Panoramic perception data conversion method for Internet of Things

Abstract

The invention discloses an Internet of Things panoramic perception data conversion method. The method comprises the following steps: acquiring a multi-source heterogeneous Internet of Things perception data set; performing space-time alignment preprocessing on the multi-source heterogeneous Internet of Things sensing data set to generate a standardized space-time data set; inputting the standardized spatio-temporal data set into a feature fusion model, and generating a panoramic perception feature vector set through a dynamic weight distribution algorithm; performing adaptive compressed encoding on the panoramic sensing feature vector set to generate a compressed sensing data stream; and inputting the compressed sensing data stream into the multi-modal decision model to generate a real-time abnormal decision instruction. The method has the following advantages and effects that the problem of panoramic feature expression distortion caused by dynamic value imbalance of the multi-source heterogeneous sensing data is effectively solved, and then the reliability of real-time decision making is improved.

Description

Technical Field

[0001] The present invention relates to the technical field of data processing, and particularly relates to a method for converting panoramic perception data of the Internet of Things. Background Art

[0002] The fundamental technical bottleneck faced by the current Internet of Things intelligent decision-making system is that the dynamic value imbalance of multi-source heterogeneous perception data leads to the distortion of panoramic feature expression, thereby undermining the reliability of downstream real-time decision-making.

[0003] Due to the differences in physical principles and design objectives of various Internet of Things terminal devices, raw data streams with fragmented spatio-temporal benchmarks, different structures, and significant quality fluctuations are generated. In the time dimension, the local clock drift of the device causes a non-linear offset of the timestamp sequence; in the space dimension, when converting the local coordinate system to the geodetic coordinate system, the terrain atmospheric refraction effect is ignored, resulting in the distortion of elevation coordinates. More critically, there is a serious value mismatch between the dynamic quality decay of the data stream and the key feature requirements of the decision-making scenario, and the existing technology uses a static processing framework to further exacerbate the systematic imbalance. The disconnection between this static processing framework and the dynamic data value change leads to high-value features being submerged by low-quality data during the fusion and compression process. For example, in an industrial scenario, the weight of data from aging temperature sensors is not dynamically adjusted downward, and its abnormal readings dilute the fault features of high-confidence vibration sensors; or key displacement features are truncated by lossy compression, resulting in the spatio-temporal graph convolutional network missing the detection of device displacement anomalies. Summary of the Invention

[0004] The purpose of the present invention is to provide a method for converting panoramic perception data of the Internet of Things to solve the problems proposed in the background art.

[0005] The above technical purpose of the present invention is achieved through the following technical solutions: A method for converting panoramic perception data of the Internet of Things includes the following steps: S100. Obtain a multi-source heterogeneous Internet of Things perception data set; wherein, the multi-source heterogeneous Internet of Things perception data set includes at least two types of perception data streams with different data types, the perception data streams respectively originate from different types of Internet of Things perception devices, and the perception data streams include timestamp sequences, spatial coordinate sequences, and device identifiers; S200. Perform spatio-temporal alignment preprocessing on the multi-source heterogeneous Internet of Things perception data set to generate a standardized spatio-temporal data set; wherein, the spatio-temporal alignment preprocessing includes: synchronously calibrating the timestamp sequences of all perception data streams based on a preset time benchmark, and performing coordinate transformation on the spatial coordinate sequences of all perception data streams based on a unified spatial coordinate system; S300. Input the standardized spatio-temporal data set into the feature fusion model to generate a panoramic perception feature vector set through a dynamic weight allocation algorithm. Among them, the dynamic weight allocation algorithm dynamically adjusts the feature fusion weight coefficient according to the device type, data sampling frequency, and data confidence index of the perception data stream. S400. Perform adaptive compression coding on the panoramic perception feature vector set to generate a compressed sensing data stream. Among them, the adaptive compression coding dynamically selects a lossless compression mode or a lossy compression mode according to the feature importance grading result. S500. Input the compressed sensing data stream into the multi-modal decision model to generate a real-time anomaly decision instruction. Among them, the multi-modal decision model fuses the spatio-temporal rule reasoning and the deep learning prediction result to generate a decision instruction.

[0006] By adopting the above technical solutions, panoramic information fusion is realized by integrating multi-source heterogeneous Internet of Things perception data. First, a diverse perception data stream containing a timestamp sequence, a spatial coordinate sequence, and a device identifier is obtained to ensure the comprehensiveness and diversity of the data source, laying a global data foundation for subsequent analysis. Secondly, spatio-temporal alignment preprocessing is performed to unify different device data into a standardized spatio-temporal framework, completely eliminating the spatio-temporal benchmark misalignment problem caused by device clock differences and inconsistent coordinate systems, and constructing a computable data environment in the spatio-temporal dimension. Then, the adaptive fusion of multi-source features is realized through a dynamic weight allocation algorithm, and the weight coefficient is dynamically optimized according to the inherent reliability of the device type, the timeliness of the sampling frequency, and the confidence level, breaking through the limitations of traditional static weighted fusion, and significantly improving the accuracy and environmental adaptability of feature expression. Next, the adaptive compression coding technology is used to intelligently select a lossless or lossy compression mode according to the feature importance grading, maximizing the reduction of data transmission load while absolutely ensuring the integrity of key information. Finally, the multi-modal decision model fuses the real-time nature of spatio-temporal rule reasoning and the generalization ability of deep learning prediction to generate a highly reliable real-time anomaly decision instruction, forming a complete closed loop from data collection to intelligent decision-making, comprehensively improving the perception accuracy, response efficiency, and decision-making intelligence level of the Internet of Things system, and providing core support for situation awareness in complex scenarios.

[0007] A further setting is that obtaining the multi-source heterogeneous Internet of Things perception data set in S100 includes: S110. Receive the original perception data packet stream through a distributed device gateway interface cluster. Among them, the original perception data packet stream at least includes an encrypted data segment and a device type identifier. S120. Perform matching retrieval in the parsing template library according to the device type identifier in the original perception data packet stream, and load the corresponding data parsing template. Among them, the data parsing template includes a byte order definition rule and a data field offset mapping table. S130. Perform a decryption operation on the encrypted data segments in the original perception data packet stream to generate binary data segments; S140. Based on the byte order definition rules and the data field offset mapping table in the data parsing template, perform structured parsing on the binary data segments to extract the timestamp sequence, the spatial coordinate sequence, and the device perception value sequence; S150. Perform multi-level data integrity verification, including: The first-level verification: Calculate the cyclic redundancy check value of the binary data segments; The second-level verification: According to the data validity verification rules in the data parsing template, detect the monotonic increase of the timestamp sequence, the physical range rationality of the spatial coordinate sequence, and the mutation threshold of the device perception value sequence; S160. Perform dimensional standardization conversion on the device perception value sequence verified by S150: Based on the dimensional conversion coefficient matrix in the data parsing template, convert the device perception value sequence into a physical quantity value sequence in the International System of Units; S170. Aggregate the timestamp sequence, the spatial coordinate sequence, and the physical quantity value sequence according to the device identifier to generate the multi-source heterogeneous Internet of Things perception data set.

[0008] By adopting the above technical solutions, a standardized data collection and preprocessing system is established, and the original perception data packet stream is efficiently received through the distributed device gateway interface cluster; based on the intelligent matching mechanism of the dynamic parsing template library, accurate parsing of device-differentiated data formats is realized, and the byte order definition rules and the data field offset mapping table are used to ensure lossless conversion from binary data to structured data, solving the problem of semantic parsing consistency of multi-source data; implement a data integrity verification system, the first-level cyclic redundancy check ensures the integrity of packet transmission at the bit level, and the second-level verifies the timestamp monotonicity, the physical range rationality of the spatial coordinates, and the numerical mutation threshold based on predefined validity rules; unify the original device values through the dimensional standardization conversion matrix, eliminate the problem of data comparability caused by dimensional differences, and ensure the comparability of physical quantities across devices; finally, aggregate the spatio-temporal and physical quantity data according to the device identifier to form a standardized set, providing a data basis with high consistency and high integrity for subsequent processing, and ensuring the quality of panoramic perception data from the source.

[0009] A further setting is that in S200, spatio-temporal alignment preprocessing is performed on the multi-source heterogeneous Internet of Things perception data set to generate a standardized spatio-temporal data set, including: S210. Based on the preset Coordinated Universal Time benchmark, dynamically calibrate the timestamp sequences of the respective perception data streams in the multi-source heterogeneous Internet of Things perception data set; S220. Convert the spatial coordinate sequences of each perception data stream from the device native coordinate system to a unified spatial coordinate system; S230. Perform joint verification on the timestamp sequence calibrated in S210 and the standardized spatial coordinate sequence in the unified spatial coordinate system generated in S220, including verifying the physical rationality of the time-space trajectory and verifying the topological consistency of the spatial coordinates, and generate spatio-temporal metadata with verification marks. The generated spatio-temporal metadata includes drift error values, coordinate conversion confidence levels, and abnormal trajectory segment position indexes; S240. Recombine the verified timestamp sequence, the standardized spatial coordinate sequence in the unified spatial coordinate system, and the physical quantity numerical sequence generated in S160 according to the following rules: Construct a spatio-temporal data cube with the device identifier as the primary key and the timestamp sequence as the temporal index; Embed the spatio-temporal metadata generated in S230 into the spatio-temporal data cube as a quality description field; Slice the recombined spatio-temporal data cube according to a preset time window and store it as a standardized spatio-temporal data set in a columnar storage format.

[0010] By adopting the above technical solutions, a high-precision spatio-temporal reference unified system is constructed, the timestamp sequences of multiple devices are dynamically clock-calibrated based on the global coordinated time reference, and the network time protocol compensation mechanism is used to solve the millisecond-level temporal misalignment problem caused by clock drift between devices; the device native coordinate system is unified to a preset spatial coordinate system through a coordinate conversion engine to ensure the geometric comparability and topological computability of cross-device spatial data; perform joint verification on the physical rationality and topological consistency of the time-space trajectory, quantitatively annotate the data quality through metadata such as drift error values and coordinate conversion confidence levels, and establish a visual label for data credibility; slice according to a preset time window and adopt a columnar storage format to optimize the storage efficiency and query performance, forming a standardized data set with spatio-temporal consistency, quality assessability, and efficient access, meeting the real-time analysis requirements of large-scale spatio-temporal data.

[0011] A further setting is that the conversion process in S220 includes three levels of processing, namely primary conversion, secondary correction, and tertiary alignment; The original perception data packet stream received in S110 further includes structured device metadata, and the structured device metadata includes device geographical coordinates; In the primary conversion process, based on the device geographical coordinates in the device metadata, the local relative coordinates are converted to geodetic coordinate system coordinates through an affine transformation matrix; In the secondary correction process, according to the device type identifier, the corresponding terrain compensation parameters are loaded to compensate for the atmospheric refraction error of the elevation coordinates; During the three-level alignment process, the nearest neighbor matching algorithm is adopted to bind the geodetic coordinate system coordinates after the first-level transformation and the second-level correction to the preset spatial grid index, generating a standardized spatial coordinate sequence in the unified spatial coordinate system.

[0012] By adopting the above technical solution, through the three-level spatial coordinate transformation and alignment processing, the accuracy and consistency of spatial data standardization are greatly improved; the device geographic coordinate information is introduced in the first-level transformation to ensure the accurate mapping from the local relative coordinates to the global geodetic coordinate system; the second-level correction compensates the atmospheric refraction error for the elevation data, effectively reducing the spatial error caused by terrain differences or environmental factors; the three-level alignment ensures the consistency and standardization of the spatial positions of all sensing devices through the nearest neighbor spatial grid index binding mechanism, providing an accurate basis for subsequent spatial clustering, path analysis, and anomaly detection, and significantly improving the overall spatio-temporal recognition ability and processing accuracy of the system.

[0013] A further setting is that the S300 inputs the standardized spatio-temporal data set into the feature fusion model, and generates a panoramic perception feature vector set through the dynamic weight allocation algorithm, including: S310. Based on the device identifier in the standardized spatio-temporal data set, load the feature extractor matching the device type in the feature fusion model; the feature extractor includes a time feature extraction module, a spatial feature extraction module, and a physical quantity feature extraction module; specifically including: The time feature extraction module performs periodic pattern analysis on the timestamp sequence to generate a time dimension feature vector; The spatial feature extraction module performs spatial clustering analysis on the standardized spatial coordinate sequence in the unified spatial coordinate system to generate a spatial distribution feature vector; The physical quantity feature extraction module performs statistical feature calculation on the physical quantity numerical sequence to generate a physical quantity statistical feature vector; S320. Generate an optimized dynamic weight coefficient based on the device type priority factor, the data sampling frequency adjustment coefficient, and the data confidence compensation factor; S330. For each sensing data stream, respectively splice its time dimension feature vector, spatial distribution feature vector, and physical quantity statistical feature vector by channel to generate a single-device feature matrix, multiply the single-device feature matrix by the corresponding optimized dynamic weight coefficient to generate a weighted feature matrix; then, aggregate the weighted feature matrices of all sensing data streams, and generate a panoramic perception feature vector set through the gated fusion unit.

[0014] By adopting the above technical solutions, by introducing feature extraction modules in three dimensions of time, space, and physical quantities, the deep information in the standardized spatio-temporal data set can be fully mined, and the expression ability of the perception data can be improved; adopting the strategy of adaptively loading feature extractors for different device types enhances the adaptability and generalization ability of the system to process data of multiple devices; the dynamic weight allocation algorithm dynamically adjusts the fusion weights according to the device characteristics and data quality, avoiding the deviation problem caused by the fixed weighting method; through the generation of weighted feature matrices and the gated fusion mechanism, the redundant information of multi-device perception features can be effectively compressed, and representative panoramic perception feature vectors can be extracted, providing high-quality and low-redundancy input data for subsequent model decision-making, and improving the integrity and accuracy of overall perception.

[0015] A further setting is that after the S330, the following steps are further included: S340. For each perception data stream, obtain its data quality decay factor, and calculate the overall quality confidence score of the panoramic perception feature vector set based on the data quality decay factor; If the overall quality confidence score is less than the preset quality threshold, trigger the full-set feature reconstruction mechanism; the full-set feature reconstruction mechanism includes: According to the position index of the abnormal trajectory segment, delete the data slice corresponding to the corresponding time period from the standardized spatio-temporal data set, and mark it as a missing data point; Complete the missing data points to generate a completed standardized spatio-temporal data set; Input the newly generated standardized spatio-temporal data set into the feature fusion model to generate an updated panoramic perception feature vector set; if the overall quality confidence of the panoramic perception feature vector set still does not meet the standard, trigger the manual review process.

[0016] By adopting the above technical solutions, by introducing the data quality decay factor to conduct an overall confidence evaluation of the panoramic perception feature vector set, the system is equipped with self-diagnosis ability and can real-time perceive the downward trend of data quality; if the overall quality confidence score is less than the preset quality threshold, immediately start the full-set feature reconstruction mechanism. By locating abnormal trajectory segments, deleting missing data, and completing standardized data operations, the fault tolerance ability of the system to abnormal data and missing information is significantly enhanced; this reconstruction process can ensure that the system perception stability remains high under the condition of abnormal data fluctuations, effectively preventing incorrect data from interfering with the analysis results.

[0017] A further setting is that the S320 specifically includes the following steps: S320.1. Based on the device identifier in the standardized spatio-temporal data set, obtain the device type priority factor; including: Query the device type identifier corresponding to the device identifier in the preset device credibility level table; Obtain the basic value of the priority weight according to the device type identifier; Multiply the basic value of the priority weight by the device type correction coefficient to generate the device type priority factor; S320.2. Based on the structured device metadata in the standardized spatio-temporal data set, obtain the data sampling frequency adjustment coefficient; including: Extract the sampling frequency of the current sensed data stream from the structured device metadata ; Compare the sampling frequency with the preset standard frequency ; If , then take as the data sampling frequency adjustment coefficient; If , then take as the data sampling frequency adjustment coefficient; where and are the upper threshold and lower threshold of the preset frequency respectively; S320.3. Based on the structured device metadata in the standardized spatio-temporal data set, obtain the data confidence compensation factor; including: Extract the coordinate transformation confidence and drift error value from the spatio-temporal metadata, calculate the basic quality score based on the coordinate transformation confidence and drift error value, and then generate the data confidence compensation factor based on the basic quality score; S320.4. Multiply the device type priority factor, the data sampling frequency adjustment coefficient and the data quality attenuation factor to generate the original dynamic weight coefficient; S320.5. Normalize the original dynamic weight coefficients of all sensed data streams to generate the optimized dynamic weight coefficients.

[0018] By adopting the above technical solution, through the comprehensive weighted calculation mechanism of the device type priority factor, the data sampling frequency adjustment coefficient and the data confidence compensation factor, a comprehensive dynamic weight generation framework is constructed, effectively enhancing the system's accurate evaluation ability of the contribution degree of different device data; this mechanism can automatically sense multi-dimensional quality factors such as the signal-to-noise ratio, update frequency and spatial accuracy of device data, dynamically generate weight coefficients, and improve the refinement level of feature fusion; the normalization process ensures the coordination and numerical stability of each weight in the overall fusion process, further improving the stability of the fusion result and the robustness of the model, and effectively adapting to the data differences and uncertainties in various complex environments.

[0019] A further setting is that in S400, adaptive compression coding is performed on the panoramic perception feature vector set to generate a compressed sensing data stream, including: S410. Based on the spatio-temporal metadata and feature vectors in the panoramic perception feature vector set, calculate a feature importance grading matrix; wherein, the generation process of the feature importance grading matrix includes: Extract the coordinate transformation confidence and drift error value from the spatio-temporal metadata; query the basic weight value in a preset feature sensitivity table according to the feature vector type; the feature vector type includes time dimension feature vectors, spatial distribution feature vectors, and physical quantity statistical feature vectors; generate a feature importance score through a weighted fusion method for the coordinate transformation confidence, drift error value, and feature vector type. S420. Trigger dual-mode compression according to the relationship between the feature importance score and a preset compression threshold: If the feature importance score is greater than or equal to the preset compression threshold, activate the lossless compression mode; If the feature importance score is less than the preset compression threshold, activate the lossy compression mode; S430. Perform streaming encapsulation on the compressed time dimension feature vectors, spatial distribution feature vectors, and physical quantity statistical feature vectors, and aggregate them into a compressed sensing data stream.

[0020] By adopting the above technical solution, through the adaptive compression coding mechanism, differential compression processing is performed on the panoramic perception feature vector set according to the feature importance, significantly improving the efficiency of data transmission and storage; in the feature grading process, the coordinate transformation confidence, drift error, and feature sensitivity information are fused, making the compression decision highly refined and target-oriented; the dual-mode compression strategy combines lossless and lossy methods, which can minimize the data volume to the greatest extent without sacrificing the integrity of key information; the finally generated compressed sensing data stream is not only structurally compact but also retains key feature information, providing efficient and reliable data support for subsequent real-time processing and analysis.

[0021] A further setting is that the lossless compression mode in S420 includes: Perform dictionary coding on the time dimension feature vectors, spatial distribution feature vectors, and physical quantity statistical feature vectors by using a sliding window-based dictionary coding algorithm, and append a cyclic redundancy check code; The lossy compression mode includes: Perform discrete cosine transform on the physical quantity statistical feature vectors to generate a frequency domain coefficient matrix; Load a quantization table matching the frequency domain coefficient matrix according to the device type identifier, and perform truncation quantization on the high-frequency components.

[0022] By adopting the above technical solution, the lossless compression mode adopts a sliding window dictionary coding strategy and combines a cyclic redundancy check mechanism to ensure that information is not lost during the compression process of feature vectors and the results are verifiable, which is especially suitable for the data fidelity transmission of key features; the lossy compression mode extracts the main components of the data through frequency domain conversion, truncates the high-frequency components to reduce redundancy, and combines a quantization table for compression, which can greatly improve the compression ratio on the basis of ensuring the integrity of the overall information structure and effectively meet the actual requirements in an environment with limited bandwidth or storage; this combined compression strategy realizes the dynamic balance between compression quality and efficiency, and improves the adaptability and practical value of the system to different scenarios.

[0023] A further setting is that in S500, the compressed sensing data stream is input into a multi-modal decision-making model to generate real-time anomaly decision-making instructions and a set of device control parameters, including: S510. Load a spatio-temporal decision rule library and a deep learning prediction model in the multi-modal decision-making model; among them, the deep learning prediction model is an anomaly prediction model based on a spatio-temporal graph convolutional network, and its input is the panoramic perception feature vector decoded from the compressed sensing data stream. S520. Perform multi-level decision fusion, including: The first-level decision: Perform real-time inference on the panoramic perception feature vector based on the spatio-temporal decision rule library to generate a set of rule-triggered events. The second-level decision: Input the panoramic perception feature vector into the spatio-temporal graph convolutional network to output an anomaly probability prediction value and a potential anomaly type label. The third-level decision: Perform logical fusion on the set of rule-triggered events and the anomaly probability prediction value, and generate an output decision, where the output decision includes an immediate action instruction, a delayed action instruction, and a no-action instruction. S530. If the output decision is an immediate action instruction, generate a set of device control parameters including a device identifier, a control instruction code, and a parameter threshold, and send them to the target Internet of Things perception device through a device control interface. If the output decisions are a delayed action instruction and a no-action instruction, record the panoramic perception feature vector in the historical decision buffer pool.

[0024] By adopting the above technical solutions, it is proposed to input the compressed sensing data stream into the multi-modal decision-making model to further exert the value of the panoramic sensing data in intelligent decision-making; the multi-modal decision-making model integrates rule reasoning and deep learning technologies, and has both interpretability and predictability. The spatio-temporal decision rule base quickly identifies known patterns in specific scenarios and realizes instant response to rule triggers; the deep learning prediction model based on the spatio-temporal graph convolutional network can mine complex patterns and potential risks and realize forward-looking anomaly prediction; the three-level fusion mechanism in the decision-making process ensures that the results of rule reasoning and model prediction confirm each other, avoiding misjudgment or missed judgment caused by a single path, and improving the accuracy and stability of the decision-making system; the finally output instructions are divided into three categories: immediate action, delayed action, and no action, realizing intelligent response management based on the degree of risk; when the decision result is an immediate action instruction, the system can automatically generate a set of device control parameters and send them to the terminal device to realize automatic intervention and closed-loop control; the delayed action and no action instructions are recorded in the historical buffer pool for subsequent evaluation and model training, reflecting the system's cumulative learning ability of historical information.

[0025] In summary, the present invention has the following beneficial effects: Effectively solve the problem that the dynamic value imbalance of multi-source heterogeneous sensing data leads to the distortion of panoramic feature expression, thereby improving the reliability of real-time decision-making. BRIEF DESCRIPTION OF THE DRAWINGS

[0026] Figure 1 It is the main process block diagram of the embodiment; Figure 2 It is the process schematic diagram of S100 in the embodiment; Figure 3 It is the process schematic diagram of S200 in the embodiment; Figure 4 It is the process schematic diagram of S300 in the embodiment; Figure 5 It is the process schematic diagram of S320 in the embodiment; Figure 6 It is the process schematic diagram of S340 in the embodiment; Figure 7 It is the process schematic diagram of S400 in the embodiment; Figure 8 It is the process schematic diagram of S500 in the embodiment. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0027] The present invention will be further described in detail below with reference to the accompanying drawings.

[0028] As shown in the Figures 1 to 8 accompanying drawings; This embodiment discloses an Internet of Things panoramic sensing data conversion method, including the following steps: S100. Obtain a multi-source heterogeneous Internet of Things (IoT) perception data set. Among them, the multi-source heterogeneous IoT perception data set includes at least two types of perception data streams with different data types. The perception data streams are respectively from different types of IoT perception devices, and the perception data streams contain a time stamp sequence, a spatial coordinate sequence, and a device identifier. S200. Perform spatio-temporal alignment preprocessing on the multi-source heterogeneous IoT perception data set to generate a standardized spatio-temporal data set. Among them, the spatio-temporal alignment preprocessing includes: synchronously calibrating the time stamp sequences of all perception data streams based on a preset time reference, and performing coordinate transformation on the spatial coordinate sequences of all perception data streams based on a unified spatial coordinate system. S300. Input the standardized spatio-temporal data set into a feature fusion model, and generate a panoramic perception feature vector set through a dynamic weight allocation algorithm. Among them, the dynamic weight allocation algorithm dynamically adjusts the feature fusion weight coefficient according to the device type, data sampling frequency, and data confidence index of the perception data stream. S400. Perform adaptive compression coding on the panoramic perception feature vector set to generate a compressed perception data stream. Among them, the adaptive compression coding dynamically selects a lossless compression mode or a lossy compression mode according to the feature importance grading result. S500. Input the compressed perception data stream into a multi-modal decision model to generate a real-time anomaly decision instruction. Among them, the multi-modal decision model fuses spatio-temporal rule reasoning and deep learning prediction results to generate a decision instruction.

[0029] In a possible implementation manner, obtaining the multi-source heterogeneous IoT perception data set in S100 includes: S110. Receive an original perception data packet stream through a distributed device gateway interface cluster. Among them, the original perception data packet stream at least includes an encrypted data segment and a device type identifier. More specifically, the original perception data packet stream includes an encrypted data segment, a cyclic redundancy check code, and structured device metadata. The structured device metadata includes a device type identifier, a device geographical coordinate, a device sampling frequency parameter, and a device health status code. S120. Perform matching retrieval in a parsing template library according to the device type identifier in the original perception data packet stream, and load the corresponding data parsing template. Among them, the data parsing template includes a byte order definition rule and a data field offset mapping table. More specifically, the data parsing template includes a byte order definition rule, a data field offset mapping table, a dimension conversion coefficient matrix, and a data validity check rule. S130. Perform a decryption operation on the encrypted data segments in the original sensed data packet stream to generate binary data segments. Preferably, an asynchronous pipeline processing architecture or a multi-core collaborative processing architecture can be used for processing. Among them, the decryption operation includes: selecting an asymmetric decryption algorithm or a lightweight symmetric decryption algorithm according to the device type identifier, and accelerating the decryption process through a hardware security module. S140. Based on the byte order definition rules and the data field offset mapping table in the data parsing template, perform structured parsing on the binary data segments to extract the timestamp sequence, the spatial coordinate sequence, and the device sensed value sequence. Among them, the structured parsing includes: identifying the data packet header identification bit, separating multi-channel sensor data, and reorganizing fragmented transmission data packets. S150. Perform multi-level data integrity verification, including: The first-level verification: Calculate the cyclic redundancy check value of the binary data segment. When the cyclic redundancy check value is inconsistent with the cyclic redundancy check code in the original sensed data packet stream, trigger a data retransmission request and mark the device exception log. The second-level verification: According to the data validity verification rules in the data parsing template, detect the monotonic increase of the timestamp sequence, the physical range rationality of the spatial coordinate sequence, and the mutation threshold of the device sensed value sequence. S160. Perform dimension standardization conversion on the device sensed value sequence verified by S150: Based on the dimension conversion coefficient matrix in the data parsing template, convert the device sensed value sequence into a physical quantity value sequence in the International System of Units. S170. Aggregate the timestamp sequence, the spatial coordinate sequence, and the physical quantity value sequence according to the device identifier to generate a multi-source heterogeneous Internet of Things sensed data set. The multi-source heterogeneous Internet of Things sensed data set is stored in a distributed cache cluster in the form of a time partition table.

[0030] Embodiment 1 Receive the original sensed data packet stream from the temperature and humidity sensor with the device type identifier THS-01 and the GPS locator with the device type identifier GPS-05 through the distributed device gateway interface cluster. In the first-level verification, calculate the CRC check value 0x7D3F of the binary data segment. If it is consistent with the original packet, it passes. In the second-level verification, detect that the timestamp is monotonically increasing, the error is less than 1 ms, and the longitude value of the spatial coordinate is within [-180, 180]. Aggregate the timestamp sequence, the spatial coordinate sequence, and the physical quantity value sequence according to the device identifier and store them as a time partition table. An example time partition table is shown in Table 1. Table 1 Example Time Partition Table

[0031] In one possible implementation, S200 performs spatiotemporal alignment preprocessing on a multi-source heterogeneous IoT sensing data set to generate a standardized spatiotemporal data set, including: S210, based on a preset global coordinated time reference, dynamically calibrating the timestamp sequence of each perception data stream in a multi-source heterogeneous IoT perception data set; including: Calculate the drift error between the local clock of each IoT sensing device and the global coordinated time reference; It should be noted that the drift error value is generated as follows: ; in, is the drift error value; is the stability attenuation factor related to the equipment type, obtained by looking up the table; The local timestamp of the device; is the globally coordinated time reference; is the ambient temperature influence coefficient; is the inherent clock variance of the device; A linear interpolation algorithm is used to reconstruct the missing time periods in the timestamp sequence. More specifically, the linear interpolation algorithm interpolates and reconstructs the missing time periods based on the time intervals between adjacent valid timestamps to generate a continuous timestamp sequence synchronized with the global coordinated time reference. When a timestamp jump exceeding a preset threshold is detected, a timing smoothing correction based on the historical sampling frequency is triggered; more specifically, the drift error calculation needs to integrate the clock stability coefficient corresponding to the device type identifier, and the timing smoothing correction needs to retain the time partition table structure of the original data.

[0032] S220, converting the spatial coordinate sequence of each perception data stream from the native coordinate system of the device to the unified spatial coordinate system; S230, performing joint verification on the timestamp sequence calibrated in S210 and the standardized spatial coordinate sequence in the unified spatial coordinate system generated in S220, including verifying the physical rationality of the time-space trajectory and verifying the topological consistency of the spatial coordinates, and generating spatiotemporal metadata with verification marks, the generated spatiotemporal metadata including drift error value, coordinate conversion confidence and abnormal trajectory segment position index; Preferably, in verifying the physical rationality of the time-space trajectory, the displacement speed of adjacent time stamps is calculated, and when it exceeds the maximum movement speed corresponding to the device type identifier, it is marked as an abnormal trajectory segment; For each device identifier, calculate the displacement velocity of adjacent spatial coordinates in timestamp order : ; in, is the displacement velocity; The X-axis coordinate difference between the th time point and the th time point; The Y-axis coordinate difference between the th time point and the th time point; The Y-axis coordinate difference between the th time point and the th time point; Indicates the time difference between the th time point and the th time point, in seconds; If , then mark the time interval , 1] as an abnormal trajectory segment; Is the maximum moving speed corresponding to the device type identifier; Record the start timestamp index and the end timestamp index of the abnormal trajectory segment, and associate the device identifier; Preferably, in verifying the topological consistency of spatial coordinates, based on the topological relationship of IoT sensing devices, detect whether the spatial coordinate sequence violates the device deployment constraint rules; Preferably, the generation process of the drift error value in spatio-temporal metadata is as follows: ; Where is the drift error value; is the number of sampling points of the current device within the time window; and are the start and end timestamps of the window respectively; represents the drift error value corresponding to the th timestamp.

[0033] Preferably, in the first-level conversion process, calculate the sum of squared residuals of the affine transformation matrix, which reflects the conversion accuracy of converting local relative coordinates to geodetic coordinate system coordinates; In the second-level correction process, obtain the calibration error of the terrain compensation parameter, based on the compensation model accuracy level corresponding to the device type identifier; In the third-level alignment process, calculate the grid matching distance output by the nearest neighbor matching algorithm, which represents the deviation between the geodetic coordinate system coordinates after the first-level conversion and the second-level correction and the preset spatial grid; Generate the coordinate conversion confidence through weighted fusion: ; Among them, , and are the error weights for primary conversion, secondary calibration, and tertiary alignment respectively, and are pre-configured according to the device type identifier.

[0034] S240. Recombine the verified timestamp sequence, the standardized spatial coordinate sequence in the unified spatial coordinate system, and the physical quantity numerical sequence generated by S160 according to the following rules: Construct a spatio-temporal data cube with the device identifier as the primary key and the timestamp sequence as the temporal index; Embed the spatio-temporal metadata generated by S230 in the spatio-temporal data cube as a quality description field; Slice the recombined spatio-temporal data cube according to a preset time window and store it as a standardized spatio-temporal data set in columnar storage format.

[0035] Embodiment 2 Calculate the GPS device clock drift error: ; When the data at the 2nd second is missing, reconstruct the missing period through linear interpolation: , ; , ; When ; During the primary conversion process, the sum of squared residuals ; During the secondary calibration process, load the GPS device terrain compensation parameter to correct the atmospheric refraction error of the elevation coordinate to 0.3 m; During the tertiary alignment process, match the preset spatial grid index, and the grid matching distance m; The corresponding coordinate conversion confidence .

[0036] Displacement speed , for the device type GPS-05 , no exception is triggered.

[0037] In a possible implementation, the conversion process in S220 includes three levels of processing, namely primary conversion, secondary calibration, and tertiary alignment; The original perception data packet stream received in S110 also contains structured device metadata, and the structured device metadata contains device geographical coordinates; During the first-level conversion process, based on the device geographical coordinates in the device metadata, the local relative coordinates are converted into coordinates in the geodetic coordinate system through an affine transformation matrix; the coordinates in the geodetic coordinate system include longitude, latitude, and elevation coordinates; During the second-level calibration process, according to the device type identifier, the corresponding terrain compensation parameters are loaded to compensate for the atmospheric refraction error of the elevation coordinates; the elevation coordinates here are derived from the coordinates in the geodetic coordinate system; During the third-level alignment process, the nearest neighbor matching algorithm is used to bind the coordinates in the geodetic coordinate system after the first-level conversion and the second-level calibration to the preset spatial grid index, generating a standardized spatial coordinate sequence in the unified spatial coordinate system.

[0038] In a possible implementation, S30 Inputs the standardized spatio-temporal data set into the feature fusion model, and generates a panoramic perception feature vector set through a dynamic weight allocation algorithm, including: S310. Based on the device identifier in the standardized spatio-temporal data set, a feature extractor matching the device type is loaded in the feature fusion model; the feature extractor includes a time feature extraction module, a spatial feature extraction module, and a physical quantity feature extraction module; specifically including: The time feature extraction module performs periodic pattern analysis on the time stamp sequence to generate a time dimension feature vector; The spatial feature extraction module performs spatial clustering analysis on the standardized spatial coordinate sequence in the unified spatial coordinate system to generate a spatial distribution feature vector; The physical quantity feature extraction module performs statistical feature calculation on the physical quantity numerical sequence to generate a physical quantity statistical feature vector; S320. Based on the device type priority factor, the data sampling frequency adjustment coefficient, and the data confidence compensation factor, an optimized dynamic weight coefficient is generated; S330. For each perception data stream, its time dimension feature vector, spatial distribution feature vector, and physical quantity statistical feature vector are respectively concatenated by channels to generate a single-device feature matrix, and the single-device feature matrix is multiplied by the corresponding optimized dynamic weight coefficient to generate a weighted feature matrix; then, the weighted feature matrices of all perception data streams are aggregated, and a panoramic perception feature vector set is generated through a gated fusion unit.

[0039] Generate a single-device feature matrix for each perception data stream and calculate the weighted feature matrix : ; Among them, is the optimized dynamic weight coefficient after normalization; Preferably, the gated fusion unit performs the following operations: ; in, is the generated panoramic perception feature vector; is the Sigmoid function; Indicates that The weighted feature matrices of each device are concatenated along the feature dimension; is the hyperbolic tangent function; , , is a trainable parameter; is the weighted sum of the feature matrices of each device.

[0040] In a possible implementation, after S330, the following steps are further included: S340: For each perception data stream, obtain its data quality attenuation factor, and calculate the overall quality confidence score of the panoramic perception feature vector set based on the data quality attenuation factor; ; in, Provide an overall quality confidence score; Indicates the The data quality degradation factor of each sensory data stream; Indicates the total number of sensory data streams included; If the overall quality confidence score is less than the preset quality threshold, , Indicates the preset quality threshold, triggering the full feature reconstruction mechanism; the full feature reconstruction mechanism includes: According to the abnormal trajectory segment position index, the data slice of the corresponding period is deleted from the standardized spatiotemporal data set and marked as missing data points; Complete the missing data points to generate a completed standardized spatiotemporal data set; More specifically: If the missing data points are in the timestamp sequence, linear interpolation or Kalman filtering algorithm is used to reconstruct the timestamp sequence; if the missing data points are in the spatial distribution feature vector, a spatial coordinate sequence is generated based on the device topology relationship and historical trajectory data; if the missing data points are in the physical quantity statistical feature vector, the device type characteristics and the physical quantity numerical sequence of the adjacent devices are used for collaborative completion.

[0041] The newly generated standardized spatiotemporal data set is input into the feature fusion model to generate an updated panoramic perception feature vector set; if the overall quality confidence of the panoramic perception feature vector set still does not meet the standard, the manual review process is triggered.

[0042] In a possible implementation, S320 specifically includes the following steps: S320.1. Obtain the device type priority factor based on the device identifier in the standardized spatio-temporal data set, including: Query the device type identifier corresponding to the device identifier in the preset device credibility level table; Obtain the basic value of the priority weight according to the device type identifier; Multiply the basic value of the priority weight by the device type correction coefficient to generate the device type priority factor; More specifically: Query the preset device credibility level table, match the device type according to the device identifier, return the basic value of the priority weight, and apply the device type correction coefficient; ; Among them, is the device type priority factor; is the basic value of the priority weight; is the device type correction coefficient.

[0043] An example table of the device credibility level table is shown in Table 2; Table 2 Example table of the device credibility level table

[0044] S320.2. Obtain the data sampling frequency adjustment coefficient based on the structured device metadata in the standardized spatio-temporal data set, including: Extract the sampling frequency of the current perception data stream from the structured device metadata ; Compare the sampling frequency with the preset standard frequency ; If , then take as the data sampling frequency adjustment coefficient; If , then take as the data sampling frequency adjustment coefficient; Among them, and are the upper threshold and lower threshold of the preset frequency respectively; S320.3. Obtain the data confidence compensation factor based on the structured device metadata in the standardized spatio-temporal data set, including: Extract the coordinate transformation confidence and drift error value from the spatio-temporal metadata, calculate the basic quality score based on the coordinate transformation confidence and drift error value, and then generate the data confidence compensation factor based on the basic quality score; The generation process of the basic quality score is as follows: ; Among them, is the basic quality score; is the coordinate transformation confidence; is the drift error value; is the maximum clock drift allowed for the device type, , is the preset weight coefficient; Next, divide by to generate the data quality attenuation factor ; S320.4. Multiply the device type priority factor, the data sampling frequency adjustment coefficient, and the data quality attenuation factor to generate the original dynamic weight coefficient; ; Among them, represents the original dynamic weight coefficient of the th perception data stream; S320.5. Normalize the original dynamic weight coefficients of all perception data streams to generate the optimized dynamic weight coefficient.

[0045] ; Among them, is the optimized dynamic weight coefficient after normalization; generates an independent for each perception data stream; among them, each corresponds to a perception data stream respectively, represents the total number of perception data streams included.

[0046] Embodiment 3 The device type priority factor corresponding to the GPS locator ; The sampling frequency , the preset standard frequency ; Since , , where the takes the value of 1.5; The data quality attenuation factor is 0.79 after calculation; The optimized dynamic weight coefficient generated by the normalization process is 0.42; The overall quality confidence score has a value of 0.8.

[0047] In a possible implementation manner, in S400, performing adaptive compression coding on the panoramic perception feature vector set to generate a compressed perception data stream includes: S410. Based on the spatio-temporal metadata and feature vectors in the panoramic perception feature vector set, calculate the feature importance grading matrix. The generation process of the feature importance grading matrix includes: Extract the coordinate transformation confidence and drift error value from the spatio-temporal metadata; query the basic weight value in the preset feature sensitivity table according to the feature vector type. The feature vector types include time dimension feature vectors, spatial distribution feature vectors, and physical quantity statistical feature vectors; generate the feature importance score by weighted fusion of the coordinate transformation confidence, drift error value, and feature vector type. The weighted fusion process is as follows: ; Among them, is the importance score of the th type of feature; is the coordinate transformation confidence; is the drift error value; is the maximum clock drift allowed by the device type; is the th type of feature's basic weight value, obtained through the preset feature sensitivity table; represents the inherent importance of the feature; , , is the preset fusion coefficient.

[0048] S420. Trigger dual-mode compression according to the relationship between the feature importance score and the preset compression threshold: If the feature importance score is greater than or equal to the preset compression threshold, activate the lossless compression mode; If the feature importance score is less than the preset compression threshold, activate the lossy compression mode; S430. Perform streaming encapsulation on the compressed time dimension feature vectors, spatial distribution feature vectors, and physical quantity statistical feature vectors, and aggregate them into a compressed sensing data stream.

[0049] More specifically: Add a compression mode identifier, feature type metadata, and time partition index header; Aggregate the encapsulated data packets into a compressed sensing data stream according to the preset time window, and send them to the edge computing node through the Internet of Things transmission protocol stack.

[0050] In a possible implementation, the lossless compression mode in S420 includes: Perform dictionary encoding on the time dimension feature vectors, spatial distribution feature vectors, and physical quantity statistical feature vectors using a sliding window-based dictionary encoding algorithm, and append a cyclic redundancy check code; Preferably, the sliding window-based dictionary encoding algorithm adopts the LZ77 algorithm; The lossy compression mode includes: Perform discrete cosine transform on the statistical feature vector of the physical quantity to generate a frequency domain coefficient matrix; Load a quantization table matching the frequency domain coefficient matrix according to the device type identifier, and perform truncation quantization on the high-frequency components.

[0051] Embodiment 4 The preset compression threshold is set to 0.7; When , activate lossless compression: compress the speed feature vector using the LZ77 algorithm → compression rate 70%; When , activate lossy compression: truncate the high-frequency components after DCT transformation of the coordinate sequence, and retain 85% of the energy; Add the following header information: "Compression mode identifier": "Lossless - LZ77", "Feature type metadata": "Physical quantity statistical feature", "Time partition index": "1620000000 - 1620003600"; Send the compressed sensing data stream to the edge node through the MQTT protocol.

[0052] In a possible implementation, in S500, input the compressed sensing data stream into the multi-modal decision model to generate real-time anomaly decision instructions and a set of device control parameters, including: S510. Load the spatio-temporal decision rule library and the deep learning prediction model in the multi-modal decision model; among them, the spatio-temporal decision rule library contains device deployment constraint rules, physical kinematic rules, and environmental threshold rules; the deep learning prediction model is an anomaly prediction model based on the spatio-temporal graph convolutional network, and its input is the panoramic perception feature vector decoded from the compressed sensing data stream; S520. Perform multi-level decision fusion, including: The first-level decision: Perform real-time inference on the panoramic perception feature vector based on the spatio-temporal decision rule library to generate a set of rule-triggered events; Specifically include: Verify the spatial topological relationship according to the device deployment constraint rules, verify the displacement speed overrun event according to the physical kinematic rules, and verify the physical quantity value out-of-bounds event according to the environmental threshold rules; The second-level decision: Input the panoramic perception feature vector into the spatio-temporal graph convolutional network, and output the anomaly probability prediction value and the potential anomaly type label; The third-level decision: Perform logical fusion on the set of rule-triggered events and the anomaly probability prediction value, and generate an output decision. The output decision includes an immediate action instruction, a delayed action instruction, and a no-action instruction; Its corresponding mathematical expression is as follows: ; Wherein, corresponds to the output decision; corresponds to the immediate action instruction; corresponds to the delayed action instruction; corresponds to the no-action instruction; represents the abnormal probability prediction value output by the spatio-temporal graph convolutional network; and respectively represent the preset high probability threshold and low probability threshold; indicates that a high-risk rule event has been triggered. The high-risk rule event includes at least one of an event of exceeding the displacement speed limit, an event of exceeding the displacement acceleration limit, or an event of exceeding the physical quantity value boundary; More specifically, the high-risk rule event includes: Event of exceeding the displacement speed limit: Triggered when the displacement speed in S230 wherein, is the maximum moving speed corresponding to the device type identifier; Event of exceeding the displacement acceleration limit: Parse the standardized spatial coordinate sequence and time stamp sequence from the panoramic perception feature vector, and then calculate the spatial displacement acceleration within adjacent time windows : ; Wherein, , are respectively and the displacement speeds at moments; When it is triggered, wherein, is the maximum acceleration threshold corresponding to the device type identifier; Event of exceeding the physical quantity value boundary: Triggered when the physical quantity value sequence exceeds the range preset by the environmental threshold rule.

[0053] S530, if the output decision is an immediate action instruction, generate a device control parameter set including the device identifier, control instruction code, and parameter threshold, and send it to the target Internet of Things perception device through the device control interface; If the output decisions are delayed action instructions and no-action instructions, record the panoramic perception feature vector in the historical decision buffer pool.

[0054] Embodiment 5 The temperature sensor value of 35°C is greater than the threshold of 30°C, triggering an event of exceeding the physical quantity value boundary; The spatio-temporal graph convolutional network inputs the panoramic feature vector and outputs the abnormal probability prediction value , the label "temperature anomaly" generates an immediate action instruction and sends it through the Modbus protocol.

[0055] This specific embodiment is only an interpretation of the present invention and is not a limitation thereof. After reading this specification, those skilled in the art can make modifications to this embodiment that do not contribute creatively as needed, but as long as they are within the scope of the claims of the present invention, they are protected by the patent law.

Claims

1. An Internet of Things panoramic perception data conversion method, characterized in that, It includes the following steps: S100. Obtain a multi-source heterogeneous Internet of Things perception data set; wherein, the multi-source heterogeneous Internet of Things perception data set includes at least two types of perception data streams with different data types, the perception data streams are respectively from different types of Internet of Things perception devices, and the perception data streams contain timestamp sequences, spatial coordinate sequences and device identifiers; S200. Perform spatio-temporal alignment preprocessing on the multi-source heterogeneous Internet of Things perception data set to generate a standardized spatio-temporal data set; wherein, the spatio-temporal alignment preprocessing includes: synchronously calibrating the timestamp sequences of all perception data streams based on a preset time benchmark, and performing coordinate transformation on the spatial coordinate sequences of all perception data streams based on a unified spatial coordinate system; S300. Input the standardized spatio-temporal data set into a feature fusion model to generate a panoramic perception feature vector set through a dynamic weight allocation algorithm; wherein, the dynamic weight allocation algorithm dynamically adjusts the feature fusion weight coefficients according to the device type, data sampling frequency and data confidence index of the perception data stream; S400. Perform adaptive compression coding on the panoramic perception feature vector set to generate a compressed perception data stream; wherein, the adaptive compression coding dynamically selects a lossless compression mode or a lossy compression mode according to the result of feature importance grading; S500. Input the compressed perception data stream into a multi-modal decision model to generate a real-time anomaly decision instruction; wherein, the multi-modal decision model fuses spatio-temporal rule reasoning and deep learning prediction results to generate a decision instruction.

2. The method for converting Internet of Things panoramic perception data according to claim 1, wherein: In S100, obtaining the multi-source heterogeneous Internet of Things perception data set includes: S110. Receive the original perception data packet stream through a distributed device gateway interface cluster; wherein, the original perception data packet stream at least contains an encrypted data segment and a device type identifier; S120. Perform matching retrieval in the parsing template library according to the device type identifier in the original perception data packet stream, and load the corresponding data parsing template; wherein, the data parsing template contains a byte order definition rule and a data field offset mapping table; S130. Perform a decryption operation on the encrypted data segment in the original perception data packet stream to generate a binary data segment; S140. Based on the byte order definition rule and the data field offset mapping table in the data parsing template, perform structured parsing on the binary data segment to extract a timestamp sequence, a spatial coordinate sequence and a device perception value sequence; S150. Perform multi-level data integrity verification; including: The first-level verification: Calculate the cyclic redundancy check value of the binary data segment; The second-level verification: According to the data validity verification rule in the data parsing template, detect the monotonic increase of the timestamp sequence, the physical range rationality of the spatial coordinate sequence and the mutation threshold of the device perception value sequence; S160. Perform dimension standardization conversion on the device perception value sequence verified by S150: Based on the dimension conversion coefficient matrix in the data parsing template, convert the device perception value sequence into a physical quantity value sequence in the international standard unit system; S170. Aggregate the timestamp sequence, spatial coordinate sequence, and physical quantity value sequence according to the device identifier to generate the multi-source heterogeneous Internet of Things perception data set.

3. A method for converting Internet of Things panoramic perception data according to claim 2, characterized in that: In S200, perform spatio-temporal alignment preprocessing on the multi-source heterogeneous Internet of Things perception data set to generate a standardized spatio-temporal data set, including: S210. Dynamically calibrate the timestamp sequence of each perception data stream in the multi-source heterogeneous Internet of Things perception data set based on a preset Coordinated Universal Time (UTC) benchmark. S220. Convert the spatial coordinate sequence of each perception data stream from the device native coordinate system to a unified spatial coordinate system. S230. Perform joint verification on the timestamp sequence calibrated in S210 and the standardized spatial coordinate sequence in the unified spatial coordinate system generated in S220, including verifying the physical rationality of the time-space trajectory and verifying the topological consistency of the spatial coordinates, and generate spatio-temporal metadata with verification marks. The generated spatio-temporal metadata includes drift error values, coordinate transformation confidence levels, and abnormal trajectory segment position indexes. S240. Recombine the verified timestamp sequence, the standardized spatial coordinate sequence in the unified spatial coordinate system, and the physical quantity value sequence generated in S160 according to the following rules: Construct a spatio-temporal data cube with the device identifier as the primary key and the timestamp sequence as the temporal index. Embed the spatio-temporal metadata generated in S230 into the spatio-temporal data cube as quality description fields. Slice the recombined spatio-temporal data cube according to a preset time window and store it as a standardized spatio-temporal data set in columnar storage format.

4. A method for converting Internet of Things panoramic perception data according to claim 3, characterized in that: The conversion process in S220 includes three levels of processing, namely primary conversion, secondary correction, and tertiary alignment. The original perception data packet stream received in S110 also includes structured device metadata, and the structured device metadata includes device geographical coordinates. In the primary conversion process, based on the device geographical coordinates in the device metadata, convert the local relative coordinates to geodetic coordinates through an affine transformation matrix. In the secondary correction process, load the corresponding terrain compensation parameters according to the device type identifier to compensate for the atmospheric refraction error of the elevation coordinates. In the tertiary alignment process, use the nearest neighbor matching algorithm to bind the geodetic coordinates after primary conversion and secondary correction to a preset spatial grid index to generate a standardized spatial coordinate sequence in the unified spatial coordinate system.

5. The method for converting Internet of Things panoramic perception data according to claim 3, wherein: S300 inputs the standardized spatio-temporal data set into a feature fusion model and generates a panoramic perception feature vector set through a dynamic weight allocation algorithm, including: S310. Based on the device identifier in the standardized spatio-temporal data set, load a feature extractor matching the device type in the feature fusion model. The feature extractor includes a time feature extraction module, a spatial feature extraction module, and a physical quantity feature extraction module. Specifically, it includes: The time feature extraction module performs periodic pattern analysis on the timestamp sequence to generate a time dimension feature vector. The spatial feature extraction module performs spatial clustering analysis on the standardized spatial coordinate sequence in the unified spatial coordinate system to generate a spatial distribution feature vector. The physical quantity feature extraction module performs statistical feature calculation on the physical quantity numerical sequence to generate a physical quantity statistical feature vector; S320, generating an optimized dynamic weight coefficient based on the device type priority factor, the data sampling frequency adjustment coefficient, and the data confidence compensation factor; S330. For each perception data stream, its time dimension feature vector, spatial distribution feature vector, and physical quantity statistical feature vector are spliced by channel to generate a single-device feature matrix. The single-device feature matrix is multiplied by the corresponding optimized dynamic weight coefficient to generate a weighted feature matrix. Then, the weighted feature matrices of all perception data streams are aggregated, and a set of panoramic perception feature vectors is generated through a gated fusion unit.

6. A method for converting Internet of Things panoramic perception data according to claim 5, characterized in that: After S330, the following steps are also included: S340: For each perception data stream, obtain its data quality attenuation factor, and calculate the overall quality confidence score of the panoramic perception feature vector set based on the data quality attenuation factor; If the overall quality confidence score is less than the preset quality threshold, the full set feature reconstruction mechanism is triggered; The full set feature reconstruction mechanism includes: According to the position index of the abnormal trajectory segment, the data slice of the corresponding time period is deleted from the standardized spatiotemporal data set and marked as a missing data point; Completing the missing data points to generate a completed standardized spatiotemporal data set; The newly generated standardized spatiotemporal data set is input into the feature fusion model to generate an updated panoramic perception feature vector set; if the overall quality confidence of the panoramic perception feature vector set still does not meet the standard, the manual review process is triggered.

7. A method for converting Internet of Things panoramic perception data according to claim 5, characterized in that: The S320 specifically includes the following steps: S320.

1. Obtaining a device type priority factor based on a device identifier in the standardized spatiotemporal data set; including: Querying a preset device credibility level table for a device type identifier corresponding to the device identifier; Obtaining a priority weight base value according to the device type identifier; Multiplying the priority weight base value by the device type correction coefficient to generate the device type priority factor; S320.

2. Obtaining a data sampling frequency adjustment coefficient based on the structured device metadata in the standardized spatiotemporal data set; including: Extract the sampling frequency of the current sensed data stream from the structured device metadata ; Compare the sampling frequency with a preset standard frequency for comparison; If , then take as the data sampling frequency adjustment coefficient; If , then take as the data sampling frequency adjustment coefficient; where and are the upper threshold and lower threshold of the preset frequency respectively; S320.

3. Obtaining a data confidence compensation factor based on the structured device metadata in the standardized spatiotemporal data set; including: Extracting a coordinate transformation confidence and a drift error value from the spatiotemporal metadata, calculating a basic quality score based on the coordinate transformation confidence and the drift error value, and then generating a data confidence compensation factor based on the basic quality score; S320.

4. Multiply the device type priority factor, the data sampling frequency adjustment coefficient, and the data quality attenuation factor to generate an original dynamic weight coefficient; S320.

5. Normalize the original dynamic weight coefficients of all perception data streams to generate optimized dynamic weight coefficients.

8. A method for converting Internet of Things panoramic perception data according to claim 1, characterized in that: The step S400 performs adaptive compression coding on the panoramic perception feature vector set to generate a compressed perception data stream, including: S410. Based on the spatio-temporal metadata and feature vectors in the panoramic perception feature vector set, calculate the feature importance grading matrix. The generation process of the feature importance grading matrix includes: Extract the coordinate transformation confidence and drift error value from the spatio-temporal metadata; query the basic weight value in the preset feature sensitivity table according to the feature vector type. The feature vector types include time dimension feature vectors, spatial distribution feature vectors, and physical quantity statistical feature vectors; generate the feature importance score through weighted fusion of the coordinate transformation confidence, drift error value, and feature vector type. S420. Trigger dual-mode compression according to the relationship between the feature importance score and the preset compression threshold: If the feature importance score is greater than or equal to the preset compression threshold, activate the lossless compression mode. If the feature importance score is less than the preset compression threshold, activate the lossy compression mode. S430. Perform streaming encapsulation on the compressed time dimension feature vectors, spatial distribution feature vectors, and physical quantity statistical feature vectors, and aggregate them into a compressed sensing data stream.

9. The method for converting Internet of Things panoramic perception data according to claim 8, wherein: The lossless compression mode in S420 includes: Perform dictionary coding on the time dimension feature vectors, spatial distribution feature vectors, and physical quantity statistical feature vectors using a sliding window-based dictionary coding algorithm, and append a cyclic redundancy check code. The lossy compression mode includes: Perform discrete cosine transform on the physical quantity statistical feature vectors to generate a frequency domain coefficient matrix. Load a quantization table matching the frequency domain coefficient matrix according to the device type identifier, and perform truncation quantization on the high-frequency components.

10. A method for converting Internet of Things panoramic perception data according to claim 1, characterized in that: In S500, input the compressed sensing data stream into a multi-modal decision model to generate real-time anomaly decision instructions and a device control parameter set, including: S510. Load a spatio-temporal decision rule library and a deep learning prediction model in the multi-modal decision model. The deep learning prediction model is an anomaly prediction model based on a spatio-temporal graph convolutional network, and its input is the panoramic perception feature vector decoded from the compressed sensing data stream. S520. Perform multi-level decision fusion, including: The first-level decision: Perform real-time inference on the panoramic perception feature vector based on the spatio-temporal decision rule library to generate a set of rule-triggered events. The second-level decision: Input the panoramic perception feature vector into the spatio-temporal graph convolutional network, and output the anomaly probability prediction value and the potential anomaly type label. The third-level decision: Perform logical fusion on the set of rule-triggered events and the anomaly probability prediction value, and generate an output decision. The output decision includes an immediate action instruction, a delayed action instruction, and a no-action instruction. S530. If the output decision is an immediate action instruction, generate a device control parameter set including the device identifier, control instruction code, and parameter threshold, and send it to the target Internet of Things perception device through the device control interface. If the output decision is a delayed action instruction and a no-action instruction, record the panoramic perception feature vector in the historical decision buffer pool.

Images