Real-time marking method and system for data credibility of smart home Internet of Things
By combining dynamic filtering and multi-head attention LSTM model, the problems of unstable sensor data quality and high misjudgment rate in smart home systems are solved, real-time credibility assessment and self-adaptation capabilities are achieved, and the robustness and processing efficiency of the system are improved.
Patent Information
- Application Number
- CN202510756158.X
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-06-08
- Publication Date
- 2025-10-10
AI Technical Summary
In existing smart home systems, the quality of sensor data is unstable, data conflicts frequently occur when multiple sources work together, traditional data processing methods lack real-time performance, have a high misjudgment rate, and are difficult to adapt to dynamic environmental changes. In addition, machine learning models suffer from catastrophic forgetting during incremental learning.
It adopts the combination of dynamic filtering technology, multi-head attention layer and bidirectional LSTM layer to perform real-time data feature extraction and credibility prediction. It optimizes the model through dynamic threshold optimization and incremental learning algorithm to achieve real-time credibility assessment and self-adaptation of data.
It realizes real-time and accurate credibility assessment of multi-source sensor data, reduces the misjudgment rate, enhances the robustness and adaptability of the system, reduces operation and maintenance costs, and adapts to changes in complex home environments.
Smart Images
Figure CN120763578A_ABST
Abstract
Description
Technical Field
[0001] The present invention belongs to the field of Internet of Things and smart home technologies, and specifically relates to a real-time sensor data quality assessment method and system based on feedforward data processing and deep learning models, which is particularly suitable for smart home automation control scenarios that require high real-time performance and high reliability. Background Art
[0002] With the rapid development of smart home IoT technology, multi-source sensor networks have been widely used in scenarios such as environmental monitoring, security warnings, and energy management. Various sensors continuously generate high-frequency time-series data, providing real-time information for intelligent decision-making. However, due to sensor hardware limitations and environmental interference, the raw data often suffers from high noise and high reliability fluctuations. Due to cost and size constraints, home sensors typically use low-precision modules, and their signals are susceptible to electromagnetic interference and physical environmental influences, resulting in unstable data quality. Furthermore, when multiple sensors work together, data conflicts often arise due to differences in installation location or sampling timing, further complicating the assessment of data credibility.
[0003] Data processing technology in smart home systems currently faces numerous challenges. Traditional a posteriori data cleaning methods require the accumulation of data over a long time window for effective processing, resulting in insufficient real-time response performance. This is particularly evident in scenarios with high timeliness requirements, such as security. Credibility assessment methods based on fixed thresholds experience a significant increase in false positives when environmental parameters suddenly change, making them unable to adapt to dynamically changing environmental demands. Furthermore, existing systems lack effective cross-sensor correlation modeling capabilities. When conflicting sensor data emerges, it is difficult to quickly and accurately identify the source of the fault, increasing the risk of system malfunction.
[0004] In terms of data processing algorithms, current solutions have obvious limitations. While high-precision sensor deployment solutions at the hardware level can improve data quality, they will significantly increase system costs, making them unsuitable for large-scale promotion and application. While offline batch processing methods at the algorithm level can guarantee processing results, they cannot meet real-time requirements; static rule engines, on the other hand, struggle to cope with complex and changing home environments. More notably, existing machine learning models generally suffer from catastrophic forgetting during incremental learning. When the system needs to adapt to new environmental patterns or sensor types, it often requires complete retraining, resulting in high operational and maintenance costs.
[0005] These issues severely restrict the reliability and practicality of smart home systems. Data quality issues directly impact the accuracy of automated decisions, insufficient real-time performance limits the system's responsiveness in emergency scenarios, and deficiencies in multi-source coordination reduce the system's overall robustness. With the continuous expansion of smart home application scenarios and increasing user demands, the development of data credibility assessment technologies that balance real-time performance, accuracy, and adaptability is crucial. Innovative solutions are urgently needed to overcome current technical bottlenecks, especially in the face of multiple challenges such as sensor network heterogeneity, dynamic environmental changes, and resource constraints. Summary of the Invention
[0006] In response to the shortcomings of the above-mentioned technologies, the present invention proposes a real-time labeling method and system for the credibility of smart home Internet of Things data based on feedforward data quality control, aiming to solve the problems of delayed data quality assessment, high misjudgment rate, and insufficient multi-sensor collaboration capabilities in the existing technologies.
[0007] A method for real-time credibility marking of smart home IoT data in the present invention comprises the following steps:
[0008] Data preprocessing and feature extraction: Receives multi-source heterogeneous data streams from multiple sensors, dynamically filters the raw data, and automatically adjusts the filter window length based on the sensor sampling rate to suppress high-frequency noise and preserve true signal characteristics. Multidimensional feature vectors are extracted from the filtered data. These feature vectors include the current sensor reading, time series statistical features, correlation features with neighboring sensors, and device health features.
[0009] Feedforward credibility prediction: The multidimensional feature vector is input into a credibility prediction model, which includes a multi-head attention layer, a bidirectional long short-term memory (LSTM) layer, and a fully connected layer. The multi-head attention layer is used to capture the spatial correlation between sensors, the bidirectional LSTM layer is used to model temporal dependencies, and the fully connected layer is used to output a confidence score. The confidence score comprehensively reflects the reliability of the data in multiple aspects such as time and space dimensions and device status.
[0010] Dynamic threshold optimization: Based on the receiver operating characteristic (ROC) curve analysis of the validation set, the optimal and worst classification thresholds are determined, and the confidence scores are divided into three levels: high confidence, medium confidence, and low confidence;
[0011] Anomaly detection and processing: When the confidence score is detected to be lower than the preset worst classification threshold, the backup sensor data source or manual review process is triggered to ensure the reliability of the data;
[0012] Feedback-driven model update: Feedback signals are collected through user feedback and / or multi-sensor data conflict results, and the parameters of the credibility prediction model are updated using an incremental learning algorithm. The incremental learning algorithm is used to prevent catastrophic forgetting during the model update process and ensure the long-term stability of the model;
[0013] Among them, the dynamic filtering processing adopts Savitzky-Golay filter or other filtering technology with dynamically adjusted window; the multi-head attention layer converts the input features into query matrix, key matrix and value matrix through a learnable parameter matrix, and calculates the weighted feature vector; the bidirectional LSTM layer includes forward LSTM and backward LSTM, which process data in chronological order and reverse order respectively to capture time series features; the incremental learning algorithm adopts elastic weight consolidation (EWC) algorithm or other incremental learning algorithms to prevent catastrophic forgetting.
[0014] A real-time credibility marking system for smart home IoT data in the present invention includes:
[0015] Multi-protocol access module, used to receive multi-source heterogeneous data streams from multiple sensors, supporting Zigbee, Wi-Fi, and Bluetooth protocols;
[0016] Edge computing unit, equipped with an NPU acceleration chip, is used to perform filtering, feature extraction, and credibility prediction in real time;
[0017] Incremental learning platform, deployed on cloud servers, for model parameter updates and version management;
[0018] Trusted tag database, used to store timestamped trustworthy tags and associated sensor metadata;
[0019] Among them, the system adopts the real-time marking method for the credibility of smart home Internet of Things data.
[0020] The technical effects achieved by the present invention are as follows:
[0021] (1) Real-time dynamic evaluation to improve data credibility: Traditional smart home systems usually use fixed thresholds or posterior analysis to evaluate data quality, resulting in delayed evaluation results and poor adaptability. The present invention combines dynamic window filtering technology with a spatiotemporal feature fusion model to achieve real-time credibility evaluation of multi-source sensor data. The dynamic window Savitzky-Golay filter automatically adjusts the window length according to the sensor sampling rate, effectively suppressing noise while retaining the true signal characteristics; the fusion model of the multi-head attention mechanism and the bidirectional LSTM can simultaneously capture the spatial correlation and temporal dependency between sensors, significantly improving the accuracy of data credibility scoring. Dynamic threshold optimization based on the ROC curve further ensures the adaptability of the evaluation results to changes in the environment and device status.
[0022] (2) Reduce the misjudgment rate and enhance the robustness of the system: the existing method is prone to misjudgment when the environmental parameters mutate or the sensor is abnormal, affecting the reliability of the system. The present application effectively reduces the risk of misjudgment through three levels of trusted label classification (highly trusted, medium trusted, and low trusted) and a backup sensor triggering mechanism. When the data is marked as low trusted, the system can automatically switch to backup sensor data or trigger an artificial review process, ensuring the reliability of critical decisions such as security warnings. At the same time, the introduction of the device health degree coding value (C dev ) enables the system to consider factors such as device aging, further reducing data misjudgment caused by hardware failure.
[0023] (3) Adaptive incremental learning, optimize long-term performance: traditional machine learning models need to be completely retrained when adapting to new environments or new sensors, resulting in high operating costs and possible loss of historical knowledge. The present application uses the Elastic Weight Consolidation (EWC) algorithm to achieve incremental learning, quantifying parameter importance through the Fisher information matrix, and protecting the memory of key parameters for historical tasks when updating the model. This mechanism enables the system to continuously optimize based on user feedback and multi-sensor conflict results, adapting to environmental changes and new device access while avoiding catastrophic forgetting problems, significantly reducing system maintenance costs.
[0024] (4) Edge-cloud collaboration, improve processing efficiency: Compared with traditional centralized processing solutions, the present application realizes distributed optimization of data processing through the collaborative design of multi-protocol access modules and edge computing units. The edge computing unit is equipped with an NPU acceleration chip, which can perform real-time computation-intensive tasks such as filtering, feature extraction, and trustworthiness prediction; the cloud incremental learning platform is responsible for model updating and version management, forming an efficient edge-cloud collaborative architecture. This design not only reduces the network transmission burden, but also protects data privacy through localized processing, making it particularly suitable for smart home application scenarios that require high real-time performance. BRIEF DESCRIPTION OF DRAWINGS
[0025] In order to more clearly illustrate the technical solutions in the embodiments of the present application, the following will briefly introduce the drawings needed in the embodiment description. Obviously, the drawings in the following description are only some embodiments of the present application, and other drawings can be obtained by those skilled in the art without creative labor.
[0026] Figure 1 : System architecture diagram, showing the complete process of data acquisition, processing, and feedback;
[0027] Figure 2 : Trustworthiness prediction model structure diagram, showing the connection relationship of attention layer, bidirectional LSTM layer, and full connection layer;
[0028] Figure 3 : Dynamic threshold optimization flow chart, including ROC curve analysis steps;
[0029] Figure 4 : Schematic diagram of parameter update of incremental learning algorithm. DETAILED DESCRIPTION
[0030] In order to describe the present invention more specifically, the technical solution of the present invention is described in detail below in conjunction with specific implementation methods.
[0031] like Figure 1 As shown, the present application provides a method for real-time labeling of the credibility of smart home IoT data, comprising the following steps:
[0032] Step (1) accepts multi-source heterogeneous data streams from temperature, humidity, light, and door magnetic sensors.
[0033] Step (2) uses Savitzky-Golay filter to perform dynamic denoising on the original signal S(t), and the window length W satisfies where f s is the sensor sampling rate (unit: Hz), Indicates rounding down;
[0034] Step (3) Construct a multidimensional feature vector from the filtered data include:
[0035] The current sensor reading S(t);
[0036] The average value in the sliding window of the past N minutes and variance
[0037] Pearson correlation coefficient with neighboring sensors Where cov(·) represents the covariance operation;
[0038] Equipment health code value C dev =w1×age+w2×error_count, where age represents the service life of the equipment and error_count represents the number of failures. The weights w1 and w2 are determined based on historical data analysis of equipment aging factors and failure counts or the experience of domain experts.
[0039] Step (4) inputs the feature vector X(t) into the credibility prediction model, such as Figure 2 As shown, the model includes:
[0040] The multi-head attention layer uses a parallel attention mechanism to process input features, specifically through Q = X(t)WQ ,K=X(t)W K ,V=X(t)W V , calculate the weighted eigenvector Where Q, K, and V are query matrix, key matrix, and value matrix, respectively. They are the query weight matrix, key weight matrix, and value weight matrix, respectively. They are all parameter matrices automatically learned through model training. d is the dimension of X(t), and d k W k The Softmax function is a normalization function that converts any real vector into a probability distribution.
[0041] In the bidirectional LSTM layer, the forward LSTM processes data in chronological order to learn the impact of historical information on the current state; the backward LSTM processes the sequence in reverse to capture the dependency of future information on the current moment. The hidden states in the two directions are fused by splicing to form a feature representation h containing the complete temporal context. t Specifically: the forward and backward LSTM passes through the output of the multi-head attention layer Calculate the hidden state separately The final output time series feature vector Where [·] is a vector concatenation operation.
[0042] The fully connected layer transforms the temporal features h t Map to the interval [0,1] and output the confidence score p(t). This score comprehensively reflects the reliability of the data in multiple aspects such as time and space dimensions and device status. Specifically: The confidence score p(t) is output by the Sigmoid function = sigmoid(W f h t +b f )∈[0,1], where W f is the weight matrix of the fully connected layer, b f is the bias term of the fully connected layer.
[0043] Step (5) Determine the dynamic optimal classification threshold T based on the validation set ROC curve high and T low , map p(t) into three levels of labels: high confidence, medium confidence, and low confidence, see Figure 3 ; Among them, high confidence (p≥T high ), Zhongxin (T low ≤p <T high ) and low confidence (p <T low ).
[0044] Step (6) When the data is marked as low confidence, the backup sensor data source or manual review process is triggered to ensure data reliability.
[0045] Step (7) collects feedback signals based on user feedback (such as APP marking data anomalies) and implicit feedback (such as more than N conflict = 3 sensor data conflicts in the same decision cycle), updates model parameters using the Elastic Weight Consolidation (EWC) algorithm, see Figure 4 When multiple sensor conflicts are detected, update the model parameters through the EWC algorithm:
[0046]
[0047] Where L EWC is the loss function of the EWC algorithm, L CE is the cross-entropy loss, N is the number of samples in the current batch, y k is the true value of the kth sample label (0 for normal, 1 for abnormal), p k (t) is the prediction probability of the model for the kth sample at time t, F i the Fisher information of the i-th parameter θ i , is the gradient solution, x m represents the mth sample, θ old is the parameter obtained from the old training task, and λ is the weight hyperparameter that balances the new and old tasks.
[0048] The application adopts dynamic window filtering technology: automatically adjust the filtering window according to the sensor sampling rate, while retaining the characteristics of the real signal and suppressing high-frequency noise; Spatiotemporal feature fusion model: capture the spatial correlation between sensors through multi-head attention mechanism, and combine bidirectional LSTM to model the temporal dependence relationship; Incremental learning mechanism: use the Elastic Weight Consolidation algorithm to prevent catastrophic forgetting during model updating, and ensure long-term stability.
[0049] Further, in step (4), for each time step t, the calculation steps of the bidirectional LSTM are as follows:
[0050] (a) Gating calculation (take forward LSTM as an example, the same for backward)
[0051] (Forgotten door)
[0052] (Input door)
[0053] (Output door)
[0054] (Candidate memory)
[0055] Where, is the output of the current time step multi-head attention layer; Represents the hidden state of the previous moment t-1, that is, the past information of LSTM; is the activation value of the forget gate, is the activation value of the input gate, is the activation value of the output gate; It is a candidate memory unit, representing new information that LSTM may store; W i → , are the weight matrices of the forget gate, input gate, output gate, and candidate memory respectively; are the forget gate, input gate, output gate, and candidate memory bias matrix respectively; tanh(·) is the hyperbolic tangent function, which is used here as the activation function.
[0056] (b) Memory unit update (forward LSTM is used as an example, and the backward LSTM is the same)
[0057] in, Represents the state of the memory unit at the current moment; Represents the state of the memory unit at the previous moment t-1; ⊙ represents element-by-element multiplication; Represents the candidate memory at the current moment, which is calculated from the current input and the hidden state at the previous moment; Represents the input gate output, control The update ratio.
[0058] (c) Output of hidden state (forward LSTM is used as an example, and the backward is the same)
[0059]
[0060] (d) Bidirectional state fusion
[0061] According to the obtained and Splicing to form the final output
[0062]
[0063] Where [·] is a vector concatenation operation.
[0064] Furthermore, the threshold optimization in step (5) specifically includes:
[0065] (a) Calculate the confidence scores of all samples on the validation set;
[0066] (b) Traverse the threshold T∈[0,1] with a step size of 0.01;
[0067] (c) Select the threshold T that maximizes the Youden index J = (TPR-FPR) high , where TPR is the true positive rate and FPR is the false positive rate;
[0068] (d) Set T low =T high -ΔT, where ΔT represents the threshold interval and is an adjustable parameter.
[0069] This application also provides a smart home IoT data credibility real-time marking system for implementing the above method, comprising:
[0070] (a) Multi-protocol access module, supporting sensor data reception of Zigbee, Wi-Fi, and Bluetooth protocols;
[0071] (b) Edge computing unit, equipped with an NPU acceleration chip, for real-time filtering, feature extraction, and credibility prediction;
[0072] (c) Incremental learning platform, deployed on cloud servers, for model parameter updates and version management;
[0073] (a) Trusted tag database, which stores timestamped credibility tags and associated sensor metadata.
[0074] The system collects environmental data in real time through a multi-source sensor network in the smart home environment, including temperature sensors (DS18B20, sampling rate 1Hz) deployed on the ceiling of each room, humidity sensors (DHT22,
[0075] 0.5Hz), light sensors near windows (BH1750, 2Hz), and door magnetic sensors (reed switches) installed on door and window frames. All data is transmitted to the edge computing gateway (NVIDIA Jetson Xavier NX) via the Zigbee 3.0 protocol.
[0076] In the preprocessing stage, a dynamic window Savitzky-Golay filter (second-order polynomial) is used for signal denoising. The window length is automatically calculated according to the sensor sampling rate (e.g., W = 3 for temperature sensor and W = 5 for light sensor). At the same time, a dynamic window Savitzky-Golay filter (second-order polynomial) is constructed, which includes time series features (mean μ = 24.8°C and standard deviation σ = 0.5°C of a 5-minute sliding window), spatial features (Pearson correlation coefficient ρ = 0.92 for adjacent sensors), and device features (health score C dev =w1×age+
[0077] w2×error_count) multidimensional feature vector.
[0078] The credibility prediction model adopts an improved spatiotemporal attention mechanism, where the query matrix Q and key matrix K are generated by learnable parameters (Q = XW Q , W Q ∈R (5×64)), the bidirectional LSTM layer (hidden state dimension 64) captures the temporal dependence, and then outputs the confidence score p(t) = 0.75 through the Sigmoid function. Based on the ROC curve analysis of the validation set, the dynamic threshold (T high = 0.82, T low = 0.62) is determined to classify it as a medium confidence level.
[0079] When explicit feedback (user labeled "26℃" as abnormal) or implicit conflict (kitchen sensor |ΔT| > 3℃ for 5 minutes) is detected, the system triggers the incremental learning process:
[0080] First, the Fisher information matrix of historical tasks is calculated Then, the elastic weight consolidation algorithm is used Update the model parameters, the whole update process takes 1.8 seconds, while ensuring the recognition accuracy of the new task (F1 = 0.93), the performance of the historical task is maintained (accuracy decrease < 3%). Actual deployment tests show that when the bedroom sensor is abnormal (p(t) = 0.58 < T low ), the system can switch to the backup sensor data and push the alarm within 500ms, while automatically recording the change of device health (C dev + = 0.5), which realizes the closed-loop processing flow from data collection, confidence evaluation to model optimization.
[0081] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present application and not to limit it. Although the present application has been described in detail with reference to the preferred embodiments, those skilled in the art should understand that the technical solutions of the present application can be modified or replaced by equivalents without departing from the spirit and scope of the present application, and they should be covered in the scope of the claims of the present application.
Claims
1. A real-time credibility marking method for smart home Internet of Things data, characterized in that: The following steps are involved: Data preprocessing and feature extraction: Receives multi-source heterogeneous data streams from multiple sensors, dynamically filters the raw data, and automatically adjusts the filter window length based on the sensor sampling rate to suppress high-frequency noise and preserve true signal characteristics. Extracting a multidimensional feature vector from the filtered data, the feature vector including the current sensor reading, time series statistical features, correlation features with neighboring sensors, and device health features; Feedforward credibility prediction: The multidimensional feature vector is input into a credibility prediction model, which includes a multi-head attention layer, a bidirectional long short-term memory network layer, and a fully connected layer. The multi-head attention layer is used to capture the spatial correlation between sensors, the bidirectional long short-term memory network layer is used to model temporal dependencies, and the fully connected layer is used to output a confidence score. The confidence score comprehensively reflects the reliability of the data in multiple aspects such as time and space dimensions and device status. Dynamic threshold optimization: Based on the receiver operating characteristic curve analysis of the validation set, the optimal classification threshold and the worst classification threshold are determined, and the confidence score is divided into three levels: high confidence, medium confidence, and low confidence; Anomaly detection and processing: When the confidence score is detected to be lower than the preset and worst classification threshold, the backup sensor data source or manual review process is triggered to ensure the reliability of the data; Feedback-driven model update: Feedback signals are collected through user feedback and / or multi-sensor data conflict results, and the parameters of the credibility prediction model are updated using an incremental learning algorithm. The incremental learning algorithm is used to prevent catastrophic forgetting during the model update process and ensure the long-term stability of the model.
2. The method for real-time marking of the credibility of smart home Internet of Things data according to claim 1, characterized in that: In the dynamic filtering process, the filter window length is calculated according to the sensor sampling rate using the following formula: where f s is the sensor sampling rate (unit: Hz), Indicates rounding down.
3. The method for real-time marking of the credibility of smart home Internet of Things data according to claim 1, characterized in that: The multidimensional feature vector includes the following contents: The sensor reading at the current moment; The mean and variance of the sliding window over the past N minutes; Pearson correlation coefficient with neighboring sensors; The equipment health code value is calculated based on the equipment's service life and the number of failures.
4. The method for real-time marking of the credibility of smart home Internet of Things data according to claim 1, characterized in that: The multi-head attention layer calculates the weighted feature vector through the following steps: Convert the input features into query matrix, key matrix and value matrix respectively through the learnable parameter matrix; Calculate the dot product of the transpose of the query matrix and the key matrix to obtain the first calculation result, divide the first calculation result by the arithmetic square root of the dimension of the key matrix to obtain the second calculation result, and normalize the second calculation result to the attention weight through the Softmax function; Multiply the attention weights by the value matrix to obtain the weighted feature vector.
5. The method for real-time marking of the credibility of smart home Internet of Things data according to claim 1, characterized in that: The forward LSTM and backward LSTM of the bidirectional LSTM layer calculate the hidden state according to the following steps: The forward LSTM processes data in chronological order, calculates the forget gate, input gate, output gate, and candidate memory cells, and updates the memory cells and hidden states. The backward LSTM processes the data in reverse, calculates the forget gate, input gate, output gate and candidate memory cells, and updates the memory cells and hidden states; The hidden states of the forward and backward LSTMs are fused by splicing to form a feature representation that contains the complete temporal context.
6. The method for real-time marking of the credibility of smart home Internet of Things data according to claim 1, characterized in that: The dynamic threshold optimization comprises the following steps: Calculate the confidence scores of all samples on the validation set; Traverse the thresholds with a preset step size and select the threshold that maximizes the Youden index as the optimal classification threshold; Calculate the worst classification threshold using the optimal classification threshold with preset parameters; The confidence scores are divided into three levels: high confidence, medium confidence and low confidence according to the optimal threshold and the worst classification threshold.
7. The method for real-time marking of the credibility of smart home Internet of Things data according to claim 1, characterized in that: The incremental learning algorithm adopts the elastic weight consolidation algorithm, and its loss function is: L EWC =L CE +λ∑ i F i (i i -θ i,old ) 2 L EWC is the loss function of the elastic weight consolidation algorithm, L CE is the cross entropy loss, F i The i-th parameter θ i Fisher information, θ i,old is the parameter obtained from the old training task, and λ is the weight hyperparameter for balancing the new and old tasks.
8. The method for real-time marking of the credibility of smart home Internet of Things data according to claim 1, characterized in that: The fully connected layer of the credibility prediction model maps the time series features to the [0, 1] interval through the Sigmoid function and outputs the confidence score.
9. The method for real-time marking of the credibility of smart home Internet of Things data according to claim 1, characterized in that: In the feedback-driven model update, user feedback includes explicit feedback and implicit feedback. Explicit feedback refers to users marking data anomalies through the APP, and implicit feedback refers to sensor data inconsistencies exceeding a preset number within the same decision cycle.
10. A real-time credibility marking system for smart home IoT data, characterized in that: include: Multi-protocol access module, used to receive multi-source heterogeneous data streams from multiple sensors, supporting Zigbee, Wi-Fi, and Bluetooth protocols; Edge computing unit, equipped with an NPU acceleration chip, is used to perform filtering, feature extraction, and credibility prediction in real time; Incremental learning platform, deployed on cloud servers, for model parameter updates and version management; Trusted tag database, used to store timestamped trustworthy tags and associated sensor metadata; Wherein, the system adopts the real-time marking method for the credibility of smart home Internet of Things data as described in any one of claims 1 to 9.