Intelligent water leakage detection system and detection method based on multi-modal sensing and cloud cooperation

Abstract

The invention discloses an intelligent water leakage detection system based on multi-modal sensing and cloud cooperation, which comprises a multi-modal sensing layer, an edge computing node and a cloud analysis platform, and belongs to the technical field of intelligent detection. Through multi-modal sensing and cloud artificial intelligence processing capability, the water leakage detection efficiency and accuracy are improved, and the problems that traditional acoustic detection is susceptible to environmental noise interference, the accuracy is low, and a vibration sensor cannot distinguish water leakage and non-water leakage impact are solved.

Description

Technical Field

[0001] The present application relates to the field of intelligent detection technology, and specifically to an intelligent water leakage detection system based on multimodal sensing and cloud collaboration. Background Art

[0002] Existing water leakage detection systems mainly include two types of data collection: acoustic detection and vibration sensors. Acoustic detection, such as listening rods, is easily interfered by environmental noise and has low accuracy, while vibration sensors cannot distinguish between water leakage and non-water leakage impact.

[0003] In terms of signal processing, traditional CNNs have low sensitivity to weak leak signals, while single classifiers have poor generalization capabilities in complex pipe scenarios. In actual pipeline detection, the leak signals collected by sensors include not only valid leak signals but also some noise signals with high amplitude and complex frequencies. These noise signals include stationary noise signals such as environmental and background noise from water impact within the pipe, as well as sudden non-stationary noise signals from external sources such as motor vehicle traffic and construction sites. The presence of these noise signals can affect the leak signal detection process. Traditional acoustic detection has a false alarm rate of >40% in 30dB ambient noise, while vibration sensors lose 60% sensitivity in non-metallic pipes. Therefore, achieving accurate leak alarms requires both data acquisition and data processing, specifically by leveraging multimodal sensing and cloud-based AI processing capabilities. Summary of the Invention

[0004] An embodiment disclosed in the present application provides an intelligent water leakage detection system based on multimodal sensing and cloud collaboration, including: a multimodal sensing layer, an edge computing node, and a cloud analysis platform; the multimodal sensing layer includes piezoelectric sensors and acoustic sensors; the acoustic sensor uses a multi-microphone array; optionally, the multimodal sensing layer also includes an ultrasonic flow meter and an infrared thermal imager for collecting pipeline vibration, water flow status and surface temperature data.

[0005] In one embodiment disclosed in the present application, the piezoelectric sensor has a frequency response of 100-2500Hz; a multi-microphone array is deployed on the outer wall of the pipe at a spacing of ≤50m, supports LoRa wireless transmission, and the sensor housing is made of IP68-rated waterproof and corrosion-resistant material to collect ambient sound. It integrates an STM32F103 chip and a DSP processor, extracts MFCC features, and uses a lightweight CNN model to classify water leakage sounds in real time. The parameters of the lightweight CNN model are less than 1MB; the IoT platform alarm is triggered only when the probability is >90%. The model is pre-trained using the UrbanSound8K dataset and compressed to 0.8MB using TensorRT INT8 quantization.

[0006] In one embodiment disclosed in this application, the edge computing node includes a main control module, a timing trigger module, a sound preprocessing module, a sound recognition module, a feature fusion module, a dynamic weight distribution module, and an Internet of Things communication module; the main control module is connected to the ReSpeaker 4-Mic array sensor, the sensor pickup radius is ≥3 meters, and the signal-to-noise ratio is ≥60dB. The timing trigger module controls the device to start detection only during the preset period at night. The sound preprocessing module uses Python The librosa library performs the following operations: normalizing the audio sampling rate to 6400Hz, trimming the duration to 4s, extracting 40-dimensional MFCC features, and outputting the leakage probability (0 / 1 classification) after inputting the MFCC features. The feature fusion module concatenates the acoustic MFCC features (40 dimensions) with the vibration signal's time-domain features (peak value / variance) into a joint feature vector. The IoT communication module uploads the results to the platform via HTTP. Edge computing nodes retain only binary classification capabilities (leakage / non-leakage), and full feature vector and probability calculation permissions are locked to the cloud analysis platform. When a decision conflicts between the edge node and the cloud, cloud commands overwrite the local results and trigger incremental model learning. The cloud analysis platform includes a hybrid classifier and a model that uses transfer learning.

[0007] In one embodiment disclosed herein, the sound recognition module is based on an improved Squeeze_Net convolutional neural network. The improved Squeeze_Net structure includes: input layer → convolution layer (96 7×7 convolution kernels) → maximum pooling layer → Fire module × 4 (channels 64 / 128 / 128 / 256) → global average pooling layer → Softmax output layer.

[0008] In one embodiment disclosed in this application, edge computing nodes also include noise reduction mechanisms, including hardware noise reduction and software noise reduction. The hardware noise reduction uses beamforming technology from a ReSpeaker 4-Mic array, while the software noise reduction uses MFCC feature filtering to filter ambient noise below 100Hz. Field measurements in municipal pipe networks have shown that this system's false alarm rate has dropped to 8.2% and its false alarm rate is less than 1.5% under comparable noise conditions.

[0009] An embodiment disclosed in this application, a 1DCNN model example: S1: input layer: 8000 points; S2: Convolutional layer 1: 16 filters, kernel size 32, stride 8 -> output size: (8000-32) / 8+1 = 996 points, 16 channels; S3: Pooling layer 1: max pooling, pooling size 4 -> 996 / 4 = 249 points, 16 channels; S4: 1D convolution layer 2: 32 filters, kernel size 16, stride 4 -> (249-16) / 4+1 ≈ 59 points, 32 channels; S5: Pooling layer 2: max pooling, pooling size 2 -> 59 / 2 ≈ 30 points, 32 channels; S6: Flattened: 30X32=960 features.

[0010] In one embodiment disclosed in the present application, the XGBoost parameter configuration is as follows: number of decision trees = 100, learning rate = 0.1, regularization parameters λ = 1.0, γ = 0.2; objective function: binary logistic regression loss + L2 regularization term; The cascade training process involves freezing the weights of the 1DCNN convolutional layer, training the fully connected layer with a water leakage dataset, using the 1DCNN output features as XGBoost input, and jointly optimizing the classification threshold. The XGBoost classifier receives the pipeline vibration frequency domain features extracted by the 1DCNN and aligns them with the acoustic MFCC features in time and space. The 1DCNN effectively captures the time domain transient characteristics of the water leakage signal, while XGBoost suppresses pipeline resonance noise through feature importance screening. The two work together to solve the echo interference problem of metal pipes.

[0011] In one embodiment disclosed in the present application, a hybrid classifier is provided with an additional SVM classification branch: The 256-dimensional features extracted by 1DCNN are input into the SVM classifier (RBF kernel function, C=1.0, γ=0.01); When the prediction difference between XGBoost and SVM is greater than 0.3, the expert review mechanism is activated; Dynamic selection of SVM kernel function: The iron pipe data uses the RBF kernel (γ = 0.01); The polynomial kernel (degree=3) was used for the PVC pipe data; The mixed pipe data uses the SIGMOID kernel.

[0012] The cloud-based analysis platform also includes multi-model fusion decision-making: Final classification probability = 0.6 × XGBoost probability + 0.4 × SVM probability; A high confidence alarm is triggered when the fusion probability is >0.85 and the consistency rate of the two models is >90%.

[0013] Support Vector Machine (SVM) is a machine learning algorithm used for classification and regression analysis. Its goal is to find a hyperplane that effectively separates samples of different categories, keeping samples of different categories as far away from the hyperplane as possible. The core concept of SVM is model selection based on the principle of structural risk minimization. It also uses kernel functions to map low-dimensional feature spaces to high-dimensional feature spaces to handle nonlinear classification problems.

[0014] In one embodiment disclosed in the present application, a method for optimizing a model using transfer learning includes: Use the SMOTE algorithm to enhance minority class samples; Optimize SVM hyperparameters (C, γ) through grid search; 5-fold cross validation was used to prevent overfitting.

[0015] In one embodiment disclosed in the present application, a dynamic weight allocation module: When the signal-to-noise ratio is greater than 30 dB, the acoustic feature weight is 0.7; When the pipeline vibration amplitude is greater than 0.5g, the vibration characteristic weight is 0.8.

[0016] An embodiment disclosed in this application is an intelligent water leakage detection method based on multimodal sensing and cloud collaboration, including: S1: Collects raw signals through piezoelectric sensors and multi-microphone arrays; S2: Extract signal depth features through 1DCNN; S3: Input XGBoost and SVM to calculate classification probability and feature importance score; S4: When the probability is > 90% and the importance of the water leakage sound feature is greater than the vibration feature, the water pipe alarm is triggered.

[0017] In one embodiment disclosed in the present application, when a combination of 1DCNN+XGBoost and SVM is used, the accuracy is improved;

[0018] In one embodiment disclosed in this application, feature importance analysis includes: Calculate the SHAP value to quantify the contribution of each dimension of MFCC; When the SHAP value of low-frequency energy (100-300 Hz) is > 0.3, it is marked as a drip event.

[0019] The terms used in the following embodiments are only for the purpose of describing specific embodiments and are not intended to limit the present application. As used in the specification of this application and the appended claims, the singular expressions "a", "an", "said", "above", "the" and "this" are intended to also include expressions such as "one or more", unless there is a clear contrary indication in the context. It should also be understood that in the following embodiments of the present application, "multiple" refers to two or more. "At least four" refers to four or more. The term "and / or" is used to describe the association relationship of associated objects, indicating that three relationships can exist; for example, A and / or B can mean: A exists alone, A and B exist at the same time, and B exists alone, where A and B can be singular or plural. The character " / " generally indicates that the previous and subsequent associated objects are in an "or" relationship.

[0020] References to "one embodiment" or "some embodiments" in this specification mean that a particular feature, structure, or characteristic described in conjunction with that embodiment is included in one or more embodiments of the present application. Thus, phrases such as "in one embodiment," "in some embodiments," "in other embodiments," and "in yet other embodiments" appearing in various places in this specification do not necessarily refer to the same embodiment, but rather mean "one or more but not all embodiments," unless otherwise specifically emphasized. The terms "including," "comprising," "having," and variations thereof mean "including but not limited to," unless otherwise specifically emphasized.

[0021] In the description of this application, it should be noted that, unless otherwise specified and limited, all materials used without indicating the manufacturer are conventional products that can be purchased commercially.

[0022] It should be understood that the technical features involved in the different embodiments described in this application can be combined with each other as long as they do not conflict with each other.

Claims

1. An intelligent water leakage detection system based on multimodal sensing and cloud collaboration, characterized in that: include: Multimodal sensing layer, edge computing nodes, and cloud-based analysis platform; The multimodal sensing layer includes a piezoelectric sensor and an acoustic sensor; The acoustic sensor adopts a multi-microphone array; The edge computing node includes a main control module, a timing trigger module, a sound preprocessing module, a sound recognition module, a feature fusion module, a dynamic weight allocation module, and an Internet of Things communication module; The cloud-based analysis platform includes a hybrid classifier and a model using transfer learning. The hybrid classifier is composed of a cascade of a 1DCNN feature extractor with three convolutional layers and an XGBoost classifier. The edge computing node performs acoustic signal preprocessing and preliminary classification, and uploads the feature vector to the cloud analysis platform for joint decision-making only when the confidence level is greater than a threshold. The threshold is the leakage probability output by the sound recognition module in the edge computing node greater than 90%.

2. The intelligent water leakage detection system based on multimodal sensing and cloud collaboration according to claim 1 is characterized in that: The sound recognition module is based on an improved Squeeze_Net convolutional neural network. The improved Squeeze_Net structure includes: input layer → convolution layer → maximum pooling layer → Fire module → global average pooling layer → Softmax output layer.

3. The intelligent water leakage detection system based on multimodal sensing and cloud collaboration according to claim 1 is characterized in that: The edge computing node also includes a noise reduction mechanism including: Hardware noise reduction and software noise reduction. The hardware noise reduction uses the beamforming technology of the ReSpeaker 4-Mic array, and the software noise reduction uses the MFCC feature to filter ambient noise below 100Hz.

4. The intelligent water leakage detection system based on multimodal sensing and cloud collaboration according to claim 1 is characterized in that: The 1DCNN structure includes: input layer → Conv1D →MaxPooling → Conv1D →GlobalAvgPooling → 256-dimensional feature vector; the 1DCNN includes two levels of convolution and pooling layers, and finally outputs a 960-dimensional feature vector.

5. The intelligent water leakage detection system based on multimodal sensing and cloud collaboration according to claim 1 is characterized in that: The XGBoost classifier is deployed on an Alibaba Cloud ECS GPU instance and receives 1DCNN feature data through the PCIe bus. The XGBoost classifier, deployed on a cloud-based GPU accelerator, uses input features derived from the frequency domain conversion output of the piezoelectric sensor analog signal by 1DCNN. The number of decision trees is configured to the upper limit of parallel processing of 50-200 physical computing cores.

6. The intelligent water leakage detection system based on multimodal sensing and cloud collaboration according to claim 1 is characterized in that: The hybrid classifier adds an SVM classification branch; when the difference between XGBoost and SVM predictions is greater than 0.3, an expert review mechanism is activated; the cloud analysis platform also includes multi-model fusion decision-making: Final classification probability = 0.6 × XGBoost probability + 0.4 × SVM probability; The SVM kernel function is dynamically selected: The RBF kernel is used for the iron pipe data; The polynomial kernel was used for the PVC pipe data; The mixed pipe data uses the SIGMOID kernel.

7. The intelligent water leakage detection system based on multimodal sensing and cloud collaboration according to claim 1 is characterized in that: The cloud-based analysis platform is configured with a SHAP value analysis module, which calls the shap.Explainer library to calculate feature contributions. The SHAP value analysis module analyzes the acoustic diffraction characteristics of the multi-microphone array, associates the MFCC 1st-5th dimension coefficients with the physical quantity of the pipeline crack size, and generates a crack location map when the low-frequency energy SHAP value is greater than 0.

3.

8. The intelligent water leakage detection system based on multimodal sensing and cloud collaboration according to claim 1 is characterized in that: The multimodal sensing layer further includes an ultrasonic flow meter for collecting pipeline water flow status data.

9. An intelligent water leakage detection method based on multimodal sensing and cloud collaboration, characterized in that: include: S1: Collects raw signals through piezoelectric sensors and multi-microphone arrays; S2: Dynamically assign acoustic / vibration sensor weights through edge computing nodes; when the acoustic signal-to-noise ratio is greater than 30dB, assign an acoustic feature weight ≥ 0.7; when the pipeline vibration amplitude is greater than 0.5g, assign a vibration feature weight ≥ 0.8; S3: Extract signal depth features through 1DCNN; S4: Calculate the importance scores of acoustic and vibration features based on the SHAP value analysis module of the cloud analysis platform; S5: When the cloud fusion probability is greater than 90% and the acoustic feature SHAP value is greater than the vibration feature SHAP value, an alarm is triggered; S6: When the classification results of the cloud analysis platform and the edge node conflict, the expert review mechanism is activated. The expert review mechanism pushes the alarm screenshot to the operation and maintenance personnel terminal through the WeChat API.

10. The intelligent water leakage detection method based on multimodal sensing and cloud collaboration according to claim 9 is characterized in that: S2 adds a dynamic weight distribution module, which is configured as follows: When the acoustic signal-to-noise ratio is in the range of 30-50dB, the acoustic feature weight is assigned to 0.7±0.1; When the pipeline vibration amplitude is >0.5g and ≤2.0g, the vibration characteristic weight is assigned as 0.8±0.1; The weight distribution strategy has been verified by experiments to reduce the missed detection rate of metal pipelines by ≥25%.

Images