A temperature and humidity sensor based on intelligent calibration of the end-side AI chip

Abstract

The present invention relates to the field of sensors, and discloses a temperature and humidity sensor based on intelligent calibration of an end-side AI chip, including: a temperature sensor for collecting measurement data of ambient temperature; a humidity sensor for collecting measurement data of ambient humidity; an operational amplifier connected to the temperature sensor and the humidity sensor, and used to amplify and process the temperature and humidity measurement signals; a calibration register for storing initial calibration parameters and providing a calibration reference to an end-side AI computing processing unit; an analog-to-digital converter connected to the operational amplifier, and used to convert the analog signal collected by the temperature and humidity sensor into a digital signal; and a serial digital interface connected to the analog-to-digital converter, and used to transmit digitized temperature and humidity measurement data. By integrating the temperature and humidity sensor with the end-side AI chip, it is possible to support the operation of a neural network algorithm model with higher computational complexity, thereby improving the measurement accuracy of the temperature and humidity sensor.

Description

Technical Field

[0001] The present invention relates to the field of sensor technology, and specifically to a temperature and humidity sensor based on intelligent calibration of an end-side AI chip. Background Art

[0002] Temperature and humidity sensors have been widely used in agricultural planting, industrial production, weather forecasting, material storage and other scenarios. Measurement accuracy is one of the important indicators of temperature and humidity sensors. Current temperature and humidity sensors generally use dielectric parameter variation type capacitive sensors to measure humidity. The humidity sensitive unit is easily affected by external factors such as temperature, light intensity, atmospheric pressure, wind speed, dust and oil pollution, thereby reducing the accuracy of humidity measurement. Among them, the humidity sensitive unit is more sensitive to temperature. Changes in ambient temperature will cause obvious temperature drift in the humidity measurement results of the temperature and humidity sensor, which is the main factor affecting the humidity measurement accuracy.

[0003] In order to improve the measurement accuracy of temperature and humidity sensors, the industry generally adopts temperature compensation solutions. Temperature compensation includes two solutions: internal compensation and external compensation. Internal compensation is to reduce the influence of temperature by improving the structure, process and materials of the sensor, which has relatively large limitations. External compensation includes two methods: hardware compensation and software compensation. The hardware compensation method generally adds a hardware circuit with a negative temperature coefficient to the acquisition circuit system of the temperature and humidity sensor, and achieves the stability of the entire system through negative feedback. However, since the electronic components of the newly added hardware circuit will also produce new temperature drift, it is difficult to achieve full-scale compensation. The introduction of nonlinear correction circuits will also bring problems with system reliability, so it is difficult to implement and promote in practical applications. The software compensation method combines the temperature and humidity sensor with the operation processing unit, and corrects the temperature error of the sensor through a software algorithm.

[0004] Compared with the hardware compensation method, the software compensation method can give full play to the microcontroller's computing control and digital processing capabilities, and can ensure that the accuracy of the temperature and humidity sensor is met. Commonly used software compensation methods include piecewise linear interpolation table lookup method, multidimensional regression analysis method, least squares polynomial curve fitting method, BP neural network method, support vector machine (SVM) method and other algorithms.

[0005] At present, integrated temperature and humidity sensors use semiconductor integration technology to integrate functional devices such as humidity sensitive elements, temperature sensitive elements, A / D conversion units, and signal processing units into a single chip, thereby realizing the software compensation function of the temperature and humidity sensor. The temperature and humidity sensor chips in the industry generally use calibration registers to realize the calibration of humidity measurement. When measuring temperature and humidity, the system software will automatically read the calibration coefficients pre-stored in the register to calibrate the temperature and humidity sensor, and finally output the final calibrated digital quantity through the signal processing unit. Due to the limited computing power of the signal processing unit, most temperature and humidity sensor manufacturers currently use interpolation table lookup method and curve fitting method. However, the compensation accuracy of the interpolation method is related to the number of calibration data points, and the number of data points is difficult to achieve sufficient measurement accuracy; the least squares polynomial curve fitting rule requires solving a large-scale matrix equation, and when there are many calibration data points, the matrix equation required is easy to become an ill-conditioned equation or a singular equation, which makes it impossible to solve.

[0006] It can be seen that the calibration scheme based on the on-chip calibration register for software compensation used in existing temperature and humidity sensors has certain defects and cannot meet the actual needs of high-precision temperature and humidity measurement. Summary of the invention

[0007] In view of the shortcomings of the prior art, the present invention provides a temperature and humidity sensor based on intelligent calibration of an end-side AI chip, which solves the problem of insufficient calibration accuracy of existing temperature and humidity sensors under ambient temperature, humidity drift and interference from multiple external parameters.

[0008] To achieve the above objectives, the present invention is implemented through the following technical solutions: a temperature and humidity sensor based on intelligent calibration of an end-side AI chip, comprising:

[0009] Temperature sensor, used to collect measurement data of ambient temperature;

[0010] Humidity sensor, used to collect measurement data of ambient humidity;

[0011] An operational amplifier, connected to the temperature sensor and the humidity sensor, and used to amplify the temperature and humidity measurement signals;

[0012] A calibration register, used to store initial calibration parameters and provide a calibration reference to the end-side AI computing processing unit;

[0013] An analog-to-digital converter, connected to the operational amplifier, for converting the analog signal collected by the temperature and humidity sensor into a digital signal;

[0014] A serial digital interface, connected to the analog-to-digital converter, for transmitting digitized temperature and humidity measurement data;

[0015] An end-side AI computing processing unit is connected to the serial digital interface. The end-side AI computing processing unit is integrated with a deep neural network model, which is used to receive temperature and humidity measurement data, perform intelligent calibration based on the calibration parameters in the calibration register, and output the calibrated measurement results.

[0016] Preferably, the end-side AI computing processing unit includes:

[0017] Model storage unit, used to store weight files and intermediate representation files of deep neural network models;

[0018] A neural network accelerator, connected to the model storage unit, for loading a weight file of the deep neural network model and performing model inference calculations;

[0019] The data interface unit is used to receive the calibrated temperature and humidity measurement data and output it through the digital interface.

[0020] Preferably, the initial calibration parameters stored in the calibration register are generated by combining hardware compensation and software compensation, wherein:

[0021] Hardware compensation reduces the impact of temperature drift through the negative feedback circuit of the operational amplifier;

[0022] Software compensation is completed by the end-side AI computing processing unit after calibrating the temperature and humidity measurement values ​​based on the deep neural network model.

[0023] Preferably, the deep neural network model is a radial basis function neural network, and the radial basis function neural network includes:

[0024] The input layer is used to receive the digital measurement data from the temperature sensor and the humidity sensor;

[0025] The hidden layer includes a plurality of radial basis function units for realizing the characteristic nonlinear mapping of the measurement data;

[0026] The output layer is used to output the calibrated temperature and humidity measurement results.

[0027] Preferably, the deep neural network model is deployed through the following steps:

[0028] Performing full-precision training on the deep neural network model based on the training data set to obtain calibration weights;

[0029] According to the bit width requirement supported by the AI ​​computing processing unit on the end side, fixed-point quantization training is performed on the weight file to adjust the weight accuracy;

[0030] The quantized weight file and intermediate representation file are loaded into the model storage unit of the end-side AI computing processing unit.

[0031] Preferably, the input data received by the end-side AI computing processing unit includes:

[0032] Digital measurements from temperature and humidity sensors;

[0033] External parameter data, including but not limited to light intensity, atmospheric pressure, wind speed, dust and oil pollution, etc.;

[0034] The deep neural network model achieves accurate calibration of temperature and humidity measurement values ​​by fusing multiple parameters.

[0035] Preferably, the quantization accuracy of the analog-to-digital converter is 24 bits, and the digital interface is an I2C or SPI bus interface.

[0036] Preferably, the end-side AI computing processing unit adopts system-level packaging technology and is integrated with the temperature sensor, humidity sensor, operational amplifier and calibration register in the same package.

[0037] Preferably, the training data set of the deep neural network model is generated by a temperature and humidity standard box, and the data set covers the following range:

[0038] The temperature range is -40℃-85℃, and the sampling interval is 0.1℃;

[0039] The humidity range is 10%RH-95%RH, and the sampling interval is 1%RH.

[0040] A method for intelligently calibrating temperature and humidity measurement values ​​based on an end-side AI chip, comprising the following steps:

[0041] S1, obtain the measured values ​​of the temperature sensor and the humidity sensor;

[0042] S2, converting the measured value into a digital signal through an analog-to-digital converter;

[0043] S3. Transmitting the digital signal to the end-side AI computing processing unit through a serial digital interface;

[0044] S4. Loading a deep neural network model into the end-side AI computing processing unit, performing inference calculations, and generating calibrated temperature and humidity measurement results;

[0045] S5. Output the calibrated result through the digital interface.

[0046] The present invention provides a temperature and humidity sensor based on intelligent calibration of an end-side AI chip. It has the following beneficial effects:

[0047] The present invention integrates the temperature and humidity sensor with the end-side AI chip, and deploys the temperature and humidity calibration algorithm model based on the deep neural network to the end-side AI chip through the model lightweight technology. The end-side AI chip calibrates the temperature and humidity sensor measurement value and then outputs it. Compared with the traditional temperature and humidity sensor solution based on the internal calibration register, the end-side AI chip calibration solution adopted by the present invention can support the neural network algorithm model with higher computational complexity, thereby improving the measurement accuracy of the temperature and humidity sensor. At the same time, the end-side AI chip of the present invention can support measurement data and calculation parameters other than the temperature and humidity sensor as the input of the algorithm model, such as external parameters such as light intensity, atmospheric pressure, wind speed, dust and oil pollution, and realizes higher-precision calibration of the temperature and humidity sensor through multi-parameter fusion, which has high engineering application value. BRIEF DESCRIPTION OF THE DRAWINGS

[0048] Figure 1 This is a functional block diagram of the temperature and humidity chip system based on the intelligent calibration of the end-side AI chip of the present invention;

[0049] Figure 2 It is the RBF network structure diagram of the present invention;

[0050] Figure 3 It is a result comparison diagram of the calibration prediction value and the true value of the present invention;

[0051] Figure 4 This is a flow chart of the temperature and humidity chip operation based on the intelligent calibration of the end-side AI chip of the present invention. DETAILED DESCRIPTION

[0052] The technical solution of the present invention will be described clearly and completely below in conjunction with the accompanying drawings of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0053] Embodiment 1:

[0054] Please see attached Figure 1 The embodiment of the present invention provides a temperature and humidity sensor based on intelligent calibration of a terminal-side AI chip, including:

[0055] Temperature sensor, used to collect measurement data of ambient temperature;

[0056] Humidity sensor, used to collect measurement data of ambient humidity;

[0057] An operational amplifier, connected to the temperature sensor and the humidity sensor, for amplifying the temperature and humidity measurement signals;

[0058] A calibration register, used to store initial calibration parameters and provide a calibration reference to the end-side AI computing processing unit;

[0059] An analog-to-digital converter, connected to the operational amplifier, for converting the analog signal collected by the temperature and humidity sensor into a digital signal;

[0060] A serial digital interface, connected to an analog-to-digital converter, is used to transmit digital temperature and humidity measurement data;

[0061] The end-side AI computing processing unit is connected to the serial digital interface. A deep neural network model is integrated in the end-side AI computing processing unit to receive temperature and humidity measurement data, perform intelligent calibration based on the calibration parameters in the calibration register, and output the calibrated measurement results.

[0062] Specifically, first, the temperature and humidity standard box is used to calibrate the measurement values ​​of the original temperature and humidity chip, and a temperature and humidity data calibration data set is built. During the test, data is collected once every 0.1°C and 1%RH change in the temperature range of -40°C to 85°C and the humidity range of 10%RH to 95%RH, and a data set containing temperature calibration values ​​and measurement values, humidity calibration values ​​and measurement values ​​is obtained, and divided into training sets and test sets. The data in the data set is normalized to improve the accuracy of model training.

[0063] Then, a temperature and humidity calibration algorithm model based on a neural network is constructed. In this embodiment, an RBF (radial basis function) neural network is used to construct the calibration algorithm model. RBF (radial basis function) neural network is an artificial neural network, and its hidden layer uses radial basis function as the activation function. The RBF network has a three-layer structure: input layer, radial basis function hidden layer, and linear output layer. Taking temperature calibration as an example, the input of the RBF neural network is 2-dimensional, the hidden layer has 2 neurons, and the output layer has 1 neuron. For the constructed network structure diagram, please refer to Figure 2 .

[0064] The propagation speed of the radial basis function is set to 1000, and the RBF neural network is trained with full precision. After the training is completed, the relevant errors such as the root mean square value (RMSE) and the mean square error (MSE) are calculated based on the test data set.

[0065] Select the end-side AI neural network acceleration chip and integrate the temperature and humidity sensor chip with the end-side neural network accelerator chip for chip-level system integration.

[0066] In this embodiment, the end-side neural network accelerator chip supports 4-bit or 8-bit quantization bit widths for weights and feature maps. Therefore, an 8-bit quantization bit width is selected to perform fixed-point quantization training on the RBF neural network model, and the RBF neural network model is deployed on the end-side AI neural network accelerator chip using a tool chain. When reading the temperature and humidity sensor measurement data, the end-side AI neural network accelerator chip will first read the measurement value of the temperature and humidity chip, and then output it after calibration through the RBF neural network.

[0067] After completing the above operations, the temperature and humidity sensor integrating the end-side AI chip and the calibration algorithm model is tested and verified based on the temperature and humidity standard box. The true value of temperature calibration and the calibrated measurement prediction value are collected, and the root mean square error (RMSE) between the calibration prediction value and the true value is calculated to be 0.11567, which can verify the effectiveness of the calibration of the temperature and humidity chip based on the end-side AI chip proposed in the present invention.

[0068] Please refer to the attached for comparison of the calibration predictions and true values. Figure 3 .

[0069] Embodiment 2:

[0070] Please see attached Figure 4 , an embodiment of the present invention provides a method for intelligently calibrating temperature and humidity measurement values ​​based on a terminal-side AI chip, comprising the following steps:

[0071] S1, obtain the measured values ​​of the temperature sensor and the humidity sensor;

[0072] S2, converting the measured value into a digital signal through an analog-to-digital converter;

[0073] S3, transmitting the digital signal to the end-side AI computing processing unit through a serial digital interface;

[0074] S4. Load the deep neural network model in the end-side AI computing processing unit, perform inference calculations, and generate calibrated temperature and humidity measurement results;

[0075] S5. Output the calibrated result through the digital interface.

[0076] Specifically, the steps for calibrating the temperature and humidity sensor based on the end-side AI chip are as follows:

[0077] (1) Constructing training and test data sets: Measure and record the temperature and humidity calibration points at different temperature and humidity standard points and the temperature and humidity measurement values ​​of the temperature and humidity sensors to construct training and test data sets;

[0078] (2) Build a temperature and humidity calibration algorithm model based on a neural network: Use the network operators (such as preprocessing, convolution, deep separation convolution, fully connected layer, global average pooling, post-processing, etc.) supported by the end-side AI neural network accelerator chip to build a neural network algorithm model (including but not limited to BP neural network, CNN neural network, radial basis function neural network, ResNet neural network, Autoencoder neural network, etc.);

[0079] (3) Full-precision training of the algorithm model: Use full-precision training of the neural network model based on the training data set and record the calibration accuracy of the model;

[0080] (3) Fixed-point quantization training of the algorithm model: The specific quantization bit width is selected based on the weight bit width supported by the end-side AI neural network accelerator chip, actual requirements, and the accuracy results of quantization training. Fixed-point quantization training is performed on the neural network model to obtain the quantized weight file and the intermediate representation file of the network model.

[0081] (4) Algorithm model deployment: The weights in the weight file are converted into the required weight storage format. At the same time, the neural network model is parsed layer by layer according to the content of the intermediate representation file to generate calculation instructions. The converted weight file is downloaded to the Flash and the calculation instructions are loaded into the instruction memory of the accelerator to complete the deployment of the neural network model on the end-side AI neural network accelerator chip.

[0082] (5) Algorithm model reasoning: When reading the temperature and humidity sensor measurement data, the end-side AI neural network accelerator chip first writes the data to be inferred (including characteristic data such as temperature, humidity, etc.) into the accelerator's feature map memory through the bus, and starts the accelerator for calculation. After the calculation is completed, the calibrated calculation result is read from the result register and the result is output through the digital interface.

[0083] Although embodiments of the present invention have been shown and described, it will be appreciated by those skilled in the art that various changes, modifications, substitutions and variations may be made to the embodiments without departing from the principles and spirit of the present invention, and that the scope of the present invention is defined by the appended claims and their equivalents.

Claims

1. A temperature and humidity sensor based on intelligent calibration of an end-side AI chip, characterized in that: include: Temperature sensor, used to collect measurement data of ambient temperature; Humidity sensor, used to collect measurement data of ambient humidity; An operational amplifier, connected to the temperature sensor and the humidity sensor, and used to amplify the temperature and humidity measurement signals; A calibration register, used to store initial calibration parameters and provide a calibration reference to the end-side AI computing processing unit; An analog-to-digital converter, connected to the operational amplifier, for converting the analog signal collected by the temperature and humidity sensor into a digital signal; A serial digital interface, connected to the analog-to-digital converter, for transmitting digitized temperature and humidity measurement data; An end-side AI computing processing unit is connected to the serial digital interface. The end-side AI computing processing unit is integrated with a deep neural network model, which is used to receive temperature and humidity measurement data, perform intelligent calibration based on the calibration parameters in the calibration register, and output the calibrated measurement results.

2. According to claim 1, a temperature and humidity sensor based on intelligent calibration of a terminal AI chip is characterized in that: The end-side AI computing processing unit includes: Model storage unit, used to store weight files and intermediate representation files of deep neural network models; A neural network accelerator, connected to the model storage unit, for loading a weight file of the deep neural network model and performing model inference calculations; The data interface unit is used to receive the calibrated temperature and humidity measurement data and output it through the digital interface.

3. According to claim 1, a temperature and humidity sensor based on intelligent calibration of a terminal AI chip is characterized in that: The initial calibration parameters stored in the calibration register are generated by combining hardware compensation and software compensation, wherein: Hardware compensation reduces the impact of temperature drift through the negative feedback circuit of the operational amplifier; Software compensation is completed by the end-side AI computing processing unit after calibrating the temperature and humidity measurement values ​​based on the deep neural network model.

4. According to claim 1, a temperature and humidity sensor based on intelligent calibration of a terminal AI chip is characterized in that: The deep neural network model is a radial basis function neural network, and the radial basis function neural network includes: The input layer is used to receive the digital measurement data from the temperature sensor and the humidity sensor; The hidden layer includes a plurality of radial basis function units for realizing the characteristic nonlinear mapping of the measurement data; The output layer is used to output the calibrated temperature and humidity measurement results.

5. According to claim 1, a temperature and humidity sensor based on intelligent calibration of a terminal AI chip is characterized in that: The deep neural network model is deployed through the following steps: Performing full-precision training on the deep neural network model based on the training data set to obtain calibration weights; According to the bit width requirement supported by the AI ​​computing processing unit on the end side, fixed-point quantization training is performed on the weight file to adjust the weight accuracy; The quantized weight file and intermediate representation file are loaded into the model storage unit of the end-side AI computing processing unit.

6. According to claim 1, a temperature and humidity sensor based on intelligent calibration of a terminal AI chip is characterized in that: The input data received by the end-side AI computing processing unit includes: Digital measurements from temperature and humidity sensors; External parameter data, including but not limited to light intensity, atmospheric pressure, wind speed, dust and oil pollution, etc.; The deep neural network model achieves accurate calibration of temperature and humidity measurement values ​​by fusing multiple parameters.

7. According to claim 1, a temperature and humidity sensor based on intelligent calibration of a terminal AI chip is characterized in that: The quantization accuracy of the analog-to-digital converter is 24 bits, and the digital interface is an I2C or SPI bus interface.

8. According to claim 1, a temperature and humidity sensor based on intelligent calibration of a terminal AI chip is characterized in that: The end-side AI computing processing unit adopts system-level packaging technology and is integrated with the temperature sensor, humidity sensor, operational amplifier and calibration register in the same package.

9. According to claim 1, a temperature and humidity sensor based on intelligent calibration of a terminal AI chip is characterized in that: The training data set of the deep neural network model is generated by a temperature and humidity standard box, and the data set covers the following ranges: The temperature range is -40℃-85℃, and the sampling interval is 0.1℃; The humidity range is 10%RH-95%RH, and the sampling interval is 1%RH.

10. A method for intelligently calibrating temperature and humidity measurement values ​​based on an end-side AI chip, according to a temperature and humidity sensor based on intelligent calibration of an end-side AI chip according to any one of claims 1-9, characterized in that: The following steps are involved: S1, obtain the measured values ​​of the temperature sensor and the humidity sensor; S2, converting the measured value into a digital signal through an analog-to-digital converter; S3. Transmitting the digital signal to the end-side AI computing processing unit through a serial digital interface; S4. Loading a deep neural network model into the end-side AI computing processing unit, performing inference calculations, and generating calibrated temperature and humidity measurement results; S5. Output the calibrated result through the digital interface.

Images