Edge computing data processing method and system for time domain feature analysis and physical constraint verification

Abstract

This invention relates to a method and system for edge computing data processing, encompassing temporal feature analysis and physical constraint verification, and pertains to the field of edge computing technology. The method includes: acquiring time-series data and image data; calculating the temporal variance and comparing it with a dispersion threshold; activating a second processing unit if the temporal variance exceeds the dispersion threshold; extracting semantic features from the image data to obtain semantic confidence, which is then used in conjunction with the temporal variance for hierarchical filtering and transmission; constructing a prompt word context and inputting it into a generative model deployed at the edge to generate a control parameter vector; inputting the control parameter vector into a digital simulation engine for physical simulation and calculating a comprehensive physical residual score based on the simulation results; comparing the comprehensive physical residual score with a verification threshold to determine the validity of the control parameter vector; transmitting the control parameter vector to a cloud data processing center if valid; and triggering a prompt word correction mechanism based on physical residuals if invalid.

Description

Technical Field

[0001] This invention relates to the field of edge computing technology, and more specifically, to an edge computing data processing method and system that performs temporal feature analysis and physical constraint verification. Background Technology

[0002] In industrial IoT edge computing scenarios, there is an urgent need for efficient data processing systems. These systems need to intelligently schedule computing resources to reduce energy consumption, ensure the logical correctness of generative data output, and prevent the loss of critical anomaly information during data transmission, thereby improving the overall system reliability and energy efficiency.

[0003] In existing technologies, edge computing data processing mainly employs two types of methods. One is a time-based processing mode, where the data acquisition terminal wakes up a high-performance processor at a fixed frequency to perform data analysis and uploading operations. The other is an end-to-end processing mode based on neural networks, which directly generates control commands by deploying deep learning models or physical information neural networks at the edge.

[0004] However, existing technologies have significant drawbacks. Computational resource scheduling is inefficient; high-performance computing units continue to run even when data shows no significant changes, resulting in energy waste. Generative data lacks real-time logical verification mechanisms, failing to intercept erroneous outputs that violate physical laws during the inference phase, leading to insufficient data confidence. Data filtering strategies rely on single-dimensional features, easily overlooking hidden anomalies with significant physical fluctuations but low semantic recognizability, causing the loss of crucial information. Summary of the Invention

[0005] This invention provides an edge computing data processing method and system for temporal feature analysis and physical constraint verification, in order to improve at least one of the above-mentioned technical problems.

[0006] Firstly, this invention provides an edge computing data processing method for temporal feature analysis and physical constraint verification, applicable to water conservancy monitoring scenarios. The edge computing data processing method includes the following steps.

[0007] Acquire time-series data and image data in edge computing scenarios.

[0008] Calculate the time-domain variance of the time series data and compare it with the dispersion threshold. If the time-domain variance is greater than the dispersion threshold, the second processing unit, which is in a dormant state, is awakened.

[0009] Semantic features are extracted from image data to obtain semantic confidence, which is then combined with temporal variance to perform hierarchical filtering and transmission of the data.

[0010] Based on the sensor's temporal characteristics, environmental constraints, and control objectives at the current moment, a contextualized prompt word is constructed and input into a generative model deployed at the edge to generate a control parameter vector.

[0011] The control parameter vector is input into the digital simulation engine for physical simulation and deduction, and the physical residual comprehensive score is calculated based on the deduction results.

[0012] The physical residual comprehensive score is compared with a validation threshold to determine whether the control parameter vector is valid. If valid, the control parameter vector is transmitted to the cloud data processing center. If invalid, a prompt word correction mechanism based on physical residuals is triggered, generating feedback prompt words and appending them to the prompt word context to drive the generative model to regenerate the control parameter vector until the validation threshold is met.

[0013] As a further aspect of the present invention, semantic feature extraction is performed on the image data to obtain semantic confidence, specifically including:

[0014] A lightweight convolutional neural network is used to perform operations on the image data, and a Softmax layer is used at the network output layer to calculate the probability distribution of the current image belonging to a predefined key physical state category. The lightweight convolutional neural network is a pruned and quantized MobileNet or YOLO-Nano architecture. The key physical state categories include valve opening status and / or instrument pointer reading range and / or pipe surface crack level and / or personnel intrusion behavior.

[0015] The maximum probability value in the probability distribution is selected as the semantic confidence level Pconf.

[0016] As a further aspect of the present invention, semantic confidence and temporal variance are used together to perform hierarchical filtering and transmission of data, specifically including:

[0017] Constructing time-domain variance This is a two-dimensional decision-making logic with the first dimension being the physical fluctuation threshold and the semantic confidence threshold being the semantic confidence threshold.

[0018] like If the current data is determined to be hidden abnormal data, a forced data retention policy is implemented, marking the original data as high priority and transmitting it to the cloud data processing center.

[0019] like If the current data is determined to be valid and normal, it will be transmitted to the cloud data processing center according to the usual priority.

[0020] like The current data is determined to be invalid and redundant, and is discarded. Here, Tvar2 is the physical fluctuation threshold, and Tconf is the semantic reliability threshold.

[0021] As a further aspect of the present invention, the generative model is a quantized version of the Transformer model deployed at the edge. The cue word context is a structured cue word context. The control parameter vector is a standardized decision suggestion.

[0022] The context of the prompt word includes at least:

[0023] The sensor's time-domain feature field is used to characterize the time-domain variance or its derived statistics at the current moment.

[0024] The environment state constraint field is used to characterize the current environment state constraints.

[0025] The control target field is used to characterize the control intent of the target.

[0026] The control parameter vector includes at least: target object identification parameter, target action parameter, and action duration parameter.

[0027] As a further aspect of the present invention, the control parameter vector is input into a digital simulation engine for physical simulation deduction, and a comprehensive score of physical residuals is calculated based on the deduction results, specifically including:

[0028] The control parameter vector is input into the digital simulation engine to perform physical simulation and obtain the simulation output results.

[0029] Calculate the mass conservation residual based on the simulation output results. and boundary constraint residuals .

[0030] based on and Calculate the comprehensive score of physical residuals .

[0031] As a further aspect of the present invention, the mass conservation residual for:

[0032] .

[0033] In the formula Input quality to the system, For system output quality, This represents the change in system storage quality.

[0034] Boundary constraint residuals for:

[0035] ;

[0036] In the formula This indicates taking the maximum value among the multiple values ​​in parentheses. This is the upper limit threshold for safety. This is the lower safety threshold. For tolerance. This is the simulation output result.

[0037] Physical residual comprehensive score for:

[0038] .

[0039] In the formula This is the normalization factor for the mass conservation residual. for The weight. for The weight.

[0040] As a further aspect of the present invention, the control parameter vector is input into a digital simulation engine for physical simulation deduction to obtain simulation output results, specifically including:

[0041] The control parameter vector is vectorized to obtain input parameters that the digital simulation engine can recognize.

[0042] Based on the input parameters, a simulation is performed in the physical simulation environment built into the digital simulation engine to obtain the simulation output results. The state equation for the simulation is: In the formula This is the simulation output result. This is the current system state vector. This is the simulation derivation function.

[0043] As a further aspect of the present invention, the physical residual comprehensive score is... Comparison with verification threshold And determine whether the control parameter vector is valid, specifically including:

[0044] exist The control parameter vector is then determined to be valid.

[0045] exist > The system determines that the control parameter vector is invalid and triggers the prompt word correction mechanism.

[0046] As a further aspect of the present invention, a prompt word correction mechanism based on physical residuals is triggered to generate feedback prompt words and append them to the prompt word context to drive the generative model to regenerate the control parameter vector until the verification threshold is met, specifically including:

[0047] Compare the weighted mass conservation residual term with the boundary constraint residual term.

[0048] like If so, the violation is determined to be due to a failure to satisfy the law of conservation of mass, and based on... The positive or negative value is further analyzed into mass accumulation or mass deficit, generating feedback prompts that include suggestions for mass conservation correction.

[0049] like If so, the violation is determined to be due to a state exceeding the boundary, and based on... Items with values ​​greater than zero are further analyzed as either exceeding the upper or lower limit of the state, generating feedback prompts that include suggestions for correcting the safety boundary.

[0050] The feedback prompt is appended to the original prompt context to constrain the generative model to regenerate the control parameter vector.

[0051] Repeat the closed-loop process of generation, simulation, score calculation, and threshold comparison until... > Obtain the control parameter vector that meets the requirements.

[0052] As a further aspect of the present invention, the time-domain variance of the time series data is calculated and compared with a dispersion threshold. If the time-domain variance is greater than the dispersion threshold, the second processing unit in a dormant state is awakened, specifically including:

[0053] Segmented statistics are performed on time series data using a sliding window.

[0054] Calculate the time-domain variance of the sampled sequence within any sliding window. The calculation model for the time-domain variance is as follows:

[0055] .

[0056] In the formula Let V be the time-domain variance. This represents the number of data sampling points within the sliding window. For the first in the window One sampled data. This is the average value of all sampled data within the window.

[0057] Define the dispersion threshold as Tvar1, and then use the time-domain variance... Compare with the dispersion threshold.

[0058] like >Tvar1, if it is determined that there is a significant fluctuation in the current data, the first processing unit generates an interrupt signal, which activates the second processing unit to enter the working state.

[0059] like If the value is less than or equal to Tvar1, it is determined that there is no significant fluctuation in the current data, and the first processing unit controls the second processing unit to remain in a dormant state.

[0060] Secondly, this invention provides an edge computing data processing system for temporal feature analysis and physical constraint verification, applicable to water conservancy monitoring scenarios. The edge computing data processing system includes a data acquisition module, a first processing unit communicatively connected to the data acquisition module, a second processing unit communicatively connected to both the data acquisition module and the first processing unit, a digital simulation engine communicatively connected to the second processing unit, and a cloud data processing center communicatively connected to both the second processing unit and the digital simulation engine.

[0061] The edge computing data processing device is used to execute the edge computing data processing method for temporal feature analysis and physical constraint verification as described in any part of the first aspect.

[0062] By adopting the above technical solution, the present invention can achieve the following technical effects:

[0063] This invention achieves significant technical results through a three-stage processing flow of "temporal variance pre-screening - multimodal data fusion filtering - physical simulation verification": In water conservancy monitoring scenarios, the system's energy efficiency ratio is greatly improved (the proportion of high-performance computing unit running time is reduced from 100% to 35%, and energy consumption is reduced by 62%), the logical reliability of generative data is significantly enhanced (error rate is reduced from 28% to 3%), and the forced retention mechanism of the two-dimensional decision matrix successfully avoids the loss of 32 sets of key "hidden abnormal data", thus solving the core problems of low resource scheduling efficiency, poor data reliability and missing key information in edge computing. Attached Figure Description

[0064] To more clearly illustrate the technical solution of the present invention, the accompanying drawings used in the specific embodiments of the present invention will be briefly introduced below. It should be understood that the following drawings only show some specific embodiments of the present invention and should not be regarded as a limitation on the scope. For those skilled in the art, other related drawings can be obtained from these drawings without creative effort.

[0065] Figure 1 This is a flowchart illustrating edge computing data processing methods.

[0066] Figure 2 This is a hardware architecture diagram of a data processing system.

[0067] Figure 3 This is a flowchart of hierarchical data processing.

[0068] Figure 4 This is a flowchart of data generation based on simulation verification.

[0069] The diagram is labeled as follows: 1-Data acquisition module, 2-First processing unit, 3-Second processing unit, 4-Digital simulation engine, 5-Cloud data processing center. Detailed Implementation

[0070] To address the technical problems of low computing resource scheduling efficiency, poor reliability of generative data logic, and easy loss of key information during data filtering in existing technologies, an edge computing data processing method based on temporal feature analysis and physical constraint verification is proposed.

[0071] This embodiment provides an edge computing data processing method and system based on temporal feature analysis and physical constraint verification, which is used to realize on-demand scheduling of edge computing resources, real-time compliance verification of generative data, and multi-dimensional accurate data filtering, thereby improving the energy efficiency, data processing reliability, and information integrity of the edge computing system.

[0072] Please see Figures 1 to 4 The first embodiment of the present invention provides an edge computing data processing method based on temporal feature analysis and physical constraint verification, which can be executed by an edge computing data processing system based on temporal feature analysis and physical constraint verification (hereinafter referred to as: edge computing data processing device).

[0073] This embodiment of the edge computing data processing system based on temporal feature analysis and physical constraint verification includes a data acquisition module 1, a first processing unit (low power) 2, a second processing unit (high performance) 3, a digital simulation engine 4, and a cloud data processing center 5. The connections between the components are as follows: the data acquisition module 1 is connected to both the first processing unit 2 and the second processing unit 3. The first processing unit 2 is connected to the second processing unit 3. The second processing unit 3 is connected to both the digital simulation engine 4 and the cloud data processing center 5. The digital simulation engine 4 is connected to the cloud data processing center 5.

[0074] The data acquisition module is used to collect multimodal data (including time series signal data and image data) in edge computing scenarios and transmit the collected data to the first processing unit.

[0075] The first processing unit (low power, such as MCU) 2 is communicatively connected to the data acquisition module 1 and the second processing unit 3, and is used to perform time-domain feature pre-screening and computing power wake-up determination, and output interrupt signals to control the working state of the second processing unit.

[0076] The second processing unit (high-performance, such as NPU) 3 is connected to the first processing unit 2, the digital simulation engine 4, and the cloud data processing center 5 to perform multimodal feature fusion analysis and data filtering, and output the data to be verified or the effective data after filtering.

[0077] The digital simulation engine 4 has a built-in physical simulation environment that communicates with the second processing unit 3 to perform simulation deduction and physical constraint verification on the control parameters output by the generative model.

[0078] The cloud data processing center 5 is communicatively connected to the second processing unit 3, and is used to receive and store the valid data transmitted by the second processing unit 3, while also receiving the verification results feedback from the digital simulation engine.

[0079] The edge computing data processing method based on time-domain feature analysis and physical constraint verification in this embodiment is based on the core principle of using low-computing-power units to pre-screen time-domain signal features to trigger high-computing-power units to work on demand, while simultaneously constructing physical simulation verification logic to perform compliance cleaning on the output data of the generative model. The data processing flow of this invention is divided into three stages, which sequentially realize on-demand wake-up of computing power, multimodal data filtering, and physical verification of generative data.

[0080] The hierarchical data processing flow of the edge computing data processing method includes steps S1 to S6.

[0081] S1. Acquire time-series data and image data in edge computing scenarios.

[0082] Specifically, time-series data and image data are collected by the data acquisition module. For example, in a reservoir water level monitoring system, the data acquisition module uses a water level sensor and an image sensor to collect time-series water level data (sampling frequency 10Hz) from the water conservancy monitoring points, as well as image data of the monitoring area.

[0083] S2. Calculate the time-domain variance of the time series data and compare it with the dispersion threshold. If the time-domain variance is greater than the dispersion threshold, wake up the second processing unit that is in a dormant state.

[0084] Step S2 is executed by the first processing unit. The core of S2 is to solve the energy efficiency problem of edge computing nodes. By performing lightweight time-domain feature analysis through the first processing unit, on-demand scheduling of high-performance computing resources can be achieved.

[0085] First, the first processing unit receives the time series data S transmitted by the data acquisition module and performs segmented statistics on the time series data using a sliding window.

[0086] Then, the time-domain variance is calculated for the sampled sequence within any sliding window.

[0087] The calculation model for time-domain variance is as follows:

[0088] .

[0089] In the formula Let V be the time-domain variance. This represents the number of data sampling points within the sliding window. For the first in the window One sampled data. This is the average value of all sampled data within the window.

[0090] Finally, the dispersion threshold is defined as Tvar1. The first processing unit will process the calculated time-domain variance. Compare with the preset dispersion threshold Tvar1.

[0091] like >Tvar1 determines that the current data has significant fluctuations and requires high-performance computing processing. The first processing unit generates an interrupt signal, which activates the second processing unit to enter the working state.

[0092] like If the value is less than or equal to Tvar1, it is determined that the current data has no significant fluctuations and no high-performance computing processing is required. The first processing unit controls the second processing unit to remain in a dormant state to reduce system energy consumption.

[0093] S3. Extract semantic features from the image data to obtain semantic confidence, and then use this confidence score along with the temporal variance to perform hierarchical filtering and transmission of the data. Preferably, S3 includes S31 to S36.

[0094] S31. A lightweight convolutional neural network is used to perform operations on the image data, and a Softmax layer is used in the network output layer to calculate the probability distribution of the current image belonging to a predefined key physical state category. The lightweight convolutional neural network is a pruned and quantized MobileNet or YOLO-Nano architecture. The key physical state categories include valve opening status and / or instrument pointer reading range and / or pipe surface crack level and / or personnel intrusion behavior.

[0095] S32. Select the maximum probability value in the probability distribution as the semantic confidence level Pconf.

[0096] The core of S3 is to solve the problem of data transmission optimization in weak network environments. It achieves multi-dimensional data filtering by fusing temporal and semantic features through the second processing unit, thus avoiding the loss of key information.

[0097] First, multimodal data is acquired. After the second processing unit is activated, it synchronously receives the temporal variance transmitted by the first processing unit. Image data transmitted with the data acquisition module.

[0098] Then, semantic features are extracted. The second processing unit uses a lightweight convolutional neural network (such as a pruned and quantized MobileNet or YOLO-Nano architecture) to perform operations on the image data. Identification objects and methods: For industrial site images, the network output layer predefines key physical state categories (e.g., valve opening status, instrument pointer reading range, pipe surface crack level, personnel intrusion behavior). The second processing unit calculates the probability distribution of the current image belonging to the above categories through a Softmax layer, and selects the maximum probability value as the semantic confidence Pconf.

[0099] Specifically, if the object to be controlled is a valve, the valve's state needs to be identified. For example, if the probability of identifying "valve fully open" is 0.92, then Pconf = 0.92. The controlled object can be set to other objects; this invention does not specifically limit this.

[0100] S33, Constructing a time-domain variance This is a two-dimensional decision-making logic with the first dimension being the physical fluctuation threshold and the semantic confidence threshold being the semantic confidence threshold.

[0101] S34, if If the current data is determined to be hidden abnormal data, a forced data retention policy is implemented, marking the original data as high priority and transmitting it to the cloud data processing center.

[0102] S35, if If the current data is determined to be valid and normal, it will be transmitted to the cloud data processing center according to the usual priority.

[0103] S36, if The current data is determined to be invalid and redundant, and is discarded. Here, Tvar2 is the physical fluctuation threshold, and Tconf is the semantic reliability threshold.

[0104] Specifically, the two-dimensional decision logic is based on time-domain variance. The logical mapping plane has the horizontal axis (representing physical signal fluctuations) and the semantic confidence level Pconf (representing AI cognitive certainty) as the vertical axis.

[0105] The construction method is as follows: set the physical fluctuation threshold Tvar2 and the semantic credibility threshold Tconf in the plane, and divide the decision space into the following four logical quadrants.

[0106] Invalid region ( ≤Tvar2,Pconf≤Tconf): Noisy data.

[0107] Regular area ( ≤Tvar2,Pconf>Tconf): Stable running data.

[0108] explicit fault area ( >Tvar2,Pconf>Tconf): Known fault data.

[0109] Hidden anomaly area ( >Tvar2,Pconf≤Tconf): An unknown dangerous condition where there are violent physical fluctuations but the AI ​​cannot recognize its pattern.

[0110] Then, the data is graded, filtered, and manipulated according to the four logical quadrants.

[0111] like( >Tvar2)∩(Pconf>Tconf) (Tconf is the preset semantic trust threshold): Determines the data as valid and normal, and transmits it to the cloud data processing center according to the normal priority.

[0112] like( ≤Tvar2)∩(Pconf≤Tconf): The data is determined to be invalid and redundant, and is discarded directly without performing a transmission operation.

[0113] like( >Tvar2)∩(Pconf≤Tconf): If the data is determined to be "hidden abnormal data", a forced data retention strategy is executed to bypass the conventional confidence filtering threshold and mark the original data packet as "high priority" for transmission to the cloud data processing center to avoid the loss of critical abnormal information.

[0114] S4. Based on the sensor's temporal characteristics, environmental constraints, and control objectives at the current moment, construct the context of the prompt words and input it into the generative model deployed on the edge side to generate the control parameter vector.

[0115] Preferably, the generative model is a quantized version of the Transformer model deployed at the edge. The cue word context is a structured cue word context. The control parameter vector is a standardized decision suggestion.

[0116] The control parameter vector includes at least: target object identification parameter, target action parameter, and action duration parameter.

[0117] The context of the prompt word is at least which of the following fields?

[0118] The sensor's time-domain feature field is used to characterize the time-domain variance or its derived statistics at the current moment.

[0119] The environment state constraint field is used to characterize the current environment state constraints.

[0120] The control target field is used to characterize the control intent of the target.

[0121] The core of S4 is to solve the problem of logical correctness in generative data. It uses a digital simulation engine to build physical constraint verification barriers to intercept erroneous data that violates physical laws.

[0122] Specifically, the second processing unit encapsulates the current sensor temporal feature sequence, environmental state constraints (such as "current water level is too high"), and control objectives (such as "rapid flood discharge") into a structured prompt context, and inputs it into the generative model (such as a quantized version of Transformer) deployed on the edge side.

[0123] The generation process is as follows: the model analyzes the context based on the attention mechanism, predicts and outputs a series of standardized control parameter vectors. (For example: =[valve ID, target opening degree, action duration]), as a decision suggestion to be executed.

[0124] S5. Input the control parameter vector into the digital simulation engine to perform physical simulation and calculate the comprehensive score of the physical residuals based on the simulation results. Preferably, S5 includes S51 to S53.

[0125] S51, Transfer the control parameter vector The data is input into a digital simulation engine for physical simulation and deduction, yielding simulation output results. Specifically, the control parameter vector is... Vectorization mapping is performed, mapping the data to input parameters that the digital simulation engine can recognize. Then, based on these input parameters, simulation derivation is performed in the physical simulation environment built into the digital simulation engine to obtain simulation output results.

[0126] Specifically, the first step is to perform vectorization mapping, transforming the control parameter vector... The data is vectorized and mapped to input parameters that the digital simulation engine can recognize. Then, a simulation is performed, substituting the mapped input parameters into the physical simulation environment built into the digital simulation engine to execute the simulation and obtain the simulation output. .

[0127] The state equation derived from the simulation is as follows: In the formula This is the simulation output result (i.e., the predicted state of the simulation output, which is used for subsequent physical constraint verification). This is the current system state vector. This is the simulation derivation function.

[0128] S52. Calculate the mass conservation residual based on the simulation output results. and boundary constraint residuals .

[0129] Specifically, construct the physical constraint loss function. The residual between the simulation output and the physical law constraints is calculated. At least include mass conservation residuals With boundary constraint residuals .

[0130] Mass conservation residual for:

[0131] .

[0132] In the formula Input quality to the system, For system output quality, This represents the change in system storage quality.

[0133] Boundary constraint residuals for:

[0134] ;

[0135] In the formula This indicates taking the maximum value among the multiple values ​​in parentheses. This is the upper limit threshold for safety. This is the lower safety threshold. For tolerance. This is the simulation output result.

[0136] Physical constraint loss function for:

[0137] .

[0138] .

[0139] In the formula for The weight. for The weight.

[0140] S53, based on and Calculate the comprehensive score of physical residuals .

[0141] Specifically, the base system calculates the comprehensive score of the physical residuals based on the laws of physical conservation. .

[0142] .

[0143] In the formula This is the mass conservation residual (i.e., the difference between fluid input and output). This is the normalization factor for the mass conservation residual. This refers to the boundary constraint residual (also known as the state-space saturated nonlinear residual). for The weight. for The weight.

[0144] Physical residual comprehensive score Used for validity determination. Specifically, it involves calculating... With the preset verification threshold Comparisons are made. The normalization factor for the mass conservation residual is usually taken as the average mass flow rate or design value of the system under normal operating conditions, in order to eliminate the influence of dimensions (for example, the rated flow rate in a hydraulic system).

[0145] S6. Comprehensive scoring of physical residuals With verification threshold The control parameter vector is compared to determine its validity. If valid, the control parameter vector is transmitted to the cloud data processing center. If invalid, a prompt word correction mechanism based on physical residuals is triggered, generating feedback prompt words and appending them to the prompt word context to drive the generative model to regenerate the control parameter vector until the verification threshold is met.

[0146] Preferably, step S6 includes S61 to S67.

[0147] S61, in If the control parameter vector is deemed valid, it will be transmitted to the cloud data processing center.

[0148] The proactive optimization function in the cloud: The cloud data processing center not only stores data but also includes a model training module. It focuses on extracting "hidden anomaly data" (i.e., data with large physical fluctuations but unknown to AI) marked by the edge as high-value negative samples, and uses incremental learning techniques to fine-tune the generative model. The fine-tuned model parameters are periodically sent back to the second processing unit at the edge via OTA technology, enabling the edge system to identify previously "unknown" anomalies and achieve self-evolution of the system.

[0149] S62, in > The system determines that the control parameter vector is invalid and triggers the prompt word correction mechanism.

[0150] S63. Compare the weighted mass conservation residual term with the boundary constraint residual term.

[0151] S64, if If so, the violation is determined to be due to a failure to satisfy the law of conservation of mass, and based on... The positive or negative value is further analyzed into mass accumulation or mass deficit, generating feedback prompts that include suggestions for mass conservation correction.

[0152] S65, if If so, the violation is determined to be due to a state exceeding the boundary, and based on... Items with values ​​greater than zero are further analyzed as either exceeding the upper or lower limit of the state, generating feedback prompts that include suggestions for correcting the safety boundary.

[0153] S66. The feedback prompt word is appended to the original prompt word context to constrain the generative model to regenerate the control parameter vector.

[0154] S67. Repeat the closed-loop process of generation, simulation, scoring calculation, and threshold comparison until... > Obtain the control parameter vector that meets the requirements.

[0155] The prompt word correction mechanism based on physical residuals is as follows.

[0156] First, error analysis is performed: the system analyzes the physical residual comprehensive score and generates feedback prompts.

[0157] Specifically, comparative and In overall score The contribution of [the organization / entity].

[0158] like If so, the primary reason for the violation is determined to be "failure to satisfy the law of conservation of mass". The system will then conduct further analysis. The symbol and size.

[0159] like This indicates that the system has a mass surplus (or excessive input). The interpretation is: "The generated control strategy has led to mass accumulation in the system (or a continuous rise in the liquid level), violating the law of conservation of mass."

[0160] like This indicates that the system has a mass deficit (or excessive output). The explanation is: "The generated control strategy has led to a mass deficit in the system (or a continuous drop in liquid level), violating the law of conservation of mass."

[0161] like If so, the primary violation is determined to be "state out of bounds". The system will then conduct further analysis. The two in the formula item.

[0162] examine Which of the following is greater than zero?

[0163] like This indicates that the predicted system state exceeds the safety limit. Therefore, it can be interpreted as: "The generated control strategy will cause the system state..." Exceeding the safety limit ".

[0164] like This indicates that the predicted system state is below the safety lower limit. Therefore, it can be interpreted as: "The generated control strategy will cause the system state..." Below the safety limit ".

[0165] Then, closed-loop generation is performed: the feedback prompt is appended to the original input, commanding the generative model: "The control strategy generated last time resulted in a quality loss. Please regenerate the strategy to correct the error."

[0166] The feedback prompt template is: "Warning: The control parameter vector u=[specific value] generated last time based on [insert original context here, such as 'current water level is too high, rapid flood discharge is required'] failed the physical simulation verification. The reason is [insert specific reason parsed from step B here, such as 'this strategy leads to mass loss' or 'this strategy will cause the water level to exceed the warning line']. Please regenerate a new control strategy that ensures the physical conservation law (mass conservation) and the system safety boundary (water level between ... meters and ... meters) are satisfied."

[0167] Finally, under the constraints of the feedback information, the model outputs a new control parameter vector. Until the physical constraints are met.

[0168] To facilitate understanding of the present invention, a practical application scenario is provided below to illustrate the application of this embodiment. Specifically, edge computing data processing in a water conservancy monitoring scenario is taken as an example. It is understood that the scope of protection of the present invention is not limited to this embodiment and can be extended to all industrial IoT edge computing scenarios.

[0169] The system configuration for this application scenario example is as follows.

[0170] The data acquisition module uses a water level sensor and an image sensor to collect water level time series data (sampling frequency 10Hz) from water conservancy monitoring points, as well as image data of the monitoring area.

[0171] The first processing unit uses an STM32L4 series MCU (low power) with a preset dispersion threshold Tvar1=0.05.

[0172] The second processing unit uses NVIDIA Jetson Xavier NX (high performance) with a preset semantic trust threshold Tconf=0.8.

[0173] The digital simulation engine has a built-in physical simulation model of the hydraulic system (including mass conservation equations and water level boundary constraint equations) and preset verification thresholds. =0.03.

[0174] The cloud-based data processing center uses Alibaba Cloud ECS servers to store valid water level data and abnormal data records.

[0175] The processing flow for application scenario examples is as follows.

[0176] The first stage involves waking up the computing power. The data acquisition module collects water level time series data S (100 sampling points in total, N=100) and transmits it to the STM32L4 MCU. The MCU calculates the time-domain variance using a sliding window algorithm. =0.08. Since 0.08 > Tvar1 (0.05), the MCU generates an interrupt signal, activating the NVIDIA Jetson Xavier NX to enter the working state.

[0177] The second stage involves multimodal data filtering. The NVIDIA Jetson Xavier NX receives data transmitted from the MCU. =0.08 and the monitoring area image data transmitted by the image sensor are convolved to obtain a semantic confidence score Pconf = 0.75. Since (0.08>0.05)∩(0.75≤0.8), the system determines that the data is "hidden abnormal data" and implements a forced retention strategy, marking the original water level data and image data as "high priority" for transmission to the cloud data processing center.

[0178] The third stage involves data verification processing.

[0179] Generative models output control parameter vectors for water conservancy regulation. =[0.2,0.5,0.3], transferred to NVIDIA Jetson Xavier NX. NX will... After vectorization and mapping, the data is input into the digital simulation engine and substituted into the state equations of the hydraulic system. Perform simulation and deduction.

[0180] Calculate the mass conservation residual =|100-98-1.5|=0.5, boundary constraint residual =0.01, Overall Score =0.26.

[0181] Because 0.26> (0.03), system determination Invalid data triggers a feedback mechanism to control the generative model to resample and generate a new control parameter vector. =[0.18,0.45,0.27], repeat the verification process until... ≤0.03 will be effective The data is transmitted to the cloud data processing center.

[0182] This embodiment calculates the temporal variance of time series data using a low-power first processing unit. ,Will Compared with the preset threshold Tvar1, only when >When Tvar1 is reached, the high-performance secondary processing unit is awakened, enabling on-demand allocation of computing resources. Through the first-stage computing power wake-up process, the runtime percentage of the high-performance computing unit is reduced from 100% in the prior art to 35%, and system energy consumption is reduced by 62%.

[0183] This embodiment incorporates temporal variance. A two-dimensional decision matrix is ​​constructed using (physical features) and semantic confidence Pconf (semantic features), targeting ( A mandatory retention mechanism for hidden anomaly data (>Tvar2)∩(Pconf≤Tconf) is designed to achieve hierarchical and classified data transmission. Through the second-stage multimodal filtering process, 32 sets of "hidden anomaly data" were successfully retained, avoiding the loss of key monitoring information and verifying the effectiveness of the invention.

[0184] This embodiment constructs a loss function that includes physical constraints. The online verification system uses a digital simulation engine to process the control parameter vectors output by the generative model. Simulations were performed to calculate the physical constraint residuals and obtain a comprehensive score. ,according to With verification threshold The comparison results determine the validity of the data. Through the third-stage physical constraint verification process, the error rate of the generative model output data is reduced from 28% in the existing technology to 3%.

[0185] This embodiment presents an edge computing data processing method that combines temporal feature analysis and physical constraint verification. The method uses a first processing unit to perform lightweight temporal variance calculation for data pre-screening. It only activates the high-performance computing unit when significant data fluctuations occur, avoiding unnecessary operation of the high-performance computing unit, significantly reducing overall system energy consumption, and improving computing resource scheduling efficiency. This effectively improves the energy efficiency ratio of the edge computing system.

[0186] This embodiment innovatively transforms physical laws into computable physical constraint loss functions, constructs an online verification barrier through a digital simulation engine, and intercepts erroneous data that violates physical laws in real time. This solves the "logical illusion" problem in the output data of generative AI models and improves the confidence and reliability of data processing results.

[0187] This embodiment constructs a two-dimensional decision matrix by fusing the time-domain signal discreteness and image semantic confidence, and designs a mandatory retention mechanism for "hidden abnormal data", which avoids the loss of key information caused by single-dimensional data filtering, while reducing the transmission and storage overhead of invalid data and improving the accuracy and efficiency of data transmission.

[0188] Obviously, the embodiments described above are only some embodiments of the present invention, and not all embodiments. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without inventive effort are within the scope of protection of the present invention.

[0189] In the several embodiments provided in this invention, it should be understood that the disclosed apparatus and methods can also be implemented in other ways. The apparatus and method embodiments described above are merely illustrative. For example, the flowcharts and block diagrams in the accompanying drawings illustrate the architecture, functionality, and operation of possible implementations of apparatus, methods, and computer program products according to various embodiments of the present invention. In this regard, each block in a flowchart or block diagram may represent a module, segment, or portion of code containing one or more executable instructions for implementing a specified logical function. It should also be noted that in some alternative implementations, the functions marked in the blocks may occur in a different order than those marked in the drawings. For example, two consecutive blocks may actually be executed substantially in parallel, and they may sometimes be executed in reverse order, depending on the functions involved. It should also be noted that each block in a block diagram and / or flowchart, and combinations of blocks in block diagrams and / or flowcharts, can be implemented using a dedicated hardware-based system that performs the specified function or action, or using a combination of dedicated hardware and computer instructions.

[0190] In addition, the functional modules in the various embodiments of the present invention can be integrated together to form an independent part, or each module can exist independently, or two or more modules can be integrated to form an independent part.

[0191] If the aforementioned functions are implemented as software functional modules and sold or used as independent products, they can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention, or the part that contributes to the prior art, or a part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, electronic device, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of the present invention. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory, random access memory, magnetic disks, or optical disks. It should be noted that, in this document, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitations, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes said element.

[0192] The terminology used in the embodiments of this invention is for the purpose of describing particular embodiments only and is not intended to be limiting of the invention. The singular forms “a,” “the,” and “the” used in the embodiments of this invention are also intended to include the plural forms unless the context clearly indicates otherwise.

[0193] It should be understood that the term "and / or" used in this article is merely a description of the relationship between related objects, indicating that three relationships can exist. For example, A and / or B can represent: A existing alone, A and B existing simultaneously, and B existing alone. Additionally, the character " / " in this article generally indicates that the preceding and following related objects have an "or" relationship.

[0194] Depending on the context, the word "if" as used here can be interpreted as "when," "when," "in response to determination," or "in response to detection." Similarly, depending on the context, the phrase "if determination" or "if detection (of the stated condition or event)" can be interpreted as "when determination," "in response to determination," "when detection (of the stated condition or event)," or "in response to detection (of the stated condition or event)."

[0195] The terms "first" and "second" used in the embodiments are merely to distinguish similar objects and do not represent a specific ordering of objects. It is understood that "first" and "second" can be interchanged in a specific order or sequence where permitted. It should be understood that the objects distinguished by "first" and "second" can be interchanged where appropriate so that the embodiments described herein can be implemented in an order other than those illustrated or described herein.

[0196] The above description is merely a preferred embodiment of the present invention and is not intended to limit the invention. Various modifications and variations can be made to the present invention by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A method for edge computing data processing with temporal feature analysis and physical constraint verification, characterized in that, Applicable to water conservancy monitoring scenarios; edge computing data processing methods include: S1. Acquire time-series data and image data in edge computing scenarios; S2. Calculate the time-domain variance of the time series data and compare it with the dispersion threshold; if the time-domain variance is greater than the dispersion threshold, wake up the second processing unit that is in a dormant state. S3. Extract semantic features from the image data to obtain semantic confidence, and then use it together with the temporal variance to perform hierarchical filtering and transmission of the data; S4. Based on the sensor temporal characteristics, environmental constraints, and control objectives at the current moment, construct the context of the prompt words and input them into the generative model deployed on the edge side to generate the control parameter vector; S5. Input the control parameter vector into the digital simulation engine to perform physical simulation and calculate the physical residual comprehensive score based on the simulation results. S6. Compare the physical residual comprehensive score with the verification threshold to determine whether the control parameter vector is valid. If it is valid, transmit the control parameter vector to the cloud data processing center. If it is invalid, trigger the prompt word correction mechanism based on physical residual, generate feedback prompt words and add them to the prompt word context to drive the generative model to regenerate the control parameter vector until the verification threshold is met. S5 specifically includes: The control parameter vector is input into the digital simulation engine to perform physical simulation and obtain the simulation output results. Calculate the mass conservation residual based on the simulation output results. and boundary constraint residuals ; based on and Calculate the comprehensive score of physical residuals ; Mass conservation residual for: ； In the formula Input quality to the system, For system output quality, This represents the change in system storage quality. Boundary constraint residuals for: ; In the formula This indicates taking the maximum value among the multiple values ​​in parentheses; This is the upper limit threshold for safety. This is the lower safety threshold. For tolerance; The simulation output results; Physical residual comprehensive score for: ； In the formula This is the normalization factor for the mass conservation residual; for The weights; for The weights; S6 includes: exist The control parameter vector is then determined to be valid. exist > The control parameter vector is determined to be invalid, and the prompt word correction mechanism is triggered. S6 includes the following when determining invalidity: Compare the weighted mass conservation residual term with the boundary constraint residual term; like If so, the violation is determined to be due to a failure to satisfy the law of conservation of mass, and based on... The positive or negative value is further analyzed into mass accumulation or mass deficit, generating feedback prompts containing suggestions for mass conservation correction; like If so, the violation is determined to be due to a state exceeding the boundary, and based on... Items with values ​​greater than zero are further parsed as states exceeding the upper or lower limits, generating feedback prompts that include suggestions for safety boundary correction. The feedback prompt is appended to the original prompt context to constrain the generative model to regenerate the control parameter vector; Repeat the closed-loop process of generation, simulation, score calculation, and threshold comparison until... > Obtain the control parameter vector that meets the requirements.

2. The edge computing data processing method for temporal feature analysis and physical constraint verification according to claim 1, characterized in that, Semantic feature extraction is performed on image data to obtain semantic confidence, specifically including: A lightweight convolutional neural network is used to perform operations on the image data, and a Softmax layer is used in the network output layer to calculate the probability distribution of the current image belonging to a predefined key physical state category; wherein, the lightweight convolutional neural network is a pruned and quantized MobileNet or YOLO-Nano architecture; the key physical state categories include valve opening status and / or instrument pointer reading range and / or pipe surface crack level and / or personnel intrusion behavior; The maximum probability value in the probability distribution is selected as the semantic confidence level Pconf.

3. The edge computing data processing method for temporal feature analysis and physical constraint verification according to claim 1, characterized in that, Semantic confidence and temporal variance are used together to perform hierarchical filtering and transmission of data, specifically including: Constructing time-domain variance A two-dimensional decision logic is defined with the first dimension and the semantic confidence Pconf as the second dimension; Tvar2 is defined as the physical fluctuation threshold and Tconf as the semantic confidence threshold. like If the current data is determined to be hidden abnormal data, a forced data retention strategy is implemented, the original data is marked as high priority and transmitted to the cloud data processing center; like If the current data is determined to be valid and normal, it will be transmitted to the cloud data processing center according to the usual priority. like The current data is determined to be invalid and redundant, and a discard operation is performed; where Tvar2 is the physical fluctuation threshold and Tconf is the semantic credibility threshold.

4. The edge computing data processing method for temporal feature analysis and physical constraint verification according to claim 1, characterized in that, The generative model is a quantized version of the Transformer model deployed at the edge; the cue word context is a structured cue word context. The control parameter vector is a standardized decision recommendation; The context of the prompt word includes at least: Sensor time-domain feature fields are used to characterize the time-domain variance or its derived statistics at the current moment; The environment state constraint field is used to characterize the current environment state constraints. The control target field is used to characterize the control intent of the target; The control parameter vector includes at least: Target object identifier parameters; Target action parameters; Action duration parameter.

5. The edge computing data processing method for temporal feature analysis and physical constraint verification according to claim 1, characterized in that, The control parameter vector is input into the digital simulation engine for physical simulation derivation, and the simulation output results are obtained, including: The control parameter vector is vectorized and mapped to obtain input parameters that the digital simulation engine can recognize; Based on the input parameters, a simulation is performed in the physical simulation environment built into the digital simulation engine to obtain the simulation output results; wherein the state equation of the simulation is: In the formula The simulation output results; This represents the current system state vector; This is the simulation derivation function.

6. The edge computing data processing method for temporal feature analysis and physical constraint verification according to claim 1, characterized in that, Calculate the time-domain variance of the time series data and compare it with the dispersion threshold; If the time-domain variance is greater than the dispersion threshold, the second processing unit, which is in a dormant state, is awakened, specifically including: Segmented statistics are performed on time series data using a sliding window. Calculate the time-domain variance of the sampled sequence within any sliding window; the calculation model for the time-domain variance is as follows: ； In the formula For time-domain variance; This represents the number of data sampling points within the sliding window. For the first in the window One sampled data; This is the average value of all sampled data within the window; Define the dispersion threshold as Tvar1, and then use the time-domain variance... Compare with the dispersion threshold; like >Tvar1, if it is determined that there is a significant fluctuation in the current data, the first processing unit generates an interrupt signal, which activates the second processing unit to enter the working state; like If the value is less than or equal to Tvar1, it is determined that there is no significant fluctuation in the current data, and the first processing unit controls the second processing unit to remain in a dormant state.

7. An edge computing data processing system for temporal feature analysis and physical constraint verification, characterized in that, Applicable to water conservancy testing scenarios; the edge computing data processing system includes a data acquisition module, a first processing unit connected to the data acquisition module, a second processing unit connected to the data acquisition module and the first processing unit, a digital simulation engine connected to the second processing unit, and a cloud data processing center connected to the second processing unit and the digital simulation engine; The edge computing data processing device is used to execute the edge computing data processing method for temporal feature analysis and physical constraint verification as described in any one of claims 1 to 6.

Images