CN-121561370-B - Edge computing data processing method and system for time domain feature analysis and physical constraint verification

CN121561370BCN 121561370 BCN121561370 BCN 121561370BCN-121561370-B

Abstract

A method and a system for processing edge calculation data of time domain feature analysis and physical constraint verification relate to the technical field of edge calculation. The method includes acquiring time series data and image data. The time domain variance is calculated and compared to a dispersion threshold. If the time domain variance is greater than the dispersion threshold, the second processing unit is awakened. And extracting semantic features of the image data to obtain semantic confidence, and then carrying out hierarchical screening transmission on the data together with the time domain variance. And constructing the context of the prompt word, inputting the context into a generating model deployed at the edge side, and generating a control parameter vector. And inputting the control parameter vector into a digital simulation engine to carry out physical simulation deduction, and calculating a physical residual error comprehensive score according to a deduction result. The physical residual composite score is compared to a verification threshold to determine whether the control parameter vector is valid. And when the control parameter vector is judged to be effective, transmitting the control parameter vector to a cloud data processing center. And triggering a prompt word correction mechanism based on the physical residual error when the judgment is invalid.

Inventors

JIANG XIAOFENG
YU SHANSHAN
CHEN XU
HUANG ZHIKE
WU SHIFENG
LIANG CHENG
Wen Guanshui

Assignees

厦门四信物联网科技有限公司

Dates

Publication Date: 20260505
Application Date: 20260121

Claims (7)

1. The edge calculation data processing method for the time domain feature analysis and the physical constraint verification is characterized by being suitable for a water conservancy detection scene, and comprises the following steps: S1, acquiring time sequence data and image data in an edge computing scene; S2, calculating the time domain variance of the time sequence data, and comparing the time domain variance with a dispersion threshold value, and waking up a second processing unit in a dormant state if the time domain variance is larger than the dispersion threshold value; S3, extracting semantic features of the image data to obtain semantic confidence, and then carrying out hierarchical screening transmission on the data together with the time domain variance; S4, constructing a prompt word context based on the sensor time domain characteristics, the environmental state constraint conditions and the control targets at the current moment, inputting the prompt word context into a generated model deployed at the edge side, and generating a control parameter vector; S5, inputting the control parameter vector into a digital simulation engine to carry out physical simulation deduction, and calculating a physical residual error comprehensive score according to a deduction result; S6, comparing the physical residual comprehensive score with a verification threshold value to judge whether the control parameter vector is effective, transmitting the control parameter vector to a cloud data processing center when judging that the control parameter vector is effective, triggering a prompt word correction mechanism based on the physical residual to generate a feedback prompt word when judging that the control parameter vector is ineffective, and adding the feedback prompt word into the prompt word context to drive the generated model to regenerate the control parameter vector until the verification threshold value is met; s5 specifically comprises the following steps: Inputting the control parameter vector into a digital simulation engine to carry out physical simulation deduction, so as to obtain a simulation output result; Calculating mass conservation residual error according to simulation output result And boundary constraint residual ; Based on And Calculating a physical residual composite score ; Conservation of mass residual error The method comprises the following steps: ; In the middle of The quality is input to the system and, For the quality of the output of the system, Storing a quality variation for the system; Boundary constraint residual The method comprises the following steps: ; In the middle of Representing the maximum of the values in brackets; is a safety upper threshold; is a safety lower threshold; Is the tolerance; outputting a result for simulation; Physical residual comprehensive scoring The method comprises the following steps: ; In the middle of A normalization factor for the mass conservation residual; Is that Weights of (2); Is that Weights of (2); S6 comprises the following steps: At the position of Judging that the control parameter vector is valid; At the position of > Judging that the control parameter vector is invalid and triggering the prompt word correction mechanism; s6 includes, when it is determined that the operation is invalid: Comparing the weighted mass conservation residual error item with the boundary constraint residual item; If it is Judging the cause of the violation as not meeting the conservation of mass and according to The positive and negative of (a) are further analyzed into mass accumulation or mass loss, and feedback prompt words containing mass conservation correction suggestions are generated; If it is Judging the cause of the violation is out of range and based on The item larger than zero in the list is further analyzed to be in a state exceeding upper limit or in a state exceeding lower limit, and feedback prompt words containing safety boundary correction suggestions are generated; Adding the feedback prompt word into the context of the original prompt word to restrict the generated model to regenerate the control parameter vector; The closed loop process of generating, simulating deduction, scoring calculation and threshold comparison is repeatedly performed until > And obtaining a control parameter vector meeting the requirements.
2. The edge computing data processing method for time domain feature analysis and physical constraint verification according to claim 1, wherein the method is characterized by extracting semantic features from image data to obtain semantic confidence, and specifically comprises the following steps: Performing operation on the image data by adopting a lightweight convolutional neural network, and calculating probability distribution of a current image belonging to a predefined key physical state category through a Softmax layer at a network output layer, wherein the lightweight convolutional neural network is a MobileNet or YOLO-Nano framework subjected to pruning and quantization, and the key physical state category comprises a valve opening state and/or a meter pointer reading range and/or a pipeline surface crack level and/or personnel invasion behavior; and selecting the maximum probability value in the probability distribution as the semantic confidence Pconf.
3. The edge computing data processing method for time domain feature analysis and physical constraint verification according to claim 1, wherein the semantic confidence and the time domain variance together perform hierarchical filtering transmission on data, and specifically comprises the following steps: build with time domain variance Defining Tvar2 as a physical fluctuation threshold and Tconf as a semantic credibility threshold for two-dimensional decision logic with a first dimension and semantic confidence Pconf as a second dimension; If it is Judging that the current data is hidden abnormal data, executing a forced data retention strategy, marking the original data as high priority and transmitting the high priority to a cloud data processing center; If it is Judging the current data to be effective normal data, and transmitting the effective normal data to a cloud data processing center according to the conventional priority; If it is And judging the current data as invalid redundant data, and executing discarding operation, wherein Tvar2 is a physical fluctuation threshold value, and Tconf is a semantic credibility threshold value.
4. The edge computing data processing method for time domain feature analysis and physical constraint verification according to claim 1, wherein the generated model is a quantized version of a transducer model deployed at an edge side, the context of the prompt word is a structured prompt word context, and the control parameter vector is a standardized decision suggestion; the prompt word context at least comprises: a sensor time domain feature field for characterizing a time domain variance or derivative statistic thereof at a current time; An environmental state constraint field for characterizing a current environmental state constraint condition; a control target field for characterizing a target control intent; The control parameter vector includes at least: A target object identification parameter; A target motion parameter; Action duration parameter.
5. The method for processing edge calculation data of time domain feature analysis and physical constraint verification according to claim 1, wherein the method for processing edge calculation data of time domain feature analysis and physical constraint verification is characterized by inputting a control parameter vector into a digital simulation engine to perform physical simulation deduction, and obtaining a simulation output result, and specifically comprises the following steps: Vectorizing and mapping the control parameter vector to obtain an input parameter which can be identified by the digital simulation engine; And executing simulation deduction in a physical simulation environment built in the digital simulation engine based on the input parameters to obtain a simulation output result, wherein a state equation of the simulation deduction is as follows: in the middle of Outputting a result for simulation; is the current system state vector; the function is deduced for simulation.
6. The method for processing edge computing data of time domain feature analysis and physical constraint verification according to claim 1, wherein the method for processing edge computing data of time domain feature analysis and physical constraint verification is characterized by computing time domain variance of time sequence data and comparing the time domain variance with a dispersion threshold, and waking up a second processing unit in a sleep state if the time domain variance is greater than the dispersion threshold, and specifically comprises: Carrying out subsection statistics on the time sequence data by adopting a sliding window; calculating time domain variance for the sampling sequence in any sliding window, wherein the calculation model of the time domain variance is as follows: ; In the middle of Is the time domain variance; sampling points for data in the sliding window; Is the inside of the window Sampling data; the average value of all sampling data in the window; defining a dispersion threshold as Tvar1, and using the time domain variance Comparing with a dispersion threshold; If it is Tvar1, judging that the current data has significant fluctuation, generating an interrupt signal by a first processing unit, and activating a second processing unit to enter a working state; If it is And the data is less than or equal to Tvar1, the fact that the current data has no significant fluctuation is judged, and the first processing unit controls the second processing unit to keep a dormant state.
7. The edge computing data processing system is characterized by being suitable for a water conservancy detection scene, and comprises a data acquisition module, a first processing unit, a second processing unit, a digital simulation engine and a cloud data processing center, wherein the first processing unit is connected with the data acquisition module in a communication mode, the second processing unit is connected with the data acquisition module and the first processing unit in a communication mode, the digital simulation engine is connected with the second processing unit in a communication mode, and the cloud data processing center is connected with the second processing unit and the digital simulation engine in a communication mode; the edge computing data processing device is used for executing the edge computing data processing method for time domain feature analysis and physical constraint verification according to any one of claims 1 to 6.

Description

Edge computing data processing method and system for time domain feature analysis and physical constraint verification Technical Field The invention relates to the technical field of edge calculation, in particular to an edge calculation data processing method and system for time domain feature analysis and physical constraint verification. Background In an industrial internet of things edge computing scenario, there is an urgent need for efficient data processing systems. The system needs to realize intelligent scheduling of computing resources to reduce energy consumption, ensure logic correctness of generated data output, and avoid losing key abnormal information in the data transmission process, so that reliability and energy efficiency ratio of the whole system are improved. In the prior art, two types of methods are mainly adopted for edge calculation data processing. One type is a timing-based processing mode, and the data acquisition terminal wakes up the high-performance processor according to fixed frequency to execute data analysis and uploading operation. The other is based on the end-to-end processing mode of the neural network, and the control instruction is directly generated by arranging a deep learning model or a physical information neural network at the edge. However, the prior art has significant drawbacks. The scheduling efficiency of the computing resource is low, and the high-performance computing unit still continuously operates when the data has no significant change, so that energy waste is caused. The generated data lacks a real-time logic verification mechanism, and error output violating the physical law cannot be intercepted in an inference stage, so that the data confidence is insufficient. The data filtering strategy relies on single dimension characteristics, and hidden abnormal data with obvious physical fluctuation and low semantic recognition degree is easy to ignore, so that key information is lost. Disclosure of Invention The invention provides a method and a system for processing edge calculation data of time domain feature analysis and physical constraint verification, which are used for improving at least one of the technical problems. The invention provides an edge calculation data processing method for time domain feature analysis and physical constraint verification, which is suitable for a water conservancy detection scene. The edge calculation data processing method comprises the following steps. Time series data and image data in an edge calculation scene are acquired. The time domain variance of the time series data is calculated and compared to a dispersion threshold. If the time domain variance is greater than the dispersion threshold, waking up the second processing unit in a sleep state. And extracting semantic features of the image data to obtain semantic confidence, and then carrying out hierarchical screening transmission on the data together with the time domain variance. And constructing a prompt word context based on the sensor time domain characteristics, the environmental state constraint conditions and the control target at the current moment, inputting the prompt word context into a generated model deployed at the edge side, and generating a control parameter vector. And inputting the control parameter vector into a digital simulation engine to carry out physical simulation deduction, and calculating a physical residual error comprehensive score according to a deduction result. The physical residual composite score is compared to a verification threshold to determine whether the control parameter vector is valid. And when the control parameter vector is judged to be effective, transmitting the control parameter vector to a cloud data processing center. And when the judgment is invalid, triggering a prompt word correction mechanism based on the physical residual error, generating a feedback prompt word and adding the feedback prompt word to the prompt word context to drive the generation model to regenerate the control parameter vector until the verification threshold is met. As a further scheme of the invention, the method for extracting the semantic features of the image data to obtain the semantic confidence comprises the following steps: And performing operation on the image data by adopting a lightweight convolutional neural network, and calculating probability distribution of the current image belonging to a predefined key physical state category through a Softmax layer at a network output layer. The lightweight convolutional neural network is a MobileNet or YOLO-Nano architecture subjected to pruning and quantization. The key physical state categories include valve opening state and/or meter pointer reading range and/or pipeline surface crack level and/or personnel intrusion behavior. And selecting the maximum probability value in the probability distribution as the semantic confidence Pconf. As a further scheme of the present invention, th