Chemical plant leakage monitoring method based on acoustic features and lightweight neural network

Abstract

The invention discloses a leakage monitoring system and method based on acoustic features and a lightweight neural network. The system comprises a hierarchical distributed low-power-consumption wireless monitoring system composed of a monitoring terminal, a LoRa gateway and a cloud platform, the monitoring terminal adopts a specific hardware type selection and connection structure of a high-temperature-resistant digital microphone INMP441, an ultra-low-power-consumption master control STM32F407 and a LoRa communication module SX1278, and is matched with a sound-light alarm circuit and a low-power-consumption power management circuit based on sound pressure threshold interruption awakening; during deployment, a triangular mesh layout is adopted in a high-risk area, a linear layout is adopted in a medium-risk area, and leakage detection is realized through an anti-interference method in which hardware filtering is combined with a voiceprint filtering library and transient energy and frequency band proportion feature extraction is combined with TFLite lightweight neural network reasoning. According to the invention, accurate detection can be realized in a 105dB strong noise environment, the response speed reaches a millisecond level, and the standby power consumption of the terminal is lt; the gas sensor is 0.5 W, wiring is not needed, deployment is flexible, and the problems that a traditional gas sensor is high in cost and complex in deployment and existing acoustic detection is poor in interference resistance are solved.

Description

Technical Field

[0001] This invention relates to the field of chemical plant safety monitoring technology, specifically to a leak monitoring system and method based on acoustic features and a lightweight neural network. Background Technology

[0002] Currently, leak detection in chemical plants mainly relies on gas sensors, but this type of solution has obvious drawbacks: the cost of a single device exceeds 2,000 yuan, resulting in high overall costs; the deployment process requires complex wiring, making construction difficult; and the maintenance costs are high, making it difficult to popularize in small and medium-sized chemical plants.

[0003] While existing acoustic detection methods attempt to compensate for the shortcomings of gas sensors, they suffer from low accuracy in the high-noise environment of factories and are easily affected by equipment operating noise, thus failing to meet actual monitoring needs. Therefore, there is an urgent need for a low-cost, highly interference-resistant, fast-response, and flexible chemical plant leak monitoring solution. Summary of the Invention

[0004] The purpose of this invention is to provide a leak monitoring system and method based on acoustic features and a lightweight neural network.

[0005] To achieve the above objectives, the present invention provides the following technical solution: Monitoring System This system is a hierarchical distributed low-power wireless monitoring system, consisting of monitoring terminals, LoRa gateways, and a cloud platform. Monitoring Terminal: As an independent device, it includes a core control unit, an acoustic signal acquisition unit, a communication unit, an alarm unit, and a power supply unit. The core control unit uses an STM32F407VGT6 microcontroller, responsible for signal processing, inference calculation, and device control. The acoustic signal acquisition unit uses an INMP441 digital MEMS microphone, connected to the main control unit via an I²S interface, and is equipped with preamplifier and anti-aliasing filter circuits to improve signal quality. The communication unit uses an SX1278 LoRa module, which realizes wireless data transmission via an SPI interface. The alarm unit includes RGB LEDs and a high-decibel buzzer, driven by the main control unit's GPIO ports. The power supply unit is powered by an 18650 lithium battery, equipped with a TPS63020 buck-boost chip, supports solar charging, and the main control program is configured to trigger an interrupt to wake up the device based on the sound pressure level threshold. Standby power consumption is <0.5W.

[0006] Deployment plan: In high-risk areas (such as pipelines and storage tanks), install in a triangular grid pattern with a spacing of 20 meters and an installation height of 1.5 meters; in medium- and low-risk areas, install in a linear pattern with a spacing of 30-50 meters.

[0007] Anti-interference and inference: Combining anti-interference methods such as hardware filtering, voiceprint filtering library, and transient energy and frequency band proportion feature extraction, with the TFLite lightweight neural network model deployed on the main control unit, leakage probability calculation is realized.

[0008] Monitoring methods Signal acquisition and preprocessing: The monitoring terminal acquires ambient audio through the INMP441 microphone. After pre-amplification and anti-aliasing filtering, the main controller performs 1-4kHz bandpass filtering and frame segmentation.

[0009] Energy judgment and wake-up mechanism: Calculate the signal energy of each frame. If it does not exceed the baseline +3dB, the terminal remains in low-power sleep mode; if it exceeds the baseline, it will wake up for subsequent processing.

[0010] Feature extraction and anti-interference: Extract transient energy and frequency band proportion feature vectors, and filter equipment operating noise through the acoustic signature filter library.

[0011] Neural network inference: A lightweight model is loaded through the TFLite framework, and the leakage probability is calculated by inputting feature vectors.

[0012] Alarm and data transmission: When the leakage probability is >90%, an audible and visual alarm is triggered, and the data is uploaded to the cloud platform via the LoRa gateway; otherwise, it returns to sleep mode.

[0013] The beneficial effects of this invention are as follows: The hardware cost of this invention's single-point monitoring terminal is controlled within 200 yuan, only one-tenth of that of traditional gas sensor solutions, making it suitable for widespread adoption in small and medium-sized chemical plants. Through hardware filtering, acoustic signature library software filtering, and dual-feature fusion extraction, it can accurately detect leaks in environments with strong noise levels up to 105dB, with a low false alarm rate. The entire process from signal acquisition to alarm is completed on local hardware in milliseconds, enabling early warning of leaks. The terminal's standby power consumption is <0.5W, it uses LoRa wireless communication and battery power, supports solar charging, requires no wiring, and allows for flexible deployment to adapt to different risk areas.

[0014] The above description is merely an overview of the technical solution of the present invention. In order to better understand the technical means of the present invention and to implement it in accordance with the contents of the specification, the preferred embodiments of the present invention are described in detail below with reference to the accompanying drawings. Attached Figure Description

[0015] Figure 1 This is a schematic diagram of a leak monitoring system based on acoustic features and a lightweight neural network, as shown in this invention. Figure 2 This is a schematic diagram illustrating the workflow of a leak monitoring system based on acoustic features and a lightweight neural network, as shown in this invention. Figure 3This is a flowchart illustrating the data transmission and cloud processing architecture of a leak monitoring system based on acoustic features and a lightweight neural network, as shown in this invention. Figure 4 This is a diagram of the broadband beamforming structure of a microphone array in a leak detection system based on acoustic features and a lightweight neural network, as shown in this invention. Figure 5 This is a speaker recognition system architecture diagram of a leak detection system based on acoustic features and a lightweight neural network, as shown in this invention. Detailed Implementation

[0016] The specific embodiments of the present invention will be described in further detail below with reference to the examples. These examples are for illustrative purposes only and are not intended to limit the scope of the invention.

[0017] Unless otherwise specified, the technical solutions described in this invention are conventional solutions in the field; unless otherwise specified, the materials described are all from commercial channels.

[0018] See Figures 1-5 This invention provides a leak detection system based on acoustic features and a lightweight neural network, specifically comprising: Hardware selection and connection 1. Core control unit: The STM32F407VGT6 microcontroller is selected. This chip has low power consumption and sufficient computing resources to meet the needs of lightweight neural network inference.

[0019] 2. Sound signal acquisition unit: The INMP441 digital MEMS microphone is used, which has high temperature resistance and is suitable for the harsh environment of chemical plants. It is connected to the PB15 (I2S2_WS), PB13 (I2S2_CK), and PB14 (I2S2_SD) pins of STM32F407VGT6 through the I²S interface. The preamplifier circuit uses the OPA2376 chip, and the anti-aliasing filter circuit is a 4th order Butterworth filter with a cutoff frequency of 4kHz.

[0020] 3. Communication Unit: The SX1278 LoRa module is connected to the PA5 (SPI1_SCK), PA6 (SPI1_MISO), and PA7 (SPI1_MOSI) pins of the STM32F407VGT6 via the SPI interface. The communication frequency is set to 433MHz and the transmission power is 17dBm.

[0021] 4. Alarm Unit: The three pins of the RGB LED are connected to the PC0, PC1, and PC2 GPIO ports of the STM32F407VGT6 respectively. The high-decibel buzzer is driven by an NPN transistor and connected to the PC3 GPIO port.

[0022] 5. Power Supply Unit: An 18650 lithium battery (capacity 3.7V / 2000mAh) is connected to a TPS63020 buck-boost chip to output 3.3V voltage to power each module. The chip's enable pin is connected to the PD0 pin of the STM32F407VGT6 to realize power control.

[0023] (II) Deployment and Implementation 1. High-risk areas: In areas such as pipeline interfaces and storage tank outlets of chemical plants, monitoring terminals are installed in a triangular grid layout with a spacing of 20 meters. The installation height is uniformly 1.5 meters to ensure that the sound signal is collected without obstruction.

[0024] 2. Low- to medium-risk areas: In workshop aisles, around equipment, and other areas, install monitoring terminals along walls or supports at linear intervals of 30-50 meters, avoiding strong vibration sources.

[0025] 3. LoRa Gateway Deployment: Install a LoRa gateway in the center of the factory area to ensure coverage of all monitoring terminals. The gateway communicates with the cloud platform via Ethernet.

[0026] (III) Software and Algorithm Implementation 1. Low power management: The STM32F407VGT6 is configured in STOP2 low power mode. The sound signal acquisition unit monitors the sound pressure in real time. When the sound pressure exceeds the set threshold (baseline + 3dB), an external interrupt is triggered to wake up the main controller.

[0027] 2. Signal processing flow: The audio signal sampling rate is set to 16kHz. After 1-4kHz bandpass filtering, it is divided into frames with a frame length of 20ms and a frame shift of 10ms. The energy of each frame is calculated using the short-time energy method.

[0028] 3. Feature extraction: For frames that meet the energy threshold, extract the energy proportion of four frequency bands: 0-1kHz, 1-2kHz, 2-3kHz, and 3-4kHz, and combine them with transient energy peaks to form a 12-dimensional feature vector.

[0029] 4. Neural Network Model: A 3-layer fully connected neural network is used, with 12 neurons in the input layer, 32 neurons in the hidden layer, and 1 neuron in the output layer (output leakage probability). After training, the model is converted to TFLite format and deployed in the main control Flash. The inference time is <50ms.

[0030] 5. Voiceprint Filtering Library: Pre-stores the voiceprint data of common equipment in chemical plants (pumps, fans, motors). Before the feature vector is input into the model, noise signals with a similarity > 0.8 are removed by cosine similarity comparison.

[0031] Workflow After the monitoring terminal is powered on, the power supply unit supplies power to each module. After the main control is initialized, it enters a low-power sleep mode. The sound signal acquisition unit continuously collects ambient audio. When the sound pressure exceeds the threshold, an interrupt is triggered to wake up the main control. The main control controls the sound signal acquisition unit to transmit audio data. After filtering, framing, and energy calculation, if the energy meets the standard, the feature vector is extracted and processed by the voiceprint filtering library before being input into the TFLite model for inference. If the leakage probability is >90%, the main control drives the RGB LED to flash (red) and the buzzer to sound. At the same time, the leakage location, time, and probability data are uploaded to the LoRa gateway through the SX1278 module. The gateway forwards the data to the cloud platform. After receiving the data, the cloud platform stores and displays the warning information. After completing one monitoring process, the terminal returns to the low-power sleep mode.

[0032] The technical features of the above embodiments can be combined in any way. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0033] The embodiments described above are merely illustrative of several implementations of the present invention, and while the descriptions are relatively specific and detailed, they should not be construed as limiting the scope of the invention patent. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of the present invention, and these all fall within the protection scope of the present invention. Therefore, the protection scope of this invention patent should be determined by the appended claims.

Claims

1. A leak monitoring system based on acoustic features and a lightweight neural network, characterized in that, include: The monitoring terminal is an independent device, comprising a core main control unit, an acoustic signal acquisition unit, a communication unit, an alarm unit, and a power supply unit; LoRa gateway is used to receive data transmitted by monitoring terminals and aggregate it to the cloud platform; Cloud platforms are used for data storage and remote monitoring; The core control unit uses an STM32F407VGT6 microcontroller; the acoustic signal acquisition unit uses an INMP441 digital MEMS microphone, which is connected via I... 2 The S-interface connects to the core control unit and is equipped with a preamplifier and anti-aliasing filter circuit. The communication unit uses an SX1278 LoRa module and connects to the core control unit via an SPI interface. The alarm unit includes an RGB LED and a high-decibel buzzer, driven by the GPIO port of the core control unit. The power supply unit is powered by an 18650 lithium battery and equipped with a TPS63020 buck-boost chip for power management. The core control unit is programmed to trigger an interrupt based on the sound pressure threshold and wake up from a low-power sleep mode. The deployment scheme for the monitoring terminals is as follows: in high-risk areas, they are installed in a triangular grid pattern with a spacing of 20 meters and an installation height of 1.5 meters; in medium- and low-risk areas, they are installed in a linear spacing of 30-50 meters. The system also includes a voiceprint filtering library and a lightweight neural network model. Leak detection is achieved through an anti-interference method that combines hardware filtering with the voiceprint filtering library and extraction of transient energy and frequency band proportion features.

2. The leakage monitoring system according to claim 1, characterized in that, The power supply unit supports solar charging, and the terminal standby power consumption is <0.5W.

3. The leakage monitoring system according to claim 1, characterized in that, The lightweight neural network model is deployed in the core control unit and uses the TFLite inference framework.

4. The leakage monitoring system according to claim 1, characterized in that, In the anti-interference method, the frequency range for extracting transient energy and frequency band proportion features is 1-4kHz.

5. A leak detection method based on acoustic features and a lightweight neural network, characterized in that, The monitoring system applied to any one of claims 1-4 includes the following steps: S1: The monitoring terminal collects ambient audio signals through the sound signal acquisition unit, and transmits them to the core main control unit after pre-amplification and anti-aliasing filtering. S2: The core control unit performs bandpass filtering and frame segmentation on the filtered audio signal and calculates the energy of each frame. S3: Determine if the signal energy is greater than the baseline +3dB. If not, the monitoring terminal returns to low-power sleep mode; if so, extract the transient energy and frequency band proportion feature vector of the signal. S4: The core control unit loads a lightweight neural network model through the TFLite inference framework, inputs feature vectors for inference, and calculates the leakage probability; S5: If the leakage probability is >90%, the core main control unit drives the alarm unit to trigger an audible and visual alarm, and transmits the data to the cloud platform via the LoRa gateway through the LoRa communication module; if the leakage probability is ≤90%, the monitoring terminal returns to low-power sleep mode.

6. The leakage monitoring method according to claim 5, characterized in that, In step S2, the baseline is updated in real time using a dynamic noise baseline learning algorithm.

7. The leakage monitoring method according to claim 5, characterized in that, In step S3, after feature vector extraction, noise filtering is performed through a voiceprint filtering library to remove interference signals from equipment operation noise.

Images