Smoke sensor monitoring sensing method and system

Abstract

The invention relates to the technical field of intelligent sensors, and particularly discloses a smoke sensor monitoring sensing method and system. The system comprises a multi-source environment sensing module, a feature extraction and fusion module, a dynamic threshold decision module and an execution control module, a comprehensive environment state vector is constructed by fusing multi-dimensional environment data, and an alarm threshold is dynamically adjusted based on a deep reinforcement learning model, so that accurate fire identification and hierarchical control are realized, and the system is suitable for popularization and application. And the monitoring reliability, the environmental adaptability and the agriculture and forestry equipment operation efficiency are effectively improved.

Description

Technical Field

[0001] This invention belongs to the field of intelligent sensor technology, specifically relating to a smoke sensor monitoring and sensing method and system. Background Technology

[0002] In the field of automated agriculture and forestry management, smoke monitoring technology is a key component in ensuring operational safety and improving management efficiency. As a core sensing device, the accuracy and reliability of smoke sensors directly affect the normal operation of large agricultural and forestry equipment such as sprinkler irrigation machines and the timeliness of disaster warnings.

[0003] Among them, the smoke sensor monitoring method aims to achieve early identification and alarm of fires or abnormal combustion by detecting the concentration of smoke particles in the ambient air. The basic principle of this method is that the sensor converts the physical quantity of smoke concentration into a measurable electrical signal, and triggers corresponding control commands according to preset judgment logic.

[0004] Existing technologies generally employ fixed threshold judgment mechanisms, which are difficult to adapt to the variable weather and operating conditions in agricultural and forestry environments. Non-fire-related interference sources such as water mist and dust can easily cause false alarms from sensors, leading to unnecessary shutdowns of sprinkler irrigation machines and severely reducing operational efficiency and equipment utilization.

[0005] Meanwhile, due to a lack of adaptability to environmental complexity, existing methods cannot dynamically adjust monitoring strategies according to different times, regions, and crop growth stages, leading to a decrease in system reliability in variable scenarios. Therefore, improving the anti-interference capability and adaptive monitoring accuracy of smoke sensors in complex agricultural and forestry environments has become a key technical challenge restricting the efficient and stable operation of intelligent agricultural and forestry equipment. Summary of the Invention

[0006] The technical problem this invention aims to solve is to overcome the shortcomings of existing smoke sensors in complex agricultural and forestry environments, such as poor anti-interference ability, high false alarm rate, and insufficient adaptive monitoring accuracy due to the use of fixed threshold judgment mechanisms. This invention provides a smoke sensor monitoring and sensing method and system. By constructing a closed-loop architecture of multi-source information fusion and dynamic decision-making, this system fundamentally improves the reliability and environmental adaptability of monitoring.

[0007] This invention provides a smoke sensor monitoring system, which includes a multi-source environmental sensing module, a feature extraction and fusion module, a dynamic threshold decision module, and an execution control module. The multi-source environmental sensing module is used to collect real-time data on smoke concentration, ambient temperature, ambient humidity, ambient air pressure, ambient wind speed and direction, and ambient optical image data. The feature extraction and fusion module is used to preprocess and extract features from the raw data collected by the multi-source environmental sensing module, and to perform spatiotemporal alignment and fusion of the extracted multi-dimensional features to generate a comprehensive environmental state vector.

[0008] The dynamic threshold decision module receives the comprehensive environmental state vector and dynamically calculates the smoke concentration alarm threshold under the current operating conditions based on the built-in adaptive decision model. The execution control module compares the real-time smoke concentration data with the alarm threshold output by the dynamic threshold decision module in real time, and generates and outputs control commands to the sprinkler or other controlled agricultural and forestry equipment based on the comparison results.

[0009] Furthermore, the multi-source environmental sensing module specifically includes a smoke concentration sensing unit, a meteorological parameter acquisition unit, and an image acquisition unit. The smoke concentration sensing unit employs a particulate matter sensor based on the laser scattering principle, with a measurement accuracy better than 0. mg / m³. 3 The sampling frequency is greater than 1Hz. The meteorological parameter acquisition unit integrates a temperature sensor, a humidity sensor, a barometric pressure sensor, and an ultrasonic anemometer and wind direction sensor to simultaneously acquire accurate environmental meteorological parameters. The image acquisition unit uses an industrial camera with near-infrared spectral response capability, and its deployment position ensures coverage of the main field of view of the sprinkler irrigation machine's operating area. The frame rate can be configured from 1 to 5 fps.

[0010] Furthermore, the workflow of the feature extraction and fusion module is as follows. First, the smoke concentration data is processed by moving average filtering to suppress transient impulse noise.

[0011] Secondly, meteorological parameter data should be standardized to eliminate the influence of different units.

[0012] Third, the ambient optical images acquired by the image acquisition unit are processed by grayscale conversion and region segmentation, and the pre-trained convolutional neural network model is used to extract the depth features related to smoke texture and motion features in the image.

[0013] Finally, the preprocessed time-series features of smoke concentration, the standardized multidimensional meteorological parameters, and the depth features extracted from the image are spliced ​​and dimensionality reduced under a unified timestamp to generate a comprehensive environmental state vector with fixed dimensions.

[0014] Furthermore, the adaptive decision-making model built into the dynamic threshold decision-making module is a policy network based on deep reinforcement learning. This policy network takes the aforementioned comprehensive environmental state vector as input and outputs the optimal smoke concentration alarm threshold for the current moment. The training process of the policy network is carried out on a simulation platform simulating various typical agricultural and forestry environments, and its reward function design comprehensively considers three key performance indicators: false alarm rate, false negative rate, and system response delay. This model can learn and memorize the optimal decision threshold corresponding to different combinations of environmental characteristics, thereby achieving dynamic adaptive adjustment of the monitoring strategy.

[0015] Furthermore, the logical decision-making process of the execution control module is as follows: The module continuously receives real-time smoke concentration data and dynamic alarm thresholds.

[0016] When the real-time smoke concentration remains above the dynamic alarm threshold for a preset duration of 3 seconds, the module generates a fire alarm signal and triggers an emergency stop command for the sprinkler. When the real-time smoke concentration is between 70% and 100% of the dynamic alarm threshold, the module generates a warning signal and initiates a speed reduction command for the sprinkler.

[0017] When the real-time smoke concentration value is below 70% of the dynamic alarm threshold, the module maintains normal system operation. All control commands are sent to the actuators in real time via the industrial bus protocol.

[0018] In one embodiment of the present invention, the system further includes an offline model optimization module. The offline model optimization module periodically collects historical operating data and corresponding real-world operating condition labels from deployed terminal devices, and uses this data to incrementally learn and fine-tune the parameters of the adaptive decision-making model in the dynamic threshold decision-making module, ensuring that the model can adapt to the slow drift of environmental characteristics during long-term operation.

[0019] Furthermore, the sensor units in the multi-source environmental sensing module adopt a unified power management and data synchronization mechanism. The power management unit dynamically adjusts the sensor power consumption according to the system operating mode, entering a low-power state during inactive periods. The data synchronization mechanism is coordinated by a high-precision real-time clock chip, ensuring that the timestamp deviation of all sensor data is less than 10ms, providing an accurate time reference for subsequent data fusion.

[0020] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0021] 1. This invention integrates a multi-source environmental perception module and constructs a feature extraction and fusion module, expanding the single smoke concentration information into a comprehensive environmental state vector that integrates meteorological, image, and other multi-dimensional information, greatly enriching the information dimensions of the decision-making basis. This fundamentally changes the limitation of traditional methods that rely on a single physical quantity for judgment, enabling the system to effectively distinguish between real fires and common interference sources such as water mist and dust, reducing the probability of false alarms.

[0022] 2. This invention employs an adaptive decision-making model based on deep reinforcement learning as the core of the dynamic threshold decision-making module, enabling the alarm threshold to be dynamically adjusted according to environmental conditions. Through extensive training in a simulation environment, this model can autonomously learn and optimize the best decision-making strategy under various complex working conditions, thereby endowing the system with strong environmental adaptability. The system does not require manual pre-setting of multiple fixed threshold modes and can automatically adapt to changes in different times, regions, and crop growth stages, improving monitoring accuracy and system reliability.

[0023] 3. This invention designs a refined execution control logic and introduces a multi-level response mechanism for early warning and alarm. This mechanism avoids the control abruptness problem caused by traditional binary judgment. When a potential risk is detected, the system can take mitigating measures such as slowing down operation, which not only ensures equipment safety but also maintains the continuity of operation to the maximum extent, effectively improving the overall operating efficiency and equipment utilization rate of large agricultural and forestry equipment such as sprinkler irrigation machines.

[0024] 4. This invention establishes a closed loop for continuous system performance optimization by introducing an offline model optimization module. This module uses actual operating data to periodically fine-tune the core decision model, effectively combating the impact of sensor performance degradation and long-term environmental feature drift on system performance, ensuring the stability and durability of monitoring accuracy during long-term deployment. Attached Figure Description

[0025] Figure 1 This is a schematic diagram of the overall technical architecture of the smoke sensor monitoring and sensing system proposed in this invention;

[0026] Figure 2 This is a schematic diagram of the core principle framework of the adaptive decision-making model based on deep reinforcement learning in this invention;

[0027] Figure 3 This is a logical flowchart of the multi-source environment perception module and the feature extraction and fusion module in this invention;

[0028] Figure 4 This is a schematic diagram of the multi-level interaction relationship and data flow between the dynamic threshold decision module and the execution control module in this invention;

[0029] Figure 5This is a flowchart illustrating the periodic optimization process of the offline model optimization module in this invention. Detailed Implementation

[0030] This embodiment details a specific implementation of a smoke sensor monitoring system. Please refer to the appendix. Figure 1 The system consists of a multi-source environmental perception module, a feature extraction and fusion module, a dynamic threshold decision module, and an execution control module. Each module is connected to the central processing unit via an industrial bus to form a closed-loop control architecture.

[0031] The multi-source environmental sensing module is responsible for real-time acquisition of multi-dimensional environmental data. This module includes a smoke concentration sensing unit, a meteorological parameter acquisition unit, and an image acquisition unit. The smoke concentration sensing unit uses a particulate matter sensor based on the laser scattering principle. Its optical cavity houses a 650nm wavelength laser diode and a photodetector, with a detection sensitivity of 0.01mg / m³. 3 The measurement range covers 0 to 10 mg / m³ 3 The sampling frequency is fixed at 2Hz.

[0032] The sensor outputs a 4 to 20 mA analog signal, which is quantized by a 16-bit analog-to-digital converter and then transmitted to the processing unit. The meteorological parameter acquisition unit integrates a platinum resistance temperature sensor, a capacitive humidity sensor, a piezoresistive barometric pressure sensor, and an ultrasonic anemometer and wind direction sensor.

[0033] The temperature sensor measures from -40℃ to 85℃ with an accuracy of 0.3℃; the humidity sensor measures from 0% to 100% relative humidity with an accuracy of 2%; the barometric pressure sensor measures from 30 to 110 kPa with an accuracy of 0.1 kPa; and the ultrasonic anemometer measures wind speed from 0 to 60 m / s with a wind direction resolution of 1° and a data output frequency of 1 Hz.

[0034] The image acquisition unit uses a 2-megapixel global shutter industrial camera equipped with an 850nm near-infrared filter, an effective pixel size of 3.45μm, and a frame rate configurable from 1fps to 5fps. The camera transmits raw image data via a gigabit Ethernet interface. Its installation position is 3m above the ground, with a pitch angle adjusted to 15 degrees, ensuring coverage of the sprinkler irrigation machine's operating area within a 50m radius.

[0035] All sensor units are powered by a unified power management unit that supports 12V DC input and has a built-in step-down converter and load switch, which can dynamically switch the sensor's operating mode according to system commands.

[0036] In low-power mode, the sensor supply current drops to 20% of that in standard mode.

[0037] The data synchronization mechanism is driven by a high-precision real-time clock chip with a clock deviation of less than 5ppm. The chip coordinates the acquisition time of each sensor through hardware trigger signals to ensure that the timestamp deviation is controlled within 8ms.

[0038] The feature extraction and fusion module processes and fuses the raw data from the multi-source environmental perception module. Please refer to the appendix. Figure 3 This module first performs a moving average filter on the smoke concentration data. The filter window width is set to 10 sampling points, corresponding to a 5-second time window. The filtered data retains the 0.1 to 2Hz frequency band components to suppress high-frequency noise. Next, the meteorological parameters are standardized using the z-score normalization method, calculated as follows:

[0039] ;

[0040] in These are the original meteorological parameter values. The average value of this parameter in the most recent 24-hour historical data. The standard deviation is given. Normalized values ​​for the four parameters—temperature, humidity, air pressure, and wind speed—are calculated independently. In the third stage, image data is processed. The RGB images captured by the camera are first converted to grayscale images and then divided into 16x16 pixel grid cells. Each grid cell is input to a pre-trained convolutional neural network model using the ResNet18 architecture, with the final fully connected layer replaced by a 128-dimensional feature output layer. The feature vectors extracted by the network include three types of deep features: texture energy, gradient orientation histogram, and optical flow magnitude.

[0041] Finally, the preprocessed smoke concentration time-series features, standardized meteorological parameter features, and image depth features are aligned by timestamps, and the total dimension is reduced from 256 to 32 using principal component analysis to generate a comprehensive environmental state vector. This vector is updated every 2 seconds and stored in a circular buffer for subsequent modules to access.

[0042] The dynamic threshold decision module receives the comprehensive environmental state vector and calculates the dynamic alarm threshold. Please refer to the appendix. Figure 2 The core of this module is a policy network based on deep reinforcement learning. The network structure includes an input layer, three hidden layers, and an output layer. The input layer has 32 neurons, consistent with the dimension of the integrated environment state vector; the hidden layers have 64, 128, and 64 neurons respectively, using ReLU as the activation function; the output layer has one neuron, using the Sigmoid function to map the output value to the range of 0 to 1, and then linearly transforming it to 0.5 to 5 mg / m². 3 The actual threshold range. The network training phase is conducted on a simulation platform, which simulates eight typical agricultural and forestry environmental scenarios, including morning dew, dry winds, and straw burning. The reward function is designed as follows:

[0043] ;

[0044] in For false alarm rate, The false negative rate, For system response delay, weighting coefficient Set it to 0.5. Set it to 0.3. Set it to 0.2.

[0045] Training employed a proximal policy optimization algorithm with a learning rate of 0.0001 and an experience replay buffer of 10,000 samples. The network performed forward inference every 30 minutes, outputting the optimal alarm threshold based on the current environmental state.

[0046] The execution control module implements a multi-level response mechanism. Please refer to the appendix. Figure 4 This module continuously monitors real-time smoke concentration and dynamic alarm thresholds. When the real-time concentration exceeds the threshold for 3 seconds, the module sends an emergency stop command to the sprinkler controller via the CAN bus. The command includes a target speed of 0% and a highest braking priority flag. When the real-time concentration is between 70% and 100% of the threshold, the module sends a speed reduction command, setting the target speed to 50% of the rated speed and the braking priority to medium. When the real-time concentration is below 70% of the threshold, a normal operation command is sent, with a target speed of 100%. All commands are appended with a 32-bit cyclic redundancy check (CRC) code, the bus transmission rate is configured to 500kbps, and the response latency is less than 50ms.

[0047] The system also includes an offline model optimization module. Please refer to the appendix. Figure 5 This module activates every 7 days, collecting historical data from terminal devices over the past 168 hours, including comprehensive environmental state vectors, actual alarm records, and manually verified tags. After data cleaning, it is divided into training and validation sets in an 8:2 ratio. The training process employs a transfer learning strategy, freezing the weights of the first two layers of the policy network and fine-tuning only the parameters of the last two layers. Fine-tuning uses the Adam optimizer with a batch size of 32 and 10 training epochs. Updated model parameters are encrypted and digitally signed before being transmitted to each terminal, with version numbers incrementing by 0.01. A rollback mechanism retains the three most recent versions.

[0048] The system hardware platform uses an industrial-grade embedded processor with a main frequency of 1.8GHz, 4GB of memory, and 32GB of storage. The operating system is a real-time Linux kernel with a task scheduling cycle of 1ms. The chassis has an IP67 protection rating and an operating temperature range of -30℃ to 70℃. The power system supports dual inputs of 220V AC and 24V DC, and the backup battery can sustain system operation for 2 hours.

[0049] This embodiment, through the specific implementation described above, achieves closed-loop control of multi-source information fusion and dynamic decision-making. Testing in typical agricultural and forestry environments shows that the false alarm rate is lower than traditional methods, monitoring accuracy is improved, and response latency is controlled within 3 seconds. This system outperforms the fixed threshold scheme in terms of both false alarm rate and monitoring accuracy.

[0050] This embodiment provides an alternative implementation scheme for a smoke sensor monitoring system, mainly optimizing the data processing flow of the feature extraction and fusion module. In this embodiment, image feature extraction employs a lightweight convolutional neural network model based on the MobileNetV2 architecture. The number of depthwise separable convolutional layers is reduced to 8, and the feature output dimension is 64. The model parameter count is 40% of the standard scheme, and the inference speed is improved by 2 times. An adaptive histogram equalization operation is added in the image preprocessing stage to enhance the contrast of smoke textures in low-light environments.

[0051] The meteorological parameter standardization method has been changed to minimum-maximum normalization, and a sliding window statistical mechanism is used in the calculation. The window length is set to 1 hour, and the parameter extreme values ​​are updated every 5 minutes. The normalization range for temperature is adjusted to -10℃ to 50℃, humidity to 20% to 95%, air pressure to 90 to 105 kPa, and wind speed to 0 to 20 m / s. This method is more suitable for short-term meteorological fluctuation scenarios.

[0052] In the feature fusion stage, an attention mechanism is used instead of principal component analysis. The attention network includes a query vector generation layer and a weight allocation layer. The query vector has a dimension of 16, and the weight allocation layer uses the Softmax function to assign importance weights to various features. The weight coefficients for smoke concentration features range from 0.3 to 0.5, meteorological features from 0.2 to 0.4, and image features from 0.3 to 0.5, with the weights dynamically adjusted according to the environmental state. The fused comprehensive environmental state vector maintains a dimension of 32, but the feature discrimination is improved.

[0053] The number of hidden layers in the policy network of the dynamic threshold decision module was reduced to two, with 48 and 24 neurons respectively. A curriculum learning strategy was adopted during the training phase, training the model in ascending order of environmental complexity. The false alarm rate weight α in the reward function was increased to 0.6, and the false negative rate weight β was decreased to 0.2 to further suppress false alarms. The network update frequency was increased to once every 15 minutes to adapt to scenarios with rapid weather changes.

[0054] The execution control module incorporates fuzzy control logic. When the real-time smoke concentration is between 85% and 100% of the threshold, the sprinkler speed gradually decreases linearly. This smooth transition mechanism avoids the impact of sudden speed changes on the pipeline system. Control commands now include timestamp verification; the receiving end discards commands with a delay greater than 100ms to ensure precise control timing.

[0055] The offline model optimization module employs a federated learning architecture. Each terminal device trains model parameters locally, uploading only incremental parameter changes to the central server. The server aggregates parameters from 10 terminals to generate a global model, which is then compressed and distributed for updates. This process also protects terminal data privacy. The model update cycle is extended to 14 days, adapting to deployment environments with poor communication conditions.

[0056] The hardware platform has been upgraded to a low-power processor with a clock speed of 1.2GHz and 2GB of memory. The integration of the sensor units has been improved, with the meteorological parameter acquisition unit adopting a six-in-one environmental sensor chip. The image acquisition unit has been replaced with a 5-megapixel rolling shutter camera.

[0057] This embodiment optimizes the algorithm architecture and hardware configuration to reduce system cost and power consumption while maintaining core performance indicators. It is particularly suitable for large-scale distributed deployment scenarios, such as monitoring networks for contiguous farmland of tens of thousands of acres.

[0058] It should be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, method, article, or apparatus.

[0059] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A smoke sensor monitoring and sensing system, characterized in that, include: The multi-source environmental sensing module is used to collect real-time data on smoke concentration, ambient temperature, ambient humidity, ambient air pressure, ambient wind speed and direction, and ambient optical image data. The feature extraction and fusion module is used to preprocess and extract features from the raw data collected by the multi-source environment perception module, and to perform spatiotemporal alignment and fusion of the extracted multi-dimensional features to generate a comprehensive environmental state vector. The dynamic threshold decision module receives the comprehensive environmental state vector and dynamically calculates the smoke concentration alarm threshold under the current operating conditions based on the built-in adaptive decision model. The execution control module is used to compare the real-time smoke concentration data with the alarm threshold output by the dynamic threshold decision module in real time, and generate and output control commands to the sprinkler or other controlled agricultural and forestry equipment based on the comparison results.

2. The smoke sensor monitoring system according to claim 1, characterized in that, The multi-source environmental perception module specifically includes a smoke concentration sensing unit, a meteorological parameter acquisition unit, and an image acquisition unit; The smoke concentration sensing unit uses a particulate sensor based on the principle of laser scattering. The meteorological parameter acquisition unit integrates a temperature sensor, a humidity sensor, a barometric pressure sensor, and an ultrasonic wind speed and direction sensor to synchronously acquire accurate environmental meteorological parameters. The image acquisition unit uses an industrial camera with near-infrared spectral response capability, and its deployment position ensures that it can cover the main field of view of the sprinkler irrigation machine's operating area.

3. The smoke sensor monitoring system according to claim 1, characterized in that, The workflow of the feature extraction and fusion module includes: The smoke concentration data is processed by moving average filtering to suppress transient impulse noise; Meteorological parameter data are standardized to eliminate the influence of different units; The ambient optical images acquired by the image acquisition unit are processed by grayscale conversion and region segmentation, and a pre-trained convolutional neural network model is used to extract depth features related to smoke texture and motion features in the image. The preprocessed time-series features of smoke concentration, the standardized multidimensional meteorological parameters, and the depth features extracted from the image are spliced ​​and dimensionality reduced under a unified timestamp to generate a comprehensive environmental state vector with fixed dimensions.

4. The smoke sensor monitoring system according to claim 1, characterized in that, The adaptive decision-making model built into the dynamic threshold decision-making module is a policy network based on deep reinforcement learning. The strategy network takes the comprehensive environmental state vector as input and outputs the optimal smoke concentration alarm threshold at the current moment. The training process of the policy network is carried out on a simulation platform that simulates various typical agricultural and forestry environments. The design of its reward function takes into account three key performance indicators: false alarm rate, false negative rate, and system response delay.

5. A smoke sensor monitoring system according to claim 1, characterized in that, The logical decision-making process of the execution control module includes: The module continuously receives real-time smoke concentration data and dynamic alarm thresholds; When the real-time smoke concentration continues to exceed the dynamic alarm threshold for a preset duration, the module generates a fire alarm signal and triggers an emergency shutdown command for the sprinkler. When the real-time smoke concentration value is within the specified value of the dynamic alarm threshold, the module generates an early warning signal and initiates a command to reduce the speed of the sprinkler. When the real-time smoke concentration value is lower than the dynamic alarm threshold, the module maintains the normal operation of the system.

6. The smoke sensor monitoring system according to claim 1, characterized in that, It also includes an offline model optimization module; The offline model optimization module periodically collects historical operating data and corresponding real-world operating condition labels from the deployed terminal devices, and uses this data to incrementally learn and fine-tune the parameters of the adaptive decision-making model in the dynamic threshold decision-making module.

7. A smoke sensor monitoring and sensing system according to claim 2, characterized in that, The sensor units in the multi-source environmental sensing module adopt a unified power management and data synchronization mechanism. The power management unit dynamically adjusts the sensor power consumption according to the system operating mode, and enters a low power consumption state during inactive periods. The data synchronization mechanism is coordinated by a high-precision real-time clock chip.

8. A smoke sensor monitoring system according to claim 3, characterized in that, The standardization process employs the z-score normalization method. The calculation formula for the z-score normalization method is as follows: ； in These are the original meteorological parameter values. This is the average value of this parameter over the most recent 24 hours of historical data. The standard deviation is denoted as .

9. A smoke sensor monitoring system according to claim 4, characterized in that, The policy network structure includes an input layer, a hidden layer, and an output layer; The input layer neurons have the same dimension as the integrated environment state vector. The number of neurons in the hidden layer is determined by the ReLU activation function. The output layer uses the Sigmoid function to map the output values ​​to a range, and then linearly transforms them into the actual threshold range.

10. A smoke sensor monitoring and sensing method, characterized in that, The smoke sensor monitoring sensing system described in any one of claims 1-9 is used to achieve the sensing.

Images