Real-time adaptive temperature monitoring intelligent sensor based on deep learning algorithm

Abstract

The invention discloses a real-time adaptive temperature monitoring intelligent sensor based on a deep learning algorithm, and belongs to the technical field of environmental parameter measurement and artificial intelligence. The sensor is composed of a main temperature measuring unit, an auxiliary environment sensing unit, a data acquisition circuit, a microcontroller, a deep learning processing module, a memory, a communication module and a power supply module. Multi-modal data are obtained through the auxiliary environment sensing unit, and the multi-modal data are preprocessed and then input into the self-adaptive temperature sensing network. In the adaptive temperature sensing network, a multi-mode encoder extracts time sequence characteristics of each channel, a convolution auto-encoder reconstructs a normal temperature measurement mode and performs anomaly detection, and a fusion decoder fuses output of each sensor into a corrected temperature value based on a dynamic weight formula. The method has the advantages of high real-time performance, high measurement precision, timely anomaly detection, high self-learning capability, low power consumption and the like.

Description

Technical Field

[0001] This invention relates to the fields of environmental parameter measurement and artificial intelligence technology, specifically to a real-time adaptive temperature monitoring smart sensor based on a deep learning algorithm. Background Technology

[0002] Temperature monitoring is a fundamental task in fields such as industrial control, smart buildings, vaccine cold chain, food safety, and biohealth. It requires sensors to not only possess high accuracy but also maintain consistent measurement results under different environments and installation methods. Existing technologies include research using deep learning to analyze camera images to estimate animal body temperature. For example, a patent from Northeast Agricultural University uses an infrared thermal imager to continuously capture images of a pig's face from the front and from above, employing a YOLO-v4-like model to identify the pig's face, forehead, and ear roots, extracting the temperature of the corresponding areas, and issuing an alert to management when abnormal body temperature is detected. This solution overcomes the contact and cross-infection problems of traditional rectal temperature measurement; however, its core idea of ​​inferring the temperature of specific areas through image processing is not suitable for general environmental temperature monitoring.

[0003] In the field of building environmental control, a patent utilizes deep learning to detect the number of people in office areas and automatically adjust the air conditioning temperature. This solution uses cameras to collect images, builds a dataset, and trains a YOLO-v3 detection model. When an empty office space is detected, the air conditioning is automatically turned off, thus reducing energy waste. This vision-based control method improves energy efficiency, but its sensing object is human activity rather than the ambient temperature itself.

[0004] Researchers have also explored using deep learning to improve the anomaly detection capabilities of temperature sensors. In vaccine cold chain monitoring, researchers proposed a system based on a semi-supervised convolutional autoencoder (CAE). This system reconstructs temperature sensor data on a low-power ESP32 microcontroller, identifying anomalies when the reconstruction error exceeds a threshold. The system trains its model using real temperature data and can identify temperature faults with approximately 92% accuracy, while consuming only 0.05W and requiring only 1.2MB of memory, making it suitable for resource-constrained environments. This method emphasizes anomaly detection through learning normal patterns, but its output is an anomaly flag, failing to directly improve real-time measurement accuracy and neglecting multi-sensor fusion calibration.

[0005] Furthermore, modern low-cost sensors are significantly affected by environmental factors such as temperature and humidity. Researchers have used machine learning to calibrate air quality sensors. Related literature has constructed acquisition systems for PM2.5, CO2, temperature, and humidity, using algorithms such as decision trees, random forests, k-nearest neighbors, and gradient boosting to map sensor outputs to reference instruments. Results show that gradient boosting and k-nearest neighbors significantly improve the accuracy of temperature and humidity sensors; the coefficient of determination for the temperature sensor increased to 0.976 after calibration. This demonstrates that machine learning can be used for the calibration of low-cost sensors, but existing methods mainly target static or batch data and lack real-time adaptive capabilities in complex environments.

[0006] In summary, current technologies have not proposed an integrated intelligent sensor that combines real-time measurement, multi-sensor fusion, self-learning, and environmental adaptability. Especially in complex industrial or public environments, traditional temperature sensors are susceptible to environmental noise, airflow, personnel movement, and sensor drift, lacking adaptive compensation mechanisms and unable to provide high-precision data while ensuring low power consumption. Summary of the Invention

[0007] Technical Objective: To address the shortcomings of existing technologies, such as low measurement accuracy, poor environmental adaptability, and untimely response, this invention discloses a real-time adaptive temperature monitoring smart sensor based on deep learning algorithms. This sensor achieves high-precision temperature measurement through multi-source temperature information fusion and self-learning algorithms, and can correct itself in real time according to environmental factors, promptly detecting abnormal states.

[0008] Technical solution: To achieve the above technical objectives, the present invention adopts the following technical solution:

[0009] A real-time adaptive temperature monitoring smart sensor based on deep learning algorithms includes a main housing, a main temperature measurement unit, an auxiliary environmental sensing unit, a data acquisition circuit, a microcontroller, a deep learning processing module, a memory, a communication module, and a power supply module.

[0010] The main body housing contains a main temperature measurement unit, an auxiliary environmental sensing unit, a data acquisition circuit, a microcontroller, a deep learning processing module, a memory, a communication module, and a power supply module.

[0011] The main temperature measurement unit is connected to the microcontroller via a data acquisition circuit to acquire raw temperature signals;

[0012] The auxiliary environmental sensing unit includes humidity, airflow, light, vibration, sound, and human presence sensors to collect environmental information;

[0013] The microcontroller has a built-in deep learning processing module that integrates the raw temperature signal and environmental information by running an adaptive temperature sensing network, outputs the corrected temperature value, and sends it through the communication module.

[0014] Preferably, the deep learning processing module includes a multimodal encoder, a convolutional autoencoder, and an adaptive fusion decoder; the multimodal encoder is used to extract the time and frequency features of each sensing channel and output a feature matrix, the convolutional autoencoder is used to reconstruct the feature matrix and calculate the reconstruction error to determine anomalies, and the adaptive fusion decoder is used to calculate the corrected temperature value based on dynamic weights.

[0015] Preferably, an incremental learning mechanism is established between the microcontroller and the memory. When incremental calibration data is acquired, the fusion weights and bias parameters are updated using the mini-batch gradient descent method, while abnormal samples are removed using a convolutional autoencoder to achieve model self-calibration.

[0016] Preferably, the multimodal encoder includes a one-dimensional convolutional layer and a gated recurrent unit for each sensing channel. The one-dimensional convolutional layer is used to extract local features, and the gated recurrent unit is used to capture the time dependence of temperature changes.

[0017] Preferably, the convolutional autoencoder is obtained through pre-training and minimizes the reconstruction error using normal temperature measurement data. When the real-time reconstruction error exceeds a set threshold, an abnormality flag is output.

[0018] Preferably, the adaptive fusion decoder calculates the corrected temperature value using the following formula:

[0019] ,

[0020] in, Let be the measurement value of the i-th sensor at time t. and Let be the mean and standard deviation of the i-th sensor within a preset time window, respectively. Here, b represents the dynamic weight, b is the bias, and N is the total number of sensors.

[0021] The dynamic weights are updated according to the following formula:

[0022] ,

[0023] in, For learning rate, For loss function, For environmental adaptability coefficient, This is the environmental error measure at time t.

[0024] Preferably, the loss function The weighted mean square error is adjusted based on real-time environmental noise. Calculate using the following formula:

[0025] ,

[0026] in Let be the measurement value of the j-th auxiliary sensor at time t. Let j be the reference value for the j-th auxiliary sensor under standard conditions. Let K be the environmental impact coefficient of the j-th auxiliary sensor, and K be the number of auxiliary sensors.

[0027] This invention also discloses a real-time adaptive temperature monitoring method, characterized by applying the aforementioned real-time adaptive temperature monitoring smart sensor based on a deep learning algorithm, comprising the following steps:

[0028] The raw data from the main temperature measurement unit and the auxiliary environmental sensing unit are acquired in parallel.

[0029] The collected raw data is normalized, filtered, and feature extracted to generate multimodal feature vectors;

[0030] The multimodal feature vectors are fed into the multimodal encoder to obtain the encoded feature matrix;

[0031] The coded feature matrix is ​​reconstructed using a convolutional autoencoder, the reconstruction error is calculated, and an alarm is output when there is an anomaly.

[0032] The adaptive fusion decoder fuses signals from various sensors based on a dynamic weighting formula and calculates the corrected temperature value.

[0033] Online learning is performed based on the collected calibration data to update the model weights and biases;

[0034] The corrected temperature value is output or sent to the host computer or cloud via the communication module.

[0035] Preferably, the dynamic weight update includes an environmental error compensation term to enhance adaptability to changes in humidity, airflow, light, and human activity environment.

[0036] Preferably, the sampling frequency of the main temperature measurement unit is 1Hz, and the sampling frequency of the auxiliary environmental perception unit is 0.5Hz; online learning only adjusts some parameters of the adaptive fusion decoder while keeping the encoder structure unchanged.

[0037] Beneficial Effects: The real-time adaptive temperature monitoring smart sensor based on deep learning algorithm provided by this invention has the following beneficial effects:

[0038] 1. This invention utilizes an adaptive temperature sensing network to perform deep fusion and dynamic calibration of multi-source sensor data, achieving real-time, high-precision measurement of ambient temperature. It does not rely on a single sensor or image recognition, overcoming the problem of traditional temperature measurement devices being greatly affected by environmental interference. The integrated convolutional autoencoder enables real-time detection of sensor and environmental anomalies. Combined with an online learning mechanism, it can automatically correct model parameters based on calibration data, extending the lifespan of the sensor. Attached Figure Description

[0039] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the accompanying drawings used in the description of the embodiments or the prior art will be briefly introduced below.

[0040] Figure 1 This is a schematic diagram of the overall structure of the intelligent sensor of the present invention;

[0041] Figure 2 This is a flowchart of the real-time adaptive temperature monitoring method of the present invention;

[0042] Figure 3 This is a structural diagram of the adaptive temperature sensing network of the present invention.

[0043] In the diagram: 1. Main temperature measurement unit; 2. Auxiliary environmental sensing unit; 3. Data acquisition circuit; 4. Microcontroller; 5. Deep learning processing module; 6. Memory; 7. Communication module; 8. Power supply module. Detailed Implementation

[0044] The present invention will now be described more clearly and completely by way of a preferred embodiment in conjunction with the accompanying drawings, but this does not limit the invention to the scope of the described embodiment.

[0045] This invention discloses a real-time adaptive temperature monitoring smart sensor based on deep learning algorithms, comprising a main housing, a main temperature measurement unit, an auxiliary environmental sensing unit, a data acquisition circuit, a microcontroller, a deep learning processing module, a memory, a communication module, and a power supply module;

[0046] The main body housing contains a main temperature measurement unit, an auxiliary environmental sensing unit, a data acquisition circuit, a microcontroller, a deep learning processing module, a memory, a communication module, and a power supply module.

[0047] The main temperature measurement unit is connected to the microcontroller through a data acquisition circuit to acquire raw temperature signals. It uses a high-precision thermistor or platinum resistance thermometer and is equipped with a multi-channel analog-to-digital converter.

[0048] The auxiliary environmental sensing unit includes humidity, airflow, light, vibration, sound, and human presence sensors to collect environmental information;

[0049] The microcontroller has a built-in deep learning processing module that integrates the raw temperature signal and environmental information by running an adaptive temperature sensing network, outputs the corrected temperature value and sends it through the communication module.

[0050] The communication module can use Wi-Fi, Bluetooth, or LoRa for remote data transmission;

[0051] The power module can be powered by a battery or an external power source and includes energy management circuitry.

[0052] The sensors collect raw temperature data from the main temperature measurement unit and environmental information from the auxiliary environmental sensing unit in parallel according to a set sampling frequency, and normalize, filter, and suppress noise in the collected data. Simultaneously, short-time Fourier transform is used to generate the frequency domain features of the temperature sequence to enrich the data dimensions.

[0053] This invention proposes an adaptive temperature sensing network, comprising three parts:

[0054] Multimodal encoders: These encoders use convolutional neural networks (CNNs) to extract features from signals from different sensors and input temporal information into gated recurrent units (GRUs) or long short-term memory networks (LSTMs) to capture the time dependence of environmental temperature changes. Unlike traditional LSTMs, lightweight convolutional gate structures with parameter sharing are introduced to reduce computational load, making them suitable for low-power chips.

[0055] Convolutional autoencoder: A pre-trained convolutional autoencoder is embedded at the edge to reconstruct the normal temperature measurement pattern and calculate the reconstruction error. When the error exceeds the threshold, an abnormal label is output to promptly indicate sensor or environmental abnormalities.

[0056] Adaptive fusion decoder: The high-dimensional features from the encoder are input into the decoder for multi-source data fusion, generating a corrected temperature output and dynamic weights. The decoder internally designs a dynamic weight update equation:

[0057] ,

[0058] in, For learning rate, For loss function, For environmental adaptability coefficient, This represents the environmental error measure at time t, such as the impact of humidity and wind speed changes on temperature measurement. The weights in this formula are not only updated based on the loss gradient but also include an environmental error increment compensation term to ensure the model can quickly adapt to environmental changes.

[0059] The corrected temperature value output by the adaptive fusion decoder is:

[0060] ,

[0061] in, Let be the measurement value of the i-th sensor at time t. and These are the mean and standard deviation of the i-th sensor within a preset time window, respectively, used to eliminate dimensional differences between different sensors. is the dynamic weight, b is the bias term obtained through training, and N is the total number of sensors.

[0062] To address long-term drift and environmental changes, this invention employs an incremental learning mechanism. When the sensor is running, a portion of the sampled data is calibrated manually or using a high-precision reference instrument and used as incremental data. The model is then fine-tuned online using the aforementioned weight update formula to avoid forgetting old knowledge. A convolutional autoencoder is used to filter outomaly samples, ensuring that the model is updated only on reliable data.

[0063] The deep learning processing module outputs the corrected temperature value to a display or host computer, and can also upload it to the cloud or local area network via the communication module. If the convolutional autoencoder detects an anomaly, it triggers an alarm interface or automatically executes a self-test program. The sensor supports remote firmware upgrades and model updates to optimize the algorithm.

[0064] like Figure 2 As shown, this invention also discloses a real-time adaptive temperature monitoring method, characterized by applying the aforementioned real-time adaptive temperature monitoring smart sensor based on a deep learning algorithm, comprising the following steps:

[0065] The raw data from the main temperature measurement unit and the auxiliary environmental sensing unit are acquired in parallel.

[0066] The collected raw data is normalized, filtered, and feature extracted to generate multimodal feature vectors;

[0067] The multimodal feature vectors are fed into the multimodal encoder to obtain the encoded feature matrix;

[0068] The coded feature matrix is ​​reconstructed using a convolutional autoencoder, the reconstruction error is calculated, and an alarm is output when there is an anomaly.

[0069] The adaptive fusion decoder fuses signals from various sensors based on a dynamic weighting formula and calculates the corrected temperature value.

[0070] Online learning is performed based on the collected calibration data to update the model weights and biases;

[0071] The corrected temperature value is output or sent to the host computer or cloud via the communication module.

[0072] Example 1

[0073] like Figure 1As shown, the main housing of the intelligent sensor in this embodiment adopts a sealed and waterproof structure, and internally houses a main temperature measurement unit 1, an auxiliary environmental sensing unit 2, a data acquisition circuit 3, a microcontroller 4, a deep learning processing module 5, a memory 6, a communication module 7, and a power supply module 8. The main temperature measurement unit 1 uses a high-precision PT1000 platinum resistance thermometer, connected to the data acquisition circuit 3 via a four-wire connection. The auxiliary environmental sensing unit 2 includes a relative humidity sensor, an airflow sensor, a photoresistor, a microphone, and a human presence sensor, used to detect parameters such as environmental humidity, airflow, light intensity, noise, and human activity. The data acquisition circuit 3 consists of a multiplexer, an anti-aliasing filter, and a 24-bit analog-to-digital converter, capable of simultaneously acquiring data from each channel at different sampling frequencies. The microcontroller 4 uses a low-power 32-bit RISC chip, integrating a neural network acceleration unit, responsible for executing deep learning algorithms and controlling data flow. The deep learning processing module 5 is a software model configured in the memory of the microcontroller 4. The memory 6 is used to store algorithm weights, temporary data, and calibration data; the communication module 7 supports both Wi-Fi and Bluetooth dual-mode to transmit data to the host computer or cloud; the power module 8 includes a rechargeable battery, a voltage regulator chip, and an energy management circuit, and supports a low-power sleep mode.

[0074] Example 2

[0075] This embodiment details the data acquisition and preprocessing of the sensor of the present invention. After the sensor is started, the microcontroller 4 controls the data acquisition circuit 3 to read the signals from each sensor according to the set sampling period. Considering that the temperature changes slowly, this embodiment sets the sampling frequency of the main temperature measurement unit to 1Hz and the sampling frequency of the auxiliary environmental sensing unit to 0.5Hz. After the acquisition is completed, the microcontroller 4 performs the following preprocessing:

[0076] Normalization: Calculate the mean for each channel of data using a real-time sliding window. Sum and standard deviation The original data is normalized to ensure that the dimensions of each channel are consistent.

[0077] Noise filtering: Kalman filtering is used to remove high-frequency noise from temperature data; median filtering is used to suppress abrupt interference from humidity, airflow and other data; low-frequency energy is extracted from microphone data to determine the level of environmental noise.

[0078] Feature extraction: Short-time Fourier transform is used to calculate the frequency distribution of the temperature sequence in the last 60 seconds to reflect the rate of temperature change; moving average is applied to the human body sensor output to obtain activity level features; these features, together with normalized data, constitute a multimodal input vector.

[0079] The multimodal data obtained after the above preprocessing is sent to the deep learning processing module 5.

[0080] Example 3

[0081] like Figure 3 As shown, the adaptive temperature sensing network of the present invention consists of three parts: a multimodal encoder, a convolutional autoencoder, and an adaptive fusion decoder.

[0082] Multimodal encoder:

[0083] Each sensing channel corresponds to a set of one-dimensional convolutional layers used to extract local temporal features. The kernel size is selected based on the physical characteristics of each sensor. For example, the kernel size for the temperature channel is 5, for the humidity channel it is 3, and for the airflow channel it is 7. After the output of the convolutional layers, the feature sequence is obtained through batch normalization and the ReLU activation function.

[0084] The feature sequences of each channel are concatenated at time steps and fed into a lightweight gated recurrent unit (GRU) network. The GRU contains two gating structures, capturing long-range dependencies while reducing computational cost. Compared to traditional LSTM, the GRU requires fewer parameters, making it suitable for embedded deployments.

[0085] The feature matrix H(t) output by the GRU is used as the output of the encoder and passed to the fusion decoder and autoencoder module.

[0086] Convolutional autoencoder:

[0087] A convolutional autoencoder consists of a symmetric encoder and decoder. The encoder part comprises several one-dimensional convolutional and pooling layers, compressing the input feature matrix H(t) into a low-dimensional latent vector z. The decoder part reconstructs the features using upsampling and deconvolutional layers. Calculate the reconstruction error. .when Exceeding the empirical threshold When an anomaly is detected, the sensor is deemed to have malfunctioned. During the model training phase, the model learns by minimizing the reconstruction error using a normal dataset. During the testing phase, the reconstruction error is calculated in real-time, and an anomaly flag is output. This module draws inspiration from the efficiency of convolutional autoencoders in temperature anomaly detection used in vaccine cold chain monitoring.

[0088] Adaptive fusion decoder:

[0089] The feature matrix and fused weight vector w(t) output by the multimodal encoder are fed into the decoder. The decoder consists of a fully connected layer and a linear combination module, and is used to output the corrected temperature value. The update of the weight vector w(t) follows the aforementioned formula:

[0090]

[0091] Where the loss function Weighted mean square error is selected:

[0092]

[0093] M represents the number of samples in the batch, and m represents the sample index. The weights are adaptively adjusted based on the level of environmental noise. Reference temperature value. Environmental error. Calculate using the following formula:

[0094] ,

[0095] in Let be the measurement value of the j-th auxiliary sensor at time t. Let j be the reference value for the j-th auxiliary sensor under standard conditions. Let K be the environmental impact coefficient of the j-th auxiliary sensor, and K be the number of auxiliary sensors. This formula is used to measure the impact of environmental changes on temperature measurement.

[0096] The adaptive temperature output is calculated using the following formula:

[0097] ,

[0098] in, Let be the measurement value of the i-th sensor at time t. and These are the mean and standard deviation of the i-th sensor within a preset time window, respectively, used to eliminate dimensional differences between different sensors. denoted by , b is the dynamic weight, is the bias term obtained through training, and N is the total number of sensors. This fusion strategy not only considers the statistical characteristics of each sensor within the historical window, but also reflects the reliability of each sensor in the current environment through dynamic weights, thereby improving measurement accuracy.

[0099] Example 4

[0100] Offline training phase:

[0101] A large amount of temperature data under different environments was collected, and real temperatures were obtained using a high-precision reference thermometer to form training, validation, and test sets. The data includes multimodal information such as indoor and outdoor temperature, humidity, and airflow, covering complex operating conditions such as low temperature, high temperature, air vents, and obstructions.

[0102] The multimodal encoder, convolutional autoencoder, and fusion decoder are trained using stochastic gradient descent. During training, the learning rate η, environment adaptation coefficient γ, and regularization coefficient are progressively adjusted to achieve optimal accuracy on the validation set.

[0103] After the model training is completed, the weight parameters are stored in memory 6, and the anomaly detection threshold θ is determined based on the test set results.

[0104] Online learning phase:

[0105] After the sensor is put into use and runs for a period of time, the temperature is verified using high-precision instruments under specific scenarios to obtain incremental calibration data. The model uses a small amount of incremental data to update some weights through mini-batch gradient descent, while using an anomaly detection module to remove abnormal samples.

[0106] Online learning adjusts only the fusion weights and bias parameters, maintaining the stability of the encoder's core structure to avoid overfitting. This strategy, combined with a dynamic weight update formula, ensures the model maintains accuracy and stability over long-term operation.

[0107] The above description is only a preferred embodiment of the present invention. It should be noted that for those skilled in the art, several improvements and modifications can be made without departing from the principle of the present invention, and these improvements and modifications should also be considered within the scope of protection of the present invention.

Claims

1. A real-time adaptive temperature monitoring smart sensor based on a deep learning algorithm, characterized in that, It includes a main housing, a main temperature measurement unit, an auxiliary environmental sensing unit, a data acquisition circuit, a microcontroller, a deep learning processing module, a memory, a communication module, and a power supply module; The main body housing contains a main temperature measurement unit, an auxiliary environmental sensing unit, a data acquisition circuit, a microcontroller, a deep learning processing module, a memory, a communication module, and a power supply module. The main temperature measurement unit is connected to the microcontroller via a data acquisition circuit to acquire raw temperature signals; The auxiliary environmental sensing unit includes humidity, airflow, light, vibration, sound, and human presence sensors to collect environmental information; The microcontroller has a built-in deep learning processing module that integrates the raw temperature signal and environmental information by running an adaptive temperature sensing network, outputs the corrected temperature value, and sends it through the communication module.

2. The real-time adaptive temperature monitoring smart sensor based on deep learning algorithm according to claim 1, characterized in that, The deep learning processing module includes a multimodal encoder, a convolutional autoencoder, and an adaptive fusion decoder. The multimodal encoder is used to extract the time and frequency features of each sensing channel and output a feature matrix. The convolutional autoencoder is used to reconstruct the feature matrix and calculate the reconstruction error to determine anomalies. The adaptive fusion decoder is used to calculate the corrected temperature value based on dynamic weights.

3. The real-time adaptive temperature monitoring smart sensor based on deep learning algorithm according to claim 2, characterized in that, An incremental learning mechanism is established between the microcontroller and the memory. When incremental calibration data is acquired, the fusion weights and bias parameters are updated using the mini-batch gradient descent method. At the same time, an abnormal sample is removed using a convolutional autoencoder to achieve model self-calibration.

4. The real-time adaptive temperature monitoring smart sensor based on deep learning algorithm according to claim 2, characterized in that, The multimodal encoder includes a one-dimensional convolutional layer and a gated recurrent unit for each sensing channel. The one-dimensional convolutional layer is used to extract local features, and the gated recurrent unit is used to capture the time dependence of temperature changes.

5. A real-time adaptive temperature monitoring smart sensor based on a deep learning algorithm according to claim 2, characterized in that, The convolutional autoencoder is obtained through pre-training and minimizes the reconstruction error using normal temperature measurement data. When the real-time reconstruction error exceeds a set threshold, an abnormality flag is output.

6. The real-time adaptive temperature monitoring smart sensor based on deep learning algorithm according to claim 2, characterized in that, The adaptive fusion decoder calculates the corrected temperature value using the following formula: ， in, Let be the measurement value of the i-th sensor at time t. and Let be the mean and standard deviation of the i-th sensor within a preset time window, respectively. Here, b represents the dynamic weight, b is the bias, and N is the total number of sensors. The dynamic weights are updated according to the following formula: ， in, For learning rate, For loss function, For environmental adaptability coefficient, This is the environmental error measure at time t.

7. A real-time adaptive temperature monitoring smart sensor based on a deep learning algorithm according to claim 6, characterized in that, The loss function The weighted mean square error is adjusted based on real-time environmental noise. Calculate using the following formula: ， in Let be the measurement value of the j-th auxiliary sensor at time t. Let j be the reference value for the j-th auxiliary sensor under standard conditions. Let K be the environmental impact coefficient of the j-th auxiliary sensor, and K be the number of auxiliary sensors.

8. A real-time adaptive temperature monitoring method, characterized in that, The application of a real-time adaptive temperature monitoring smart sensor based on a deep learning algorithm as described in any one of claims 1-7 includes the following steps: The raw data from the main temperature measurement unit and the auxiliary environmental sensing unit are acquired in parallel. The collected raw data is normalized, filtered, and feature extracted to generate multimodal feature vectors; The multimodal feature vectors are fed into the multimodal encoder to obtain the encoded feature matrix; The coded feature matrix is ​​reconstructed using a convolutional autoencoder, the reconstruction error is calculated, and an alarm is output when there is an anomaly. The adaptive fusion decoder fuses signals from various sensors based on a dynamic weighting formula and calculates the corrected temperature value. Online learning is performed based on the collected calibration data to update the model weights and biases; The corrected temperature value is output or sent to the host computer or cloud via the communication module.

9. The real-time adaptive temperature monitoring method according to claim 8, characterized in that, The dynamic weight update includes an environmental error compensation term to enhance adaptability to changes in humidity, airflow, lighting, and human activity environment.

10. A real-time adaptive temperature monitoring method according to claim 8, characterized in that, The main temperature measurement unit has a sampling frequency of 1Hz, and the auxiliary environmental sensing unit has a sampling frequency of 0.5Hz; online learning only adjusts some parameters of the adaptive fusion decoder while keeping the encoder structure unchanged.

Images