AI-based Internet of Things multi-source data anomaly perception analysis method

Abstract

The invention discloses an Internet of Things multi-source data anomaly perception analysis method based on AI, and relates to the field of data anomaly perception analysis, and the main scheme is that the method comprises the steps: carrying out the analysis and calculation of communication quality data, and obtaining a signal-to-noise ratio dynamic entropy value and beam forming distortion degree of a physical layer channel of a communication system; the Doppler contour similarity and the point cloud structure entropy of the base station signal data are obtained by analyzing and calculating the sensing parameter data; calculating a hardware anomaly index of the base station equipment according to the equipment state data; the method comprises the following steps: analyzing and calculating environment interference data to obtain a multipath interference factor of a channel; further analyzing to obtain a comprehensive feature fusion coefficient and a comprehensive abnormal probability; the historical fault data and the comprehensive abnormal probability are analyzed, and the fault type is judged, so that the multi-source data fusion capability of the analysis method is improved, the dynamic adaptability of the evaluation threshold is enhanced, and the diagnosis efficiency is improved.

Description

Technical Field

[0001] The present invention relates to the technical field of data anomaly perception analysis, and specifically to an AI-based method for anomaly perception analysis of multi-source data in the Internet of Things. Background Art

[0002] With the rapid development of Internet of Things technology, the scale of devices and the dimension of data in scenarios such as intelligent industry and smart city have increased exponentially. The real-time anomaly perception of multi-source heterogeneous data has become the core requirement for ensuring system reliability.

[0003] Traditional anomaly detection methods mostly focus on a single data dimension, such as device status or communication quality, and it is difficult to cope with the dynamic coupling effect of multi-source data in complex environments. For example, in the industrial Internet of Things, device hardware anomalies may be superimposed with environmental interference, resulting in communication link fluctuations and perception data distortion. Threshold judgment in a single dimension is prone to false positives or false negatives.

[0004] Therefore, the current technology has the following defects: First, the multi-source data fusion ability is insufficient. Existing methods usually analyze communication quality or device operation parameters independently, lacking the associated modeling of perception data and environmental interference, resulting in one-sided extraction of anomaly features. Second, the dynamic adaptability is poor. Most solutions rely on static thresholds or fixed rules and cannot effectively capture non-linear changes such as channel dynamic entropy and beam distortion, resulting in a high false alarm rate in complex electromagnetic environments or mobile scenarios. Third, the diagnostic logic lags. Existing technologies mostly make isolated decisions based on real-time data, lacking the associated analysis with historical fault libraries, and it is difficult to achieve intelligent inference of fault types and root cause location. Summary of the Invention

[0005] (I) Technical Problems to be Solved Aiming at the deficiencies of the existing technology, the present invention provides an AI-based method for anomaly perception analysis of multi-source data in the Internet of Things, which at least solves the problems of insufficient multi-source data fusion ability, poor dynamic adaptability, and lagging diagnostic logic in the existing technology.

[0006] (II) Technical Solutions To achieve the above objectives, the present invention is realized through the following technical solutions: An AI-based method for anomaly perception analysis of multi-source data in the Internet of Things, including: Step 1: Obtain the communication quality data, perception parameter data, device status data, and environmental interference data of the Internet of Things, and perform preprocessing; Step 2: By analyzing and calculating the communication quality data, obtain the dynamic entropy value of the signal-to-noise ratio of the physical layer channel of the communication system and the beamforming distortion degree; by analyzing and calculating the sensing parameter data, obtain the Doppler profile similarity and the point cloud structure entropy of the base station signal data; by calculating the device status data, obtain the hardware anomaly index of the base station device; by analyzing and calculating the environmental interference data, obtain the multipath interference factor of the channel; Step 3: Through comprehensive analysis of the Doppler profile similarity, the point cloud structure entropy, and the multipath interference factor, obtain the comprehensive feature fusion coefficient; and calculate the comprehensive anomaly probability based on the comprehensive feature fusion coefficient; Step 4: Set the comprehensive evaluation threshold, analyze and evaluate the comprehensive anomaly probability based on the comprehensive evaluation threshold, and then determine whether to initiate the diagnostic instruction; Step 5: By analyzing the historical fault data and the comprehensive anomaly probability, determine the fault type.

[0007] In the preferred solution of the above AI-based multi-source data anomaly perception analysis method for the Internet of Things, by analyzing and calculating the communication quality data, obtain the dynamic entropy value of the signal-to-noise ratio of the subcarrier. The formula is as follows: ; Where, SND t represents the dynamic entropy value of the signal-to-noise ratio of the subcarrier; SN c represents the instantaneous signal-to-noise ratio of the c th subcarrier; c represents the serial number of the subcarrier, N represents the number of subcarriers; SN avg represents the average value of the instantaneous signal-to-noise ratios of all subcarriers; Calculate the beamforming distortion degree. The formula is as follows: ; Where, BFD t represents the beamforming distortion degree; W id,ab represents the theoretical beam weight matrix, W re,ab represents the actual beam weight matrix; a and b respectively represent the number of rows and columns of the matrix, m and n respectively represent the maximum values of the number of rows and columns of the matrix.

[0008] In the preferred solution of the above AI-based multi-source data anomaly perception analysis method for the Internet of Things, by analyzing the sensing parameter data, obtain the Doppler profile similarity. The calculation formula is as follows: ; Among them: DPS t represents the Doppler profile similarity; Dref(v) represents the reference spectrum of the Doppler velocity distribution generated by training historical normal data; Dobs(v) represents the Doppler velocity distribution function of real-time observation; v represents the moving speed of the target object; The formula for calculating the point cloud structure entropy is as follows: ; Among them, PSE t represents the point cloud structure entropy; P k represents the probability density of the three-dimensional point cloud in the spatial bin k, k represents the serial number of the spatial bin, and K represents the total number of spatial bins.

[0009] In the preferred scheme of the above-mentioned AI-based multi-source data anomaly perception and analysis method for the Internet of Things, by analyzing the device status data, the hardware anomaly index is obtained, and the formula is as follows: ; Among them, HAL t represents the hardware anomaly index; T J represents the device chip junction temperature; T th represents the temperature threshold of the chip; T mac represents the maximum temperature allowed for the device chip under normal working conditions; M u represents the memory usage; M to represents the total memory capacity; ɑ represents the weight coefficient of the temperature state; β represents the weight coefficient of the memory state, and ɑ + β = 1.

[0010] In the preferred scheme of the above-mentioned AI-based multi-source data anomaly perception and analysis method for the Internet of Things, by analyzing the environmental interference data, the multipath interference factor is obtained, and the formula is as follows: ; Among them, MIF t represents the multipath interference factor; SY l represents the delay of the l th multipath component; h l represents the channel coefficient of the l th multipath; L represents the number of multipaths.

[0011] In the preferred solution of the above AI-based multi-source data anomaly perception and analysis method for the Internet of Things, the convolutional neural network is used to extract features from the signal-to-noise ratio dynamic entropy value, beamforming distortion degree, Doppler profile similarity, and point cloud structure entropy to obtain the first feature vector; then the recurrent neural network is used to process the hardware anomaly index and multipath interference factor to obtain the second feature vector, and then the two feature vectors are fused to obtain the comprehensive feature fusion coefficient. The formula is as follows: ; Among them, Conv1d represents one-dimensional convolution, which is a convolution operation used to process one-dimensional sequence data; GRU is a type of recurrent neural network RNN, which is used to process time series data; represents the continuous state of the hardware anomaly index and multipath interference factor within a time series segment. t-g:t represents a time series segment, indicating the time period from time t-g to time t, and g represents the size of the time window; ⊕ represents the feature fusion operation, which is used to splice or add two feature vectors element by element to generate a joint feature vector.

[0012] In the preferred solution of the above AI-based multi-source data anomaly perception and analysis method for the Internet of Things, the comprehensive anomaly probability is calculated based on the comprehensive feature fusion coefficient. The formula is as follows: ; Among them, Pan t represents the comprehensive anomaly probability; σ represents the Sigmoid function, which is used to map the output value to the range of 0 to 1; λ r The attention weight of the different-dimensional feature vectors r of the comprehensive feature fusion coefficient is calculated through the softmax function. Refer to the formula: , where wa is the attention weight matrix; r represents the serial number of the different-dimensional feature vectors of the comprehensive feature fusion coefficient, and the values are 1, 2, 3, or 4, corresponding to the communication dimension, perception dimension, device dimension, and environment dimension respectively; MLP r is a frontal neural network, which has layer perceptrons corresponding to different-dimensional feature vectors r and is used to process the information of the corresponding feature dimensions.

[0013] In the preferred solution of the above AI-based multi-source data anomaly perception and analysis method for the Internet of Things, the comprehensive evaluation threshold is calculated based on the comprehensive anomaly probability in the historical time period. The formula is as follows: ; Among them, TH t represents the comprehensive evaluation threshold; μ t-1 represents the moving average of the comprehensive anomaly probability in the previous time period; τ t-1represents the standard deviation of the comprehensive anomaly probability in the previous time period, ρ is a constant, and multiplying by τ t-1 represents the multiple of the standard deviation; The method for analyzing and evaluating the comprehensive anomaly probability based on the comprehensive evaluation threshold to determine whether to initiate a diagnostic instruction is as follows: When Pan t > TH t , initiate the diagnostic process; when Pan t ≤TH t , do not initiate the diagnostic process.

[0014] In the preferred solution of the above AI-based IoT multi-source data anomaly perception analysis method, by analyzing the comprehensive anomaly probability, the posterior probability of different fault causes is obtained, and the formula is as follows: ; Among them, P(X i |Pan t ) represents the posterior probability of the occurrence of the fault cause X t when the observed comprehensive anomaly probability is Pan i ; P(Pan t-z |X i ) represents the probability of the historical comprehensive anomaly probability Pan i observed in the historical fault log when the fault cause X t-z occurs; P(X i ) represents the prior probability of the occurrence of the fault cause X i ; i represents the serial number of the fault cause, and the value is a positive integer.

[0015] In the preferred solution of the above AI-based IoT multi-source data anomaly perception analysis method, the method for judging the fault type is as follows: Traverse the values of P(X i |Pan t ) for different fault causes Xi, and arrange the values of P(X i |Pan t ) in descending order of priority. At this time, the fault cause X i |Pan t ) corresponding to the maximum value is the first-priority fault cause and is used as the fault type determined this time. The priorities of the remaining fault causes decrease in order from large to small according to the values of P(X i |Pan i |Pan t ); If the first-priority fault cause is excluded, then check according to the priorities of the remaining fault causes in sequence.

[0016] (III) Beneficial effects The present invention provides an AI-based method for abnormal perception analysis of multi-source data in the Internet of Things, which has the following beneficial effects: (1) Through preprocessing, communication quality data, sensing parameters, device status, and environmental interference data are mapped to a unified numerical interval, eliminating the interference of dimensional differences on the fusion model, enabling the "computability" of multi-source data, laying a foundation for subsequent cross-modal feature fusion, avoiding feature deviation caused by data heterogeneity, and solving the problem in the prior art that due to the lack of a unified preprocessing and alignment mechanism for cross-modal data (communication, sensing, device, environment), it is difficult to eliminate data heterogeneity, thereby causing feature fusion failure or misjudgment.

[0017] (2) Analyze and calculate various types of data to obtain features such as the dynamic entropy value of the signal-to-noise ratio of the physical layer channel of the communication system, the beamforming distortion degree, the Doppler profile similarity and point cloud structure entropy of base station signal data, the hardware abnormality index of base station devices, and the multipath interference factor of the channel. Extracting these targeted features can accurately describe the status and potential abnormal situations in different aspects of the Internet of Things system. The comprehensive application of these features makes the abnormal perception analysis more in-depth and detailed, capable of effectively capturing various potential abnormal situations, and improving the accuracy and sensitivity of abnormal perception.

[0018] (3) Through comprehensive analysis of the Doppler profile similarity, point cloud structure entropy, and multipath interference factor, obtain the comprehensive feature fusion coefficient, and calculate the comprehensive abnormal probability based on this. This comprehensive feature fusion method gives full play to the advantages of each feature, organically integrates information from different aspects, and can more comprehensively reflect the overall operating state of the Internet of Things system. The calculation of the comprehensive abnormal probability quantifies the influence of multiple features into a specific probability value, intuitively representing the likelihood of the system having an abnormality, providing a direct basis for subsequent evaluation and decision-making, helping to achieve early warning and timely handling of abnormal situations in the Internet of Things system, and enhancing the reliability and stability of the system.

[0019] (4) Analyze and evaluate the comprehensive abnormal probability based on the comprehensive evaluation threshold, and then determine whether to initiate a diagnostic instruction. Setting a reasonable comprehensive evaluation threshold can effectively control false alarms and missed alarms, ensuring that the diagnostic instruction is only initiated when the abnormal probability reaches a certain level, avoiding unnecessary resource waste and system interference. In this way, the automatic and intelligent evaluation and decision-making of abnormal situations in the Internet of Things system are realized, improving the operation and maintenance efficiency and management level of the system, being able to respond to potential failure risks in a timely manner, and reducing the probability and impact of failures.

[0020] (5) By analyzing historical fault data and comprehensive anomaly probabilities, the fault type is determined. Utilizing the experience and knowledge in historical fault data and combining with the current comprehensive anomaly probability can more accurately identify the specific fault type. This helps the operation and maintenance personnel to prepare corresponding maintenance plans and resources in advance, quickly locate and solve fault problems, reduce fault repair time and system downtime, and improve the availability and user experience of the Internet of Things system. At the same time, through the statistics and analysis of fault types, it can also provide a strong basis for system optimization and improvement, further enhancing the performance and reliability of the Internet of Things system. BRIEF DESCRIPTION OF THE DRAWINGS

[0021] Figure 1 It is a schematic diagram of the steps of the AI-based multi-source data anomaly perception and analysis method for the Internet of Things of the present invention. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0022] Next, the technical solutions in the embodiments of the present invention will be clearly and completely described in conjunction with the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative efforts shall fall within the protection scope of the present invention.

[0023] Embodiment 1 Please refer to Figure 1 , the present invention provides an AI-based multi-source data anomaly perception and analysis method for the Internet of Things, including: Step 1: Obtain the communication quality data, perception parameter data, device status data, and environmental interference data of the Internet of Things, and perform preprocessing.

[0024] It should be noted that after obtaining the communication quality data, perception parameter data, device status data, and environmental interference data, the obtained data are respectively normalized and preprocessed to eliminate the dimension differences of different data, and then subsequent formula calculations are performed, which can facilitate the fusion calculation of multi-data.

[0025] Step 2: By analyzing and calculating the communication quality data, obtain the signal-to-noise ratio dynamic entropy value and beamforming distortion degree of the physical layer channel of the communication system; by analyzing and calculating the perception parameter data, obtain the Doppler profile similarity and point cloud structure entropy of the base station signal data; calculate the hardware anomaly index of the base station device through the device status data; by analyzing and calculating the environmental interference data, obtain the multipath interference factor of the channel.

[0026] Step 201: The communication quality data includes obtaining the instantaneous signal-to-noise ratio (SNR) of each subcarrier from a network base station or access point. The instantaneous SNR data of all the collected subcarriers is aggregated into a data processing unit, the number of subcarriers is counted, the aggregated instantaneous SNR data is summed up, and then divided by the number of subcarriers to obtain the average value of the instantaneous SNR of all subcarriers. Then, based on the instantaneous SNR of the subcarriers and the average value of the instantaneous SNR of all subcarriers, the dynamic entropy value of the SNR of the subcarriers is calculated through an entropy value calculation model. SND t , and the calculation formula of the value calculation model is as follows: ; Where, SND t represents the dynamic entropy value of the SNR of the subcarriers; SN c represents the instantaneous SNR of the c th subcarrier; c represents the serial number of the subcarrier, taking positive integer values, N represents the number of subcarriers; SN avg represents the average value of the instantaneous SNR of all subcarriers.

[0027] It should be noted that this formula is used to quantify the degree of channel quality fluctuation. The higher the entropy value, the more unstable the channel state, which is conducive to accurately evaluating the state of the channel.

[0028] Obtain the channel state information of each subcarrier from the physical layer channel state information of the Internet of Things base station, and obtain the beam control parameters from the real-time beam control parameters of the base station antenna array. It is obtained through the physical layer channel estimation module. The physical layer channel estimation module is an important part of the communication system and is used to estimate the characteristics of the signal during channel transmission. The purpose of channel estimation is to predict the channel transmission situation for correct processing of the received signal; in the communication system, the spectrum is divided into multiple subcarriers, and each subcarrier can be modulated and demodulated independently.

[0029] Step 202: The communication quality data also includes the theoretical beam weight matrix extracted from the design document or configuration file of the beamforming algorithm; and the actual channel response matrix calculated through channel measurement between the Internet of Things device and the base station, by sending pilot signals and receiving the channel state information CSI feedback, using channel estimation algorithms such as minimum mean square error (MMSE), least squares (LS), etc., as the actual beam weight matrix; by inputting the theoretical beam weight matrix and the actual beam weight matrix into the calculation model of the beamforming distortion degree, the beamforming distortion degree BFD is obtained. t , and the calculation formula of the beamforming distortion degree calculation model is as follows: ; Among them, BFD t represents the beamforming distortion degree; W id,ab represents the theoretical beam weight matrix, W re,ab represents the actual beam weight matrix, which can be read from the base station digital intermediate frequency unit DSP chip; a and b respectively represent the number of rows and columns of the matrix, m and n respectively represent the maximum values of the number of rows and columns of the matrix.

[0030] It should be noted that the numerator part of the formula calculates the square root of the sum of the squares of the differences between the elements of the theoretical and actual beam weight matrices, which reflects the deviation degree between the actual beamforming and the theoretical design; the denominator part is the square root of the sum of the squares of the elements of the theoretical beam weight matrix, which is used to normalize the numerator, making the distortion degree result relative and able to more objectively reflect the distortion situation of beamforming, avoiding comparison deviations caused by different scales of the theoretical weight matrix itself.

[0031] In the Internet of Things communication system, beamforming technology is crucial for improving the directivity and efficiency of signal transmission. However, the actual beamforming effect may deviate from the theoretical design due to various factors, such as hardware deviation, environmental interference, signal reflection, etc., resulting in problems such as a decline in signal transmission quality and unstable communication links. Traditional monitoring methods are difficult to accurately quantify this deviation degree, so they cannot timely and effectively detect abnormal situations in the beamforming link; by calculating the beamforming distortion degree through this formula, the difference between the actual beamforming and the theoretical design can be accurately quantified, providing an intuitive and reliable indicator for monitoring the accuracy and stability of beamforming. During the abnormal perception and analysis process, once the beamforming distortion degree exceeds the normal range, a warning can be issued in a timely manner to prompt the operation and maintenance personnel to pay attention to potential problems of beamforming-related equipment or parameters. This helps to discover abnormal factors that may affect communication quality in advance, take corresponding adjustment and optimization measures, ensure the normal operation of beamforming technology, and then improve the signal transmission quality and link stability of the Internet of Things communication system, reduce problems such as communication interruption or data transmission errors caused by abnormal beamforming, and ensure the efficient and reliable operation of the entire Internet of Things system.

[0032] Step 203: The sensed parameter data includes using a millimeter-wave radar signal processing module to collect a large amount of historical normal data, and through Fourier transform, converting the time-domain signal into a frequency-domain signal to extract the Doppler velocity distribution characteristics, generating a Doppler velocity distribution reference spectrum, and a real-time Doppler velocity distribution function obtained by real-time acquisition of radar echo signals; by inputting the Doppler velocity distribution reference spectrum generated by training historical normal data and the real-time observed Doppler velocity distribution function into the Doppler profile similarity calculation model, the Doppler profile similarity DPS is calculated. t The calculation formula of the Doppler profile similarity calculation model is as follows: ; where: DPS t represents the Doppler profile similarity; Dref(v) represents the Doppler velocity distribution reference spectrum generated by training historical normal data; Dobs(v) represents the real-time observed Doppler velocity distribution function, which can be extracted by the radar signal processing module; v represents the moving speed of the target object.

[0033] It should be noted that both Dref(v) and Dobs(v) are functions representing the Doppler velocity distribution; Dref(v) generates a reference spectrum by collecting and analyzing a large amount of historical normal data (for example, the radar echo signals of normal moving targets), using machine learning or statistical analysis methods, such as obtaining historical data through a millimeter-wave radar signal processing module and using data analysis tools such as Python, MATLAB, etc. for training to generate a reference spectrum; Dobs(v) can collect radar echo signals in real time through the millimeter-wave radar signal processing module and perform signal processing (such as Fourier transform) to extract the Doppler velocity distribution. The Doppler profile similarity is used to detect abnormal behaviors of moving targets, such as sudden acceleration or change of direction. If the DPS value is less than 0.6, it indicates that the currently observed Doppler velocity distribution is quite different from the reference spectrum, and an abnormality may occur.

[0034] This formula calculates the Doppler profile similarity by comparing the difference between the Doppler velocity distribution reference spectrum Dref(v) generated by training historical normal data and the real-time observed Doppler velocity distribution function Dobs(v). Specifically, the numerator part of the formula calculates the sum of the squares of the differences between the reference spectrum and the observed spectrum, reflecting the degree of difference between the two; the denominator part calculates the sum of the squares of the sum of the reference spectrum and the observed spectrum, which is used to normalize the numerator so that the similarity result falls between 0 and 1. The closer the similarity is to 1, the more similar the real-time observed Doppler profile is to the historical normal data, and vice versa, the greater the difference.

[0035] In the Internet of Things communication system, the echo signal of the base station is affected by various factors, such as environmental changes, equipment aging, signal interference, etc., resulting in changes in the Doppler velocity distribution. Traditional monitoring methods are difficult to effectively quantify the difference between the real-time Doppler profile and the normal situation, and cannot detect signal abnormalities in a timely manner, thus affecting the accurate judgment of the base station signal status; calculating the Doppler profile similarity through this formula can accurately quantify the similarity between the real-time observed Doppler velocity distribution and the historical normal data, providing an intuitive and reliable indicator for monitoring the stability of the base station signal.

[0036] Step 204: The sensed parameter data also includes the point cloud data of the target scene collected by the three-dimensional sensor, and preprocess the collected point cloud data; divide the three-dimensional space into multiple equal-sized spatial bins, calculate the number of three-dimensional point clouds in each spatial bin by calculating the index of the point cloud coordinates in the bin coordinate system, and then calculate the probability density of the three-dimensional point cloud in each spatial bin; input the probability density in the spatial bin into the point cloud structure entropy calculation model to obtain the point cloud structure entropy PSE t , and the calculation formula of the point cloud structure entropy calculation model is as follows: ; where PSE t represents the point cloud structure entropy; P k represents the probability density of the three-dimensional point cloud in the spatial bin k, k represents the serial number of the spatial bin, and the value is a positive integer. K represents the total number of spatial bins, which is determined by the maximum detection distance of the lidar.

[0037] This formula obtains the point cloud structure entropy by calculating the probability density of the three-dimensional point cloud in different spatial bins. Specifically, first divide the maximum detection space of the lidar into multiple bins, then count the frequency of the point cloud appearance in each bin to obtain the probability density pk. Finally, use the formula of information entropy to calculate the weighted sum of the probability densities of all bins to obtain the point cloud structure entropy. The point cloud structure entropy reflects the complexity and uncertainty of the spatial distribution of the point cloud data. The higher the entropy value, the more complex and disordered the spatial distribution of the point cloud data.

[0038] In the Internet of Things (IoT) sensing environment, lidar point cloud data is vulnerable to environmental factors such as weather changes and obstacle movements, as well as device states such as changes in scanning accuracy and wear of mechanical components. These factors can cause changes in the spatial distribution of point cloud data. Traditional point cloud data processing methods are difficult to effectively quantify such changes in spatial distribution and cannot detect anomalies in point cloud data in a timely manner, thus affecting the accuracy of the perception of the surrounding environment. By calculating the point cloud structure entropy using this formula, the complexity and changes in the spatial distribution of point cloud data can be accurately quantified, providing an intuitive and reliable indicator for monitoring the stability and consistency of point cloud data. This ensures the quality and reliability of point cloud data, thereby improving the environmental perception ability and decision-making accuracy of the IoT sensing system, reducing problems such as perception errors or system misjudgments caused by abnormal point cloud data, and ensuring the efficient and stable operation of the entire IoT system.

[0039] Step 205: Analyze the device status data to obtain the hardware anomaly index, based on the following formula: ; where HAL t represents the hardware anomaly index; T J represents the device chip junction temperature, which is collected by a temperature sensor; T th represents the temperature threshold of the chip, defined according to the chip's specification sheet, usually 85 degrees Celsius; T mac represents the maximum temperature allowed for the device chip under normal operating conditions, which can refer to the parameters defined in the device specification sheet to ensure that the chip operates within a safe temperature range; M u represents the memory usage, which can be read from the device driver; M to represents the total memory capacity. This is a parameter in the device hardware configuration used to evaluate the memory usage; ɑ represents the weight coefficient of the temperature state; β represents the weight coefficient of the memory state, and ɑ + β = 1. Here, ɑ can take a value of 0.7 and β can take a value of 0.3.

[0040] It should be noted that the device chip here can be the core processing chip or key functional chip in the IoT device, such as one of the central processing unit (CPU) chip, graphics processing unit (GPU) chip, baseband chip, or application-specific integrated circuit (ASIC) chip, or select the chip with the highest chip junction temperature as the device chip junction temperature.

[0041] This formula calculates the hardware anomaly index by comprehensively considering two key factors: the junction temperature of the device chip and the memory usage. Specifically, first, calculate the difference between the junction temperature of the device chip and the chip temperature threshold, and then divide it by the difference between the maximum allowable temperature of the chip and the temperature threshold to obtain a temperature-related normalized index, which reflects the degree to which the temperature exceeds the normal range. At the same time, calculate the ratio of the memory usage to the total memory capacity to obtain a memory-usage-related index, which reflects the tightness of the memory usage. Then, assign corresponding weight coefficients to these two indexes respectively, and finally add the two to get the hardware anomaly index. The setting of the weight coefficients can be adjusted according to the actual situation.

[0042] During the operation of Internet of Things devices, the anomalies in the hardware state may be caused by various factors, such as overheating of the chip, excessive memory usage, etc. Traditional hardware monitoring methods often only focus on a single index, making it difficult to comprehensively and accurately reflect the overall operating state of the hardware, and prone to false negatives or false positives. In addition, the hardware parameters and operating conditions of different devices vary greatly, lacking a general and effective method for quantifying hardware anomalies. By calculating the hardware anomaly index through this formula, it is possible to comprehensively consider two key factors, chip temperature and memory usage, to comprehensively and accurately quantify the degree of hardware anomalies, providing an intuitive and reliable index for monitoring the hardware state of Internet of Things devices.

[0043] Step 206: Analyze the environmental interference data to obtain the multipath interference factor, and the formula is as follows: ; where MIF t represents the multipath interference factor; SY l represents the delay of the l th multipath component, which can be obtained by analyzing the channel impulse response CIR; h l represents the channel coefficient of the l th multipath, which can be extracted from the CIR data; L represents the number of multipaths, which can be determined by the peak detection algorithm of the CIR.

[0044] This formula calculates the multipath interference factor by analyzing the channel impulse response data. Specifically, first, obtain the delay and channel coefficient of each multipath component. The delay reflects the time difference of the signal transmission on different paths, and the channel coefficient reflects the attenuation and phase change of the signal on different paths. Then, calculate the product of the delay of each multipath component and the square of the modulus value of its channel coefficient, and sum up the product values of all multipath components. Finally, divide by the number of multipaths to obtain the multipath interference factor. This factor quantifies the degree of signal interference caused by the multipath effect. The larger the factor value, the more serious the multipath interference.

[0045] In Internet of Things (IoT) communication, when signals propagate in complex environments such as those with buildings and terrains in cities where signals can be reflected, scattered, and refracted, multipath effects occur. This leads to multiple signal copies from different paths in the received signal, and these copies interfere with each other, causing problems such as signal distortion and increased bit error rate. Existing methods are difficult to accurately quantify the degree of multipath interference, thus unable to effectively evaluate and improve communication quality. By calculating the multipath interference factor using this formula, the interference degree of multipath effects on signals can be accurately quantified, providing an intuitive and specific indicator for evaluating channel quality and communication reliability.

[0046] Step 3: Through comprehensive analysis of the Doppler profile similarity, point cloud structure entropy, and multipath interference factor, obtain the comprehensive feature fusion coefficient; and calculate the comprehensive anomaly probability based on the comprehensive feature fusion coefficient.

[0047] Step 301: Use a convolutional neural network to extract features from the signal-to-noise ratio dynamic entropy value, beamforming distortion degree, Doppler profile similarity, and point cloud structure entropy to obtain the first feature vector; then use a recurrent neural network to process the hardware anomaly index and multipath interference factor to obtain the second feature vector, and then fuse the two feature vectors to obtain the comprehensive feature fusion coefficient. The formula is as follows: ; where, F t represents the comprehensive feature fusion coefficient; Conv1d represents one-dimensional convolution, which is a convolution operation used to process one-dimensional sequence data; it extracts local features by sliding a convolution kernel over the input data; represents the joint distribution of the signal-to-noise ratio dynamic entropy value, beamforming distortion degree, Doppler profile similarity, and point cloud structure entropy at time t; GRU is a type of recurrent neural network (RNN) used to process time series data. represents the continuous state of the hardware anomaly index and multipath interference factor from time t - g to time t within a time series segment, where g is the size of the time window used to capture the change trends of HAI and MIF over a period of time; ⊕ represents the feature fusion operation, which is used to splice or add two feature vectors element by element to generate a joint feature vector.

[0048] This formula processes different types of data through a convolutional neural network and a recurrent neural network respectively, and then fuses the two obtained feature vectors to obtain a comprehensive feature fusion coefficient. Specifically, first, one-dimensional convolution is used to extract features from the joint distribution over time of four features, namely the dynamic entropy value of the signal-to-noise ratio, the beamforming distortion degree, the Doppler profile similarity, and the point cloud structure entropy, to obtain the first feature vector and capture the local features of these features. At the same time, a recurrent neural network is used to process the continuous states of the hardware anomaly index and the multipath interference factor from time t - g to time t, to obtain the second feature vector and capture the changing trends of these two features in the time series. Finally, through a feature fusion operation, the two feature vectors are concatenated or added element-wise to generate a comprehensive feature fusion coefficient, realizing the deep fusion of multi-source heterogeneous data and making full use of the information in each feature.

[0049] In the abnormal perception analysis of multi-source data in the Internet of Things, different types of feature data (such as the dynamic entropy value of the signal-to-noise ratio, the beamforming distortion degree, the Doppler profile similarity, the point cloud structure entropy, the hardware anomaly index, and the multipath interference factor) have different characteristics and time dimensions. Traditional feature processing methods are difficult to process static features and time series features simultaneously, unable to make full use of the information in each feature, resulting in incomplete and inaccurate feature expressions and affecting the performance of abnormal perception. By calculating the comprehensive feature fusion coefficient using this formula, it is possible to effectively fuse the feature information from different data sources, including the local features of the dynamic entropy value of the signal-to-noise ratio, the beamforming distortion degree, the Doppler profile similarity, and the point cloud structure entropy extracted by the convolutional neural network, as well as the time series features of the hardware anomaly index and the multipath interference factor processed by the recurrent neural network. This fusion method gives full play to the advantages of the convolutional neural network in processing local features and the recurrent neural network in processing time series data, enabling the obtained comprehensive feature fusion coefficient to more comprehensively and accurately reflect the operating state of the Internet of Things system, providing a more powerful feature expression for subsequent abnormal perception analysis, improving the accuracy and timeliness of abnormal perception, reducing the missed reports and false alarms of abnormal situations, enhancing the performance and reliability of the abnormal perception analysis method of the entire Internet of Things system, and enabling it to more effectively handle abnormal situations in a complex Internet of Things environment.

[0050] It should be noted that the number of convolutional kernels of Conv1D can be set to 2, corresponding to the sensing dimension and the communication dimension; select an appropriate convolutional kernel size, such as kernel_size = 3, to extract local features; output dimension: the output of Conv1D is a two-dimensional matrix with a shape of (T, 2), where T is the length of the input sequence and 2 is the number of convolutional kernels. The two dimensions correspond to the sensing dimension such as DPS and PSE, and the communication dimension such as SNDE and BFD; the number of hidden units of GRU is set to 2, corresponding to the device dimension and the environment dimension; the input of GRU is ; Output Dimension: The output of the GRU is a two-dimensional matrix with a shape of (T, 2), where T is the length of the time series and 2 is the number of hidden units. The two dimensions correspond to the device dimension such as HAI and the environmental dimension such as MIF; The input to the feature fusion (⊕) is the output of Conv1D and GRU, and its output dimension is the concatenated F t The dimension of is 2 + 2 = 4, and each dimension corresponds to the communication dimension, the perception dimension, the device dimension, and the environmental dimension.

[0051] Step 302: Calculate the comprehensive anomaly probability based on the comprehensive feature fusion coefficient, and the formula is as follows: ; Among them, Pan t represents the comprehensive anomaly probability; σ represents the Sigmoid function, which can be expressed as , used to map the output value to the interval from 0 to 1, λ r represents the attention weight of the different-dimensional feature vectors r of the comprehensive feature fusion coefficient, which is used to measure the importance of each-dimensional feature vector in anomaly detection and can be calculated by the softmax function. Refer to the formula: , where, wa is the attention weight matrix, which is used to learn the importance of each feature dimension. Usually, a random initialization method such as Gaussian distribution or Xavier initialization is used and learned through the known backpropagation method and optimizer such as Adam; r represents the serial number of the different-dimensional feature vectors of the comprehensive feature fusion coefficient, and the values are 1, 2, 3, or 4, corresponding to the communication dimension, the perception dimension, the device dimension, and the environmental dimension respectively; MLP r is a frontal neural network with four groups of layer perceptrons corresponding to different-dimensional feature vectors r, which is used to process the information of the corresponding feature dimension and can extract the contribution of each-dimensional feature to the anomaly probability.

[0052] This formula calculates the comprehensive anomaly probability through a multi-layer perceptron (MLP) and an attention mechanism. Specifically, first, the multi-layer perceptron is used to process the four feature dimensions of the comprehensive feature fusion coefficient respectively to extract the contribution of each-dimensional feature to the anomaly probability. Then, the importance of each feature dimension in anomaly detection is measured by the attention weight, and the features of different dimensions are weighted. Finally, the weighted features are mapped to the interval from 0 to 1 through the Sigmoid function to obtain the comprehensive anomaly probability. The attention weight is calculated by the softmax function, which reflects the relative importance of each feature dimension in anomaly detection under the current feature fusion coefficient, enabling the model to automatically learn which features are more critical in anomaly detection. The multi-layer perceptron is used to perform a non-linear transformation on each feature dimension to extract deeper feature representations In the abnormal perception analysis of multi-source data in the Internet of Things, different features contribute differently to abnormalities, and there are complex non-linear relationships between features. Traditional methods are difficult to effectively measure the importance of each feature and cannot fully explore the deep associations between features, resulting in inaccurate calculation of abnormal probabilities and inability to meet the requirements of precise anomaly detection. By calculating the comprehensive abnormal probability through this formula, different contributions of different features to anomaly detection can be fully considered, and the attention mechanism is used to dynamically adjust the weights of each feature, enabling the model to pay more attention to features related to abnormalities. At the same time, the non-linear transformation ability of the multi-layer perceptron can deeply explore the complex relationships between features and extract deeper feature representations. This method based on the attention mechanism and multi-layer perceptron effectively improves the accuracy of comprehensive abnormal probability calculation, provides a more reliable basis for subsequent anomaly assessment, enhances the performance of abnormal perception analysis, enables the system to more accurately identify abnormal situations, reduces false alarms and missed alarms, improves the reliability and stability of the Internet of Things system, and ensures timely discovery and handling of potential abnormal problems.

[0053] Step Four: Set the comprehensive evaluation threshold, analyze and evaluate the comprehensive abnormal probability based on the comprehensive evaluation threshold, and then determine whether to initiate the diagnostic instruction.

[0054] Step 401: Calculate the comprehensive evaluation threshold according to the comprehensive abnormal probability in the historical time period, and the formula is as follows: ; where, TH t represents the comprehensive evaluation threshold; μ t-1 represents the moving average of the comprehensive abnormal probability in the previous time period; τ t-1 represents the standard deviation of the comprehensive abnormal probability in the previous time period, ρ is a constant, and multiplying by τ t-1 represents a multiple of the standard deviation, which can be adjusted according to actual needs. Here, it can take values from 1 to 3.

[0055] This formula calculates the comprehensive evaluation threshold through the comprehensive abnormal probability in the historical time period. Specifically, first calculate the moving average of the comprehensive abnormal probability in the previous time period to reflect the average level of historical abnormal probabilities; at the same time, calculate the standard deviation of the comprehensive abnormal probability in the previous time period to reflect the fluctuation of historical abnormal probabilities. Then, multiply the standard deviation by a constant and add it to the moving average to obtain the comprehensive evaluation threshold. The constant ρ can be adjusted according to actual needs and usually takes values from 1 to 3, which is used to control the looseness of the threshold. The larger the ρ value, the looser the threshold, and vice versa.

[0056] In the abnormal perception analysis of multi-source data in the Internet of Things, there is a lack of an adaptive evaluation threshold setting method. Traditional methods often use fixed thresholds, which are difficult to adapt to the dynamic characteristics of the abnormal probability changing over time, and are prone to false alarms or missed alarms. The abnormal probability distributions in different scenarios vary greatly, and fixed thresholds cannot meet the diverse requirements. By calculating the comprehensive evaluation threshold using this formula, the threshold can be dynamically adjusted according to the average level and fluctuation of the historical abnormal probability, enabling it to adaptively reflect the current abnormal probability distribution characteristics. During the abnormal evaluation process, once the comprehensive abnormal probability exceeds this dynamic threshold, an alarm can be issued in a timely manner, effectively improving the accuracy and timeliness of abnormal evaluation, and reducing the situations of false alarms and missed alarms. This method of dynamically adjusting the threshold enhances the flexibility and adaptability of abnormal perception analysis, enabling the system to better cope with different scenarios and the dynamically changing Internet of Things environment, improving the performance and reliability of the entire abnormal perception system, and providing strong support for the timely discovery and handling of abnormalities.

[0057] Step: 402: The method for analyzing and evaluating the comprehensive abnormal probability based on the comprehensive evaluation threshold to determine whether to initiate a diagnostic instruction is as follows: When Pan t >TH t , initiate the diagnostic process. When Pan t ≤TH t , do not initiate the diagnostic process.

[0058] This step determines whether to initiate the diagnostic process by comparing the comprehensive abnormal probability and the comprehensive evaluation threshold. Specifically, when the comprehensive abnormal probability is greater than the comprehensive evaluation threshold, it is considered that there may be an abnormality in the system, and the diagnostic process is initiated; conversely, when the comprehensive abnormal probability is less than or equal to the comprehensive evaluation threshold, it is considered that the system is operating normally, and the diagnostic process is not initiated. This threshold-based judgment method can effectively respond to and handle abnormal situations quickly.

[0059] In the abnormal perception analysis of multi-source data in the Internet of Things, there is a lack of an effective abnormal evaluation and decision-making mechanism, making it difficult to accurately judge when to initiate the diagnostic process. Traditional methods often rely on fixed thresholds or manual experience, unable to adapt to the dynamic changes of the abnormal probability, and are prone to false alarms or missed alarms; through the abnormal evaluation method of this step, it can automatically and accurately determine whether to initiate the diagnostic process based on the comparison result of the comprehensive abnormal probability and the comprehensive evaluation threshold. This effectively improves the accuracy and timeliness of abnormal evaluation, reduces the situations of false alarms and missed alarms, ensures that the system can take measures in a timely manner when an abnormality occurs, enhances the reliability and practicality of the entire abnormal perception analysis system, and provides strong guarantee for the stable operation of the Internet of Things system.

[0060] Step Five: Judge the type of fault by analyzing the historical fault data and the comprehensive abnormal probability.

[0061] Step 501: By analyzing the comprehensive anomaly probability, obtain the posterior probability of different failure causes, based on the following formula: ; where, P(X i |Pan t ) represents the posterior probability of the occurrence of failure cause X t when the comprehensive anomaly probability Pan i is observed; P(Pan t-z |X i ) represents the probability of the historical comprehensive anomaly probability Pan i observed in the historical failure log when the failure cause X t-z occurs; P(X i ) represents the prior probability of the occurrence of failure cause X i ; i represents the serial number of the failure cause, taking positive integer values.

[0062] It should be noted that P(Pan t-z |X i ) can collect a large number of failure records from the historical failure log or historical failure case library, including the failure cause X i and the corresponding historical comprehensive anomaly probability Pan t-z ; the failure cause X i can include AAU hardware failure, CU software error, core network congestion, etc., and the historical comprehensive anomaly probability Pan t-z is the comprehensive anomaly probability calculated by the calculation method of Step 1 to Step 3 in the historical situation, and the calculation method is P(Pan t-z |X i ) = the number of times Pan i is observed under the failure cause X t-z / the total number of occurrences of the failure cause X i . P(X i ) can collect a large number of failure records from the historical failure log or historical failure case library, and the calculation method is: P(X i ) = the number of occurrences of the failure cause X i / the total number of failures.

[0063] In the abnormal perception analysis of multi-source data in the Internet of Things, traditional methods are difficult to accurately identify the specific failure causes leading to anomalies, cannot effectively utilize historical failure data to assist in judgment, and are prone to problems such as inaccurate fault diagnosis and low efficiency. When multiple failure causes exist simultaneously, it is even more difficult for traditional methods to determine the most likely root cause of the failure; Through this step, the posterior probability of different fault causes under the currently observed comprehensive anomaly probability can be accurately calculated, so as to quickly and accurately locate the most likely fault cause. This not only improves the accuracy of fault diagnosis, but also greatly enhances the diagnosis efficiency and reduces the time and workload of manual fault troubleshooting. For example, in a complex Internet of Things system, if multiple abnormal situations such as decreased communication quality, device hardware anomalies, and increased environmental interference occur simultaneously, this method can comprehensively consider these abnormal manifestations, combine historical fault data, and accurately determine which fault cause or causes lead to the current abnormal situation, enabling the operation and maintenance personnel to quickly take targeted maintenance measures, promptly restore the normal operation of the system, reduce the impact of faults on system performance and business operations, and improve the availability and reliability of the entire Internet of Things system.

[0064] Step 502: The method for judging the fault type is as follows: Traverse the values of P(X i |Pan t ) for different fault causes Xi, arrange the values of P(X i |Pan t ) in descending order of priority. At this time, the fault cause X i |Pan t ) corresponding to the maximum value of P(X i |Pan i |Pan t ) is the first-priority fault cause and is used as the fault type determined this time. The priorities of the remaining fault causes decrease successively from large to small according to the values of P(X i |Pan t ).

[0065] If the first-priority fault cause is excluded, then the remaining fault causes are checked in the order of their priorities.

[0066] The above embodiments can be implemented in whole or in part by software, hardware, firmware, or any other combination. When implemented using software, the above embodiments can be implemented in whole or in part in the form of a computer program product. Those of ordinary skill in the art can realize that the units and algorithm steps of each example described in combination with the embodiments disclosed herein can be implemented by electronic hardware or a combination of computer software and electronic hardware. Whether these functions are executed in a hardware or software manner depends on the specific application and design constraints of the technical solution.

[0067] The units described as separate components may or may not be physically separated, and the components shown as units may or may not be physical units, that is, they may be located in one place or distributed to multiple network units. Some or all of the units can be selected according to actual needs to achieve the purpose of the solution of this embodiment.

[0068] As described above, it is only the specific implementation manner of the present application, but the protection scope of the present application is not limited thereto. Any person skilled in the art within the technical scope disclosed by the present application can easily think of changes or substitutions, which should all be covered within the protection scope of the present application.

Claims

1. An AI-based method for abnormal perception and analysis of multi-source data in the Internet of Things, characterized in that, Including: Step 1: Obtain the communication quality data, perception parameter data, device status data, and environmental interference data of the Internet of Things, and perform preprocessing; Step 2: Through the analysis and calculation of the communication quality data, obtain the signal-to-noise ratio dynamic entropy value and beamforming distortion degree of the physical layer channel of the communication system; through the analysis and calculation of the perception parameter data, obtain the Doppler profile similarity and point cloud structure entropy of the base station signal data; calculate the hardware anomaly index of the base station device through the device status data; through the analysis and calculation of the environmental interference data, obtain the multipath interference factor of the channel; Step 3: Through the comprehensive analysis of the Doppler profile similarity, point cloud structure entropy, and multipath interference factor, obtain the comprehensive feature fusion coefficient; and calculate the comprehensive anomaly probability based on the comprehensive feature fusion coefficient; Step 4: Set the comprehensive evaluation threshold, analyze and evaluate the comprehensive anomaly probability based on the comprehensive evaluation threshold, and then determine whether to start the diagnostic instruction; Step 5: Judge the fault type by analyzing the historical fault data and the comprehensive anomaly probability.

2. The AI-based abnormal perception and analysis method for multi-source data in the Internet of Things according to claim 1, wherein The communication quality data includes obtaining the instantaneous signal-to-noise ratio of each subcarrier from a network base station or access point, aggregating the instantaneous signal-to-noise ratio data of all the collected subcarriers into a data processing unit, counting the number of subcarriers, summing up the aggregated instantaneous signal-to-noise ratio data, and then dividing by the number of subcarriers to obtain the average value of the instantaneous signal-to-noise ratios of all subcarriers; then, based on the instantaneous signal-to-noise ratio of the subcarriers and the average value of the instantaneous signal-to-noise ratios of all subcarriers, the dynamic entropy value of the signal-to-noise ratio of the subcarriers is calculated through an entropy value calculation model SND t ; The communication quality data also includes a theoretical beam weight matrix extracted from the design document or configuration file of the beamforming algorithm; and an actual channel response matrix calculated by using a channel estimation algorithm through channel measurement between the IoT device and the base station, sending pilot signals and receiving feedback channel state information, which is used as the actual beam weight matrix; by inputting the theoretical beam weight matrix and the actual beam weight matrix into the calculation model of the beamforming distortion degree, the beamforming distortion degree BFD is obtained t 。 3. The AI-based IoT multi-source data anomaly perception and analysis method according to claim 1, wherein The perception parameter data includes a Doppler velocity distribution reference spectrum generated by using a millimeter-wave radar signal processing module to collect a large amount of historical normal data, converting the time-domain signal into a frequency-domain signal through Fourier transform to extract the Doppler velocity distribution characteristics, and a real-time Doppler velocity distribution function obtained by collecting radar echo signals in real time; by inputting the Doppler velocity distribution reference spectrum generated by training historical normal data and the real-time observed Doppler velocity distribution function into a Doppler profile similarity calculation model, the Doppler profile similarity DPS is calculated t ; The sensed parameter data further includes point cloud data of the target scene collected by a three-dimensional sensor, and preprocesses the collected point cloud data; divides the three-dimensional space into multiple spatial bins of equal size, calculates the index of the point cloud coordinates in the bin coordinate system to count the number of three-dimensional point clouds in each spatial bin, and further calculates the probability density of the three-dimensional point cloud in each spatial bin; inputs the probability density in the spatial bin into the point cloud structure entropy calculation model to obtain the point cloud structure entropy PSE t 。 4. The AI-based abnormal perception analysis method for multi-source data in the Internet of Things according to claim 1, characterized in that, By analyzing the device status data, the hardware anomaly index is obtained, and the basis formula is as follows: ； Among them, HAL t represents the hardware exception index; T J represents the junction temperature of the device chip; T th represents the temperature threshold of the chip; T mac represents the maximum temperature allowed for the device chip under normal operating conditions; M u represents the memory usage; M to represents the total memory capacity; α represents the weight coefficient of the temperature state; β represents the weight coefficient of the memory state, and α + β = 1.

5. The AI-based abnormal perception analysis method for multi-source data in the Internet of Things according to claim 1, wherein By analyzing the environmental interference data, the multipath interference factor is obtained, and the basis formula is as follows: ； Among them, MIF t represents the multipath interference factor; SY l represents the time delay of the l -th multipath component; h l represents the channel coefficient of the l -th multipath; L represents the number of multipaths.

6. The AI-based IoT multi-source data abnormal perception and analysis method according to claim 5, wherein In Step 3, through the convolutional neural network, feature extraction is performed on the signal-to-noise ratio dynamic entropy value, beamforming distortion degree, Doppler profile similarity, and point cloud structure entropy to obtain the first feature vector; then, through the recurrent neural network, the hardware anomaly index and multipath interference factor are processed to obtain the second feature vector, and then the two feature vectors are fused to obtain the comprehensive feature fusion coefficient, and the basis formula is as follows: ； Among them, F t represents the comprehensive feature fusion coefficient; Conv1d represents one-dimensional convolution, which is a convolution operation used to process one-dimensional sequence data; GRU is a type of recurrent neural network RNN and is used to process time series data; represents the continuous state of the hardware anomaly index and the multipath interference factor within a time series segment. t - g: t represents a time series segment, indicating the time period from time t - g to time t, and g represents the size of the time window; ⊕ represents the feature fusion operation, which is used to splice or element-wise add two feature vectors to generate a joint feature vector.

7. The AI-based abnormal perception and analysis method for multi-source data in the Internet of Things according to claim 6, characterized in that, Calculate the comprehensive anomaly probability based on the comprehensive feature fusion coefficient, and the basis formula is as follows: ； Among them, Pan t represents the comprehensive anomaly probability; σ represents the Sigmoid function, which is used to map the output value to the interval from 0 to 1; λ r The attention weight of the different-dimensional feature vectors r of the comprehensive feature fusion coefficient is calculated by the softmax function, referring to the formula: , where wa is the attention weight matrix; r represents the serial number of the different-dimensional feature vectors of the comprehensive feature fusion coefficient, and the values are 1, 2, 3, or 4, corresponding to the communication dimension, perception dimension, device dimension, and environment dimension respectively; MLP r is a front neural network, which has layer perceptrons corresponding to different-dimensional feature vectors r and is used to process information corresponding to the feature dimensions.

8. The AI-based abnormal perception and analysis method for multi-source data in the Internet of Things according to claim 7, characterized in that Calculate the comprehensive evaluation threshold according to the comprehensive anomaly probability of the historical time period, and the basis formula is as follows: ； Among them, TH t represents the comprehensive evaluation threshold; μ t-1 represents the moving average of the comprehensive anomaly probability in the previous time period; τ t-1 represents the standard deviation of the comprehensive anomaly probability in the previous time period. ρ is a constant, and multiplying it by τ t-1 represents the multiple of the standard deviation; The method for analyzing and evaluating the comprehensive anomaly probability based on the comprehensive evaluation threshold and determining whether to start the diagnostic instruction is as follows: When Pan t > TH t , start the diagnostic process. When Pan t ≤ TH t , do not start the diagnostic process.

9. The AI-based abnormal perception analysis method for multi-source data in the Internet of Things according to claim 7, characterized in that, By analyzing the comprehensive anomaly probability, the posterior probability of different fault causes is obtained, and the basis formula is as follows: ； Among them, P(X i |Pan t ) represents the posterior probability of the occurrence of the fault cause X t when the observed comprehensive anomaly probability Pan i occurs; P(Pan t-z |X i ) represents the probability of the historical comprehensive anomaly probability Pan i observed in the historical fault log when the fault cause X t-z occurs; P(X i ) represents the prior probability of the occurrence of the fault cause X i ; i represents the serial number of the fault cause, and the value is a positive integer.

10. The AI-based abnormal perception analysis method for multi-source data in the Internet of Things according to claim 9, wherein The method for judging the fault type is as follows: traverse the values of P(X i |Pan t ) under different fault causes Xi, arrange the values of P(X i |Pan t ) in descending order of priority. At this time, the fault cause X i |Pan t corresponding to the maximum value of P(X i |Pan i |Pan t ) is the fault cause of the first priority and is used as the fault type for this judgment. The priorities of the remaining fault causes decrease in turn from large to small according to the values of P(X i |Pan t ); If the first-priority fault cause is excluded, then the remaining fault causes are checked in order of their priorities.

Images