Real-time multi-source data fusion processing method and system for mobile intelligent device

Abstract

The application provides a real-time fusion processing method and system of multi-source data of a movable intelligent device, and relates to the technical field of data processing.The method comprises the following steps: in step 1, a multi-dimensional original data stream is collected in real time through a heterogeneous sensor array integrated by the movable intelligent device, and a time-space synchronization mechanism is applied to align the data streams of different sensors to generate a time-space consistent original data set; in step 2, a dynamic interpolation compensation operation is performed on the original data set, and a dynamic calibration framework is constructed based on the internal topological relationship of the data stream to form a dynamic perception domain; an evolution sequence is generated according to the data unit evolution behavior of the domain boundary, a space correction value is generated through the offset characteristics of the evolution sequence and a preset reference, and a preprocessed data of the fusion space correction value is generated in combination with real-time data correlation analysis.The application realizes real-time and efficient fusion processing of multi-source data of the movable intelligent device by triggering dynamic adjustment and updating the fusion parameters through a sliding window according to abnormal detection.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of data processing, in particular to a multi-source data real-time fusion processing method and system of a movable intelligent device. BACKGROUND

[0002] In smart homes, there are some defects in real-time fusion processing of multi-source data collected by movable intelligent devices in traditional technologies. Movable intelligent devices in smart homes are usually equipped with multiple sensors, such as infrared sensors, cameras, temperature and humidity sensors, etc. For example, when a smart security inspection robot is inspecting indoors, an infrared sensor detects human heat source information, and a camera captures a human image at the same time. In the fusion of these two types of data, traditional technologies may not be able to quickly establish an effective association due to differences in data format and feature dimension. This compatibility problem makes it difficult for multi-source data to fully play a synergistic role, affecting the accuracy of scene state judgment.

[0003] In addition, traditional technologies have deficiencies in real-time data processing and dynamic adjustment. Movable intelligent devices need to adjust the operation strategy in a timely manner according to real-time data. During the cleaning process of a smart sweeper robot, an ultrasonic sensor detects low furniture in front, and a laser radar environment map shows that there is a certain space for detouring in this area. Traditional data fusion processing technology may not be able to efficiently fuse and calculate these two types of real-time data in a short time, resulting in a robot that is stuck or makes a wrong judgment when adjusting the travel route, either colliding with furniture or detouring unnecessary distances, reducing cleaning efficiency. This lack of real-time problem makes it difficult for movable intelligent devices to flexibly adapt to dynamically changing home environments. SUMMARY

[0004] The technical problem to be solved by the present application is to provide a multi-source data real-time fusion processing method and system of a movable intelligent device, which improves the fusion processing of multi-source heterogeneous data through time and space synchronization calibration, dynamic reliability evaluation and adaptive weighted fusion, and edge layer processing.

[0005] To solve the above technical problems, the technical solution of the present application is as follows:

[0006] In a first aspect, a multi-source data real-time fusion processing method of a movable intelligent device, the method comprising:

[0007] Step 1: Real-time collection of multi-dimensional raw data streams by a heterogeneous sensor array integrated by a movable intelligent device, and alignment of data streams of different sensors by applying a time and space synchronization mechanism to generate a time and space consistent raw data set;

[0008] Step 2, performing a dynamic interpolation compensation operation on the original data set, and constructing a dynamic calibration framework based on the inherent topological relationship of the data stream to form a dynamic perception domain; generating an evolution sequence according to the evolution behavior of the data units of the domain boundary, generating a spatial correction value through the offset characteristics of the evolution sequence and the preset reference, and generating preprocessed data fused with the spatial correction value in combination with real-time data correlation analysis;

[0009] Step 3, performing multi-dimensional feature extraction on the preprocessed data, real-time evaluating the confidence quality of each data source, and generating a feature set with real-time confidence evaluation;

[0010] Step 4, based on the feature set with real-time confidence evaluation, dynamically allocating the fusion weight of each data source through an adaptive weighting strategy, and combining a layered processing mechanism of an edge computing architecture to perform real-time fusion calculation, and generating a low-latency fusion result;

[0011] Step 5, sending the low-latency fusion result to an edge node via a lightweight transmission protocol, executing a real-time decision algorithm at the node to generate a control instruction and feeding back to a movable intelligent device;

[0012] Step 6, based on the generated full-process data, dynamically adjusting the fusion strategy parameters through real-time anomaly detection and device computing power adaptation, and configuring computing resources to key function data processing links to realize real-time fusion processing of multi-source data.

[0013] The second aspect is a multi-source data real-time fusion processing system of a movable intelligent device, comprising:

[0014] A data acquisition module is configured to collect multi-dimensional original data streams in real time through a heterogeneous sensor array integrated in the movable intelligent device, and align the data streams of different sensors by applying a space-time synchronization mechanism to generate a space-time consistent original data set;

[0015] A dynamic processing module is configured to perform a dynamic interpolation compensation operation on the original data set, and construct a dynamic calibration framework based on the inherent topological relationship of the data stream to form a dynamic perception domain, generate an evolution sequence according to the evolution behavior of the data units of the domain boundary, generate a spatial correction value through the offset characteristics of the evolution sequence and the preset reference, and generate preprocessed data fused with the spatial correction value in combination with real-time data correlation analysis;

[0016] A feature evaluation module is configured to perform multi-dimensional feature extraction on the preprocessed data, real-time evaluate the confidence quality of each data source, and generate a feature set with real-time confidence evaluation;

[0017] A fusion calculation module is configured to generate a low-latency fusion result based on the feature set with real-time confidence evaluation, dynamically allocate the fusion weight of each data source through an adaptive weighting strategy, and combine a layered processing mechanism of an edge computing architecture to perform real-time fusion calculation.

[0018] an instruction feedback module, configured to send the low-delay fusion result to the edge node via a lightweight transmission protocol, and generate control instructions by the node executing a real-time decision algorithm and feed back to the movable intelligent device;

[0019] a resource scheduling module, configured to dynamically adjust fusion strategy parameters and configure computing resources to key function data processing links based on the generated whole-process data, through real-time anomaly detection and device computing power adaptation, to realize efficient fusion processing of multi-source data.

[0020] In a third aspect, a computing device includes:

[0021] one or more processors;

[0022] a storage device configured to store one or more programs, when the one or more programs are executed by the one or more processors, so that the one or more processors implement the method.

[0023] In a fourth aspect, a computer readable storage medium stores a program, which is executed by a processor to implement the method.

[0024] The above scheme of the present application at least has the following beneficial effects:

[0025] By aligning heterogeneous sensor data through a space-time synchronization mechanism, filling in data missing areas through dynamic interpolation compensation operations, and then calibrating data offset through spatial correction values, the errors and discontinuity in the original data collection process are reduced, the preprocessed data is more in line with the actual scene, the low delay and real-time of data fusion are realized, the layered processing mechanism of the edge computing architecture is adopted, the fusion weight is dynamically allocated combined with the adaptive weighting strategy, the calculation speed is accelerated while ensuring the fusion accuracy; the application of lightweight transmission protocol reduces the data transmission time, ensures that the fusion result can be quickly converted into control instructions and fed back to the device, meets the core demand of real-time response of mobile intelligent devices; by real-time evaluation of the confidence quality of each data source, dynamic confidence labels are given to different sensor data, so that the fusion process can rely on high-quality data, reducing the influence of single sensor failure or data anomaly on the overall result, at the same time, based on real-time anomaly detection and computing power adaptation mechanism of the whole process data, the fusion strategy parameters can be dynamically adjusted, and stable operation can still be maintained when the device state changes in complex environment; by dynamically adjusting the fusion strategy parameters, the calculation resources are accurately allocated to the key function data processing link, avoiding resource waste and improving the utilization efficiency of device computing power; the layered processing mechanism of edge computing architecture reduces data transmission and redundant calculation, reduces device energy consumption, and prolongs the endurance time of mobile intelligent devices; multi-dimensional data generated by different types of sensors in complex dynamic environment is processed, the dynamic perception domain is constructed and evolved through sequence analysis, the perception of the device to environmental changes is more delicate and accurate, and the combination of real-time decision algorithm and control instruction feedback mechanism enables the device to quickly respond to various situations in the environment, improving the reliability and practicality of autonomous navigation, intelligent inspection, dynamic environment interaction and the like. BRIEF DESCRIPTION OF DRAWINGS

[0026] Figure 1 FIG. 1 is a flowchart of a multi-source data real-time fusion processing method of a mobile intelligent device provided by an embodiment of the present application.

[0027] Figure 2 FIG. 2 is a schematic diagram of a multi-source data real-time fusion processing system of a mobile intelligent device provided by an embodiment of the present application. DETAILED DESCRIPTION

[0028] Exemplary embodiments of the present disclosure will be described more fully hereinafter with reference to the accompanying drawings; however, they are not limited to the embodiments set forth herein but can be implemented in various forms. The embodiments are provided so that this disclosure will be thorough, and will fully convey the scope of the disclosure to those skilled in the art.

[0029] As Figure 1As shown, the embodiment of the application proposes a real-time multi-source data fusion processing method of a mobile intelligent device, which comprises the following steps:

[0030] Step 1, through the integrated heterogeneous sensor array of the mobile intelligent device, real-time acquisition of multi-dimensional original data stream is performed, and a time-space synchronization mechanism is applied to align the data streams of different sensors to generate a time-space consistent original data set;

[0031] Step 2, a dynamic interpolation compensation operation is performed on the original data set, and a dynamic calibration framework is constructed based on the internal topological relationship of the data stream to form a dynamic perception domain; an evolution sequence is generated according to the evolution behavior of the data units of the domain boundary, a space correction value is generated through the offset characteristics of the evolution sequence and the preset reference, and a preprocessed data of the fusion space correction value is generated in combination with the real-time data correlation analysis;

[0032] Step 3, multi-dimensional feature extraction is performed on the preprocessed data, the confidence quality of each data source is evaluated in real time, and a feature set with real-time confidence evaluation is generated;

[0033] Step 4, based on the feature set with real-time confidence evaluation, the fusion weight of each data source is dynamically allocated through an adaptive weighting strategy, and real-time fusion calculation is performed in combination with the hierarchical processing mechanism of the edge computing architecture to generate a low-delay fusion result;

[0034] Step 5, the low-delay fusion result is sent to the edge node via a lightweight transmission protocol, and a control instruction is generated by executing a real-time decision algorithm at the node and fed back to the mobile intelligent device;

[0035] Step 6, based on the generated whole-process data, the fusion strategy parameters are dynamically adjusted through real-time anomaly detection and device computing power adaptation, and computing resources are configured to the key function data processing link to realize real-time fusion processing of multi-source data.

[0036] In the embodiment of the application, the time-space synchronization mechanism is used to align the heterogeneous sensor data stream, which solves the time misalignment problem caused by different sampling rates and transmission delays, and the space coordinate inconsistency problem caused by the difference in sensor installation position, improves the time-space matching accuracy of multi-source data, and the dynamic interpolation compensation operation can intelligently fill the data missing or abnormal area, in combination with the dynamic calibration framework based on the topological relationship, the internal structure characteristics of the data stream can be captured in real time, the space correction value is generated by analyzing the evolution behavior of the data units of the domain boundary, the precise correction of the data offset in the dynamic environment is realized, and the integrity and reliability of the data are enhanced; the multi-dimensional feature extraction and real-time confidence evaluation mechanism can comprehensively analyze the feature performance of each data source in the time domain, frequency domain and statistical domain, and quantify the data quality through indicators such as entropy change rate and frequency spectrum energy distribution, this refined evaluation system can dynamically identify and reduce the influence of low-quality data, and improve the accuracy of the overall data fusion.

[0037] The adaptive weighting strategy dynamically allocates fusion weights according to data confidence, combines the hierarchical processing mechanism of the edge computing architecture, and reasonably distributes the computing tasks in the terminal and the edge node. This cooperative processing mode reduces the data transmission overhead and computing redundancy, realizes the millisecond-level fusion response speed, and meets the real-time decision-making needs of the mobile intelligent device. The lightweight transmission protocol ensures efficient transmission of the fusion results to the edge node. The real-time decision algorithm quickly generates control instructions based on environmental characteristic parameters, and guarantees reliable execution of the instructions through a retransmission mechanism with conflict detection, so that the device can quickly respond in a complex environment, improving the stability of core applications such as autonomous navigation and target tracking. The real-time anomaly detection mechanism based on the whole process data can timely detect data fluctuations or system failures, and dynamically adjust the fusion strategy parameters and computing power allocation.

[0038] In a preferred embodiment of the present application, step 1, the above-mentioned step 1, through the integrated heterogeneous sensor array of the mobile intelligent device, real-time acquisition of multi-dimensional original data stream, and application of space-time synchronization mechanism to align the data stream of different sensors, generate space-time consistent original data set, which can include:

[0039] In the embodiment of the present application, in the data acquisition stage, after the mobile intelligent device is started, the integrated heterogeneous sensor array is first activated, including but not limited to laser radar, camera, inertial measurement unit, ultrasonic sensor, etc. Each sensor continuously acquires data according to its own preset sampling frequency, for example, the laser radar acquires environmental point cloud data at a frequency of 10 times per second, the camera captures image data at a rate of 30 frames per second, and the inertial measurement unit records acceleration and angular velocity data at a frequency of 100 times per second. The system allocates independent cache space for each sensor data stream, stores original data frames in real time, and adds a millisecond-level acquisition timestamp to each data frame. The timestamp is generated based on the high-precision clock built-in the device, ensuring the uniformity of time recording.

[0040] In the time synchronization process, first, the timestamps of all data frames in each sensor data stream are extracted, and a comparison table containing sensor ID, data frame number, and timestamp is established. By analyzing the timestamp distribution of different sensors near the same time in the comparison table, the average time offset of each sensor is calculated, which is the average value of the difference between the timestamp of a data frame of a sensor and the reference clock of the device. For sensors with large differences in sampling frequency, a sliding window method is used for time alignment. The time axis of the sensor with the highest sampling frequency is taken as the reference, and the data streams of other sensors are divided into time windows. In each window, data points are supplemented by linear interpolation, so that different sensors have corresponding data records at the same time node. At the same time, the time synchronization error is monitored in real time. The deviation of the current timestamp of each sensor from the reference time is calculated every 100 milliseconds. If the deviation exceeds 5 milliseconds, the time offset is recalculated and the interpolation parameters are adjusted to ensure that the time synchronization accuracy is maintained within milliseconds.

[0041] The space alignment operation is based on the physical installation parameters of the sensors, including the three-dimensional coordinates (X, Y, Z axis positions) and installation attitude angles (pitch angle, roll angle, yaw angle) of each sensor in the device coordinate system. According to these parameters, a space conversion mechanism is constructed. For the original data collected by each sensor, first, the installation position offset is eliminated by coordinate translation operation. The data point coordinates in the local coordinate system of the sensor are added to the installation position coordinates, and then converted to the origin reference of the device coordinate system. Then, through the rotation matrix operation corresponding to the attitude angle, the directional deviation caused by the difference in installation angle is corrected, so that the data direction of all sensors is unified to point to the motion coordinate system of the device. For visual sensors, distortion correction is also needed through calibration parameters to convert image pixel coordinates to actual physical space coordinates, ensuring consistent spatial scale with other sensors.

[0042] The time and space alignment verification link adopts a double verification mechanism. In the time dimension, 10 groups of multi-sensor data frames at different times are randomly selected, and the standard deviation of the timestamp difference is calculated. If the standard deviation is less than 2 milliseconds, the time synchronization is considered qualified. In the space dimension, fixed reference points around the device are selected, and the coordinate measurement results of the same reference point by different sensors are compared. The root mean square error of the coordinate deviation is calculated. If the error is less than a preset threshold (such as 5 centimeters), the space alignment is considered qualified. If either dimension fails the verification, the time offset calculation window size or the space conversion matrix parameters are adjusted, and the alignment operation is re-executed until the verification passes. The multi-sensor data that passes the verification is reordered according to the timestamp order and integrated into a unified data structure, which includes the original data, spatial coordinate information, and data integrity markers of each sensor at each time node. At the same time, a time and space alignment report is generated, recording the time offset, space conversion parameters, and alignment error values of each sensor. Finally, a raw data set with consistent time and space attributes is formed.

[0043] In a preferred embodiment of the present application, the step 2 performs a dynamic interpolation compensation operation on the original data set, and forms a dynamic perception domain based on the internal topological relationship of the data stream, which can include:

[0044] In step 220, manifold learning is performed on the original data set consistent in space and time to extract the spatial neighborhood characteristics and time continuity features of the data units, and a sparse correlation matrix including quantized correlation strength is generated;

[0045] In step 221, based on the sparse correlation matrix, the gradient modulus value of each row element is calculated, the mutation position where the gradient modulus value exceeds the set threshold is detected, and the spatial division of the internal stable area and the external change area is formed with the mutation position as the boundary point;

[0046] In step 222, in the external change area, interpolation calculation based on intensity weighting is performed on the data missing area according to the corresponding connection strength value in the sparse correlation matrix to generate a compensated continuous data stream;

[0047] In step 223, based on the spatial distribution characteristics of the continuous data stream, a dynamic perception domain of the structured mapping framework is formed.

[0048] In the embodiment of the present application, all data units are selected from the original data set consistent in space and time, each data unit includes corresponding spatial coordinates, time stamp and feature parameters. In the manifold learning analysis, the feature parameters of each data unit are first converted into high-dimensional vectors, the cosine similarity between any two data unit vectors is calculated, the closer the similarity value is to 1, the closer the correlation between the two in the feature space, and the spatial neighborhood characteristics are extracted. Taking the three-dimensional spatial coordinates of the data unit as the reference, the Euclidean distance between each two data units is calculated, and the distance value less than the preset spatial threshold is considered as a potential spatial adjacent unit. For the time continuity feature, the continuous data units of the same monitoring area are arranged in ascending order according to the time stamp, and the absolute difference value of the feature parameters of adjacent data units is calculated. The smaller the difference value, the better the time continuity. The correlation strength is calculated by combining the spatial distance and the time difference value. The spatial distance is normalized and multiplied by the spatial weight, and the time difference value is normalized and multiplied by the time weight. The sum of the two is the initial correlation strength. The initial correlation strength is then corrected by the cosine similarity, and finally the quantized correlation strength value between 0 and 1 is obtained. The larger the value, the more significant the correlation. When constructing the matrix, the rows and columns correspond to different data unit numbers respectively, and the matrix cells are filled with the correlation strength values of the corresponding two data units. The cells with correlation strength values lower than the minimum effective threshold are directly assigned a value of 0 to form a sparse correlation matrix that only retains strong correlation relationships. The proportion of non-zero elements in the matrix does not exceed the preset proportion of the total number of elements.

[0049] For each row of the sparse association matrix, the positions and corresponding association strength values of all non-zero elements are extracted in the order of column index, the association strength difference values of adjacent column non-zero elements are calculated, the square values of each difference value are accumulated and then square rooted to obtain the gradient modulus value of the row elements, which reflects the overall change amplitude of the association strength of the data units in the row; the gradient modulus value of each row is compared with a preset mutation detection threshold value, when the gradient modulus value of a row is greater than the threshold value, the column position with the most drastic gradient change in the row is further located, the local gradient value of the adjacent column elements in the row is calculated, the local gradient value is the absolute value of the association strength difference between the adjacent two columns, and the column position with the maximum local gradient value is the mutation position; the mutation positions of all rows are collected, these positions are marked in the spatial coordinate system, and adjacent mutation positions are connected to form a closed boundary line; the region inside the boundary line, whose contained data unit corresponding matrix row gradient modulus value is lower than the threshold value and the association strength between data units changes smoothly, is delimited as the internal stable region; the region outside the boundary line, which has rows with gradient modulus values exceeding the threshold value and data unit association strength fluctuation, is delimited as the external change region.

[0050] First, all data units in the external change region are scanned, and by comparing the time stamp interval of the complete data sequence and the continuity of the characteristic parameters, the missing region with data value being empty or characteristic parameter jump exceeding the allowed range is identified, the start and end time stamps and the spatial coordinate range of the missing region are recorded, the five nearest data units (spatially closest and time stamp adjacent) around the missing region are extracted from the sparse association matrix, the connection strength values of these data units and the center of the missing region are obtained, the connection strength values are normalized by dividing each connection strength value by the sum of all connection strength values to obtain the weight coefficient of each reference data unit, the sum of the weight coefficients is 1, the effective data values of each reference data unit in the corresponding time period of the missing region are collected, each data value is multiplied by the corresponding weight coefficient and then summed to obtain the interpolated data value at the first time point of the missing region, and the interpolated data at each time point of the missing region is calculated in sequence according to the time interval; for a missing region with a large spatial range, multiple sub-regions are divided according to the spatial grid, and the above weighted interpolation calculation is performed on each sub-region respectively, and all interpolated data values are filled into the missing region to form a time-continuous and spatially-complete data stream, and the smoothness error between the interpolated data and the surrounding effective data is calculated to ensure that the error value is lower than the preset standard.

[0051] The compensated continuous data stream is meshed according to spatial coordinates, each mesh unit contains all data values of the spatial position in a set time window, the data mean, variance and distribution density of each mesh unit are calculated, the mean reflects the typical data level of the region, the variance reflects the data fluctuation degree, and the distribution density reflects the data collection density, the mesh units are clustered according to the similarity of the data mean, the mesh units with a mean difference less than a set threshold are classified into the same feature region, and the core data features (such as mean range and main change trend) of each feature region are marked; the adjacent relationship and data interaction frequency between different feature regions are counted, the higher the interaction frequency, the stronger the correlation between the regions, the correlation strength between the regions is represented by a weighted connection line, the feature region is taken as a basic unit, and a structured mapping framework including region boundary coordinates, core feature parameters and inter-region correlation relationship is constructed, each element in the framework corresponds to a specific spatial region and its data characteristics, with the continuous input of new data, the core parameters and correlation relationship of each feature region are updated in real time, and when the region data features change by more than an adjustment threshold, the region boundary and clustering group are dynamically adjusted, thereby forming a dynamic perception domain that can adapt to data changes in real time.

[0052] The spatial adjacency and time continuity of the data units are captured through manifold learning, the sparse correlation matrix generated by quantifying the correlation strength is filtered to remove irrelevant data interference, only meaningful correlation relationships are retained, and the internal relationship between the data units is clearer, the gradient modulus is used to detect the mutation position and divide the region, the stable and changeable regions of the data distribution can be automatically identified, the limitation of adopting a unified processing strategy for the overall data is avoided, the internal stable region can be processed in a lightweight manner to save computing power, the external changeable region is intensively processed, the accurate allocation of data processing resources is realized, the overall processing efficiency is improved, the strength weighted interpolation calculation is used for the missing data in the external changeable region, the weight of the reference data is determined according to the correlation strength, the interpolation result is more in line with the actual data distribution law, compared with simple linear interpolation, the error caused by data missing is reduced, and the dynamic perception domain can reflect the spatial distribution change of the data in real time. The structured mapping framework clearly presents the data features and correlation relationships of each region, the perception of the complex environment is more comprehensive and accurate, and the dynamic adjustment mechanism ensures that the perception domain can adapt to data changes in time, and the adaptability to the dynamic environment is improved.

[0053] In a preferred embodiment of the present application, the step 2 of generating an evolution sequence according to the data unit evolution behavior of the domain boundary, generating a spatial correction value according to the deviation characteristics of the evolution sequence and the preset reference, and generating preprocessed data fused with the spatial correction value in combination with real-time data correlation analysis can include:

[0054] In step 224, the core perception area data unit is acquired, and a state transition vector in a continuous time window is collected in real time.

[0055] Step 225, based on the state transition vector, time stamp sorting, generating a multi-dimensional time series describing the evolution of data units;

[0056] Step 226, aligning the multi-dimensional time series with the preset reference sequence by dynamic time warping, calculating the deviation of each dimension covariance characteristic;

[0057] Step 227, according to the covariance characteristic deviation, generating a space correction coefficient through a preset conversion rule, injecting the space correction coefficient into the continuous data stream, and simultaneously dynamically adjusting the correction amplitude based on the Pearson correlation characteristics of multi-source sensor data, to generate preprocessed data fused with space correction values.

[0058] In the embodiment of the present application, from the dynamic perception domain, the core perception area is comprehensively screened according to the data distribution density, correlation strength and stability index. First, the ratio of the number of data points to the area of each grid unit is calculated to obtain the data distribution density; then the correlation strength sum of the unit and the adjacent unit is counted; finally, the fluctuation range of the data value in the unit in the past 10 time periods is analyzed, the smaller the fluctuation range, the higher the stability, and the continuous region that meets the conditions of density being higher than 1.5 times the average value, correlation strength sum being in the top 20%, and stability index being in the top 30% is selected as the core perception area.

[0059] The length of the continuous time window is set to 60 seconds, and the time accuracy is 100 milliseconds, that is, the window contains 600 sampling times. For each sampling time, the multi-dimensional feature parameters of each data unit in the core perception area are extracted, including the value size, change direction, change rate, spatial position and distribution density. The difference values of the feature parameters of the same data unit at adjacent two sampling times are calculated, for example, the value change amount is obtained by subtracting the value at the previous time from the value at the current time, and the position offset amount is obtained by subtracting the coordinate at the previous time from the coordinate at the current time. These difference values are arranged into a vector form according to the feature dimensions, such as [value change amount, direction change angle, rate change value, X-axis offset amount, Y-axis offset amount, Z-axis offset amount, density change rate], to form a state transition vector.

[0060] The validity of all generated state transition vectors is verified, and the module length (the square root of the sum of squares of each dimension value) of each vector is calculated. If the module length exceeds the normal range (3 times the historical average module length of this type of data unit), check whether there is a mutation in the feature parameters of the data unit corresponding to the vector at the adjacent time. If there is a mutation and there is no reasonable environmental change explanation, the vector is determined as an abnormal value and is excluded. If the mutation is related to environmental changes (such as suddenly entering a high temperature area), the vector is retained and marked as a special event.

[0061] All state transition vectors that pass the validity verification are collected, the exact time stamp (including millisecond information) when each vector is generated is extracted, a time index table is constructed, the time stamps are arranged in ascending order to form a continuous time axis, for each state transition vector, the corresponding position on the time axis is found according to the time stamp thereof, and each feature dimension value in the vector is split and filled into the corresponding position of the multi-dimensional time sequence, for example, the state transition vector [5, 15°, 2.3, 0.2, -0.1, 0, 0.05] corresponds to a time point t, which is split into a value of 5 of the numerical change dimension at the time t, a value of 15° of the direction change dimension at the time t, and so on. For a missing data time point (i.e., there is no valid state transition vector at the time), a linear interpolation method of adjacent time points before and after is used to fill in, the difference between the values of the corresponding dimensions of the previous valid time point and the next valid time point is calculated, and the difference is distributed to the missing point according to the time interval ratio. All feature dimension time sequences are combined to form a multi-dimensional time sequence matrix, each row of the matrix represents a time change sequence of a feature dimension, and each column represents a multi-dimensional state at a time point. The matrix is smoothed by using a moving average method, for each dimension value at each time point, the average value of the previous and next 5 time points is calculated as the new value of the point, and the influence of accidental noise is eliminated.

[0062] The preset reference sequence matching the current scene in the system is called, the reference sequence is a multi-dimensional time sequence of the same type of data unit collected in an ideal environment, contains the same feature dimensions and has a similar length (allowing a length difference of ±10%), and the multi-dimensional time sequence and the preset reference sequence are aligned by dynamic time warping (DTW). First, a distance matrix is constructed, each element of the matrix represents the feature vector distance between a time point in the multi-dimensional time sequence and a time point in the reference sequence. When calculating the distance, the difference of each dimension is weighted and summed, and the weight is preset according to the importance of the dimension to the whole data (for example, the weight of the numerical change dimension is 0.3, the weight of the position offset dimension is 0.2, and so on).

[0063] The final path is found in the distance matrix by a dynamic programming algorithm, the starting point of the path is the top left corner of the matrix, and the ending point is the bottom right corner, the path can only move right, down or right down, the cumulative distance of each possible path is calculated, and the path with the minimum cumulative distance is selected as the final alignment path, along the final path, the multi-dimensional time series and the time points of the reference sequence are one-to-one corresponding, and sequence alignment is realized; after alignment, the covariance characteristics are calculated according to the feature dimensions respectively, for each dimension, the average value of all time point values of the multi-dimensional time series in the dimension is calculated first, then the average value of all time point values of the reference sequence in the dimension is calculated, then for each time point, the difference between the value of the multi-dimensional time series and the average value thereof, and the difference between the value of the reference sequence and the average value thereof are calculated, the average value of all time points is calculated after the two difference values are multiplied, and the covariance value of the dimension is obtained, the covariance value of the multi-dimensional time sequence is subtracted from the standard covariance value of the reference sequence, and the covariance characteristic deviation of each dimension is obtained.

[0064] For the covariance characteristic deviation of each feature dimension, a preset piecewise linear conversion rule is applied to generate a space correction coefficient, a plurality of deviation threshold intervals are set, for example, [-∞, -5), [-5, -2), [-2, 2), [2, 5), [5, +∞), each interval corresponds to a different correction coefficient generation formula, when the deviation falls in the [-∞, -5) interval, the correction coefficient = deviation × (-0.3); when the deviation falls in the [-5, -2) interval, the correction coefficient = deviation × (-0.2); and so on, so as to ensure that the larger the deviation is, the larger the absolute value of the correction coefficient is, and the correction direction is opposite to the deviation direction, the generated space correction coefficient is associated with the corresponding data unit in the continuous data stream according to the feature dimensions, for each data point in the data stream, according to the time stamp and the spatial position thereof, the corresponding correction coefficient is found, the correction coefficient is superimposed on the original value of the data point, for example, the original value is V, and the corresponding correction coefficient is K, then the corrected value is V+K, and for a multi-dimensional data point, the correction operation is performed on each dimension respectively.

[0065] Meanwhile, the Pearson correlation characteristics of multi-source sensor data are calculated. For the same feature dimension, the Pearson correlation coefficient between the data collected from different sensors is calculated. First, the average values of the two sensor data sequences are calculated. Then, for each time point, the difference between the two sequence values and the respective average values is calculated. The two difference values are multiplied and summed, and then divided by the product of the standard deviations of the two sequences to obtain the correlation coefficient. The closer the correlation coefficient is to 1, the stronger the linear correlation between the two sensor data is. The closer it is to 0, the weaker the correlation is. According to the correlation coefficient, the correction amplitude is dynamically adjusted. A correlation coefficient threshold is set, for example, 0.7. When the correlation coefficient is greater than 0.7, the correction coefficient remains unchanged. When the correlation coefficient is between 0.4 and 0.7, the correction coefficient is multiplied by the correlation coefficient. When the correlation coefficient is less than 0.4, the correction coefficient is multiplied by 0.4. In this way, sufficient correction is given to sensor data with strong correlation, and the correction amplitude is appropriately reduced for sensor data with weak correlation to avoid over-correction.

[0066] The data stream after the injection of the correction coefficient and the adjustment of the amplitude is integrated. The continuity between adjacent data points is checked. If a mutation occurs (the difference between adjacent data points exceeds 5 times the normal fluctuation range), the mutation point is smoothed. The weighted average value of each of the three data points before and after the point is calculated as a new value, and the weight decreases with the distance from the point (for example, the weight of the first three points is 0.1, the weight of the first two points is 0.2, and so on). Finally, the preprocessed data that is time-continuous, spatially consistent, and fused with spatial correction values is generated.

[0067] By focusing on the core perception area and collecting state transition vectors, the data change characteristics of the key area can be accurately captured. The setting of continuous time windows ensures the time sequence integrity of data changes and avoids the interference of irrelevant area data. The state transition vectors are sorted by time to generate multi-dimensional time series, clearly presenting the evolution trajectory of different feature dimensions over time, making the change rule of data units change from disorder to order, facilitating comparison and analysis with the reference sequence. Dynamic time warping alignment solves the difference in time rhythm between the actual sequence and the reference sequence. The calculation of the covariance feature deviation quantity quantifies the deviation degree of the sequence fluctuation characteristics, accurately identifies the abnormality of the data in stability, and provides a clear quantitative basis for correction. By generating spatial correction coefficients through pre-set rules and dynamically adjusting the amplitude, accurate correction of data deviation is achieved, ensuring that the correction direction is consistent with the deviation direction, and avoiding over-correction or insufficient correction by adjusting the amplitude through the Pearson correlation characteristics. The preprocessed data fused with spatial correction values reduces the data deviation caused by environmental interference and sensor errors, improving the accuracy and reliability of the overall data processing.

[0068] In a preferred embodiment of the present application, step 3, multi-dimensional feature extraction is performed on the preprocessed data to evaluate the confidence quality of each data source in real time, and a feature set with real-time confidence evaluation is generated, which can include:

[0069] Step 330, based on the preprocessed data stream, time domain, frequency domain and statistical domain feature dimensions are separated through parallel feature extraction channels to generate a multi-dimensional feature set;

[0070] Step 331, based on the multi-dimensional feature set, the entropy value change rate of each feature dimension is independently calculated, and based on the entropy value stability in the sliding window, a preliminary confidence index of each feature dimension is generated, the spectral energy distribution of each feature dimension is extracted, and the spectral confidence coefficient of each feature dimension is calculated according to the main frequency band energy concentration degree;

[0071] Step 332, based on the preliminary confidence index and the spectral confidence coefficient, the comprehensive confidence weight of each feature dimension is generated by weighted geometric mean fusion;

[0072] Step 333, based on the comprehensive confidence weight, dynamic attenuation is applied to the comprehensive confidence weight of the feature dimension associated with the corresponding data source to generate an attenuated comprehensive confidence weight, and the comprehensive confidence weight is used as metadata to mark the corresponding feature vector to form a feature set with weight marking;

[0073] Step 334, based on the feature set with weight marking, the mutual exclusivity between feature dimensions is analyzed, and when a conflict feature dimension is detected, dynamic rebalancing adjustment is performed on the marked confidence weight to generate a feature set with real-time confidence evaluation.

[0074] In an embodiment of the present application, the preprocessed data stream is classified according to sensor type and data acquisition frequency and distributed to three independent parallel feature extraction channels, each channel processes data simultaneously without interference, ensuring feature extraction efficiency. First, the continuous preprocessed data stream is divided into multiple data segments at a fixed time interval (such as 1 second), each data segment contains all data points in that time period. The mean value of each data segment is calculated by adding all data values in the segment and dividing by the number of data points to obtain the mean value feature reflecting the average level of data in that time period. Then, the maximum value in the data segment is found as the peak value feature, and the minimum value is found as the valley value feature. The difference between the two is the fluctuation range of the data in that segment. When calculating the variance, first calculate the difference between each data value and the mean value, square the sum of these differences, and then divide by the number of data points. The larger the variance, the more dramatic the data fluctuation. At the same time, the numerical difference between adjacent data points is calculated and divided by the corresponding time interval to obtain the change rate feature, which reflects the speed of data change over time.

[0075] The frequency domain feature extraction channel first divides the preprocessed data stream into windows of 5 seconds, and the number of data points in each window is determined according to the sampling frequency (for example, if the sampling frequency is 10 Hz, then there are 50 data points in each window). The data in each window is analyzed periodically, the regularity of repeated data values is counted, the main fluctuation frequency is determined, the amplitude of data fluctuation at each frequency is calculated, the energy value of the frequency is obtained by squaring the amplitude, the total energy of all frequencies is counted, the frequency range with an energy ratio of more than 50% of the total energy is found as the main frequency band, the ratio of the total energy in the main frequency band to the total energy is calculated, and the main frequency band energy ratio is obtained. At the same time, the standard deviation of the energy value of each frequency in the main frequency band is calculated, and then divided by the average energy of the main frequency band to obtain the energy distribution uniformity. The lower the uniformity, the more concentrated the energy is on a few frequencies.

[0076] The statistical domain feature extraction channel performs overall distribution analysis on the preprocessed data stream, sorts all data points by size, and takes the data value at the middle position as the median, which reflects the medium level of the data. The upper quartile value (data value at the 75% position after sorting) and the lower quartile value (data value at the 25% position after sorting) are calculated, and the difference between the two is the interquartile range, which is used to describe the dispersion of the data. When calculating skewness, the cube of the difference between each data value and the mean is calculated, the average of these cubes is obtained to get the third central moment, and then divided by the cube of the standard deviation. Positive skewness indicates that the data distribution is right-skewed, and negative skewness indicates that the data distribution is left-skewed. For kurtosis calculation, the fourth central moment (average of the fourth power of the difference between the data value and the mean) is divided by the fourth power of the standard deviation. Kurtosis greater than 3 indicates that the data distribution is steeper, and kurtosis less than 3 indicates that the data distribution is flatter. Finally, all the features extracted by the three channels are classified and arranged according to the time domain, frequency domain, and statistical domain, forming a multi-dimensional feature set containing multiple feature parameters.

[0077] For each feature dimension in the multi-dimensional feature set, collect the feature values of all data points in that dimension, count the number of occurrences of each feature value, and divide the number of occurrences by the total number of data points to obtain the probability of occurrence of each feature value. The entropy value of the feature dimension is obtained. The higher the entropy value, the more dispersed the feature value distribution. A sliding window with a length of 10 seconds is set, and the window moves every 2 seconds. Each window contains the entropy values of 5 data segments. The standard deviation of the entropy values in the window is calculated. The smaller the standard deviation, the smaller the change in entropy within the window, indicating that the entropy is more stable. The change rate of the entropy value is calculated by subtracting the average entropy value of the previous window from the average entropy value of the current window, and then dividing the result by the time interval (2 seconds) between the two windows, to obtain the change in entropy per unit time.

[0078] When generating the preliminary confidence index, the following standards are used: when the standard deviation of the stability of the entropy value in the sliding window is less than 0.1 and the absolute value of the entropy value change rate is less than 0.05, the preliminary confidence index is 0.9-1.0; when the stability standard deviation is between 0.1-0.3 or the absolute value of the entropy value change rate is between 0.05-0.1, the index is 0.6-0.8; when the stability standard deviation is greater than 0.3 or the absolute value of the entropy value change rate is greater than 0.1, the index is 0.3-0.5. For newly enabled feature dimensions, the preliminary confidence index is temporarily set to 0.7 within the first 30 seconds, and after accumulating sufficient data, the index is calculated according to the above standards. When extracting the spectral energy distribution, for frequency domain feature dimensions, the energy value of each frequency point is recorded, the total energy of the main frequency band (the frequency interval with energy accounting for more than 50% of the total energy) is calculated, and the main frequency band energy concentration is the ratio of the total to the total energy of all frequencies. When calculating the spectral confidence coefficient, if the concentration is greater than 0.7, the coefficient is 0.8-1.0; if the concentration is between 0.5-0.7, the coefficient is 0.5-0.7; if the concentration is less than 0.5, the coefficient is 0.2-0.4. For feature dimensions without a clear main frequency band, calculate the energy fluctuation amplitude (the difference between the maximum and minimum energy values), and if the fluctuation amplitude is less than 10% of the total energy, the coefficient is 0.6-0.8, otherwise it is 0.2-0.5.

[0079] According to the type of feature dimension, the weight distribution proportion of the preliminary confidence index and the spectral confidence coefficient is determined. For time domain feature dimensions, since the entropy value change rate better reflects its time stability, the preliminary confidence index weight is set to 0.6 and the spectral confidence coefficient weight is set to 0.4. In frequency domain feature dimensions, spectral energy distribution is the core characteristic, so the weights of both are 0.5. Statistical domain feature dimensions are greatly influenced by spectral characteristics, so the preliminary confidence index weight is set to 0.4 and the spectral confidence coefficient weight is set to 0.6. When calculating the weighted geometric mean fusion, first take the corresponding weight power of the preliminary confidence index, for example, the time domain feature preliminary confidence index is 0.8 and the weight is 0.6, then calculate the 0.6 power of 0.8, which is about 0.85. Then take the corresponding weight power of the spectral confidence coefficient, for example, the spectral confidence coefficient is 0.7 and the weight is 0.4, then calculate the 0.4 power of 0.7, which is about 0.91. Then multiply the two results, 0.85×0.91≈0.77, to get the comprehensive confidence weight original value of the feature dimension.

[0080] The original values of all feature dimensions are normalized. First, calculate the sum of all original values, then divide each original value by the sum, so that the sum of the normalized comprehensive confidence weights is 1. For example, the original values of three feature dimensions are 0.77, 0.5 and 0.3, the sum is 1.57, and after normalization, they are approximately 0.49, 0.32 and 0.19 respectively. The higher the comprehensive confidence weight value, the stronger the reliability of the feature dimension data in the current environment.

[0081] Query the abnormal record of the data source corresponding to each feature dimension in the past 5 minutes, the abnormal record includes the condition that the feature value exceeds the normal range, the data transmission interruption, and the deviation from other data sources is too large, etc. The number of abnormal occurrences is counted, the abnormal frequency (abnormal times ÷ 5 minutes) is calculated, when the abnormal frequency is 1-3 times per hour (i.e. 0.08-0.25 times in 5 minutes), the attenuation coefficient is set to 0.9; the abnormal frequency is 3-5 times per hour (0.25-0.42 times in 5 minutes), the attenuation coefficient is set to 0.8; the abnormal frequency is more than 5 times per hour (more than 0.42 times in 5 minutes), the attenuation coefficient is set to 0.7; the decay is applied to the comprehensive confidence weight of all feature dimensions associated with the corresponding data source, the decayed weight = normalized comprehensive confidence weight × attenuation coefficient; for example, the normalized weight of a certain feature dimension is 0.49, the abnormal frequency of the corresponding data source is 4 times per hour, and the attenuation coefficient is 0.8, then the decayed weight is 0.49 × 0.8 = 0.39, if there is no abnormal record in the past 5 minutes or the abnormal frequency is less than 1 time per hour, the attenuation coefficient is 1.0, and the weight remains unchanged.

[0082] The decayed comprehensive confidence weight is used as metadata and marked on the corresponding feature vector in the form of key-value pair, such as feature vector {“mean”: 25, “peak”: 30, “variance”: 5}, after marking, it becomes {“mean”: 25, “peak”: 30, “variance”: 5, “confidence_weight”: 0.39}, all feature vectors with weight marking are classified according to data source type and feature dimension, feature vectors of the same data source are placed in the same subset, forming a clear structure of feature set with weight marking, each feature vector in the set clearly carries weight information reflecting its reliability.

[0083] The mutual exclusivity index between the feature dimensions is calculated. For different feature dimensions of the same data source, the change trend of the feature value in the continuous 10 time windows is collected, and the correlation coefficient of the change trend is used to measure the mutual exclusivity. If the correlation coefficient is less than -0.7, it indicates that the change trends of the two are completely opposite, and it is determined as mutually exclusive feature dimensions. For the same type of feature dimensions of different data sources (such as the distance features of two laser radars), the feature value difference of continuous 5 time points is calculated. If the average value of the difference value exceeds 3 times the average value of the normal difference value, and lasts for more than 3 time points, it is determined as a conflict feature dimension. When a conflict feature dimension is detected, dynamic rebalancing adjustment is performed. First, the weight difference of the conflicting parties is calculated. For example, the weight of feature A is 0.39, the weight of feature B is 0.32, and the difference is 0.07. The weight value of the lower weight dimension is reduced by 20% of the difference value, i.e. feature B is reduced by 0.07x20%=0.014, and the adjusted value is 0.32-0.014=0.306. The weight value of the higher weight dimension is increased by 10% of the difference value, i.e. feature A is increased by 0.07x10%=0.007, and the adjusted value is 0.39+0.007=0.397. The sum of the weights of the two after adjustment remains unchanged (0.39+0.32=0.397+0.306≈0.70).

[0084] If the conflict feature dimension comes from different data sources, compare the historical reliability scores of the data sources (number of abnormal record times ÷ total record times), and prefer to retain the feature weight of the data source with high reliability score. The feature weight of the data source with low score is reduced by 10%. After adjustment, all feature vectors and corresponding real-time confidence evaluation weights are integrated to form a feature set with real-time confidence evaluation, ensuring that each feature has a clear reliability identifier.

[0085] The multi-domain features are extracted by parallel channels, each channel focuses on the feature mining of a specific dimension, avoids the mutual interference during the extraction of different types of features, can more accurately capture the change trend of data in the time domain, the energy distribution in the frequency domain and the distribution characteristics in the statistical domain, the comprehensive extraction of the multi-domain features makes the cognition of the data more stereoscopic, and the limitations of a single feature dimension are compensated, the entropy value change rate and the spectral energy concentration degree are calculated independently for each feature dimension, the fine evaluation of the data quality is realized, the setting of the sliding window can track the stability change of the data in real time, the entropy value change rate reflects the fluctuation of the data over time, the spectral energy concentration degree reflects the effectiveness of the frequency domain features, the hierarchical generation of the preliminary confidence index and the spectral confidence coefficient converts the abstract confidence concept into a quantifiable index, and a unified standard is provided for the reliability comparison of different feature dimensions, and the weighted geometric mean fusion method fully considers the characteristic differences of different feature types, dynamically adjusts the weight proportion, so that the fusion result is more in line with the essential properties of various features, the time domain features focus on time stability, the frequency domain features focus on spectral concentration, and the statistical domain features consider multiple aspects, the differentiated fusion strategy improves the accuracy of the comprehensive confidence weight, and the normalization processing ensures the comparability of the weight, so that the feature dimension with high reliability can obtain higher attention in the fusion process.

[0086] In a preferred embodiment of the application, step 4 is based on the feature set with real-time confidence evaluation, dynamically allocates the fusion weight of each data source through an adaptive weighting strategy, and combines the hierarchical processing mechanism of the edge computing architecture to perform real-time fusion calculation and generate a low-delay fusion result, which can include:

[0087] Step 440: According to the comprehensive confidence weight corresponding to the feature set, the data sources are sorted to generate a data source sorting sequence;

[0088] Step 441: Based on the data source sorting sequence, the weight values corresponding to each data source are dynamically allocated, and the feature-level weighted fusion calculation is performed on the perception layer of the edge computing architecture to receive the weight values of each data source and the feature set, and generate a preliminary fusion result;

[0089] Step 442: The preliminary fusion result is transmitted to the decision layer of the edge computing architecture, and the result-level calculation based on the context information and the fusion result is performed in the decision layer to generate a fusion result;

[0090] Step 443: Real-time monitoring of the calculation delay of the fusion result, when the calculation delay exceeds the preset delay threshold, dynamically adjusting the weight value allocation method to obtain a fusion result that meets the preset delay threshold.

[0091] In the embodiment of the present application, from the feature set with real-time confidence evaluation, the comprehensive confidence weight value of each feature vector is extracted, the correspondence table of feature dimension and data source is established, and it is clear that each feature dimension belongs to which data source. For the same data source, the comprehensive confidence weight values of all feature dimensions under it are collected, and the arithmetic mean of these weight values is calculated, that is, all weight values are added and divided by the number of feature dimensions, to obtain the overall reliability score of the data source. For example, data source A contains 3 feature dimensions of distance, speed and direction, and the weight values are 0.82, 0.78 and 0.85 respectively. The overall reliability score is (0.82+0.78+0.85) ÷ 3 = 0.82.

[0092] The overall reliability scores of all data sources are arranged in descending order to form a preliminary data source sorting sequence. If the overall reliability scores of two or more data sources are the same (the difference is within 0.02), further comparison is made on their stability indexes. The stability index is calculated by counting the number of abnormalities of the data source in the past 10 minutes. The fewer the number of abnormalities, the higher the stability index. The calculation formula is stability index = 1-(abnormal number ÷ total data transmission number). The closer the index value is to 1, the better the stability is. The data source with high stability index is arranged in front. If the stability indexes are still the same, the data transmission delay of the data source is compared. The transmission delay is the average time interval from the data source sending data to the edge computing architecture receiving data, which is obtained by taking the average value of 5 consecutive measurements. The data source with smaller delay value is arranged in front in the sorting. After the above multi-layer comparison, the data source sorting sequence arranged in descending order of comprehensive reliability is finally generated. Each position in the sequence corresponds to a data source and its related information.

[0093] According to the data source sorting sequence, an exponential decay method is used to dynamically assign weight values. The basic weight value of the data source ranked first is set to 0.4. For each subsequent position, the weight value is multiplied by the decay coefficient 0.8. For example, the weight of the data source ranked first is 0.4, the weight of the data source ranked second is 0.4 x 0.8 = 0.32, the weight of the data source ranked third is 0.32 x 0.8 = 0.256, the weight of the data source ranked fourth is 0.256 x 0.8 = 0.2048, and so on. After assigning weights to all data sources, the sum of all weight values is calculated. If the sum is not equal to 1, normalization processing is performed, that is, each weight value is divided by the sum, so that the sum of the normalized weight values is 1, ensuring the rationality of the weight assignment.

[0094] The perception layer of the edge computing architecture deploys multiple independent receiving ports, each port corresponding to a data source, and receives feature sets and corresponding weight values sent by each data source in real time. The perception layer multiplies the feature values and weight values of different data sources under the same feature dimension to obtain the weighted contribution value of each data source in that feature dimension. For example, under a certain feature dimension, the feature value of data source A is 50 and the weight is 0.4; the feature value of data source B is 55 and the weight is 0.32, so their weighted contribution values are 50x0.4=20 and 55x0.32=17.6 respectively. The weighted contribution values of all data sources under the same feature dimension are added to obtain the fusion value of that feature dimension. In the above example, the fusion value is 20+17.6=37.6. The above calculation process is repeated for all feature dimensions to generate a feature vector containing the fusion values of all feature dimensions. The feature vector is verified for effectiveness, and the deviation of each fusion value from the standard value under the same conditions at the same time is calculated. If the deviation exceeds the allowed range (±20% of the standard value), the fusion value is marked as suspicious and is linearly interpolated and corrected using the fusion values of the adjacent two time windows. Finally, a stable and reliable preliminary fusion result is generated.

[0095] The preliminary fusion result is transmitted from the perception layer to the decision layer through the high-speed data bus inside the edge computing architecture. During transmission, the data is compressed and verified. Compression uses a lightweight algorithm to reduce data volume, and verification uses additional check codes to ensure data transmission integrity. After receiving the preliminary fusion result, the decision layer retrieves the current context information from the system database, including environmental parameters (such as light intensity, temperature, humidity), task information (such as current task type, priority, progress), and device status (such as power, computing power usage, sensor working status). According to the environmental parameters in the context information, the feature dimension weights are adjusted. For example, when the light intensity is higher than the threshold, the weight of the visual feature dimension is reduced by 15%, and the weight of the infrared feature dimension is increased by 20%. When the temperature is abnormal, the weight of the temperature-related feature dimension is increased by 25%. According to the task priority, the weight of the feature dimension related to the core target of the task is increased, such as the position and direction feature dimension weight in the navigation task, which is increased by 30%; the distance and speed feature dimension weight in the obstacle avoidance task, which is increased by 35%.

[0096] The adjusted weights are multiplied by the feature dimension fusion values in the preliminary fusion result to obtain weighted fusion values. All weighted fusion values are then normalized so that their sum is 1. The decision layer analyzes the normalized weighted fusion values according to the pre-set decision rules (such as if-else logic, threshold judgment). For example, when the weighted fusion value of a certain target feature dimension exceeds 0.6, it is determined that the target exists. The fusion value of the position feature dimension is combined to determine the target position, and finally a fusion result containing target information, environmental assessment, state suggestion, etc. is generated.

[0097] In the decision layer, a high-precision timer is set to record the time interval from receiving the preliminary fusion result to outputting the final fusion result, which is accurate to the millisecond level. This time interval is the current calculation delay. The calculation delay is compared with the preset delay threshold (set according to the application scenario, such as 50 milliseconds for real-time navigation scenarios and 100 milliseconds for environmental monitoring scenarios). If the calculation delay is less than or equal to the threshold, the current weight distribution method is maintained. If the calculation delay exceeds the threshold, the weight adjustment mechanism is triggered. The processing time of each data source is analyzed. The time from data reception to completion of feature-level fusion of each data source is recorded in the perception layer, and the average processing time of each data source is calculated. Data sources with longer processing times (more than twice the total average time) are marked and their priority in the data source sorting sequence is reduced. For example, the top 20% of data sources with the longest processing time are downgraded by 2-3 places in the sorting sequence.

[0098] The weight values are redistributed according to the adjusted sorting sequence. The weights of low-time-consumption data sources at the top of the sorting sequence are increased, and the weights of high-time-consumption data sources at the bottom of the sorting sequence are decreased. For example, the weight of the data source originally ranked first is increased from 0.4 to 0.5, the weight of the data source originally ranked second is increased from 0.32 to 0.35, and the weight of the data source originally ranked third is decreased from 0.256 to 0.15. The sum of the weights is still 1. The new weight distribution method is sent to the perception layer, and the perception layer performs feature-level fusion calculation according to the new weight to generate a new preliminary fusion result and pass it to the decision layer. The decision layer uses the new preliminary fusion result for result-level calculation while continuing to monitor the calculation delay. If the adjusted delay still exceeds the threshold, the sorting adjustment and weight distribution process is repeated to further increase the weight of low-time-consumption data sources (by 5-10% each time) and decrease the weight of high-time-consumption data sources (by 10-15% each time) until the calculation delay meets the preset threshold requirement and a fusion result that meets the real-time requirement is generated.

[0099] The data source quality is quantitatively evaluated and prioritized by comprehensively weighting the confidence of the data source, so that the high-reliability data source can obtain a higher weight in the fusion, and the multi-layer comparison of stability and transmission delay makes the sorting result more comprehensively reflect the comprehensive performance of the data source, avoiding the limitation of single index sorting; the dynamic weight distribution strategy highlights the role of high-priority data sources through exponential decay, while taking into account the contribution of other data sources, realizing the reasonable cooperation between data sources, parallel processing of the edge computing perception layer and feature-level weighted fusion, reducing the redundancy of data transmission and processing, and improving the efficiency of fusion calculation; the effectiveness verification and correction mechanism further ensures the accuracy of the preliminary fusion result, provides high-quality input data for the decision layer calculation, and combines the context information to realize result-level calculation, so that the fusion result can dynamically adapt to environmental changes and task requirements, enhance the scene adaptability of the system, and dynamically adjust the feature weight according to the environmental parameters and task priority, ensure the dominant position of key features in the fusion, and improve the matching degree of the fusion result and the actual application requirement; the application of decision rules converts complex feature data into intuitive decision information, provides clear guidance for real-time operation of the mobile intelligent device, and improves the intelligent level of the device.

[0100] In a preferred embodiment of the present application, the step 5 of transmitting the low-delay fusion result to the edge node via a lightweight transmission protocol, executing a real-time decision algorithm in the node to generate a control instruction and feeding back to the mobile intelligent device can include:

[0101] Step 550, based on the environmental feature parameters in the low-delay fusion result, querying the preset instruction generation rule library, matching the state transition strategy suitable for the current scene;

[0102] Step 551, inputting the state transition strategy into the state transition equation calculation process, combining the current device state parameters to generate the final action sequence of the mobile intelligent device;

[0103] Step 552, analyzing the execution time stamp of each action in the final action sequence, calculating the emergency level of each action; according to the difference between the emergency level and the action time stamp, using a timestamp difference-based encoding algorithm to compress the control instruction data set;

[0104] Step 553, sending the compressed control instruction data packet through wireless communication, and using a retransmission mechanism with collision detection in the transmission layer to ensure that the instruction is reliably fed back to the execution end of the mobile intelligent device.

[0105] In the embodiment of the present application, environmental feature parameters are extracted from the low-delay fusion result, which include but are not limited to the distance, number, type of obstacles around the movable intelligent device, the terrain slope of the area, the road flatness, and the motion direction, speed, etc. of dynamic targets (such as pedestrians, other mobile devices) in the environment, and then the extracted environmental feature parameters are compared one by one with the index conditions in the preset instruction generation rule library, which stores state transition strategy templates corresponding to different environmental feature combinations, for example, when the obstacle distance is less than a preset safety threshold and the number is single, the corresponding state transition strategy is "slow down and avoid"; when the terrain slope is greater than a certain value, the corresponding state transition strategy is "reduce the driving speed and adjust the power output". In the comparison process, the rule item with the highest matching degree with the key parameters (such as obstacle distance, terrain slope) in the extracted environmental feature parameters is first selected, and then the additional parameters (such as obstacle type, dynamic target speed) under the item are verified again to ensure that all parameters meet the requirements of the rule item, and finally the state transition strategy adapted to the current scene is determined.

[0106] First, the current state parameters of the movable intelligent device are collected, including the real-time driving speed, remaining power, current location coordinates, steering angle, power system output power, etc. of the device, and then the state transition strategy determined in step 550 is decomposed into specific constraints and target parameters, for example, the "slow down and avoid" strategy is decomposed into constraints such as "reduce the speed to 50% of the original speed" and "adjust the steering angle to 15 degrees away from the obstacle direction", and then these constraints and target parameters are input into the state transition equation calculation process, and the current device state parameters are also input. In the calculation process, first, according to the speed constraint in the state transition strategy, combined with the current driving speed, the speed difference that needs to be reduced or increased is calculated to determine the speed change per time unit; then, according to the steering constraint and the current steering angle, the adjustment amplitude and required time of the steering angle are calculated; for the power system output power, according to the speed change and the terrain condition (combined with the terrain slope in the environmental feature parameters), the power adjustment value required to maintain the target speed is calculated. Through the calculation of parameters in multiple dimensions such as speed, steering, and power, the action instructions are arranged in chronological order to generate the final action sequence containing "first slow down to X km / h for Y seconds, then turn Z degrees, and adjust the power output to A watts".

[0107] The execution timestamp corresponding to each action in the final action sequence is extracted, the timestamp is accurate to the millisecond level, the specific time point when each action plan starts to execute is recorded, and the urgency level of each action is calculated. The calculation method is to compare the execution timestamp of the action with the current system time to obtain a time difference value. The smaller the time difference value, the faster the action needs to be executed, and the higher the urgency level. At the same time, the importance of the action is combined, such as the action involving obstacle avoidance and emergency parking. Under the same time difference value, the urgency level of the action is one level higher than that of the ordinary steering and acceleration action. For example, if the execution timestamp of an action is 1 second different from the current system time, and the action is an obstacle avoidance action, the urgency level is set to the highest level. If the time difference value of another action is 3 seconds, and the action is an ordinary acceleration action, the urgency level is set to the middle level.

[0108] Next, the difference between the action timestamps, i.e., the difference between the execution timestamps of the adjacent two actions, is calculated to obtain time interval data. When using a timestamp difference-based encoding algorithm for compression processing, the complete timestamp of the first action is retained. For subsequent actions, the complete timestamp is no longer stored, but the difference between the timestamp and the timestamp of the previous action is stored. At the same time, the difference value data is classified and processed according to the urgency level. The timestamp difference value corresponding to the action with a high urgency level is recorded with higher accuracy to ensure the accuracy of the time information. For actions with a low urgency level, the accuracy of the difference value recording can be appropriately reduced to reduce the amount of data. In addition, repeated action instructions in the control instruction data set are identified, such as a plurality of consecutive identical steering angle adjustment instructions. Only the first instruction and the number of repetitions are recorded to further compress the data volume.

[0109] The compressed control instruction data packet is encapsulated according to the format requirements of the wireless communication protocol, and the header information of the data packet is added, including the sender address, the receiver address, the data packet serial number, the data length, etc. After encapsulation, the data packet is sent out through a wireless communication module (such as Wi-Fi, Bluetooth, 5G, etc.). During transmission, the transmission layer monitors the sending state of the data packet in real time. After sending each data packet, the sender starts a timer and waits for the acknowledgment reply signal from the receiver. If the acknowledgment signal returned by the receiver is received within the preset time threshold, it means that the data packet has been successfully received, the timer is cleared, and the next data packet is sent. If no acknowledgment signal is received, it is determined that data collision or loss may have occurred. At this time, the sender first detects the occupation of the current wireless channel. If the channel is idle, the data packet is immediately resent. If the channel is busy, the sender waits for a random time interval and then tries to send again. During the retransmission process, the number of retransmissions is counted. When the number of retransmissions reaches the preset upper limit, the data packet is marked as transmission abnormal, and the transmission state is fed back to the system. At the same time, the subsequent data packets are continuously sent to ensure that the instructions that are not abnormal can be transmitted preferentially, thereby ensuring that the control instructions can be reliably fed back to the execution end of the movable intelligent device.

[0110] By accurately extracting environmental feature parameters and performing multi-dimensional matching with a preset rule library, a state transition strategy suitable for the current scene can be quickly and accurately determined, avoiding decision deviation caused by environmental judgment errors. This rule library-based matching method reduces the decision-making time in complex scenarios, improves the relevance and reliability of the state transition strategy, and generates action sequences by combining current device state parameters and state transition strategies for calculation, making the action sequences meet the requirements of the environmental scene and fully consider the actual situation of the device itself. The analysis of the emergency level of the action and the encoding compression processing based on the timestamp difference can ensure the efficient transmission of emergency action instructions. By compressing the control instruction data set, the data transmission amount is reduced, the bandwidth occupancy rate of wireless communication is reduced, the transmission time is shortened, and the retransmission mechanism with conflict detection solves the problem of data conflict and loss in wireless transmission, improving the reliability of control instruction transmission. Through real-time monitoring and timely retransmission, the situation of device execution end action interruption or error execution caused by instruction loss is avoided. At the same time, the limitation of retransmission times and the abnormal feedback mechanism balance the transmission reliability and transmission efficiency, ensuring that the control instructions can still be stably and timely delivered to the execution end in a complex wireless communication environment, and ensuring the normal operation of the mobile intelligent device.

[0111] In a preferred embodiment of the present application, step 6, based on the generated full-process data, dynamically adjusts the fusion strategy parameters through real-time anomaly detection and device computing power adaptation, and configures computing resources to key function data processing links to realize real-time fusion processing of multi-source data, which can include:

[0112] Step 660, construct an anomaly detection matrix of full-process data, and generate a fusion strategy parameter adjustment instruction when the singular value of the matrix exceeds a preset threshold;

[0113] Step 661, based on the fusion strategy parameter adjustment instruction, obtain the real-time computing power level of the mobile intelligent device, and dynamically reallocate processor core resources to the feature extraction, fusion calculation and decision generation links according to the confidence quality coefficient proportion of each processing link to obtain a core resource reallocation result;

[0114] Step 662, based on the core resource reallocation result, update the sensor compensation coefficient, multi-source data weight and state transition decision rule using a sliding window mechanism to realize real-time fusion processing of multi-source data.

[0115] In the embodiment of the present application, when constructing the anomaly detection matrix of the whole process data, first, collect various types of data in the whole process of multi-source data, including raw data collected by sensors, intermediate data in feature extraction link, result data after fusion calculation, and output data generated by decision, etc., then fill these data into the rows and columns of the matrix in the order of time sequence and data type, forming a two-dimensional matrix containing the whole process data, i.e. the anomaly detection matrix; when singular values of the detection matrix are detected, first, the data standardization processing is performed on the constructed anomaly detection matrix to eliminate the influence of different data types and magnitudes, and then through the related calculation process of singular value decomposition, all singular values are extracted from the matrix, and each singular value extracted is compared with the preset threshold one by one, when it is found that there is a singular value exceeding the preset threshold, the system will automatically trigger the instruction generation mechanism to generate the fusion strategy parameter adjustment instruction.

[0116] When the real-time computing power level of the mobile intelligent device is obtained based on the fusion strategy parameter adjustment instruction, the built-in computing power monitoring mechanism of the instruction triggering device will real-time statistical the number of instructions completed by the processor in unit time, the number of processor cores currently in active state and the usage rate of the processor, etc. Through comprehensive analysis of these data, the current real-time computing power level of the mobile intelligent device is obtained, including the total number of available processor cores, the processing capacity of each core, etc. When allocating processor core resources according to the confidence quality coefficient proportion of each processing link, first, collect the performance data of the three processing links of feature extraction, fusion calculation and decision generation in the data processing process, including the accuracy, stability, error rate, etc. of data processing, based on these data, calculate the confidence quality coefficient of each link, the higher the confidence quality coefficient, the higher the reliability and importance of the link in data processing, then calculate the proportional relationship between the confidence quality coefficients of the three links, and then combine the total available processor core number in the obtained real-time computing power level, and allocate the total core resources to the three links according to the above proportion; for example, if the total available core number is 10, and the confidence quality coefficient proportion of the three links is 3:5:2, then 3 cores are allocated to the feature extraction link, 5 cores are allocated to the fusion calculation link, and 2 cores are allocated to the decision generation link, so as to obtain the core resource reallocation result.

[0117] When updating the sensor compensation coefficient based on the core resource reallocation result using the sliding window mechanism, first, according to the resource proportion of each link after core resource reallocation, the size of the sliding window is determined. The window corresponding to the link with high resource proportion can be appropriately increased to more comprehensively capture data changes. The sliding window slides on the data sequence according to a fixed time interval or data quantity interval. After each sliding of the window, the collected data of the sensor in the window and the actual real data (or calibrated data) are collected. By comparing the differences between the data collected by the sensor in the window and the real data, the average error of each sensor in the window is calculated. According to the size of the average error, the original sensor compensation coefficient is adjusted. The larger the error, the greater the adjustment range of the compensation coefficient. In this way, the update of the sensor compensation coefficient is completed.

[0118] When updating the multi-source data weight, the sliding window mechanism is also relied on. After the window slides, the contribution of the data of different data sources in the window in the fusion processing is collected, including the accuracy, completeness and timeliness of the data. According to these indicators, the performance of each data source is evaluated and scored. The scoring results are combined with the resource proportion of each link in the core resource reallocation to adjust the original multi-source data weight. For the data source that performs well in the window and has a high resource proportion in the corresponding processing link, the weight is appropriately increased. Conversely, the weight is reduced. When updating the state transition decision rule, the sliding window collects the state transition data in the window and the corresponding decision results. The law of state transition in these data and the effectiveness of the decision results are analyzed. Combined with the change of the processing capacity of each link after core resource reallocation, the original state transition decision rule is modified. For example, when the resource of the fusion calculation link increases and the processing capacity improves, more consideration of the details of the fusion calculation can be added to the decision rule to make the state transition decision more accurate, thereby realizing the real-time fusion processing of multi-source data.

[0119] By constructing an abnormality detection matrix of the whole process data and monitoring the singular value, the abnormality existing in the whole process of the multi-source data can be found in time, adjustment instructions are generated when the singular value exceeds a preset threshold, interference of abnormal data on fusion processing can be avoided, the accuracy of multi-source data fusion is preliminarily ensured, and the fusion processing is ensured to be performed on the premise that the data basis is reliable. Based on the adjustment instructions and the real-time computing power level, the processor core resources are dynamically redistributed according to the confidence quality coefficient proportion, so that the resources can be more reasonably utilized, more resources can be allocated to the key links with high confidence quality coefficients, the processing efficiency and quality of the links can be improved, resource waste can be avoided, and the mobile intelligent device can play the final processing performance under the condition of limited computing power. The sliding window mechanism is used to update the sensor compensation coefficient, the multi-source data weight and the state transition decision rule, so that the fusion strategy parameters can be real-time adapted to the data change and the resource allocation situation. The dynamic characteristics of the sliding window ensure the timeliness of the parameter update. The update of the sensor compensation coefficient can reduce the influence of sensor error. The adjustment of the multi-source data weight can highlight the role of high-quality data. The state transition decision rule can improve the decision accuracy, and finally realize the efficient and accurate real-time fusion processing of multi-source data.

[0120] As shown in Figure 2 Embodiments of the present application also provide a multi-source data real-time fusion processing system of a mobile intelligent device, comprising:

[0121] A data acquisition module is configured to acquire multi-dimensional original data streams in real time through a heterogeneous sensor array integrated in the mobile intelligent device, and align the data streams of different sensors by applying a space-time synchronization mechanism to generate a space-time consistent original data set.

[0122] A dynamic processing module is configured to perform a dynamic interpolation compensation operation on the original data set, and construct a dynamic calibration framework based on the internal topological relationship of the data stream to form a dynamic perception domain. An evolution sequence is generated according to the data unit evolution behavior of the domain boundary. A space correction value is generated by the offset characteristics of the evolution sequence and a preset reference, and a preprocessed data of the fusion space correction value is generated in combination with real-time data correlation analysis.

[0123] A feature evaluation module is configured to perform multi-dimensional feature extraction on the preprocessed data, and evaluate the confidence quality of each data source in real time to generate a feature set with real-time confidence evaluation.

[0124] A fusion calculation module is configured to dynamically allocate the fusion weight of each data source based on the feature set with real-time confidence evaluation by an adaptive weighting strategy, and perform real-time fusion calculation in combination with a hierarchical processing mechanism of an edge computing architecture to generate a low-delay fusion result.

[0125] An instruction feedback module is configured to send the low-delay fusion result to the edge node via a lightweight transmission protocol, and the node executes a real-time decision algorithm to generate a control instruction and feed back to the movable intelligent device;

[0126] A resource scheduling module is configured to dynamically adjust fusion strategy parameters and configure computing resources to key function data processing links based on the generated whole-process data, through real-time exception detection and device computing power adaptation, to realize efficient fusion processing of multi-source data.

[0127] It should be noted that the system is a system corresponding to the above method, and all the implementation manners in the above method embodiment are applicable to this embodiment and can achieve the same technical effects.

[0128] Embodiments of the application also provide a computing device, comprising a processor and a memory storing a computer program, wherein the computer program is executed by the processor to perform the method described above. All the implementation manners in the above method embodiment are applicable to this embodiment and can achieve the same technical effects.

[0129] Embodiments of the application also provide a computer-readable storage medium storing instructions, which, when executed on a computer, cause the computer to perform the method described above. All the implementation manners in the above method embodiment are applicable to this embodiment and can achieve the same technical effects.

[0130] The above is the preferred embodiment of the application, and it should be noted that for those skilled in the art, without departing from the principles of the application, a number of improvements and refinements can be made, and these improvements and refinements should also be considered as the protection scope of the application.

Claims

1. A method for real-time fusion processing of multi-source data from a mobile intelligent device, characterized in that, The method includes: Step 1: Through a heterogeneous sensor array integrated into a mobile smart device, multi-dimensional raw data streams are collected in real time, and a spatiotemporal synchronization mechanism is applied to align the data streams of different sensors to generate a spatiotemporally consistent raw dataset. Step 2 involves performing dynamic interpolation compensation on the original dataset and constructing a dynamic calibration framework based on the inherent topological relationship of the data flow to form a dynamic sensing domain. An evolutionary sequence is generated based on the evolutionary behavior of data units at the domain boundary. Spatial correction values ​​are generated through the offset characteristics of the evolutionary sequence and a preset benchmark. Combined with real-time data correlation analysis, preprocessed data with fused spatial correction values ​​is generated. Specifically, this includes: performing manifold learning on the spatiotemporally consistent original dataset to extract the spatial adjacency and temporal continuity characteristics of data units, generating a sparse correlation matrix including quantified correlation strength; calculating the gradient magnitude of each row element based on the sparse correlation matrix, detecting abrupt changes where the gradient magnitude exceeds a set threshold, and using these abrupt changes as boundary points to form a spatial segmentation of the internal stable region and the external change region; within the external change region, based on the sparse correlation matrix... The corresponding connection strength values ​​are used to perform strength-weighted interpolation calculations on the missing data regions to generate a compensated continuous data stream. Based on the spatial distribution characteristics of the continuous data stream, a dynamic sensing domain of a structured mapping framework is formed. Data units in the core sensing area are acquired, and state transition vectors within a continuous time window are collected in real time. Based on the state transition vectors, they are sorted by timestamps to generate a multidimensional time series describing the evolution of data units. The multidimensional time series is dynamically time-aligned with a preset benchmark series, and the deviation of the covariance features of each dimension is calculated. Based on the covariance feature deviation, spatial correction coefficients are generated through preset transformation rules. The spatial correction coefficients are injected into the continuous data stream, and the correction magnitude is dynamically adjusted based on the Pearson correlation characteristics of multi-source sensor data to generate preprocessed data with fused spatial correction values. Step 3: Perform multidimensional feature extraction on the preprocessed data, evaluate the confidence quality of each data source in real time, and generate a feature set with real-time confidence evaluation. Step 4: Based on the feature set with real-time confidence assessment, dynamically allocate the fusion weights of each data source through an adaptive weighting strategy, and perform real-time fusion calculations in conjunction with the layered processing mechanism of the edge computing architecture to generate low-latency fusion results. Step 5: Send the low-latency fusion results to the edge node via a lightweight transmission protocol. At the node, execute a real-time decision-making algorithm to generate control commands and feed them back to the mobile smart device. Step 6: Based on the generated full-process data, dynamically adjust the fusion strategy parameters through real-time anomaly detection and device computing power adaptation, and allocate computing resources to key functional data processing links to achieve real-time fusion processing of multi-source data.

2. The method for real-time fusion processing of multi-source data in a mobile intelligent device according to claim 1, characterized in that, Multidimensional feature extraction is performed on the preprocessed data, the confidence quality of each data source is evaluated in real time, and a feature set with real-time confidence evaluation is generated, including: Based on the preprocessed data stream, the time domain, frequency domain and statistical domain feature dimensions are separated through parallel feature extraction channels to generate a multi-dimensional feature set; Based on a multidimensional feature set, the entropy change rate is calculated independently for each feature dimension. Based on the entropy stability within the sliding window, a preliminary confidence index for each feature dimension is generated. The spectral energy distribution of each feature dimension is extracted. Based on the energy concentration of the main frequency band, the spectral confidence coefficient of each feature dimension is calculated. Based on the preliminary confidence index and the spectral confidence coefficient, a weighted geometric mean is used to fuse them and generate a comprehensive confidence weight for each feature dimension. Based on the comprehensive confidence weight, the comprehensive confidence weight of the feature dimension associated with the corresponding data source is dynamically decayed to generate the decayed comprehensive confidence weight. The comprehensive confidence weight is used as metadata and labeled on the corresponding feature vector to form a feature set with weight labeling. Based on the weighted feature set, the mutual exclusivity between feature dimensions is analyzed. When conflicting feature dimensions are detected, the comprehensive confidence weight of the annotation is dynamically rebalanced and adjusted to generate a feature set with real-time confidence assessment.

3. The method for real-time fusion processing of multi-source data in a mobile intelligent device according to claim 2, characterized in that, Based on a feature set with accompanying real-time confidence assessment, the fusion weights of each data source are dynamically allocated through an adaptive weighting strategy. Real-time fusion computation is then performed using a layered processing mechanism within an edge computing architecture, generating low-latency fusion results, including: Based on the comprehensive confidence weights corresponding to the feature sets, the data sources are sorted to generate a data source sorting sequence; Based on the data source sorting sequence, the weight values ​​corresponding to each data source are dynamically allocated. In the perception layer of the edge computing architecture, the weight values ​​and feature sets of each data source are received, feature-level weighted fusion calculation is performed, and a preliminary fusion result is generated. The initial fusion results are passed to the decision layer of the edge computing architecture, where result-level computation based on context information and fusion results is performed to generate the fusion result. The computational latency of the fusion results is monitored in real time. When the computational latency exceeds the preset latency threshold, the weight allocation method is dynamically adjusted to obtain a fusion result that meets the preset latency threshold.

4. The method for real-time fusion processing of multi-source data in a mobile intelligent device according to claim 3, characterized in that, The low-latency fusion results are sent to edge nodes via a lightweight transmission protocol. Real-time decision-making algorithms are executed at the nodes to generate control commands and feed them back to the mobile intelligent device, including: Based on the environmental feature parameters in the low-latency fusion results, a pre-defined instruction generation rule base is queried, and a state transition strategy adapted to the current scenario is matched. The state transition strategy is input into the state transition equation calculation process, and combined with the current device state parameters, the final action sequence of the mobile intelligent device is generated. The execution timestamps of each action in the final action sequence are analyzed to calculate the urgency level of each action; based on the difference between the urgency level and the action timestamp, a timestamp-based coding algorithm is used to compress the control instruction dataset. The compressed control command data packets are sent wirelessly, and a retransmission mechanism with collision detection is used at the transport layer to ensure that the commands are reliably fed back to the execution end of the mobile smart device.

5. The method for real-time fusion processing of multi-source data in a mobile intelligent device according to claim 4, characterized in that, Based on the generated end-to-end data, through real-time anomaly detection and device computing power adaptation, the fusion strategy parameters are dynamically adjusted, and computing resources are allocated to key functional data processing stages to achieve real-time fusion processing of multi-source data, including: Construct an anomaly detection matrix for the entire process data. When an outlier in the matrix is ​​detected to exceed a preset threshold, generate a command to adjust the fusion strategy parameters. Based on the fusion strategy parameter adjustment instructions, the real-time computing power level of the mobile intelligent device is obtained, and according to the confidence quality coefficient ratio of each processing link, the processor core resources are dynamically reallocated to the feature extraction, fusion computing and decision generation links to obtain the core resource reallocation result. Based on the results of core resource reallocation, a sliding window mechanism is used to update sensor compensation coefficients, multi-source data weights, and state transition decision rules to achieve real-time fusion processing of multi-source data.

6. A real-time fusion processing system for multi-source data from a mobile intelligent device, wherein the system implements the method as described in any one of claims 1 to 5, characterized in that, include: The data acquisition module is used to collect multi-dimensional raw data streams in real time through a heterogeneous sensor array integrated into a mobile smart device, and to apply a spatiotemporal synchronization mechanism to align the data streams of different sensors to generate a spatiotemporally consistent raw dataset. The dynamic processing module is used to perform dynamic interpolation compensation operations on the original dataset, and to build a dynamic calibration framework based on the inherent topological relationship of the data flow to form a dynamic sensing domain. It generates an evolution sequence according to the evolution behavior of data units at the domain boundary, generates spatial correction values ​​through the offset characteristics of the evolution sequence and the preset benchmark, and generates preprocessed data with fused spatial correction values ​​by combining real-time data correlation analysis. Specifically, this includes: performing manifold learning on the spatiotemporally consistent original dataset to extract the spatial adjacency and temporal continuity features of data units, generating a sparse association matrix including quantified association strength; based on the sparse association matrix, calculating the gradient magnitude of each row element, detecting abrupt changes where the gradient magnitude exceeds a set threshold, and using these abrupt changes as boundary points to form a spatial segmentation of an internal stable region and an external change region; within the external change region, performing strength-weighted interpolation calculations on the missing data regions based on the corresponding connection strength values ​​in the sparse association matrix to generate a compensated continuous data stream; and forming a structure based on the spatial distribution characteristics of the continuous data stream. The dynamic sensing domain of the mapping framework is constructed; data units in the core sensing area are acquired, and state transition vectors within a continuous time window are collected in real time; based on the state transition vectors, they are sorted by timestamps to generate a multidimensional time series describing the evolution of data units; the multidimensional time series is dynamically time-normalized and aligned with a preset benchmark series, and the deviation of the covariance features of each dimension is calculated; based on the covariance feature deviation, spatial correction coefficients are generated through preset transformation rules, and the spatial correction coefficients are injected into the continuous data stream. At the same time, the correction magnitude is dynamically adjusted based on the Pearson correlation characteristics of multi-source sensor data to generate preprocessed data with fused spatial correction values. The feature evaluation module is used to extract multidimensional features from preprocessed data, evaluate the confidence quality of each data source in real time, and generate a feature set with real-time confidence evaluation. The fusion computing module is used to dynamically allocate fusion weights for each data source based on a feature set with accompanying real-time confidence assessment through an adaptive weighting strategy, and perform real-time fusion computing in conjunction with the layered processing mechanism of the edge computing architecture to generate low-latency fusion results. The instruction feedback module is used to send the low-latency fusion results to the edge nodes via a lightweight transmission protocol, and execute real-time decision-making algorithms at the nodes to generate control commands and feed them back to the mobile smart devices. The resource scheduling module is used to dynamically adjust the fusion strategy parameters based on the generated full-process data, through real-time anomaly detection and device computing power adaptation, and to allocate computing resources to key functional data processing links, so as to achieve efficient fusion processing of multi-source data.

7. A computing device, characterized in that, include: One or more processors; A storage device for storing one or more programs, which, when executed by one or more processors, cause the one or more processors to implement the method as described in any one of claims 1 to 5.

8. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a program that, when executed by a processor, implements the method as described in any one of claims 1 to 5.

Images