Learnable multi-edge-aware device self-calibration method and apparatus, and storage medium

Abstract

The application relates to a self-calibration method, device and storage medium of an incrementally learnable multi-edge perception device. The method comprises the following steps: firstly, based on the current calibration environment, environment perception calibration data is acquired; secondly, based on the current application environment, real-time output response values of each edge perception device are acquired; finally, based on the environment perception calibration data and the real-time uploaded output response values, the output response values of the edge perception device are converted into corresponding perception physical quantities by using a calibration model based on meta-learning and incremental learning, and self-calibration is completed; the acquisition process of the calibration model comprises offline training based on meta-learning and cloud server optimization based on incremental learning. Compared with the prior art, the application has the advantages of significantly improving the calibration efficiency of a large number of sensors, self-adaptation and continuous optimization, etc.

Description

Technical Field

[0001] This invention relates to the field of robotics and sensor signal processing technology, and in particular to a self-calibration method, apparatus and storage medium for a multi-edge sensing device capable of growth learning. Background Technology

[0002] Edge sensing devices refer to a class of devices that can be deployed over a large area and possess real-time sensing data processing capabilities, widely used in fields such as the Internet of Things (IoT), smart homes, smart cities, and industrial automation. They typically include various sensors and computing environments, meaning processing occurs where the sensing data is generated. These devices can collect various environmental data (such as temperature, humidity, light, sound, motion, etc.) and process and analyze this data through built-in processors and algorithms to achieve various functions, such as environmental monitoring, security monitoring, health tracking, and intelligent control. Sensors are crucial components of edge sensing devices, and their signal calibration methods are particularly important. For example, in the field of autonomous driving, a large number of sensors are relied upon to detect information such as distance and force. Sensor data is essential for ensuring that vehicles accurately perceive their environment and quickly adjust their operation. Improper sensor calibration can lead to data distortion, affecting the vehicle's decision-making system and even endangering traffic safety. Therefore, efficient, accurate, and adaptive unified calibration of a large number of sensors is a critical issue. However, because sensor calibration involves fine and complex adjustments to intrinsic and extrinsic parameters, quickly calibrating a large number of sensors is extremely challenging. Therefore, developing a unified calibration method for a large number of sensors can not only ensure the high consistency and reliability of the sensed data, but also improve the automation efficiency of the entire system.

[0003] There are three main approaches to sensor calibration. The first is based on the sensor's physical characteristics. For example, calibrating optical distortion parameters to correct the image of an optical sensor offers high reliability and accuracy, but requires detailed information about the sensor's physical characteristics. The second approach uses validated reference standards to calibrate the sensor. For example, for temperature sensors, a standard heat source with a known temperature is used for calibration. This method is simple and reliable, but it is highly dependent on the accuracy of the reference standard and is not suitable for all types of sensors. The calibration process using these methods is complex and time-consuming, making them unsuitable for the unified calibration of a large number of sensors. Recently, the third approach, data-driven calibration, has gradually become mainstream. This type of method primarily uses machine learning techniques to correct sensor data. For example, using a multilayer perceptron to correct deviations in sensor output can handle complex nonlinear problems and is suitable for the unified calibration of a large number of sensors. However, fitting a model of sensor output using a neural network often requires collecting a large amount of data in the calibration environment for model training and may suffer from overfitting. Furthermore, some applications require sensors to have self-adjusting capabilities during use to adapt to environmental changes in real time, improving the sensor's flexibility and adaptability. Therefore, how to reduce the dependence on large amounts of data, avoid overfitting, and improve the environmental adaptability and flexibility of sensor calibration has become a problem that needs to be solved in this field. Summary of the Invention

[0004] The purpose of this invention is to overcome the shortcomings of the prior art, such as strong dependence on data volume and insufficient adaptability to environmental changes, and to provide a self-calibration method, device and storage medium for multi-edge sensing devices that can grow and learn.

[0005] The objective of this invention can be achieved through the following technical solutions:

[0006] According to a first aspect of the present invention, a self-calibration method for a multi-edge sensing device capable of growth learning is provided, comprising the following steps:

[0007] S1, based on the current calibration environment, acquire environmental perception calibration data;

[0008] S2, based on the current application environment, obtain the real-time output response values ​​of each edge sensing device;

[0009] S3, based on the environmental perception calibration data and the real-time uploaded output response values, the output response values ​​of the edge sensing device are converted into corresponding sensing physical quantities using a calibration model based on meta-learning and incremental learning, thus completing self-calibration. The process of acquiring the calibration model includes offline training based on meta-learning and cloud server optimization based on incremental learning. The offline training process includes:

[0010] S301, acquire the offline output response values ​​of each edge sensing device, and the standard reference values ​​corresponding to each offline output response value;

[0011] S302, Based on a deep learning framework, a neural network model is constructed using the offline output response value and the standard reference value;

[0012] S303, The neural network model is trained using a meta-learning algorithm to obtain the trained calibration model.

[0013] As a preferred technical solution, the cloud server optimization process specifically includes the following steps:

[0014] S311, Receive and periodically preprocess the real-time uploaded output response values, and load the trained calibration model;

[0015] S312, Define the loss function and training optimization algorithm for contrastive learning;

[0016] S313, build a comparative learning framework, train the calibration model based on the real-time uploaded output response value, and optimize the parameters of the current calibration model;

[0017] S314, repeat S313 a preset number of times to obtain new calibration model parameters, and when there is performance optimization in the calibration results, update the current calibration model parameters to obtain a new calibration model.

[0018] As a preferred technical solution, in step S2, when acquiring the real-time output response value of each edge sensing device, multiple samplings are performed, and the number of samplings is determined according to the range of the edge sensing device.

[0019] As a preferred technical solution, the measurement equipment in the calibration environment includes pre-arranged standardized test equipment and verified high-precision measurement tools.

[0020] As a preferred technical solution, the process of obtaining the offline output response value and the standard reference value includes: selecting multiple edge sensing devices, each of which actively interacts with the measurement device in the calibration environment based on target driving to obtain the corresponding offline output response value and standard reference value.

[0021] As a preferred technical solution, S302 specifically includes:

[0022] The parameters of the neural network model are initialized using a random number seed;

[0023] The input vector is constructed using the offline output response value and the standard reference value.

[0024] As a preferred technical solution, S303 specifically includes:

[0025] S3031, the input vector is divided into multiple task sets, the number of task sets is equal to the number of edge sensing devices, and each task set includes a support set and a query set in equal proportions;

[0026] S3032, input the input vector of the support set of any task set into the current neural network model, obtain the corresponding loss function, and adjust the parameters of the current neural network model based on the current loss function;

[0027] S3033, save the current neural network model parameters and load them into the current neural network model;

[0028] S3034, Input the input vector of the query set of the same task set into the current neural network model to obtain the corresponding loss function;

[0029] S3035, repeat steps S3032 to S3034 until all task sets have been calculated, obtain the final total loss function, adjust and save the current meta-learning algorithm parameters based on the total loss function, and load them into the current neural network model;

[0030] S3036, repeat steps S3032 to S3035 a preset number of times to perform meta-learning training and obtain the final neural network model parameters.

[0031] As a preferred technical solution, the total loss function is the sum of the corresponding loss functions obtained in step S3034 for all task sets.

[0032] According to a second aspect of the present invention, a self-calibration apparatus for a multi-edge sensing device capable of growth learning is provided, comprising a memory, a processor, and a program stored in the memory, wherein the processor executes the program to implement the method described therein.

[0033] According to a third aspect of the present invention, a storage medium is provided having a program stored thereon, which, when executed, implements the method described thereon.

[0034] Compared with the prior art, the present invention has the following beneficial effects:

[0035] 1. This invention models the correction characteristics of edge sensing devices based on neural networks and trains the neural network using meta-learning algorithms. The resulting meta-learning-based calibration model can significantly improve the calibration efficiency of a large number of edge sensing devices, including sensors, ensuring that a large number of edge sensing devices have uniform, effective, and accurate sensing capabilities. At the same time, the meta-learning algorithm can reduce the requirements and dependence of the neural network model on the amount of training data, improve the generalization performance of the system, avoid overfitting, and improve the accuracy of the output of edge sensing devices.

[0036] 2. This invention designs a contrastive learning algorithm to optimize a calibration model based on meta-learning, which can achieve adaptive optimization and continuous growth learning of the calibration model, enhance its adaptability in the face of unknown environmental conditions, and improve the flexibility of the calibration model. Attached Figure Description

[0037] Figure 1 A flowchart of the method provided by the present invention;

[0038] Figure 2 This is a diagram of the calibration model architecture provided in the embodiments of the present invention;

[0039] Figure 3 This is a diagram of the meta-learning process architecture provided in this embodiment of the invention;

[0040] Figure 4 This is a hardware architecture diagram of a biomimetic self-calibration device for a distributed electronic skin capable of growth learning, provided in an embodiment of the present invention. Detailed Implementation

[0041] The present invention will now be described in detail with reference to the accompanying drawings and specific embodiments. These embodiments are based on the technical solution of the present invention and provide detailed implementation methods and specific operating procedures. However, the scope of protection of the present invention is not limited to the following embodiments.

[0042] Example

[0043] This embodiment provides a self-calibration method for multi-edge sensing devices with growth learning capability, particularly incorporating biomimetic principles. In this embodiment, the edge sensing device primarily refers to a sensor. This method learns sensor error correction patterns based on a neural network model, effectively fitting the accurate sensor output, achieving rapid and unified self-calibration of a large number of sensors. It also uses a meta-learning algorithm to reduce the data volume requirement and, based on contrastive learning, enables the sensing device to achieve growth learning (i.e., incremental learning) to adapt to constantly changing external conditions.

[0044] like Figure 1 As shown, the aforementioned method includes the following steps:

[0045] Step S1: Based on the current calibration environment, acquire environmental perception calibration data. Specifically:

[0046] For pressure sensors used in electronic skin systems, an accurate and effective calibration environment was created indoors. This environment included standardized testing equipment and validated, high-precision pressure measurement tools. To create an effective and accurate calibration environment, standardized testing equipment was deployed in fixed locations to rigorously calibrate the accuracy of measurement results. Furthermore, a range of validated, high-precision measurement tools were provided in the environment to offer reliable benchmarks for evaluating sensor performance.

[0047] The pressure measuring tool has three fixed positions, located on the XOY, XOZ, and YOZ planes respectively, with the tool's normal vector being the positive normal vector of the corresponding plane. This arrangement not only rigorously calibrates the accuracy of the sensor data but also provides a comprehensive and reliable benchmark for evaluating the sensor's performance.

[0048] Step S2: Based on the current application environment, acquire the real-time output response values ​​of each edge sensing device. This requires selecting multiple representative sensors on the robot operating platform and actively interacting with measurement devices in the target-driven and calibration environment to collect their output response values ​​and reference values. Specifically:

[0049] In this embodiment, a distributed electronic skin is worn on a robotic arm, and 10 representative locations are marked on the electronic skin as marker points. The network formed by these marker points roughly covers the entire robotic arm. At each selected marker point, active interaction is performed between the target-driven measurement equipment and the calibration environment. The output response values ​​of the sensors are acquired through this active interaction, while a validated high-precision measurement tool provides the corresponding standard reference values. The comprehensiveness of sensor data acquisition is ensured through the active interaction strategy and adjustments to the calibration environment conditions. Considering the small sample size of the calibration model parameters and meta-learning, it is generally necessary to repeat sampling k times at each marker point, ensuring the sensor output covers the measurement range as much as possible, resulting in a total of n*k sets of data (where n represents the number of sensors at representative locations). In this embodiment, sampling is repeated 100 times at each marker point, while ensuring the sensor output values ​​uniformly cover the measurement range. A total of 10*100 = 1000 sets of data are obtained.

[0050] Step S3, in the application phase, based on edge computing devices and acquired environmental perception calibration data, the sensor output response values ​​without reference values ​​are uploaded to the cloud server in real time. A calibration model based on meta-learning and incremental learning is then used to convert the sensor output response values ​​into corresponding perceived physical quantities, completing the self-calibration. Specifically:

[0051] (1) Collect the raw sensor output response values ​​without processing and store them in the buffer area; carry the calibration model obtained based on meta-learning and convert it into the corresponding physical quantity to realize unified, effective and accurate pressure sensing of large-area electronic skin;

[0052] (2) Edge computing devices are deployed near the electronic skin and transmit the temporarily stored raw sensor output response values ​​wirelessly to the host computer, and then upload them to the cloud server. The cloud server stores these raw sensor data without reference values ​​for subsequent growth learning, i.e. incremental learning.

[0053] The process of obtaining the calibration model based on meta-learning and incremental learning includes offline training based on meta-learning and cloud server optimization based on incremental learning.

[0054] (1) The offline training process based on meta-learning includes:

[0055] Step S301: In the offline phase, acquire the output response values ​​of each sensor and the standard reference values ​​corresponding to each output response value.

[0056] Step S302: Based on the deep learning framework, construct a neural network model using the offline output response values ​​and standard reference values. Specifically:

[0057] S3021, Construct a multilayer perceptron as the calibration model, i.e., the structure of the initial neural network model, such as... Figure 2 As shown, the initial parameters of the model are generated from a random number seed;

[0058] In this embodiment, as Figure 3 As shown, the neural network consists of three fully connected layers. The first layer has an input dimension of 1 and an output dimension of 128, the second layer has an input dimension of 128 and an output dimension of 256, and the third layer has an input dimension of 256 and an output vector of 1.

[0059] S3022, construct the input vector based on the data obtained from active interaction, that is, construct the input vector using the offline output response value and the standard reference value. The input vector includes the response output of the sensor at each representative location and the corresponding reference value.

[0060] Step S303: Train the neural network model using a meta-learning algorithm to obtain the trained calibration model. Specifically:

[0061] S3031, for sensors at 10 representative locations, divides all input vectors into 10 task sets. Each task set is further divided into a support set and a query set in equal proportions. The ratio of the number of samples in the support set to the number of samples in the query set is typically 1:1, depending on the scenario task and the amount of available data. Considering the sample diversity of the support set and the sample representativeness of the query set, the ratio can be adjusted according to requirements to achieve the best learning efficiency and generalization performance of the model.

[0062] S3032, set the task set T i The input vector of the support set is input into the current neural network model. The loss function L(θ) is calculated by forward propagation. Based on the current loss function, the parameters θ of the current neural network model are adjusted by backpropagation using gradient descent.

[0063] S3033, save the current neural network model parameters θ and load them into the current neural network model;

[0064] S3034, set the task set T i The input vector of the query set is input into the current neural network model, and the loss function is calculated during forward propagation. i ;

[0065] S3035, repeat steps S3032 to S3034 until all 10 task sets have been calculated, obtain the final total loss function, and based on the total loss function, adjust and save the current meta-learning algorithm parameters φ using gradient descent backpropagation, and load them into the current neural network model; the process of each iteration cycle can be refined as follows:

[0066] Extract a task set T from the iterative data set i Execute S3032 to S3034 to obtain the corresponding loss function l. i Repeat this step until each task set in the iterative dataset has completed its forward propagation computation; after completing the above steps, calculate the loss l for each task set sample. i Adding them together yields the total loss function L(φ), which can be expressed as:

[0067]

[0068] Based on this, backpropagation is performed and gradients are calculated to update the parameters φ of the meta-learning algorithm. The meta-learning training process is as follows: Figure 3 As shown, where F φ The function f is the result of computation by the meta-learning algorithm. θ , which represents the function calculated by the neural network model.

[0069] S3036, Repeat steps S3032 to S3035 a preset number of times to perform meta-learning training, and obtain the final neural network model parameters, so that the model has better generalization ability and learning ability.

[0070] (2) The process of cloud server optimization based on incremental learning, that is, on the cloud server, based on comparative learning, continuously optimizes the parameters of the neural network model to improve the accuracy and adaptability of the general calibration model. Specifically, it includes the following steps:

[0071] Step S311: Standardize and preprocess the large number of unlabeled output response values ​​collected; load the calibration model trained based on meta-learning.

[0072] Step S312, define the loss function L for contrastive learning. contrast And train the optimization algorithm, and train and optimize the current calibration model.

[0073] Specifically, the loss function L for contrastive learning contrast Represented as:

[0074]

[0075] In the formula, x a x p x n These represent edge sensing device data for anchor points, positive samples, and negative samples, respectively. This represents the first two layers of a multilayer perceptron, based on the loss function L. contrast The difference between similar and dissimilar samples can be measured; then the Adam optimizer is selected for model training; in other embodiments, other suitable optimization algorithms can be selected for model training.

[0076] Step S313: Build a contrastive learning framework, input the unlabeled output response value into the optimized calibration model for forward propagation calculation, then calculate the gradient based on the backpropagation of the loss, and use the Adam optimizer to update the current calibration model parameters. Repeat this step until all unlabeled data has been calculated.

[0077] Step S314: Repeat S313 a preset number of times to obtain new calibration model parameters. When there is performance optimization in the calibration results, update the current calibration model parameters and send them back to the edge computing device to obtain a new calibration model.

[0078] Step S315: Periodically use newly collected unlabeled output response values ​​to continuously optimize the model in order to cope with changes in data distribution, achieve growth learning, and continuously improve the model's accuracy and adaptability.

[0079] Figure 4The hardware architecture system for the method provided in this embodiment is shown. Combined with the robot operation platform, the electronic skin can realize self-learning and growth learning of perception capabilities, and has good task performance and broad application prospects.

[0080] Furthermore, this embodiment also provides a self-calibration device for a multi-edge sensing device capable of growth learning, including a memory, a processor, and a program stored in the memory, which, when executed by the processor, implements the aforementioned methods. The device processor includes a central processing unit (CPU), which can perform various appropriate actions and processes according to computer program instructions stored in read-only memory (ROM) or loaded from the memory unit into random access memory (RAM). Various programs and data required for device operation can also be stored in the RAM. The CPU, ROM, and RAM are interconnected via a bus. Input / output (I / O) interfaces are also connected to the bus. Multiple components in the device are connected to the I / O interfaces, including: input units, such as a keyboard, mouse, etc.; output units, such as various types of displays, speakers, etc.; storage units, such as disks, optical disks, etc.; and communication units, such as network interface cards, modems, wireless transceivers, etc. The communication units allow the device to exchange information / data with other devices through computer networks such as the Internet and / or various telecommunications networks. The processing unit performs the various methods and processes described above, such as the steps described above. For example, in some embodiments, the foregoing steps can be implemented as a computer software program tangibly contained in a machine-readable medium, such as a storage unit. In some embodiments, part or all of the computer program can be loaded and / or installed on the device via ROM and / or a communication unit. When the computer program is loaded into RAM and executed by the CPU, one or more steps described above can be performed. Alternatively, in other embodiments, the CPU can be configured to perform the foregoing methods by any other suitable means (e.g., by means of firmware). The functions described above can be performed at least in part by one or more hardware logic components. For example, without limitation, exemplary types of hardware logic components that can be used include: Field Programmable Gate Arrays (FPGAs), Application-Specific Integrated Circuits (ASICs), Application Standard Products (ASSPs), System-on-Chip (SOCs), Complex Programmable Logic Devices (CPLDs), and so on.

[0081] Furthermore, this embodiment also provides a storage medium on which a program is stored, which, when executed, implements the aforementioned method. The program code for implementing the method of the present invention can be written in any combination of one or more programming languages. This program code can be provided to a processor or controller of a general-purpose computer, special-purpose computer, or other programmable data processing device, such that when executed by the processor or controller, the program code causes the functions / operations specified in the flowcharts and / or block diagrams to be implemented. The program code can be executed entirely on a machine, partially on a machine, partially on a machine and partially on a remote machine as a standalone software package, or entirely on a remote machine or server. In the context of this invention, a computer-readable medium can be a tangible medium that may contain or store a program for use by or in conjunction with an instruction execution system, apparatus, or device. A machine-readable medium can be a machine-readable signal medium or a machine-readable storage medium. A machine-readable medium can include, but is not limited to, electronic, magnetic, optical, electromagnetic, infrared, or semiconductor systems, apparatus, or devices, or any suitable combination thereof. More specific examples of machine-readable storage media include electrical connections based on one or more wires, portable computer disks, hard disks, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM or flash memory), optical fiber, portable compact disk read-only memory (CD-ROM), optical storage devices, magnetic storage devices, or any suitable combination of the foregoing.

[0082] The preferred embodiments of the present invention have been described in detail above. It should be understood that those skilled in the art can make numerous modifications and variations based on the concept of the present invention without creative effort. Therefore, all technical solutions that can be obtained by those skilled in the art based on the concept of the present invention through logical analysis, reasoning, or limited experimentation on the basis of existing technology should be within the scope of protection defined by the claims.

Claims

1. A self-calibration method for multi-edge sensing devices capable of growth learning, characterized in that, Includes the following steps: S1, based on the current calibration environment, acquire environmental perception calibration data; S2, based on the current application environment, obtain the real-time output response values ​​of each edge sensing device; S3, based on the environmental perception calibration data and the real-time uploaded output response values, the output response values ​​of the edge sensing device are converted into corresponding sensing physical quantities using a calibration model based on meta-learning and incremental learning, thus completing self-calibration. The process of acquiring the calibration model includes offline training based on meta-learning and cloud server optimization based on incremental learning. The offline training process includes: S301, acquire the offline output response value of each edge sensing device, and the standard reference value corresponding to each offline output response value; S302, Based on a deep learning framework, a neural network model is constructed using the offline output response value and the standard reference value; S303, The neural network model is trained using a meta-learning algorithm to obtain the trained calibration model; The cloud server optimization process specifically includes the following steps: S311, Receive and periodically preprocess the real-time uploaded output response values, and load the trained calibration model; S312, Define the loss function and training optimization algorithm for contrastive learning; S313, build a comparative learning framework, train the calibration model based on the real-time uploaded output response value, and optimize the parameters of the current calibration model; S314: Repeat S313 a preset number of times to obtain new calibration model parameters. When there is performance optimization in the calibration results, update the current calibration model parameters to obtain a new calibration model.

2. The self-calibration method for multi-edge sensing devices capable of growth learning according to claim 1, characterized in that, In step S2, when acquiring the real-time output response value of each edge sensing device, multiple samplings are performed, and the number of samplings is determined according to the range of the edge sensing device.

3. The self-calibration method for multi-edge sensing devices capable of growth learning according to claim 1, characterized in that, The measurement equipment in the calibration environment includes pre-arranged standardized test equipment and validated high-precision measurement tools.

4. The self-calibration method for multi-edge sensing devices capable of growth learning according to claim 1, characterized in that, The process of obtaining the offline output response value and the standard reference value includes: selecting multiple edge sensing devices, and each edge sensing device actively interacting with the measurement device in the calibration environment based on target driving to obtain the corresponding offline output response value and standard reference value.

5. The self-calibration method for multi-edge sensing devices capable of growth learning according to claim 1, characterized in that, Specifically, S302 includes: The parameters of the neural network model are initialized using a random number seed; The input vector is constructed using the offline output response value and the standard reference value.

6. The self-calibration method for multi-edge sensing devices capable of growth learning according to claim 5, characterized in that, Specifically, S303 includes: S3031, the input vector is divided into multiple task sets, the number of task sets is equal to the number of edge sensing devices, and each task set includes a support set and a query set in equal proportions; S3032, input the input vector of the support set of any task set into the current neural network model, obtain the corresponding loss function, and adjust the parameters of the current neural network model based on the current loss function; S3033, save the current neural network model parameters and load them into the current neural network model; S3034, Input the input vector of the query set of the same task set into the current neural network model to obtain the corresponding loss function; S3035, Repeat steps S3032 to S3034 until all task sets have been calculated, obtain the final total loss function, adjust and save the current meta-learning algorithm parameters based on the total loss function, and load them into the current neural network model; S3036, repeat steps S3032~S3035 a preset number of times to perform meta-learning training and obtain the final neural network model parameters.

7. The self-calibration method for multi-edge sensing devices capable of growth learning according to claim 6, characterized in that, The total loss function is the sum of the corresponding loss functions obtained in step S3034 for all task sets.

8. A self-calibration device for a multi-edge sensing device capable of growth learning, comprising a memory, a processor, and a program stored in the memory, characterized in that, When the processor executes the program, it implements the method as described in any one of claims 1-7.

9. A storage medium having a program stored thereon, characterized in that, When the program is executed, it implements the method as described in any one of claims 1-7.

Images