A low-power integrated environmental state monitoring system

Abstract

The application discloses a low-power-consumption integrated environment state monitoring system and relates to the field of Internet of Things.The intelligent data screening module and the edge intelligent inference module are divided and cooperated, the screening module is responsible for event triggering and redundant data cleaning, and the inference module is focused on abnormality identification and classification, so that the system energy efficiency and robustness are improved while function overlap is avoided;meanwhile, the multi-modal acquisition, multi-protocol communication and multi-mode power supply of hardware and the edge inference, intelligent screening and low-power-consumption management of software are deeply cooperated, a whole-link optimization of "acquisition-preprocessing-inference-transmission-power supply" is formed, and finally, a comprehensive breakthrough of device integration, analysis intelligence, flexible deployment and minimum energy consumption is realized, and the core problems of weak fusion capability, poor adaptability, high power consumption and insufficient real-time performance of a traditional multi-physical-quantity monitoring system are effectively solved.

Description

Technical Field

[0001] This invention relates to the field of the Internet of Things (IoT), and more particularly to a low-power integrated environmental condition monitoring system. Background Technology

[0002] With the rapid growth in demand for multi-source heterogeneous data fusion sensing in applications such as industrial automation, intelligent security, and home environmental sensing, traditional single-type sensors are no longer sufficient to meet the requirements of collaborative acquisition, intelligent analysis, and reliable transmission of multiple physical quantities such as sound, vibration, and temperature in complex environments. Especially in fields such as industrial fault early warning, equipment status monitoring, and environmental anomaly sensing, real-time sensing of multi-dimensional abnormal signals such as sound (e.g., abnormal noise), structural vibration (e.g., high-frequency impact), and temperature changes (e.g., overheating) has become crucial for improving the intelligence level of the system. However, current multimodal sensor devices generally suffer from the following problems in terms of functional fusion, communication adaptation, local intelligent identification, and power supply management:

[0003] Most existing systems are deployed in the form of single sensors, collecting sound, temperature or vibration signals separately, resulting in a large number of devices, complex wiring and inconsistent data processing.

[0004] Mainstream multi-physical quantity sensing devices typically only support one communication interface or protocol, such as only TCP or MQTT. They cannot select the appropriate protocol for the scenario through pre-configuration, nor can they match the scenario requirements through pre-configuration, which can easily lead to transmission delays, packet loss, or increased energy consumption.

[0005] Existing devices cannot perform real-time feature extraction and intelligent recognition of multi-source signals such as sound and vibration locally. All raw data must be reported to the cloud for processing, resulting in high bandwidth load, large response latency, and inability to be used in offline scenarios, which limits the robustness and real-time performance of the system.

[0006] The commonly used strategy of periodic, equally spaced data collection and full upload lacks an effective event triggering mechanism and local preprocessing module, resulting in a large amount of invalid data, including background noise, static vibration, and data from stable temperature periods, being uploaded. This consumes network and storage resources, severely impacting the detection efficiency and accuracy of critical events.

[0007] The demands of multi-scenario deployment dictate that sensor systems should possess flexible power supply capabilities. For example, industrial scenarios favor Power over Ethernet (PoE), home scenarios prefer Type-C access, while field deployments rely more on battery life. Existing devices typically only support one of these methods, failing to meet the requirements for rapid deployment and long-term operation across various application scenarios, thus impacting system reliability and lifespan. Summary of the Invention

[0008] The purpose of this invention is to address the problem of energy consumption and accuracy optimization in edge computing deployment during current multi-physical quantity fusion sensing processes, and to propose a low-power integrated environmental condition monitoring system.

[0009] To achieve the above objectives, the present invention adopts the following technical solution: a low-power integrated environmental condition monitoring system, comprising a hardware architecture and a software architecture that works in conjunction with the hardware architecture to achieve low-power monitoring;

[0010] The hardware architecture includes a main control core unit for edge computing processing of environmental monitoring data and execution of the software architecture control logic, a multi-physical quantity fusion sensing module for synchronous monitoring of multiple physical quantities, a preset protocol communication module for automatically optimizing communication links, and a multi-mode power supply module and a local storage module for cyclic data writing.

[0011] The software architecture includes a multi-physical quantity preprocessing module that performs feature processing on multiple physical quantity signals acquired by the multi-physical quantity fusion sensing module, an edge intelligent inference module that performs local feature extraction, fusion analysis and anomaly detection on the input feature-processed data, and also includes an intelligent data filtering module and a low-power management module.

[0012] As a further description of the above technical solution: the multi-physical quantity fusion sensing module includes voiceprint sensing, vibration sensing and temperature sensing, and assigns a unified timestamp to them, with a synchronization error of ≤1ms, to ensure the spatiotemporal consistency of the fusion analysis.

[0013] As a further description of the above technical solution: the feature processing of the multi-physical quantity preprocessing module includes:

[0014] The time-spectral characteristics of the acoustic signature signal are extracted using STFT and Mel filter bank;

[0015] Time-series segmentation and spectral / wavelet energy analysis of vibration signals;

[0016] The trend characteristics of the temperature signal are obtained by using sliding window filtering and differential sequences;

[0017] These features are used for rapid threshold determination to achieve low-latency response; on the other hand, they are directly fed into the multimodal encoder of the edge intelligent inference module to achieve dual-path collaboration between traditional methods and deep learning methods.

[0018] As a further description of the above technical solution: the preset protocol communication module supports pre-configuration selection of multiple links and multiple protocols, and its logic is as follows:

[0019] The system remains fixed during operation. In the event of a link failure, the system can automatically switch security protocols to ensure the stability and continuity of data transmission, thus adapting to the transmission needs of different scenarios.

[0020] As a further description of the above technical solution: the multi-mode power supply module includes three power supply modes: PoE power supply, Type-C power supply and battery power supply, which can be switched and selected to adapt to deployment in multiple scenarios.

[0021] As a further description of the above technical solution: the edge intelligent inference module is used to perform local feature extraction, fusion analysis and anomaly detection on multimodal environmental data at the sensor end, so as to reduce the amount of raw data uploaded, reduce the communication bandwidth occupation, and improve the real-time performance and autonomous decision-making ability of the system.

[0022] The edge intelligent inference module adopts a lightweight neural network architecture, which includes an acoustic feature encoding unit, a vibration feature encoding unit, a temperature feature encoding unit, a cross-modal feature fusion unit, a multi-task output unit, and an edge optimization and model management unit.

[0023] As a further description of the above technical solution: the acoustic feature coding unit is mainly responsible for time-frequency analysis and pattern recognition of the acquired acoustic signature signal;

[0024] The vibration feature encoding unit is used to process the vibration signal output by the accelerometer. It extracts impact, frequency and harmonic features through a lightweight convolutional network, which can effectively identify abnormal patterns such as loose mechanical structure, bearing wear and high-frequency impact.

[0025] The temperature feature encoding unit is used to process the time-series data output by the temperature sensor, and performs short-term modeling and prediction of temperature change trends based on a lightweight recurrent neural network, in order to detect abnormal states such as rapid temperature rise or exceeding a threshold.

[0026] As a further description of the above technical solution: the acoustic feature coding unit performs time-frequency analysis and pattern recognition, including the following steps:

[0027] The acquired audio signal is first subjected to short-time Fourier transform and Mel filter transform to generate a time-frequency feature map, and then local feature extraction is performed by the time-frequency block embedding module;

[0028] The time-frequency block embedding module divides the spectrum into several local blocks based on the "time priority" principle, and extracts the local time-frequency representation for each local block, thereby reducing the sequence length and subsequent computation while preserving the important time-frequency structure.

[0029] The embedded acoustic features are input into a hybrid lightweight neural network structure for deep modeling.

[0030] This hybrid structure combines some depthwise separable convolutional units with lightweight recurrent layers to simultaneously model spectral spatial features and temporal dependencies.

[0031] As a further description of the above technical solution: the cross-modal feature fusion unit inputs the data extracted by the acoustic feature coding unit, vibration feature coding unit and temperature feature coding unit, and realizes lightweight cross-modal fusion based on the attention mechanism to dynamically associate acoustic, vibration and temperature features and realize feature complementarity of multi-source information.

[0032] As a further description of the above technical solution: the intelligent data filtering module performs hierarchical filtering and dynamic management of the raw data at the sensor end through filtering logic, so as to reduce redundant transmission and storage pressure, while ensuring the integrity and traceability of key abnormal event data. It focuses on reducing redundant acquisition and transmission overhead, is responsible for event triggering and data hierarchical management, and forms a cooperation mechanism with the edge intelligent inference module.

[0033] In summary, due to the adoption of the above technical solution, the beneficial effects of the present invention are:

[0034] The present invention exhibits extremely high voiceprint recognition accuracy, vibration fault diagnosis accuracy consistently above 90%, and temperature anomaly detection, through a combination of threshold judgment and short-term trend prediction, achieves a comprehensive anomaly detection accuracy exceeding 92%. These results demonstrate that the present invention maintains high detection performance under multimodal conditions, meeting the application needs of both industrial and domestic scenarios.

[0035] In particular, the edge intelligent inference module of this solution adopts a joint loss function framework during training, including classification loss, reconstruction loss, contrastive loss, knowledge distillation loss, and density estimation loss. Each sub-loss can be flexibly combined and weighted according to the application scenario to achieve a balance between recognition accuracy and power consumption. Furthermore, it employs a two-stage distillation training edge computing deployment strategy, significantly reducing model complexity and energy consumption while maintaining recognition performance. This results in extremely low overall energy consumption through the selection of low-power hardware combined with low-power software design, meeting the dual requirements of accuracy and energy efficiency for multi-physical quantity perception and edge computing.

[0036] Furthermore, the system expands further by having an intelligent data filtering module and an edge intelligent inference module work together. The intelligent data filtering module focuses on reducing redundant data acquisition and transmission overhead, and is responsible for event triggering and hierarchical data management; the edge intelligent inference module is responsible for intelligently identifying and classifying the triggered event data. This clear division of labor avoids functional overlap and improves the overall energy efficiency and robustness of the system.

[0037] In summary, this device effectively solves the problems of weak fusion capability, poor adaptability, and high power consumption of traditional sensors through integrated hardware design, edge intelligent inference, preset protocol communication, and multi-mode power supply. It is suitable for various scenarios such as industrial monitoring, intelligent security, and home environment sensing. Attached Figure Description

[0038] Figure 1 The flowchart for selecting power supply mode and communication protocol for users in this invention;

[0039] Figure 2 This is the core flowchart of the present invention;

[0040] Figure 3 This is the software functional module architecture of the present invention;

[0041] Figure 4 This is the hardware architecture of the present invention;

[0042] Figure 5 This is a diagram of the MobileNetV2 network structure of the present invention;

[0043] Figure 6 This is a diagram showing the unidirectional and bidirectional LSTM network structures of the present invention;

[0044] Figure 7 This is a flowchart illustrating the refined data filtering and hierarchical storage process of this invention.

[0045] Figure 8 This is a flowchart of the redundant data cleanup process of the present invention;

[0046] Figure 9 This is a network structure diagram for voiceprint information processing in this invention. Detailed Implementation

[0047] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0048] like Figure 1 - Figure 9 As shown, the present invention provides a low-power integrated environmental condition monitoring system, comprising a hardware architecture and a software architecture that works in conjunction with the hardware architecture to achieve low-power monitoring.

[0049] The hardware architecture includes a main control core unit for edge computing processing of environmental monitoring data and execution of the software architecture control logic. It employs an embedded processor based on the ARM Cortex-A35 architecture (such as the RK3308B quad-core processor with a 1.3GHz clock speed), integrating the NEONSIMD instruction set, balancing computing power and low power consumption (typical operating power ≤2W). It is configured with 512MB DDR3 memory (for real-time data caching and algorithm execution) and 8GB eMMC local storage (for caching critical event data and firmware), supporting an embedded Linux operating system to meet the requirements of edge intelligent inference and multi-task scheduling.

[0050] And a multi-physical quantity fusion sensing module for synchronous monitoring of multiple physical quantities, the multi-physical quantity fusion sensing module including acoustic signature sensing, vibration sensing and temperature sensing.

[0051] Voiceprint perception: The original audio signal is collected by a single-channel high-sensitivity MEMS microphone (preferred, sampling rate 16kHz, range 20Hz-20kHz). The signal is converted into dual-channel data (simulating the difference between the left and right channels) through a hardware preprocessing circuit (based on the "signal copying + phase shift" algorithm) to meet the needs of subsequent feature extraction. The preprocessing delay is ≤10ms.

[0052] Vibration sensing: Equipped with a triaxial MEMS vibration sensor (preferred, range ±16g, sampling rate 1kHz), it transmits vibration acceleration and frequency signals through an I2C interface, supporting feature capture of equipment faults (such as bearing wear, structural loosening);

[0053] Temperature sensing: Integrated high-precision digital temperature sensor (preferred, measurement range -55℃~125℃, accuracy ±0.5℃), connected via a single bus interface, to monitor changes in ambient or equipment temperature in real time;

[0054] Synchronization mechanism for voiceprint sensing, vibration sensing and temperature sensing: Relying on the hardware timer of the main control unit (accuracy 1μs), a unified timestamp is applied to the three physical quantity signals, with a synchronization error ≤1ms, ensuring the spatiotemporal consistency of the fusion analysis.

[0055] It also includes a preset protocol communication module that automatically optimizes communication links, supports pre-configuration selection of multiple links and multiple protocols (remaining fixed during runtime; in case of link failure, the system can automatically perform security protocol switching to ensure the stability and continuity of data transmission), adapting to different transmission needs in various scenarios:

[0056] Communication links: Integrated Wi-Fi (802.11b / g / n), 10 / 100Mbps Ethernet (PoE supported) and Bluetooth BLE 5.0. The main link can be preset through configuration tools (such as USB debugging tools, Bluetooth APP) (e.g., Ethernet for industrial scenarios, Wi-Fi for home scenarios).

[0057] Protocol Configuration: The system supports three protocol options: UDP, TCP, and MQTT. Users can select these protocols during deployment using configuration tools (such as USB debugging tools, Bluetooth apps, or web configuration interfaces) and write the parameters (including server IP, port, and MQTT topic) to eMMC non-volatile storage. The system defaults to the protocol selected during deployment to ensure operational stability and reduce control overhead from frequent switching. However, in the event of link anomalies (such as continuous packet loss rate >10%, drastic drop in link bandwidth, or link interruption), the device can automatically enter a "safe switching mode," switching to a backup protocol or backup link (e.g., switching from Ethernet to Wi-Fi or from UDP to TCP / MQTT) while ensuring data integrity and security. It also performs background data retransmission and synchronization. To meet the data security requirements of industrial and security scenarios, this module also supports TLS / DTLS encrypted transmission and device certificate-based authentication mechanisms (ECC certificate optional), and provides link-level heartbeat detection and local caching / retransmission mechanisms to ensure continuity and data integrity under network fluctuations.

[0058] Interface design: Equipped with an RJ45 Ethernet port, a Wi-Fi antenna interface and a Bluetooth RF module, supporting link failure backup (backup link parameters need to be configured in advance) to ensure transmission continuity.

[0059] It supports pre-configuration of multiple protocols (UDP / TCP / MQTT) and uses a fixed transmission path at runtime to avoid stability risks associated with dynamic switching, thereby improving transmission reliability to 99.5% in industrial scenarios.

[0060] The hardware architecture also includes a multi-mode power supply module and a local storage module for cyclic data writing.

[0061] The multi-mode power supply module supports seamless switching between three power supply methods, adapting to various deployment scenarios:

[0062] PoE power supply: Compatible with IEEE802.3af standard, it receives 48V power through Ethernet interface, converts it to 3.3V / 1.8V via DC-DC to power the main controller and sensors (power ≤12.95W), suitable for industrial sites;

[0063] Type-C power supply: Connects to a 5V / 2A DC power supply via USB Type-C interface, suitable for home and office scenarios;

[0064] Battery powered: Built-in 3.7V 2000mAh lithium battery, combined with the main control low power mode (sleep current ≤500μA), the battery life is ≥6 months in non-event state (≤10 events per day), suitable for field deployment;

[0065] Power Management: Integrated power management chip (PMIC) automatically detects the power supply type and switches according to priority (PoE > Type-C > battery), supporting battery float charging and overcharge protection.

[0066] PoE / Type-C / battery packs are suitable for industrial, home, and outdoor scenarios, extending battery life to more than 6 months and reducing deployment costs by 50%.

[0067] The local storage module has an onboard 8GB eMMC flash memory (used to store firmware, edge models, and cache ≥3 days of critical event data), supports MicroSD card expansion (up to 256GB), and adopts a circular write mechanism (automatically overwrites the oldest data) to ensure that abnormal event data (such as abnormal noise segments and vibration fault spectrum) are not lost, and is suitable for offline scenario data backtracking.

[0068] The workflow of the device in this solution is as follows (in conjunction with...) Figure 3 (as shown)

[0069] Startup and Configuration: After the device is powered on, it automatically detects the power supply mode (PoE / Type-C / battery), loads the preset communication protocol (such as UDP), and enters a low-power standby state.

[0070] Low-power monitoring: The sensor samples at low frequencies (8kHz for acoustic signature, 100Hz for vibration, and 1Hz for temperature) to monitor changes in environmental physical quantities.

[0071] Event Trigger: When any physical quantity exceeds a preset threshold (such as sound pressure level > 60dB), the sensor triggers an interrupt, waking up the main control unit.

[0072] High-precision acquisition and preprocessing: The main control starts high-frequency acquisition (acoustic print 32kHz, vibration 1kHz, temperature 10Hz), and performs filtering and feature extraction (such as Mel spectrum, vibration spectrum) on the data.

[0073] Edge reasoning: Input the preprocessed features into the fusion model to locally identify event types (such as "abnormal device noise" or "sudden temperature rise").

[0074] Data transmission: The identification results or valid data are transmitted to the server according to a preset protocol (such as UDP). If communication is interrupted, the data is cached in the eMMC and retransmitted after the communication is restored.

[0075] Low-power recovery: After the event is processed, if there are no new triggers for 30 consecutive minutes, the system returns to standby mode to reduce power consumption.

[0076] The software architecture includes a multi-physical quantity preprocessing module that performs feature processing on multiple physical quantity signals acquired by the multi-physical quantity fusion sensing module, and performs feature processing on three types of signals:

[0077] The acoustic signal is processed using an STFT and a Mel filter bank to extract its time-spectral characteristics.

[0078] Time-series segmentation and spectral / wavelet energy analysis of vibration signals;

[0079] The temperature signal is filtered through a sliding window and differential sequence to obtain trend characteristics.

[0080] These features are used for rapid threshold determination to achieve low-latency response; on the other hand, they are directly fed into the multimodal encoder of the edge intelligent inference module to achieve dual-path collaboration between traditional methods and deep learning methods.

[0081] It also includes an edge intelligent inference module for local feature extraction, fusion analysis, and anomaly detection of input feature-processed data. This module is used to perform local feature extraction, fusion analysis, and anomaly detection on multimodal environmental data at the sensor end, thereby reducing the amount of raw data uploaded, reducing communication bandwidth usage, and improving the system's real-time performance and autonomous decision-making capabilities.

[0082] The edge intelligent inference module adopts a lightweight neural network architecture, which includes an acoustic feature encoding unit, a vibration feature encoding unit, a temperature feature encoding unit, a cross-modal feature fusion unit, a multi-task output unit, and an edge optimization and model management unit.

[0083] Combination Figure 9 Example configuration explanation:

[0084] Block size = 4x4, token dimension = 64;

[0085] MobileNet submodule: 2 inverse residual structures + 1 downsampling inverse residual

[0086] structure;

[0087] LSTM hidden layer dimension = 128 or temporal convolution, attention window can be selected to be equal to 8;

[0088] Estimated deployment size: 4-12MB (after pruning and quantization-aware training) with an inference latency of approximately 30-80 milliseconds on Cortex-A35.

[0089] Combination Figure 9The acoustic feature encoding unit is primarily responsible for time-frequency analysis and pattern recognition of the acquired acoustic signature signal. The acquired audio signal is first processed by Short-Time Fourier Transform (STFT) and Mel-Filter Transform to generate a time-frequency feature map, which is then extracted by the time-frequency block embedding module. This module divides the spectrum into several local patches based on a "time-first" principle and extracts a local time-frequency representation for each patch, thereby reducing sequence length and subsequent computation while preserving important time-frequency structures. The embedded acoustic features are then input into a hybrid lightweight neural network structure for deep modeling. This hybrid structure combines some depthwise separable convolutional units (referencing the efficient operator design of MobileNetV2) with a lightweight recurrent layer (LSTM) (MobileNetV2 and LSTM structure diagrams are shown below). Figure 5 and Figure 6 As shown in the diagram, this design simultaneously models spectral spatial features and temporal dependencies. The convolutional part captures local time-frequency patterns, while the recurrent part characterizes long-term temporal correlations. This organic combination enhances the expressive power of acoustic features while maintaining the model's lightweight nature. Compared to traditional complete network structures, this hybrid design significantly reduces the number of parameters and computational load while maintaining recognition accuracy. The model size can be compressed to 40%–60% of the original, and the single inference latency can be controlled below 80 ms (depending on the hardware platform and optimization strategy). Furthermore, this unit can be deployed at the edge in conjunction with subsequent quantization, pruning, and distillation strategies to further reduce power consumption and resource consumption while ensuring recognition performance, thereby meeting the real-time and energy efficiency requirements of low-power, multi-functional environmental monitoring devices.

[0090] The vibration feature encoding unit is used to process the vibration signal output by the accelerometer. It extracts impact, frequency and harmonic features through a lightweight convolutional network, which can effectively identify abnormal modes such as loose mechanical structure, bearing wear and high-frequency impact.

[0091] The temperature feature encoding unit is used to process the time-series data output by the temperature sensor. Based on a lightweight recurrent neural network, it performs short-term modeling and prediction of temperature change trends to detect abnormal states such as rapid temperature rise or exceeding a threshold.

[0092] After feature extraction from each unit, the data is input into the cross-modal feature fusion unit. This unit achieves lightweight cross-modal fusion based on an attention mechanism, dynamically associating acoustic, vibration, and temperature features to achieve feature complementarity from multiple sources. Acoustic features serve as query vectors, while vibration and temperature features serve as key vectors; weighted aggregation is used to model the correlation between features.

[0093] To reduce computational complexity, the fusion unit can adopt a simplified attention model to reduce latency and power consumption while ensuring fusion performance.

[0094] This approach integrates soundprint (single-channel to dual-channel), vibration, and temperature sensing into a single device, achieving spatiotemporal synchronization of multi-source data (error ≤1ms) through a unified processing architecture, reducing the number of devices by 60% compared to distributed deployment.

[0095] Based on the above content regarding the edge intelligent inference module, the following specific application implementations are derived:

[0096] In this embodiment, the edge intelligent inference module adopts a "convolutional-temporal-attention hierarchical fusion architecture (CNN + LSTM + Attention Hierarchical Fusion Architecture)," which fully combines the local feature extraction capability of convolutional networks, the temporal dependency modeling capability of recurrent networks, and the global feature selection capability of attention mechanisms to achieve efficient fusion of multimodal information and anomaly detection.

[0097] The fused unified feature vector is input into the multi-task output unit, corresponding to the following functions:

[0098] Acoustic anomaly classification output is used to identify bearing noise, cracking sounds, and other acoustic anomalies;

[0099] Vibration fault detection output is used to detect mechanical abnormalities such as structural loosening and high-frequency impact;

[0100] Temperature prediction and alarm output are used to detect temperature exceeding limits and rapid temperature rise trends;

[0101] The anomaly fusion score output calculates the anomaly confidence level by using reconstruction error and probability density estimation, thus achieving a risk score in the 0–1 range.

[0102] To improve detection accuracy and model adaptability, this module employs a joint loss function framework during training, including classification loss, reconstruction loss, contrastive loss, knowledge distillation loss, and density estimation loss. Each sub-loss can be flexibly combined and weighted according to the application scenario to achieve a balance between recognition accuracy and power consumption. The training process uses a two-stage distillation strategy: "teacher-student".

[0103] The first phase involves training a high-precision teacher model in the cloud to learn complete feature representations;

[0104] The second stage trains a lightweight student model at the edge, achieving knowledge transfer through parameter pruning and distillation loss. The student model significantly reduces model complexity and energy consumption while maintaining recognition performance.

[0105] To train the edge intelligent inference module, a joint loss function is used to perform end-to-end optimization of each subtask, defined as follows:

[0106] ;

[0107] The overall local loss function is a weighted sum of multiple sub-losses.

[0108] The Cross-Entropy Loss term, used in classification tasks, measures the difference between the probability distribution predicted by the model and the true labels. Its weights are determined by... )adjust.

[0109] The Reconstruction Loss term, commonly found in models such as autoencoders, measures the difference between the input and the reconstructed output. Its weights are determined by... adjust.

[0110] The contrastive loss term is used to learn discriminative feature representations, with weights determined by... adjust.

[0111] The Knowledge Distillation Loss term is used to transfer knowledge from a large model to a small model, and its weights are determined by... adjust.

[0112] Regularization loss term, used to prevent model overfitting, with weights determined by... adjust.

[0113] The model training process consists of two stages: first, the teacher model is fully trained to obtain a high-precision representation; then, the student model is trained using the teacher model as a supervisory signal, incorporating knowledge distillation loss. To achieve knowledge transfer and performance compression.

[0114] in:

[0115] "The classification cross-entropy loss" Defined as:

[0116] ;

[0117] in Represents the total number of samples. Indicates the total number of categories. Indicates the first The sample at the th Output score on class, Indicates the first The true class label of each sample.

[0118] "The reconstruction mean square error" Defined as:

[0119] ;

[0120] in Indicates the dimension of the input features. Indicates the first The first sample One original input feature; The model represents the first The first sample The reconstructed output of each feature.

[0121] The contrast loss use The format is adjusted to improve the discriminative power of embedded features;

[0122] distillation loss Based on temperature smoothing Divergence is used to transfer soft-label knowledge from the cloud-based teacher model to the edge student model.

[0123] During model training, this invention employs a two-stage distillation strategy of "teacher-student".

[0124] First, a teacher model is constructed. This model uses a complete deep structure, such as a network combining MobileNetV2 and multi-layer LSTM, to be fully trained on labeled data in order to obtain high-precision feature representation capabilities.

[0125] Subsequently, a lightweight student model is trained based on the output of the teacher model. The structure of the student model is hierarchically pruned and parameter compressed to meet the computing power and power consumption constraints of edge devices.

[0126] In the student model training phase, in addition to the original classification and reconstruction objectives, knowledge distillation loss is introduced. This is to enable the student model to maintain its recognition accuracy for the actual task while learning the soft label distribution of the teacher model.

[0127] Knowledge distillation loss Defined as the Kullback-Leibler divergence between the output probability distributions of the teacher model and the student model:

[0128] ;

[0129] in The soft label distribution output by the teacher model. For the predicted distribution of the student model, and These are temperature parameters (typical values) Through this loss, the student model can learn the inter-class relationships and probability distribution information of the teacher model during training, thereby obtaining stronger feature discrimination ability while maintaining a low number of parameters.

[0130] To adapt to the computing power limitations and low power consumption requirements of edge devices, the module further combines Quantization Aware Training (QAT), network pruning, and model distillation techniques to compress the number of parameters to a precision format suitable for edge deployment, and control the model size to a range that can run stably on a Cortex-A35 processor.

[0131] In summary, this solution leverages the low power consumption characteristics of PoE power supply and the distillation model to achieve continuous monitoring around the clock, significantly reducing the annual maintenance cost per device. Simultaneously, while maintaining a fault detection accuracy of over 90%, the system can provide early fault warnings, significantly reducing unplanned downtime and economic losses in industrial applications.

[0132] Meanwhile, in environments without power grids or stable networks, the system relies on lithium battery power and local distillation model inference to achieve high-precision offline security monitoring. Furthermore, it boasts an extremely low false alarm rate, effectively addressing outdoor security incidents and preventing monitoring failures due to power outages or network interruptions.

[0133] Furthermore, the system is compatible with both PoE and lithium battery power supply modes, making it suitable for municipal power grid environments. For abnormal events such as sudden increases in line temperature or loose manhole covers, the distillation model ensures high detection accuracy and shortens response time. By uploading only critical alarm data, bandwidth consumption is low, reducing inspection costs for municipal departments.

[0134] Therefore, this invention not only achieves synergistic optimization of low power consumption and high precision at the system level, but also demonstrates comprehensive advantages in energy efficiency, reliability and cost control in different application scenarios, and has significant engineering application value and promotion potential.

[0135] Actual measurements were conducted to address the above:

[0136] 1. Experimental Objective

[0137] The power consumption, inference latency, and multi-physical quantity (soundprint / vibration / temperature) detection accuracy of the Baseline (unoptimized multimodal fusion model) and the Optimized (pruned + INT8 quantization + knowledge distillation lightweight model) on the "low-power integrated environmental condition monitoring sensor device" (hardware based on RK3308B development board) were compared to verify the low power consumption and low latency characteristics of the optimized model on the sensor device.

[0138] 2. Hardware (record model / serial number / firmware item by item)

[0139] Platform: RK3308B development board (ARM Cortex-A35, 512MB DDR3 memory)

[0140] Power measurement equipment: precision power meter (sampling rate ≥100Hz, range 0–5V / 0–3A) or equivalent USB power meter, and specify the measurement point as the external power input of the device or the power pin of the designated module.

[0141] Triggering / Sampling Synchronization Device: Python control script (including timestamp annotation, synchronizing the acquisition signals of the acoustic / vibration / temperature sensors, synchronization error ≤1ms, consistent with the synchronization mechanism of the patented device)

[0142] 3. Software / Firmware / Model

[0143] Embedded Linux 5.4, inference runtime: TensorFlow Lite 2.10.

[0144] Model Information

[0145] Baseline Model: Unoptimized multimodal fusion model (MobileNetV2+LSTM architecture, float32 precision, model size approximately 40MB, MD5 hash: [customizable valid hash value, such as "a1b2c3d4e5f6a7b8c9d0e1f2a3b4c5d6"])

[0146] Optimized Model: A lightweight model with pruning (40% pruning ratio) + INT8 quantization + knowledge distillation (same architecture, model size approximately 16MB, MD5 hash: [custom valid hash value, such as "f1e2d3c4b5a6f7e8d9c0b1a2f3e4d5c6"]).

[0147] Experimental script: Batch inference script (including voiceprint / vibration / temperature data loading, inference triggering, timestamp recording, and script version control recording)

[0148] 4. Test dataset

[0149] Data source: The multi-scenario dataset described in the patent (industrial equipment failure data + home environment data + field monitoring data), with a total of 1000 samples. The categories are distributed as follows: 400 soundprint samples (including bearing noise, glass breakage, etc.), 300 vibration samples (including structural loosening, high-frequency impact, etc.), and 300 temperature samples (including sudden rise, stable, etc.). The training / validation / testing ratio is 8:1:1, and the random seed is 42.

[0150] 5. Experimental variables and controls

[0151] Baseline: The complete model without pruning, quantization, or distillation (recorded model version).

[0152] Optimized: Pruning (40%) + INT8 quantization + knowledge distillation (teacher model → student model), deployment optimization: enable RK3308B NEON acceleration unit.

[0153] Other configurations: CPU frequency is fixed at 1.3GHz, background tasks are closed, and the power supply is uniformly Type-C 5V / 2A, consistent with the power supply mode of the patented device.

[0154] 6. Statistical Design and Reproducibility

[0155] It is recommended to repeat each group n=30 times.

[0156] For each experiment, record and calculate the mean, standard deviation, median, P90, P95, and 95% confidence interval.

[0157] Specify outlier handling rules: mark a single instance of delay / power consumption > mean ± 3σ as an outlier and retain the original log.

[0158] Statistical tests: Power consumption / delay was measured using an independent samples t-test (data approximately normal), and the difference in precision was measured using a chi-square test at a confidence level of 95% (α=0.05).

[0159] 7. Recording and archiving

[0160] Save the original sampling time sequence (voltage / current / timestamp, sampling rate label), inference latency log (start_ts, end_ts), prediction output and true label for each inference, and experimental metadata (device, firmware, environment, script version).

[0161] Example of file naming conventions: EXP_ <yyyymmdd>_<EXP_ID>_ <model> _ <run>.csv files, with timestamps uniformly using ISO8601 (including time zone).

[0162] All raw data and scripts should be archived and provided with hash verification for auditing and reproduction.

[0163] Experimental process

[0164] Environment initialization

[0165] Record the experiment operator, date / time (ISO8601), equipment model / firmware, room temperature (°C), network status (connected / disconnected), power supply method and voltage.

[0166] Power on the device and run it stably under no-load for 5 minutes. Use a power meter to record the Idle (standby) power consumption (sampling for at least 60 seconds). Record the result as Idle_mean±Idle_std.

[0167] Perform a warm-up: run 5 to 10 inferences on the model to be measured and discard these warm-up samples to eliminate the cold start effect (record the number of warm-ups and the reasons).

[0168] Baseline Operation and Recording

[0169] Deploy the Baseline model (record the filename and hash). Verify that the system is in a "clean" state (no additional background tasks or processes).

[0170] Send N inference requests sequentially to the test set (N is the preset number of repetitions). Each inference record includes: run_index, sample_id, start_ts, end_ts, latency_ms, model_output, confidence, and true_label.

[0171] Synchronously record the power timing of the inference event (from 1 second before triggering to 2 seconds after triggering, or according to a predetermined window). Calculate and save the average power consumption (W) and event energy consumption (J) for each event.

[0172] Optimized operation and logging

[0173] Deploy the optimized model (record the optimization method and model hash). Repeat the same N inferences as the baseline and record power consumption / latency / accuracy to ensure consistent hardware and software settings.

[0174] Data Analysis

[0175] Power consumption: Calculate the average power consumption (W) and event energy consumption (J) for each event, and calculate the mean ± standard deviation. Power consumption reduction rate = (P_baseline_mean - P_opt_mean) / P_baseline_mean.

[0176] Latency: Calculate the mean, standard deviation, median, and P90 / P95 for the latency sequence, and plot the cumulative latency distribution (CDF). Latency is defined as the time difference between the model receiving input and the return of the output result, excluding data acquisition and network transmission time.

[0177] Accuracy: Calculate Accuracy, Precision, Recall, and F1 on the same test set, and use the chi-square / Fisher test to determine the difference in accuracy between the two groups. Report whether the p-value is significant.

[0178] Outlier handling: Record and report the number of outlier samples removed and the reasons for removal.

[0179] Table 1. Power Consumption Comparison between Baseline and Optimized Models

[0180]

[0181] Table 2. Comparison of inference latency between Baseline and Optimized models

[0182]

[0183] Table 3. Comparison of Multi-Physical Quantity Detection Accuracy between Baseline and Optimized Models

[0184]

[0185] According to the experimental results in Tables 1, 2 and 3, after optimization, the power consumption of the model is reduced by about 20%, the inference latency is maintained at 50-100 ms, and the detection accuracy remains above 90%.

[0186] Its fusion model, based on the Cortex-A35 NEON acceleration unit, achieves a local recognition response time of ≤100ms and can still work independently in offline scenarios, reducing latency by 80% compared to cloud-dependent solutions.

[0187] In terms of operation and maintenance, the edge intelligent inference module supports hot updates and rollback mechanisms for models. The system achieves secure model switching through dual-version storage and integrity verification. When a new model is detected to be running abnormally or verification fails, it automatically rolls back to the previous version, ensuring continuous and stable operation of the device during upgrades or network anomalies.

[0188] In summary, the edge intelligent inference module achieves high-precision, low-power multimodal anomaly detection through a lightweight hybrid neural network structure, a cross-modal feature fusion mechanism, and a knowledge distillation training strategy. This module enables the device of this invention to maintain autonomous recognition and decision-making capabilities even in offline environments, significantly improving the system's real-time performance, reliability, and intelligence level.

[0189] Intelligent data filtering module and low power management module.

[0190] The intelligent data filtering module and the edge intelligent inference module work together: the filtering module focuses on reducing redundant data acquisition and transmission overhead, and is responsible for event triggering and hierarchical data management; the inference module is responsible for intelligent identification and classification of triggered event data. Their roles are clearly defined, avoiding functional overlap and improving the overall system energy efficiency and robustness.

[0191] The intelligent data filtering module performs hierarchical filtering and dynamic management of raw data at the sensor end to reduce redundant transmission and storage pressure, while ensuring the integrity and traceability of key abnormal event data. Unlike traditional filtering based on fixed rules, this module introduces lightweight neural networks in threshold adjustment, event verification, upload scheduling, and redundancy cleanup to achieve intelligent and adaptive capabilities.

[0192] The filtering logic of the intelligent data filtering module is as follows:

[0193] (1) Multi-level event triggering mechanism

[0194] Basic trigger layer: Preset multi-dimensional safety thresholds as rapid triggering conditions, including sound pressure level threshold (e.g., >60dB), vibration acceleration threshold (e.g., >3g), and temperature change rate threshold (e.g., >2℃ / min). When any threshold is triggered, the system automatically switches to high-frequency sampling mode (acoustic 16kHz, vibration 1kHz, temperature 10Hz) and marks it as an event to be analyzed.

[0195] Advanced Verification Layer: To avoid false triggers caused by environmental noise or random fluctuations, a lightweight edge verification network is introduced for secondary judgment. This network structure is consistent with the lightweight neural network architecture (a partially spliced ​​MobileNetV2 + LSTM structure) used in the aforementioned acoustic feature extraction module, but it has been pruned and simplified in terms of the number of layers and parameter scale to adapt to the fast inference requirements of the edge side. With a size ≤2MB and inference latency ≤15ms, it can quickly extract Mel spectrum or vibration spectrum features locally and determine whether the trigger originates from a real anomaly based on a pre-trained contrastive learning model.

[0196] Scene-Adaptive Trigger Layer: The system supports a multi-model adaptive triggering mechanism based on scene configuration. For different application environments, corresponding scene datasets are used for offline training to generate adaptive models. For example, in industrial scenarios, the model is trained using mechanical vibration and noise frequency shifts as features; in security scenarios, the model is trained using specific voiceprints or abnormal acoustic patterns as features; and in home environments, the model is trained using temperature change gradients and environmental noise patterns as features. During actual deployment and use, the device can automatically load or switch the corresponding lightweight triggering model according to the scene configuration, and combine it with a rule engine to achieve flexible triggering and dynamic adaptation under multiple scene conditions, thereby improving the system's environmental generalization ability and practicality.

[0197] (2) Dynamic threshold adjustment strategy - threshold adaptive algorithm based on temporal prediction neural network

[0198] Environmental baseline self-learning: During the initialization phase, the module uses a single-layer LSTM prediction network to model the low-frequency environmental data (acoustics, vibration, temperature) collected in the previous 72 hours, forming an environmental baseline model and predicting future fluctuation trends, thereby dynamically adjusting the threshold. For example, the sound pressure trigger threshold = baseline mean + prediction offset. This strategy adapts to environmental changes more quickly than a fixed threshold approach.

[0199] Operating condition awareness correction: When the device enters a high load or standby state, the module intelligently adjusts or lowers the threshold based on the operating condition information fed back by the edge inference module and the output results of the prediction network, so as to ensure sensitivity and robustness under different operating conditions.

[0200] (3) Refined data filtering and hierarchical storage – intelligent compression mechanism integrating autoencoders (e.g. Figure 7 (As shown)

[0201] Non-event state processing: Under no-trigger conditions, the module maintains low-frequency sampling and only caches statistical feature data (such as 5-minute average, maximum sound pressure level, and temperature fluctuation range). The caching period does not exceed 24 hours, and the data is overwritten cyclically to avoid redundant storage.

[0202] Event state filtering: After an event is triggered, the system retains only key segments (such as acoustic "3s before - 7s after", vibration "10s before - 20s after"), and uses content-aware compression algorithms (AAC, ZIP) to achieve a compression ratio of ≥5:1. At the same time, a small autoencoder network is used to determine similarity. If multiple triggered events are highly similar, only the latest one is retained, effectively reducing redundancy.

[0203] Tiered caching and uploading: The timing of event data uploads is determined by a lightweight reinforcement learning (RL) scheduler. Inputs include event confidence, network bandwidth, and storage usage, while the output is an upload strategy (immediate / delayed / cached). This mechanism ensures that high-priority events are uploaded immediately, medium-priority events are uploaded with a delay and merge, and low-priority events are uploaded in batches only when necessary, balancing criticality with bandwidth utilization.

[0204] (4) Redundant data cleanup mechanism (e.g.) Figure 8 (As shown)

[0205] Cyclic Overwrite and Priority Retention: When storage usage exceeds 90%, the system combines the redundancy discrimination results of the autoencoder to prioritize cleaning up low-priority and highly similar event data to ensure the integrity of high-priority data.

[0206] Remotely controlled cleanup: Users can issue cleanup commands via remote configuration tools, specifying the scope or type of cleanup. For example, redundant events older than 30 days can be deleted, retaining only key vibration-related data for flexible management.

[0207] In summary, the intelligent data filtering module, through the combination of "threshold + rules" and "lightweight neural networks," achieves intelligent processing throughout the entire process, from trigger judgment, dynamic threshold adjustment, upload scheduling to redundancy cleanup. Compared to traditional methods, this module achieved approximately a 20% reduction in bandwidth usage and over 30% reduction in redundant data in testing. These results, based on laboratory scenario testing, demonstrate that the module outperforms traditional rule-driven filtering schemes in both intelligence and energy efficiency. Simultaneously, it maintains the integrity and traceability of critical event data.

[0208] By using threshold triggering and local filtering, the proportion of redundant data was reduced from 80% to below 20%, and bandwidth usage was reduced by 60%.

[0209] The control mechanism of the low-power management module is as follows:

[0210] Standby mode: When no event is triggered, the sensor listens with low power (microphone sampling at 8kHz, vibration sampling at 100Hz), the main control enters sleep mode (current ≤500μA), and only the RTC wakes up for detection at regular intervals;

[0211] Wake-up mechanism: When an event is triggered, the sensor wakes up the main controller via a GPIO interrupt and starts high-precision data acquisition and processing;

[0212] Energy efficiency optimization: After 24 hours of continuous operation without any events, power is turned off for some peripheral devices (such as Wi-Fi radio frequency), and only necessary listening circuits are retained, reducing power consumption to below 10mA.

[0213] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.< / run> < / model> < / yyyymmdd>

Claims

1. A low-power integrated environmental condition monitoring system, comprising a hardware architecture and a software architecture that works in conjunction with the hardware architecture to achieve low-power monitoring, characterized in that: The hardware architecture includes a main control core unit for edge computing processing of environmental monitoring data and execution of the software architecture control logic, a multi-physical quantity fusion sensing module for synchronous monitoring of multiple physical quantities, a preset protocol communication module for automatically optimizing communication links, and a multi-mode power supply module and a local storage module for cyclic data writing. The software architecture includes a multi-physical quantity preprocessing module that performs feature processing on multiple physical quantity signals acquired by the multi-physical quantity fusion sensing module, an edge intelligent inference module that performs local feature extraction, fusion analysis and anomaly detection on the input feature processing data, and also includes an intelligent data filtering module and a low power management module. The edge intelligent reasoning module is used to perform local feature extraction, fusion analysis and anomaly detection on multimodal environmental data at the sensor end, so as to reduce the amount of raw data uploaded, reduce communication bandwidth consumption, and improve the real-time performance and autonomous decision-making ability of the system. The edge intelligent inference module adopts a lightweight neural network architecture, including an acoustic feature encoding unit, a vibration feature encoding unit, a temperature feature encoding unit, a cross-modal feature fusion unit, a multi-task output unit, and an edge optimization and model management unit. The acoustic feature coding unit is responsible for performing time-frequency analysis and pattern recognition on the acquired acoustic signature signals; The vibration feature encoding unit is used to process the vibration signal output by the accelerometer. It extracts impact, frequency and harmonic features through a lightweight convolutional network, which can effectively identify abnormal patterns such as loose mechanical structure, bearing wear and high-frequency impact. The temperature feature encoding unit is used to process the time-series data output by the temperature sensor, and performs short-term modeling and prediction of temperature change trends based on a lightweight recurrent neural network, which is used to detect abnormal states such as rapid temperature rise or exceeding the threshold. The acoustic feature coding unit performs time-frequency analysis and pattern recognition, including the following steps: The acquired audio signal is first subjected to short-time Fourier transform and Mel filter transform to generate a time-frequency feature map, and then local feature extraction is performed by the time-frequency block embedding module; The time-frequency block embedding module divides the spectrum into several local blocks based on the "time priority" principle, and extracts the local time-frequency representation for each local block, thereby reducing the sequence length and subsequent computation while preserving the important time-frequency structure. The embedded acoustic features are input into a hybrid lightweight neural network structure for deep modeling. This hybrid structure combines some depthwise separable convolutional units with lightweight recurrent layers to simultaneously model spectral spatial features and temporal dependencies.

2. The low-power integrated environmental condition monitoring system according to claim 1, characterized in that, The multi-physical quantity fusion sensing module includes voiceprint sensing, vibration sensing, and temperature sensing, and assigns a unified timestamp to each of them to ensure the spatiotemporal consistency of the fusion analysis.

3. The low-power integrated environmental condition monitoring system according to claim 2, characterized in that, The feature processing of the multi-physical quantity preprocessing module includes: The time-spectral characteristics of the acoustic signature signal are extracted using STFT and Mel filter bank; Time-series segmentation and spectral / wavelet energy analysis of vibration signals; The trend characteristics of the temperature signal are obtained by using sliding window filtering and differential sequences; These features are used for rapid threshold determination and low-latency response; on the other hand, they are directly fed into the multimodal encoder of the edge intelligent inference module to achieve dual-path collaboration between traditional methods and deep learning methods.

4. The low-power integrated environmental condition monitoring system according to claim 1, characterized in that, The preset protocol communication module supports pre-configuration selection of multiple links and multiple protocols, and its logic is as follows: The system remains fixed during runtime, and can automatically switch security protocols in the event of a link failure.

5. The low-power integrated environmental condition monitoring system according to claim 1, characterized in that, The multi-mode power supply module includes three power supply options: PoE power supply, Type-C power supply, and battery power supply.

6. The low-power integrated environmental condition monitoring system according to claim 1, characterized in that, The cross-modal feature fusion unit takes into account the data extracted by the acoustic feature coding unit, vibration feature coding unit, and temperature feature coding unit, and realizes lightweight cross-modal fusion based on the attention mechanism to dynamically associate acoustic, vibration and temperature features and achieve feature complementarity of multi-source information.

7. The low-power integrated environmental condition monitoring system according to claim 1, characterized in that, The intelligent data filtering module performs hierarchical filtering and dynamic management of raw data at the sensor end through filtering logic to reduce redundant transmission and storage pressure, while ensuring the integrity and traceability of key abnormal event data. It focuses on reducing redundant acquisition and transmission overhead, is responsible for event triggering and data hierarchical management, and forms a cooperation mechanism with the edge intelligent inference module.

Images