Instead, various steps can be processed in reverse order or simultaneously as needed. Furthermore, other operations can be added to these processes, or one or more steps can be removed from these processes.

[0017] Figure 1 This is a schematic diagram of a low-power data fusion method for multi-protocol heterogeneous IoT terminals provided in an embodiment of this application.

[0018] Figure 2 This is a schematic diagram of a low-power data fusion system architecture for multi-protocol heterogeneous IoT terminals provided in an embodiment of this application.

[0019] Figure labeling: 10 sensor data acquisition module, 20 structured data acquisition module, 30 data dimensionality reduction processing module, 40 data fusion output module. Detailed Implementation

[0020] To further illustrate the technical means and effects adopted by the present invention in order to achieve the intended purpose, the following detailed description is provided in conjunction with the accompanying drawings and preferred embodiments, based on the specific implementation methods, structures, features and effects of the present invention.

[0021] This application provides a low-power data fusion method for multi-protocol heterogeneous IoT terminals, such as... Figure 1 As shown, the method includes: Step S100: Identify the communication protocol type used by each terminal in the multi-protocol heterogeneous IoT terminal through the converged gateway deployed at the network edge, and call the corresponding protocol parsing engine to collect the raw sensor data uploaded by each terminal respectively.

[0022] Preferably, the network edge refers to the network location between IoT terminal devices and cloud data centers, such as embedded gateway devices on the industrial field side, edge computing nodes, or access gateways on the base station side. The converged gateway integrates physical layer transceiver units for multiple communication protocols, such as LoRa RF chips, ZigBee coordinator modules, BLE Bluetooth modules, Wi-Fi modules, and RS485 wired interfaces, enabling simultaneous access to terminal devices using different physical layer transmission standards. In the same IoT application, there may be terminal devices using different communication standards. For example, temperature sensors use the ZigBee protocol to transmit data, humidity sensors use the LoRa protocol, power monitoring terminals use the Modbus RTU protocol for RS485 wired transmission, and mobile inspection terminals use the BLE protocol. After receiving a data frame from a terminal, the converged gateway parses the synchronization word, frame start delimiter, or protocol version number field in the frame header and matches it with the internally preset protocol feature library (containing standard frame structure templates for each protocol) to determine the communication protocol type to which the frame belongs. Then, the corresponding protocol parsing engine is invoked to collect the raw sensor data uploaded by each terminal. The protocol parsing engine is an internal software unit of the gateway. Each engine implements complete frame parsing functions for a specific communication protocol, including physical layer synchronization and demodulation, data link layer frame delimitation, address verification, CRC / FCS verification, and network layer / application layer payload extraction. The gateway polls or interrupts the reading of the output results of each protocol parsing engine to obtain the raw sensor data uploaded by each terminal, such as the voltage AD value output by the thermocouple and the resistance voltage AD value of the resistive humidity sensor.

[0023] Step S200: Perform noise removal and outlier cleaning on the original sensor data, and convert the cleaned data into structured data in a unified format.

[0024] Step S200 further includes performing noise removal processing on the original sensing data and obtaining the data distribution of the noise-removed data; sequentially calculating the first quartile and the third quartile of the data distribution, using the difference between the first quartile and the third quartile as the interquartile range; using the first quartile minus a first multiple of the interquartile range as a lower threshold, and the third quartile plus a second multiple of the interquartile range as an upper threshold to form a filtering constraint; determining data that does not meet the filtering constraint as outliers and removing them to obtain preprocessed data, and converting the preprocessed data into the structured data according to preset rules.

[0025] Preferably, noise removal processing refers to filtering the original sensor data, including but not limited to: smoothing filtering; using moving average or median filtering for continuously sampled time-series data to remove high-frequency random fluctuations caused by sensor thermal noise or quantization noise; trend removal; differential or high-pass filtering for slow-drift components in the data; eliminating baseline shifts introduced by environmental temperature changes or sensor aging; then, statistically analyzing the set of values ​​for all data samples after filtering, calculating the sorted sequence, minimum, maximum, and frequency distribution of this set, and obtaining the data distribution of the noise-removed data; the first quartile is the value at the 25th percentile after sorting all noise-removed data samples in ascending order, which can be calculated using the position method: position = (n+1) × 0.25, taking this position relative to... The corresponding sorting value, the third quartile is the value at the 75th percentile after sorting, position = (n+1) × 0.75. The difference between the first quartile and the third quartile is used as the interquartile range, reflecting the dispersion of the middle 50% of the samples in the dataset, and is not sensitive to extreme values ​​at both ends. The first multiple of the first quartile minus the interquartile range is used as the lower threshold. Data samples below this threshold are judged as abnormally low values, where the first multiple is a preset positive coefficient with a value of 1.5. The second multiple of the third quartile plus the interquartile range is used as the upper threshold. Data samples above this threshold are judged as abnormally high values, where the second multiple is a preset positive coefficient, which may be the same as the first multiple. The lower and upper thresholds define the numerical range that valid data should meet, forming a filtering constraint used to determine whether each data point is valid.

[0026] Preferably, data that does not meet the screening constraints is identified as outliers and marked as outliers. The causes may include sensor momentary failure, bit jumps during transmission, or strong external interference coupling into the signal link. The data sample is removed from the dataset to obtain preprocessed data, in which each sample is within the normal range of the statistical distribution. Predefined rules refer to a predefined data format conversion mapping table, containing multiple mapping relationships, physical quantity scaling conversion coefficients to convert AD readings into actual physical dimensions, such as mapping AD values ​​of 0-4095 to -40℃ to +85℃; field name mapping, unifying synonymous fields in different protocols into standardized field names, such as mapping "temp", "T", and "temperature" to "temperature_value"; timestamp standardization, synchronizing the time counting bases of different terminals to the unified UTC millisecond timestamp of the Network Time Protocol (NTP); unit system unification, converting Fahrenheit to Celsius, and unifying the relative humidity percentage to a decimal representation of 0-100, etc.; converting preprocessed data into structured data according to preset rules, usually stored in tabular form, with each row representing a sensor record and each column representing a standardized data field, such as timestamp, device ID, data type, value, unit, etc.

[0027] Furthermore, step S200 also includes obtaining the communication protocol type corresponding to the preprocessed data; determining whether the data structure complexity of the communication protocol type is higher than a preset complexity threshold; if it is higher, calling a first number of parallel computing nodes to perform format conversion on the data of the communication protocol type, and releasing more than a second number of parallel computing nodes after the conversion is completed to reduce energy consumption; if it is not higher, calling a single computing node to perform format conversion on the data of the communication protocol type.

[0028] Preferably, the protocol identification unit inside the fusion gateway maintains a mapping table between terminal identifiers and communication protocol types. When preprocessed data is ready to enter the format conversion stage, the gateway looks up the communication protocol type used by the terminal in the mapping table based on the terminal identifier carried by each piece of data. Data structure complexity refers to the computational load required to process the data payload during parsing and conversion. It is determined by the number of data fields contained in the protocol data frame, whether there are multi-level nested TLV structures or the depth of JSON object hierarchy, the degree of mixing of multiple data types, byte order conversion requirements, and whether it is necessary to recalculate CRC check or perform data integrity verification. The preset complexity threshold is a numerical critical point set by experimental calibration or theoretical calculation when the gateway is manufactured or initialized. This value is stored in the system configuration parameters in the unit of "number of CPU clock cycles required per kilobyte of data" or "number of instructions". The quantified value of the data structure complexity of the current protocol is compared with the preset complexity threshold to obtain a judgment result of being higher or lower.

[0029] Preferably, if the quantified value of the data structure complexity of the current protocol is higher than a preset complexity threshold, the parallel computing nodes are independent computing cores in the multi-core processor inside the fusion gateway, parallel logic units partitioned within the FPGA, or independently schedulable real-time task threads in an embedded system. Each node has the complete computing capability to independently execute data format conversion tasks. The first number is a preset positive integer representing the total number of parallel computing nodes actively activated when processing high-complexity protocols. It is usually set to be equal to or slightly higher than the maximum parallelism of the protocol format conversion task under theoretical analysis. For example, a protocol frame with 8 independent fields and no dependencies between fields can have its fields assigned to 8 nodes for parallel processing. The gateway's task scheduler requests the first number of computing nodes from the idle node pool, batch-allocates conversion task queues for data of this protocol type, and each node independently reads the allocated data subset, performs field extraction, word processing, and other tasks. Operations such as sequence conversion, data type conversion, and unit scaling conversion are performed. Each parallel computing node maps each field in the preprocessed data to a standardized structured field according to the conversion rules, generating structured data records. The second quantity is a preset positive integer, and the second quantity is less than the first quantity. The second quantity represents the number of resident parallel nodes that remain active to continue processing continuously arriving data after the peak period of the conversion task ends. When all the data to be converted for the high-complexity protocol has been processed, the gateway's task scheduler puts the nodes that are currently active in the parallel node state exceeding the second quantity into a dormant or idle state. For example, it sends a thread suspension signal to the operating system kernel, powers down the FPGA logic unit, or invalidates the clock gating signal of the computing core. By reducing the number of active computing nodes, the dynamic power consumption of the chip is reduced, including switching power consumption and short-circuit power consumption, thereby reducing the overall power consumption of the converged gateway. If the current protocol's data structure complexity quantification value is less than or equal to the preset complexity threshold, the gateway requests a computing node from the idle node pool to perform the format conversion task of the protocol data. It processes the preprocessed data one by one in a serial manner, and sequentially completes operations such as field extraction, data type conversion, and unit scale conversion to generate structured data records.

[0030] Step S300: An overcomplete dictionary is introduced to perform dimensionality reduction processing on the structured data to obtain dimensionality reduction data features corresponding to each terminal, and an association network reflecting the degree of correlation between trusted data sources is constructed based on the dimensionality reduction data features.

[0031] Step S300 further includes: constructing an overcomplete dictionary with the number of atoms being a first multiple of the structured data dimension; initializing the dictionary atoms and sparse coefficient matrix of the overcomplete dictionary; in each iteration, fixing the current dictionary atoms and updating the sparse coefficient matrix, then fixing the current sparse coefficient matrix and updating the dictionary atoms, alternating iterations until the reconstruction error is lower than a preset tolerance or the number of iterations reaches the upper limit; and using the finally obtained sparse coefficient matrix as the dimensionality reduction data feature.

[0032] Preferably, the structured data dimension refers to the number of numerical fields contained in each structured data record. For example, if a structured record contains six fields: timestamp, device ID, temperature, humidity, air pressure, and voltage, then the dimension is 6. The first multiple is a preset positive integer coefficient, usually 2 or 3, which is pre-calibrated according to the sparsity of the data and the requirements for reconstruction accuracy. The number of atoms is equal to the structured data dimension multiplied by the first multiple, that is, the number of atoms is greater than the data dimension. A floating-point matrix is ​​constructed as an overcomplete dictionary, with the number of rows equal to the structured data dimension and the number of columns equal to the number of atoms. Each column of the matrix is ​​called a dictionary atom. This dictionary has mathematical overcompleteness, that is, the number of atoms exceeds the minimum number of linearly independent vectors required to construct the matrix, indicating that the data sample can be represented by a variety of different combinations of atoms. Each atom in the matrix is ​​assigned an initial value, which can be achieved by randomly selecting K samples from a structured dataset as the initial values ​​of the atom, using the first K columns of the discrete cosine transform matrix as the initial atom, or generating the element values ​​using a Gaussian random matrix. After initialization, a normalization operation is performed on each atom vector, such as calculating its L2 norm and dividing each element in the vector by the norm to make the L2 norm of each atom 1. Then, the sparse coefficient matrix is ​​initialized. Each column of this matrix corresponds to a sparse representation coefficient vector of a data sample under an overcomplete dictionary. Initially, all elements of the sparse coefficient matrix are set to zero, filled with random small values, or obtained as an initial estimate through a matching pursuit algorithm.

[0033] Preferably, in each iteration, the current dictionary matrix is ​​considered known and fixed, and the sparse coefficient matrix is ​​updated. That is, for each data sample, the optimization objective is to minimize the square of the reconstruction error, with the constraint that the number of non-zero elements does not exceed the sparsity. The sparse coefficient vector corresponding to each sample is obtained through an orthogonal matching pursuit algorithm or a basis pursuit algorithm. These vectors are then combined to obtain the sparse coefficient matrix, and each atom in the dictionary matrix is ​​updated column-by-column. This alternating update process of sparse encoding and dictionary update is repeated. After each iteration, the overall reconstruction error, i.e., the sum of the squares of the reconstruction errors of all samples, is calculated. The preset tolerance is the maximum allowed reconstruction error. When the reconstruction error falls below the preset tolerance or the number of iterations reaches the upper limit, it indicates that the dictionary has sufficient data representation accuracy, and the iteration terminates. The final sparse coefficient matrix is ​​used as the dimensionality-reduced data feature.

[0034] Furthermore, step S300 also includes: obtaining the data dimension of the structured data, and using a first multiple of the data dimension as the initial number of atoms in the overcomplete dictionary; randomly initializing the initial number of dictionary atoms, and normalizing each dictionary atom to obtain an initial overcomplete dictionary; obtaining any data sample in the structured data, and using the absolute value of the inner product of the arbitrary data sample and each dictionary atom in the initial overcomplete dictionary as a dictionary atom selection constraint; selecting a target dictionary atom set from the initial overcomplete dictionary with the goal of maximizing the dictionary atom selection constraint, and using the target dictionary atom set as the overcomplete dictionary.

[0035] Preferably, the total number of numeric fields in each record of the structured dataset is read to determine the data dimension of the structured data. The first multiple of the data dimension is used as the initial number of atoms in the overcomplete dictionary; that is, the number of atomic columns in the dictionary is equal to the original dimension multiplied by a first multiple greater than 1. An overcomplete dictionary requires the number of atoms to be greater than the data dimension to ensure its overcompleteness. A matrix is ​​created, and all elements in the matrix are randomly generated and filled according to a uniform or standard normal distribution. Each dictionary atom is normalized, and the Euclidean norm of each column of the matrix is ​​calculated. Then, each element in that column is divided by the norm to obtain the initial overcomplete dictionary. An arbitrary data sample is randomly selected from the structured dataset, and the absolute value of the inner product between the arbitrary data sample and each dictionary atom in the initial overcomplete dictionary is calculated. This serves as a constraint for selecting dictionary atoms to evaluate the representational ability of each atom for the sample. The larger the absolute value of the inner product, the more consistent the direction of the atom with the sample vector, and the stronger the atom's contribution to the linear representation of the sample. For each atom, the absolute value of its inner product with the sample is calculated sequentially. The target dictionary atom set is selected in descending order. Finally, the target dictionary atom set is used as an overcomplete dictionary, which becomes the initial input dictionary for the sparse coding and dictionary update iteration process, while ensuring reduced computational overhead.

[0036] Furthermore, step S300 also includes performing clustering processing on each of the dimensionality-reduced data features to obtain clustering results; identifying and removing untrusted nodes based on the clustering results, and filtering out target feature sets corresponding to each trusted node from each of the dimensionality-reduced data features; constructing the association network based on the similarity between each feature in the target feature set and using the similarity as the association weight between each trusted node.

[0037] Preferably, K-means, density-based spatial clustering (DBSCAN), or Gaussian mixture model are used to cluster the features of each dimensionality-reduced data. Clustering is performed in a K-dimensional feature space, using the Euclidean distance between feature vectors as a similarity measure. The output clustering results include at least the cluster number assigned to each feature vector, the cluster center vector of each cluster, and a list of original terminal node indices contained in each cluster. Based on the distribution of cluster members in the clustering results, terminal nodes corresponding to certain clusters are determined to be untrusted nodes. The determination rules include, but are not limited to, clusters with fewer than a preset minimum size threshold, whose members are marked as untrusted nodes, and isolated points in the feature space that deviate from the mainstream data distribution corresponding to small clusters. When using the DBSCAN equal-density clustering algorithm, boundary points or noise points that are not assigned to any core cluster are marked as untrusted nodes. The distance between the feature vector of each node and the center of its cluster is calculated, and nodes whose distance exceeds a preset tolerance threshold are marked as outliers within the cluster. Then, terminal nodes marked as untrusted nodes are removed from the processing flow, including deleting their feature vectors and corresponding terminal identifiers from the nodes; the dimensionality-reduced feature vectors of all trusted nodes are retained to obtain the target feature set corresponding to the trusted nodes, with each vector corresponding to the feature representation of a trusted terminal node. For multiple dimensionality-reduced feature vectors in the target feature set, cosine similarity or Pearson correlation coefficient is used to calculate the pairwise similarity, and then this is directly used as the association weight between two nodes. When the weight is lower than a preset threshold, it can be set to zero to avoid noise introduced by weak associations; finally, a weighted undirected graph is constructed in the form of an adjacency matrix to determine the association network, where each node corresponds to a trusted terminal. For any pair of nodes, if its association weight is greater than a preset edge existence threshold, an edge is established between the nodes, with each edge carrying a weight value. The association network is used to constrain the weighting coefficient allocation of each data source during the fusion process.

[0038] Furthermore, step S300 also includes: extracting the center of any cluster of any cluster in the clustering result, and calculating the inter-cluster distance between any cluster and its neighboring clusters; determining the neighboring clusters whose inter-cluster distance is less than a predetermined distance threshold as overlapping clusters; marking the overlapping data nodes corresponding to the overlapping clusters as candidate abnormal nodes; obtaining the historical data deviation rate of the candidate abnormal nodes, and using the historical data deviation rate being higher than a predetermined deviation rate as a removal constraint, and removing the candidate abnormal nodes as untrusted nodes.

[0039] Preferably, the cluster center of any cluster in the clustering results is extracted, that is, the mean vector of the feature vectors of all samples in the cluster is calculated. Adjacent clusters refer to pairs of clusters that are spatially adjacent in the feature space, that is, the distance between the center vectors is less than the preset adjacent detection radius or there is a direct adjacency relationship between the two cluster boundaries. The Euclidean distance between the cluster centers of any pair of adjacent clusters is calculated, which reflects the degree of separation between the two clusters in the feature space. The smaller the distance, the closer the feature distribution of the two clusters is. The predetermined distance threshold is a positive real number pre-calibrated according to the inherent dispersion of the dataset. It represents the critical distance value for determining whether two clusters overlap in the feature space. It is usually taken as the standard deviation of the distance between all cluster centers multiplied by a coefficient. When the distance between the centers of two clusters is less than the predetermined distance threshold, it indicates that the distribution areas of the two clusters overlap in the feature space. The terminal nodes corresponding to the two clusters have highly similar behavior patterns at the dimensionality reduction feature level, and are therefore determined to be overlapping clusters.

[0040] Preferably, for each candidate abnormal node, its data deviation within a historical time window (e.g., the past 100 reporting cycles) is statistically analyzed. Specifically, the historical mean and standard deviation of the original cluster to which the node belongs are calculated, the sensor values ​​reported by the node each time within the current time window are obtained, and the deviation rate of each reported value is calculated, i.e., the degree of deviation expressed in standard deviation. The historical data deviation rate of the node is calculated, i.e., the time average of the deviation rate. The predetermined deviation rate threshold is a preset positive real number threshold, representing the maximum allowable average deviation rate, typically ranging from 2 to 3, corresponding to 2 to 3 under a normal distribution. A confidence interval of 3 standard deviations is used. For each candidate anomalous node, the historical data deviation rate is compared with a predetermined deviation rate threshold. If the historical data deviation rate is higher than the predetermined deviation rate threshold, it indicates that the node has consistently deviated from the normal data distribution within the historical time window, meeting the elimination constraint. The node is then removed from the trusted node set; that is, its terminal identifier is removed from the node list, and its dimensionality-reduced feature vector is not included in the target feature set. If the historical data deviation rate is not higher than the predetermined deviation rate threshold, it indicates that although the node falls into an overlapping cluster, the historical data deviation is within the allowable range, not meeting the elimination constraint, and is retained. This dual-criteria mechanism reduces the probability of mistakenly eliminating normal nodes due to simple cluster overlap, while verifying through historical deviation to ensure that the finally eliminated nodes do indeed have persistent data anomalies.

[0041] Step S400: The target conflict evidence is weighted and corrected using the associated network as a constraint, and weighted fusion is performed based on the corrected weights to output the integrated data after the fusion of the multi-protocol heterogeneous IoT terminals.

[0042] Preferably, the association network represents the inherent coupling relationship between trusted terminal nodes at the physical sensing level. The target conflict evidence is the complete dataset that is determined to have data conflict, specifically including the identifiers of all conflicting terminal nodes, the sensor values ​​reported by each node, and the corresponding dimensionality-reduced feature vectors. For example, in the measurement values ​​of the same physical quantity, the readings of some nodes deviate from the overall data distribution by more than a preset multiple standard deviation; the known physical constraints are violated between multiple associated physical quantities; and the time series change direction of some nodes is opposite to that of most nodes in terms of data change trends. For each data source node determined to be conflict evidence, its topological information in the association network is used as the basis for weight adjustment. The corrected weights are determined and the sensor data of all terminal nodes are weighted and fused. Before the fusion is performed, all corrected weights need to be normalized so that the sum of the weights equals 1 to ensure that the dimensions of the fusion result are consistent with the original data. Finally, the comprehensive data after the fusion of multi-protocol heterogeneous IoT terminals is output, which includes at least the fusion timestamp, fusion value, number of nodes participating in the fusion, and credibility identifier. If the proportion of the number of effective fusion nodes exceeds a preset ratio, it is marked as "high credibility" and then sent to the cloud platform through the uplink communication interface.

[0043] Furthermore, step S400 also includes: extracting the first conflict data corresponding to the first conflict data source in the target conflict evidence; analyzing the first conflict data under the association network based on a predetermined feature dimension, and constructing a first data feature parameter set of the first conflict data; performing weighted analysis on the first data feature parameter set to obtain a first weight coefficient; weighting the first conflict data with the first weight coefficient as the weight to obtain the target valid data of the target conflict evidence; and performing data fusion of the multi-protocol heterogeneous IoT terminals based on the target valid data to obtain the comprehensive data.

[0044] Step S400 further includes that the predetermined feature dimensions include sample size, distribution characteristics, and data importance.

[0045] Preferably, a terminal node is randomly selected from the target conflict evidence set as the first conflict data source. From the sensor data records reported by this first conflict data source node within the current fusion cycle, the values ​​of all its sensor fields are extracted to determine the first conflict data, including the node's terminal identifier, measured values ​​of each physical quantity, timestamp information, and the quantified value of its deviation in the current conflict detection. Then, the first conflict data is analyzed under the association network according to defined feature dimensions. These defined feature dimensions include sample size, distribution characteristics, and data importance. Sample size refers to the total number of valid data samples from the first conflict data source node participating in fusion within the current fusion cycle. The specific value depends on the terminal's sampling and reporting frequency. For example, if a report is submitted every 60 seconds, and the current cycle covers the most recent 5 reports, then the sample size is 5. The larger the sample size, the higher the confidence level of the statistical estimate. Distribution characteristics refer to the statistical distribution morphology parameters of the node's historical data, including the standard deviation of the data reported by the node within the current cycle, and the node's... The historical data includes the mean and standard deviation, as well as the standardized deviation of the current period's mean relative to the historical mean. Data importance refers to the business priority weight of the terminal node in the overall IoT sensing system. This value is a preset static parameter stored in the node configuration table of the converged gateway. The value is determined based on factors including the criticality of the physical area where the terminal is located, the criticality of the terminal's sensor type, and the redundancy of the terminal's collected data. The quantified values ​​of the predetermined feature dimensions of the first conflicting data source node are comprehensively evaluated in the context of the associated network. Specifically, the association weight vector between the node and all other trusted nodes is extracted from the adjacency matrix of the associated network. The aggregation association strength of the node is calculated, and the topological location characteristics of the node in the network (including neighborhood density and average association strength of neighbors) are calculated. Then, the distribution characteristics of the first conflicting data are compared with the average distribution characteristics of its neighboring nodes to calculate the relative deviation. Finally, a multi-dimensional feature vector is determined as the first data feature parameter set of the first conflicting data.

[0046] Preferably, each component in the first set of data feature parameters is multiplied by its corresponding preset weighting coefficient and summed to obtain a comprehensive score. The preset weighting coefficients reflect the influence of different feature dimensions on the credibility of the data source. For example, the weighting coefficient for sample size reflects the influence weight of statistical confidence, the weighting coefficient for distribution deviation reflects the influence weight of data consistency, and the weighting coefficient for aggregation correlation strength reflects the credibility support of the data source in the network. The comprehensive score is then normalized and mapped to the final first weighting coefficient, representing the relative credibility of the first conflicting data source in the conflict evidence set. Finally, the original reported values ​​of the first conflicting data source nodes are multiplied by their respective weighting coefficients. The weighted coefficient is used to obtain the weighted contribution value of the data source. Weighted operations are performed on all conflicting data source nodes in the target conflict evidence set, and all weighted contribution values ​​are summed to obtain the target effective data. Among them, the data source with higher weight has a larger contribution to the estimated value, and the data source with lower weight has a smaller contribution, thereby achieving effective suppression and compromise of conflict evidence. Finally, the target effective data is used as the representative value of the target conflict evidence subset within this fusion period to replace the original scattered conflict values. All data source nodes that have not experienced conflict retain their original values. A weighted average fusion operation is performed on the data of multi-protocol heterogeneous IoT terminals to obtain comprehensive data.

[0047] In the above text, refer to Figure 1 This paper describes in detail a low-power data fusion method for multi-protocol heterogeneous Internet of Things (IoT) terminals according to embodiments of the present invention. Next, reference will be made to... Figure 2 This invention describes a low-power data fusion system for multi-protocol heterogeneous Internet of Things (IoT) terminals.

[0048] The low-power data fusion system for multi-protocol heterogeneous IoT terminals according to embodiments of the present invention addresses the technical problems in the prior art, such as low data fusion accuracy, high power consumption, and the tendency for conflicting evidence to dominate the results at the edge. It achieves the technical effect of improving the reliability and accuracy of edge-side fused data while reducing the power consumption of terminal-gateway communication and computing. Figure 2 As shown, the low-power data fusion system for multi-protocol heterogeneous IoT terminals includes: a sensor data acquisition module 10, a structured data acquisition module 20, a data dimensionality reduction processing module 30, and a data fusion output module 40.

[0049] The sensor data acquisition module 10 is used to identify the communication protocol type adopted by each terminal in the multi-protocol heterogeneous IoT terminals through a converged gateway deployed at the network edge, and to call the corresponding protocol parsing engine to collect the raw sensor data uploaded by each terminal respectively; the structured data acquisition module 20 is used to perform noise removal and outlier cleaning on the raw sensor data, and to convert the cleaned data into structured data in a unified format; the data dimensionality reduction processing module 30 is used to introduce an overcomplete dictionary to perform dimensionality reduction processing on the structured data, to obtain the dimensionality reduction data features corresponding to each terminal, and to construct an association network reflecting the degree of correlation between trusted data sources based on the dimensionality reduction data features; the data fusion output module 40 is used to adjust the weight of the target conflict evidence with the association network as a constraint, and to perform weighted fusion based on the adjusted weights, and output the comprehensive data after fusion of the multi-protocol heterogeneous IoT terminals.

[0050] The specific configuration of the structured data acquisition module 20 will be described in detail below. The structured data acquisition module 20 further includes: performing noise removal processing on the original sensor data and obtaining the data distribution of the noise-removed data; sequentially calculating the first quartile and the third quartile of the data distribution, using the difference between the first quartile and the third quartile as the interquartile range; using the first quartile minus a first multiple of the interquartile range as a lower threshold, and the third quartile plus a second multiple of the interquartile range as an upper threshold, forming a filtering constraint; identifying data that does not meet the filtering constraint as outliers and removing them to obtain preprocessed data, and converting the preprocessed data into the structured data according to preset rules.

[0051] The specific configuration of the structured data acquisition module 20 will be described in detail below. The structured data acquisition module 20 further includes: acquiring the communication protocol type corresponding to the preprocessed data; determining whether the data structure complexity of the communication protocol type is higher than a preset complexity threshold; if higher, calling a first number of parallel computing nodes to perform format conversion on the data of the communication protocol type, and releasing any parallel computing nodes exceeding the second number after conversion to reduce energy consumption; if not higher, calling a single computing node to perform format conversion on the data of the communication protocol type.

[0052] The specific configuration of the data dimensionality reduction processing module 30 will be described in detail below. The data dimensionality reduction processing module 30 further includes: constructing an overcomplete dictionary with a number of atoms that is a first multiple of the structured data dimension; initializing the dictionary atoms and sparse coefficient matrix of the overcomplete dictionary; in each iteration, fixing the current dictionary atoms and updating the sparse coefficient matrix, then fixing the current sparse coefficient matrix and updating the dictionary atoms, alternating iterations until the reconstruction error is lower than a preset tolerance or the number of iterations reaches an upper limit; and using the finally obtained sparse coefficient matrix as the dimensionality reduction data feature.

[0053] The specific configuration of the data dimensionality reduction processing module 30 will be described in detail below. The data dimensionality reduction processing module 30 further includes: obtaining the data dimension of the structured data, and using a first multiple of the data dimension as the initial number of atoms in the overcomplete dictionary; randomly initializing the initial number of dictionary atoms, and normalizing each dictionary atom to obtain an initial overcomplete dictionary; obtaining any data sample from the structured data, and using the absolute value of the inner product of the arbitrary data sample and each dictionary atom in the initial overcomplete dictionary as a dictionary atom selection constraint; selecting a target dictionary atom set from the initial overcomplete dictionary with the goal of maximizing the dictionary atom selection constraint, and using the target dictionary atom set as the overcomplete dictionary.

[0054] The specific configuration of the data dimensionality reduction processing module 30 will be described in detail below. The data dimensionality reduction processing module 30 further includes: performing clustering processing on each of the dimensionality-reduced data features to obtain clustering results; identifying and removing untrusted nodes based on the clustering results, and filtering out target feature sets corresponding to each trusted node from each of the dimensionality-reduced data features; constructing the association network based on the similarity between features in the target feature set, and using the similarity as the association weight between each trusted node.

[0055] The specific configuration of the data dimensionality reduction processing module 30 will be described in detail below. The data dimensionality reduction processing module 30 further includes: extracting the center of any cluster from any cluster in the clustering results, and calculating the inter-cluster distance between any cluster and its neighboring clusters; identifying neighboring clusters whose inter-cluster distance is less than a predetermined distance threshold as overlapping clusters; marking the overlapping data nodes corresponding to the overlapping clusters as candidate anomalous nodes; obtaining the historical data deviation rate of the candidate anomalous nodes, and using the historical data deviation rate being higher than a predetermined deviation rate as a removal constraint, removing the candidate anomalous nodes as untrusted nodes.

[0056] The specific configuration of the data fusion output module 40 will be described in detail below. The data fusion output module 40 further includes: extracting first conflict data corresponding to the first conflict data source in the target conflict evidence; analyzing the first conflict data under the association network based on a predetermined feature dimension, and constructing a first data feature parameter set for the first conflict data; performing weighted analysis on the first data feature parameter set to obtain a first weight coefficient; weighting the first conflict data with the first weight coefficient as the weight to obtain the target valid data of the target conflict evidence; and performing data fusion of the multi-protocol heterogeneous IoT terminals based on the target valid data to obtain the comprehensive data.

[0057] The specific configuration of the data fusion output module 40 will be described in detail below. The data fusion output module 40 further includes: the predetermined feature dimensions include sample size, distribution characteristics, and data importance.

[0058] The low-power data fusion system for multi-protocol heterogeneous IoT terminals provided in this embodiment of the invention can execute the low-power data fusion method for multi-protocol heterogeneous IoT terminals provided in any embodiment of the invention, and has the corresponding functional modules and beneficial effects of executing the method.

[0059] The above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention in any way. Although the present invention has been disclosed above with reference to preferred embodiments, it is not intended to limit the present invention. Any person skilled in the art can make some modifications or alterations to the above-disclosed technical content to create equivalent embodiments without departing from the scope of the present invention. Any modifications, equivalent changes, and alterations made to the above embodiments based on the technical essence of the present invention without departing from the scope of the present invention shall still fall within the scope of the present invention.

Claims

1. A low-power data fusion method for multi-protocol heterogeneous IoT terminals, characterized in that, include: By deploying a converged gateway at the network edge, the communication protocol type used by each terminal in the multi-protocol heterogeneous IoT terminal is identified, and the corresponding protocol parsing engine is called to collect the raw sensor data uploaded by each terminal. The original sensor data is subjected to noise removal and outlier cleaning, and the cleaned data is converted into structured data in a unified format. An overcomplete dictionary is introduced to perform dimensionality reduction on the structured data to obtain dimensionality reduction data features corresponding to each terminal, and an association network reflecting the degree of correlation between trusted data sources is constructed based on the dimensionality reduction data features. The conflict evidence is weighted and corrected using the associated network as a constraint, and then weighted and fused based on the corrected weights to output the integrated data after the fusion of the multi-protocol heterogeneous IoT terminals.

2. The low-power data fusion method for multi-protocol heterogeneous IoT terminals as described in claim 1, characterized in that, The raw sensor data undergoes noise removal and outlier cleaning, and the cleaned data is converted into structured data in a unified format, including: The original sensing data is subjected to noise removal processing, and the data distribution of the noise-removed data is obtained; The first quartile and the third quartile of the data distribution are calculated sequentially, and the difference between the first quartile and the third quartile is taken as the interquartile range. The first quartile minus a multiple of the interquartile range is used as the lower threshold, and the third quartile plus a multiple of the interquartile range is used as the upper threshold to form a filtering constraint. Data that does not meet the filtering constraints is identified as outliers and removed to obtain preprocessed data. The preprocessed data is then converted into the structured data according to preset rules.

3. The low-power data fusion method for multi-protocol heterogeneous IoT terminals as described in claim 2, characterized in that, Converting the preprocessed data into structured data according to preset rules includes: Obtain the communication protocol type corresponding to the preprocessed data; Determine whether the data structure complexity of the communication protocol type is higher than a preset complexity threshold; If the number of parallel computing nodes exceeds the limit, the first number of parallel computing nodes are invoked to perform format conversion on the data of the communication protocol type, and after the conversion is completed, the excess parallel computing nodes are released to reduce energy consumption. If the value is not higher, a single computing node is invoked to perform format conversion on the data of the communication protocol type.

4. The low-power data fusion method for multi-protocol heterogeneous IoT terminals as described in claim 1, characterized in that, An overcomplete dictionary is introduced to perform dimensionality reduction on the structured data, resulting in dimensionality-reduced data features for each terminal, including: Construct an overcomplete dictionary whose number of atoms is a first multiple of the dimension of the structured data; Initialize the dictionary atoms and sparse coefficient matrix of the overcomplete dictionary; In each iteration, the current dictionary atom is fixed and the sparse coefficient matrix is ​​updated, then the current sparse coefficient matrix is ​​fixed and the dictionary atom is updated again. This process is repeated alternately until the reconstruction error is lower than the preset tolerance or the number of iterations reaches the upper limit. The sparse coefficient matrix obtained at the end is used as the feature of the dimensionality reduction data.

5. The low-power data fusion method for multi-protocol heterogeneous IoT terminals as described in claim 4, characterized in that, Construct an overcomplete dictionary with a number of atoms that is a multiple of the dimension of the structured data, including: Obtain the data dimension of the structured data, and use the first multiple of the data dimension as the initial number of atoms in the overcomplete dictionary; The initial number of dictionary atoms is randomly initialized, and each dictionary atom is normalized to obtain an initial overcomplete dictionary; Obtain any data sample from the structured data, and use the absolute value of the inner product of the arbitrary data sample and each dictionary atom in the initial overcomplete dictionary as the dictionary atom selection constraint; With the goal of maximizing the dictionary atom selection constraints, a target dictionary atom set is selected from the initial overcomplete dictionary, and the target dictionary atom set is used as the overcomplete dictionary.

6. The low-power data fusion method for multi-protocol heterogeneous IoT terminals as described in claim 4, characterized in that, Based on the aforementioned dimensionality-reduced data features, a correlation network reflecting the degree of correlation between trustworthy data sources is constructed, including: Clustering is performed on the features of each of the dimensionality-reduced data to obtain the clustering results; Based on the clustering results, untrusted nodes are identified and removed, and target feature sets corresponding to each trusted node are obtained by filtering from each of the dimensionality-reduced data features. The association network is constructed based on the similarity between features in the target feature set and using the similarity as the association weight between trust nodes.

7. The low-power data fusion method for multi-protocol heterogeneous IoT terminals as described in claim 6, characterized in that, Identifying and removing untrusted nodes based on the clustering results includes: Extract the cluster center of any cluster class from the clustering results, and calculate the inter-cluster distance between any cluster class and its neighboring cluster classes. Adjacent clusters whose inter-cluster distance is less than a predetermined distance threshold are identified as overlapping clusters. Mark the overlapping data nodes corresponding to the overlapping clusters as candidate abnormal nodes; The historical data deviation rate of the candidate abnormal node is obtained, and the candidate abnormal node is eliminated as an untrusted node by using the historical data deviation rate being higher than a predetermined deviation rate as a elimination constraint.

8. The low-power data fusion method for multi-protocol heterogeneous IoT terminals as described in claim 6, characterized in that, The conflict evidence is weighted and corrected using the aforementioned network as a constraint, and then weighted and fused based on the corrected weights to output the integrated data fused from the multi-protocol heterogeneous IoT terminals, including: Extract the first conflict data corresponding to the first conflict data source from the target conflict evidence; Based on predetermined feature dimensions, the first conflict data is analyzed under the association network, and a first data feature parameter set of the first conflict data is constructed. A weighted analysis is performed on the first set of data feature parameters to obtain the first weight coefficient; The first conflict data is weighted using the first weight coefficient as the weight to obtain the target valid data of the target conflict evidence; The comprehensive data is obtained by fusing data from the multi-protocol heterogeneous IoT terminals based on the target valid data.

9. The low-power data fusion method for multi-protocol heterogeneous IoT terminals as described in claim 8, characterized in that, The predetermined feature dimensions include sample size, distribution characteristics, and data importance.

10. A low-power data fusion system for multi-protocol heterogeneous IoT terminals, characterized in that, The system is used to implement the low-power data fusion method for multi-protocol heterogeneous IoT terminals as described in any one of claims 1 to 9, and the system comprises: The sensor data acquisition module is used to identify the communication protocol type used by each terminal in the multi-protocol heterogeneous IoT terminal through the converged gateway deployed at the network edge, and call the corresponding protocol parsing engine to collect the raw sensor data uploaded by each terminal respectively. The structured data acquisition module is used to perform noise removal and outlier cleaning on the raw sensor data, and convert the cleaned data into structured data in a unified format. The data dimensionality reduction processing module is used to introduce an overcomplete dictionary to perform dimensionality reduction processing on the structured data, obtain the dimensionality reduction data features corresponding to each terminal, and construct an association network reflecting the degree of correlation between trusted data sources based on the dimensionality reduction data features. The data fusion output module is used to correct the weights of the target conflict evidence with the associated network as a constraint, and to perform weighted fusion based on the corrected weights to output the comprehensive data after the fusion of the multi-protocol heterogeneous IoT terminals.

Images