Mechanical equipment state monitoring method and system based on multiple sensors

Abstract

The invention discloses a mechanical equipment state monitoring method and system based on multiple sensors, and the method comprises the five core steps: multi-modal data collection and preprocessing, dynamic feature fusion, adaptive threshold diagnosis, digital twin fault tracing and predictive maintenance decision. All-domain coverage of equipment is realized through a three-layer sensor network architecture, the problems of data synchronization and interference resistance are solved by utilizing a temperature and vibration integrated sensor, deep fusion and anomaly detection of multi-source data are realized in combination with an attention mechanism, a Gaussian mixture model, a three-dimensional convolutional neural network and the like, and finally a precise maintenance strategy is generated through digital twinning and reinforcement learning. The multi-sensor-based mechanical equipment state monitoring system comprises a sensor network layer, an edge computing layer, a cloud platform layer and a man-machine interaction layer, supports federated learning to protect data privacy, improves real-time diagnosis capability through edge-cloud collaboration, and enhances a reality interface to realize intelligent operation and maintenance interaction.

Description

Technical Field

[0001] The present invention relates to the field of intelligent monitoring technology, and in particular to a multi-sensor based mechanical equipment status monitoring method and system. Background Art

[0002] In the field of intelligent monitoring of industrial equipment, existing technologies primarily rely on single sensors or fixed-weight fusion solutions, making them incapable of addressing the multi-dimensional monitoring needs of complex mechanical systems. Traditional approaches suffer from the following common problems: First, single sensors monitor a single dimension, capturing only a single physical quantity such as vibration or temperature. They are unable to identify multi-parameter coupled faults, resulting in missed detection rates exceeding 30%. Second, asynchronous data acquisition by independent sensors causes timestamp errors, and significant dimensional differences. Traditional normalization methods ignore signal correlations, resulting in inefficient feature fusion. Furthermore, fixed-weight models cannot adapt to dynamic load fluctuations. Under high loads, key features are easily overwhelmed by noise, while under low loads, abnormal signal sensitivity is insufficient, leading to false alarm rates increasing to 40% with fluctuations in operating conditions. At the maintenance strategy level, reliance on manual experience or scheduled maintenance can lead to both over-maintenance and the risk of unplanned downtime due to delayed maintenance. Centralized data processing solutions also face the risk of industrial data leakage, while limited edge computing power makes real-time diagnostic capabilities difficult to meet high-precision monitoring requirements.

[0003] The core bottleneck of existing technologies lies in the dual lack of "efficient fusion of multi-source data" and "construction of an intelligent decision-making closed loop"; on the one hand, complex mechanical systems (such as CNC machine tool spindles and wind turbine gearboxes) need to synchronously process multimodal data such as vibration, temperature, and current, but traditional architectures lack effective support for sensor spatiotemporal synchronization, noise robustness fusion, and load adaptive weighting, resulting in incomplete extraction of abnormal features; on the other hand, there is a "data-decision-making" fault between fault diagnosis and maintenance strategies. It is impossible to accurately trace the source of faults through digital twin technology, and it is difficult to dynamically optimize maintenance actions based on real-time working conditions. As a result, equipment health management has long remained in the "monitoring-alarm" stage, and has not formed an intelligent closed loop of "diagnosis-prediction-maintenance". In view of this, we propose a multi-sensor-based mechanical equipment condition monitoring method and system. Summary of the Invention

[0004] The main purpose of the present invention is to provide a method and system for monitoring the condition of mechanical equipment based on multiple sensors, which can effectively solve the problems in the background technology.

[0005] To achieve the above object, the technical solution adopted by the present invention is: A method for monitoring the condition of mechanical equipment based on multiple sensors comprises the following steps: S1. Multimodal data acquisition and preprocessing: Deploy vibration sensors, temperature sensors, dust concentration sensors, and current transformers to collect vibration signals, temperature field distribution, dust concentration, and electrical parameters of the equipment's core components, transmission chain, and auxiliary systems. Perform wavelet denoising and normalization on the raw data. Wavelet denoising addresses the issue of vibration signals being susceptible to cutting fluid impact and motor electromagnetic interference. Normalization unifies the data from different sensors into a standard normal distribution using the following formula: , in, It is the normalized data value, which is used to eliminate the dimension difference of different sensor data and facilitate unified analysis. Collect data for raw sensors, is the original data mean, The above formula converts the raw data of different sensors into standard normal distribution data with a mean of 0 and a standard deviation of 1 through normalization, thereby improving the accuracy of subsequent feature fusion. S2. Dynamic multimodal feature fusion: Construct a dynamic weighted model of attention mechanism, based on the real-time load coefficient L= Adjust the weight of each modal feature , the formula is as follows: , in, For the The feature weight of the modality, It is a real-time load system, calculated by the torque sensor and defined as the ratio of the real-time torque to the rated torque. For the The a priori importance coefficient of the mode, is the total number of sensor modes, is an exponential function used to amplify the weight differences of different modes under different loads. The above formula implements dynamic weighting based on the SoftMax function, which automatically increases the weight of key features such as vibration under high load; S3, Adaptive threshold diagnosis: Use Gaussian mixture model to cluster the fusion features, combined with Mahalanobis distance Dynamic update of alarm thresholds: , in, is the Mahalanobis distance, which is used to measure the degree of deviation between the current feature vector and the normal operating condition distribution. The larger the value, the higher the abnormal probability. is the fusion feature vector at the current moment, is the cluster center vector of the Gaussian mixture model, which is calculated in real time by the GMM algorithm. is the covariance matrix, The inverse matrix is ​​used to dynamically calculate the difference between the feature vector and the normal distribution through the Mahalanobis distance. Compared with the Euclidean distance, this formula takes into account the correlation between features and can more accurately detect multi-dimensional anomalies. S4. Digital Twin Fault Tracing: Build a 3D digital twin model, integrate the fault signature library of more than 10 key components, and map real-time monitoring data to the model; S5. Predictive maintenance decision-making: Based on the long short-term memory network, the remaining life of the equipment is predicted and maintenance recommendations and spare parts lists are output. The maintenance strategy is dynamically optimized through reinforcement learning. The formula is as follows: , in, is the state-action value function, is the current device state vector, including multimodal features such as vibration amplitude, temperature, and dust concentration. Its dimension is consistent with the number of sensors. For maintenance actions, is the learning rate, For instant rewards, is the discount factor, Indicates the next state All possible maintenance actions under To perform an action The next state after For the next state All possible moves Maximum The maintenance action set includes "continue monitoring", "stop for maintenance", and "replace parts". Dynamic optimization strategy, integrating maintenance costs, downtime losses and production plans.

[0006] Preferably, the sensor deployed in step S1 includes a temperature-vibration integrated sensor, which is specifically designed as follows: Integrate MEMS accelerometers and thin-film thermocouples to replace traditional independent vibration sensors and temperature sensors; The IEEE1588 protocol is used to implement hardware synchronization clock, ensuring that the vibration and temperature data timestamp error is less than 1μs; Differential signal transmission reduces electromagnetic interference and increases transmission distance.

[0007] Preferably, in step S1, the vibration signal and the temperature signal are subjected to joint noise reduction processing, specifically including: Construct the signal cross-correlation matrix: , in, Vibration signal With temperature signal The cross-correlation function reflects the time delay between the two signals The degree of linear correlation under the condition, the larger the value, the stronger the correlation. is the numeric expectation operator, is the time domain signal collected by the vibration sensor, is the mean value of the vibration signal, is the mean value of the temperature signal, is the time delay, is the time domain signal collected by the temperature sensor. The above formula is used to calculate the time domain of the two signals. The correlation under Cross-correlation filtering based on wavelet transform: attenuates frequency components with cross-correlation values ​​below the threshold and retains strongly correlated fault feature components.

[0008] Preferably, in step S1, the sensors are deployed according to a three-layer network architecture of "core components - transmission chain - auxiliary system", specifically including: Core component layer: Integrated temperature and vibration sensors are deployed in key locations such as bearing seats and gearboxes. These sensors integrate MEMS accelerometers and thin-film thermocouples, and use the IEEE1588 precision clock protocol to achieve nanosecond-level synchronization of vibration and temperature data. This solves the feature misalignment problem caused by asynchronous acquisition by traditional independent sensors. Transmission chain layer: Deploy strain sensors and speed encoders on the transmission shaft and coupling. The strain sensors are used to monitor deformation, and the speed encoders are used to monitor the torque-speed coupling of the transmission system. Auxiliary system layer: Current transformers and pressure sensors are deployed in the motor housing and lubrication pipelines to collect motor load and lubrication status parameters.

[0009] Preferably, in step 3, an improved DS evidence theory is adopted, which specifically includes: Conflict quantification: The degree of conflict of sensor evidence is calculated using the Bhattacharyya distance, which is: , in, and The fault propositions for sensor 1 and sensor 2 are The basic probability distribution of The larger the value, the more significant the conflict of evidence; BPA construction: Generate BPA based on fuzzy naive Bayes method, through membership function To quantify the uncertainty of low signal-to-noise data, the formula is: , in, is a fuzzy membership function used to quantify the uncertainty of low signal-to-noise ratio data. The higher the value, the higher the probability that the data belongs to normal working conditions. is the signal mean, is the standard deviation, and the above membership function formula is used to process low signal-to-noise ratio data and improve the diagnostic accuracy in a noisy environment; Dynamic weighted combination: The weighted average rule is used to fuse evidence. The formula is: , in, For sensors The credibility weight is dynamically adjusted through Bhattacharyya distance and sensor accuracy.

[0010] Preferably, in step S3, a three-dimensional convolutional neural network is used, which specifically includes: Data tensor construction: Convert the Mel-spectrogram of the vibration signal and the temperature time series data into a three-dimensional tensor; Cross-modal attention mechanism: The spatiotemporal attention weights are calculated through 3D convolution and Softmax. The formula is: , Among them, represents the time ,frequency ,sensor The attention coefficient is a local slice of the three-dimensional tensor. The vibration Mel spectrum and temperature time series data are converted into a three-dimensional tensor. The spatiotemporal features are extracted through the 3D convolution kernel. The attention weight of the SoftMax output is Automatically focuses on the time and space areas related to the fault, outputs quantitative diagnostic results such as "gear crack probability 78%", and supports three-level early warning; Probability output: The fully connected layer outputs the fault probability distribution, such as "gear crack probability 78%", to support graded warnings.

[0011] Preferably, in step S3, a federated learning architecture is adopted, specifically including: Edge node training: Local device training data fusion model, only uploading gradients To the central server; Global model aggregation: The central server updates the global model parameters using the FedAvg algorithm. The formula is: , in, For the The local model parameters of the edge nodes, including the weights and biases of the neural network, is the total data weight, For the The amount of data per node; Knowledge distillation: Transfer global knowledge through loss function, the formula is: , in, is the cross entropy loss, which is used to measure the difference between the predicted value and the true label. for Divergence, used to transfer the soft label knowledge of the global model, =0.7 is the weight coefficient to balance the two.

[0012] A multi-sensor based mechanical equipment condition monitoring system, applicable to the multi-sensor based mechanical equipment condition monitoring method, comprising: a sensor network layer, an edge computing layer, a cloud platform layer, and a human-computer interaction layer; The sensor network layer includes vibration sensors, temperature sensors, dust concentration sensors and current transformers, and is arranged in three layers: "core components - transmission chain - auxiliary systems" to achieve full coverage of the equipment. The edge computing layer uses an embedded processor to perform S1-S3 preprocessing, feature fusion, and threshold diagnosis, supports 5G / Ethernet dual communication links, and realizes real-time data interaction between the edge computing layer and the cloud platform layer; The cloud platform layer deploys a digital twin engine and LSTM prediction model, achieves 3D visualization through WebGL, and integrates a reinforcement learning module to optimize maintenance strategies; The human-computer interaction layer provides PC and mobile interfaces, supporting fault warning, historical data query and maintenance work order generation.

[0013] Furthermore, the sensor network layer adopts a hierarchical calibration mechanism: The core component sensor is automatically calibrated every 24 hours, and the zero drift is corrected through the built-in standard vibration source and constant temperature reference point; The transmission chain sensor uses hardware synchronization triggering and achieves nanosecond synchronization through the IEEE1588 precise clock protocol, ensuring that the time stamp error between the transmission chain vibration and speed signals is less than 50ns. The edge computing layer deploys a lightweight neural network inference engine: TensorRT is used to accelerate the time-frequency analysis of vibration signals, shortening the computation time of short-time Fourier transforms. Dynamic sparsification technology is used to prune neural networks, reducing model parameters while maintaining accuracy and improving edge inference speed. The integrated adaptive filter estimates the temperature sensor drift in real time and dynamically adjusts the compensation coefficient to improve temperature measurement accuracy.

[0014] Furthermore, the cloud platform layer integrates a multi-dimensional health assessment model: The digital twin model is used to calculate the "three-dimensional health" of the equipment: mechanical health, thermal health, and energy efficiency health. Mechanical health is calculated based on the permutation entropy of the vibration signal, with higher values ​​indicating greater signal complexity. Thermal health is based on the temperature field gradient distribution, and energy efficiency health is based on the current-power regression model. The three are weighted using the hierarchical analysis method to generate a comprehensive health index. Using knowledge graph technology to build a fault causal relationship network, linking more than 100 historical fault cases, and supporting multi-level fault tracing from "abnormal vibration → bearing wear → insufficient lubrication"; The human-computer interaction layer provides an augmented reality operation and maintenance interface: Scan the device through the mobile app, overlay the 3D digital twin model in real time, and display the health heat map of each component; Supports voice interaction to generate maintenance work orders: "Create a bearing replacement work order, high priority", the work order information is automatically synchronized to the ERP system, and supports fault warning hierarchical display.

[0015] Compared with the prior art, the present invention has the following beneficial effects: 1. High-precision multimodal data fusion: The three-layer sensor network architecture and integrated temperature and vibration sensor design are adopted to achieve full coverage and synchronous data collection of the equipment's core components, transmission chains and auxiliary systems, solve the feature dislocation problem caused by traditional asynchronous collection, and fully capture the coupling characteristics of multiple physical quantities.

[0016] The dynamic weighted model and joint noise reduction technology automatically adjust the modal weight according to the real-time load, effectively retaining the characteristics of strongly correlated faults, improving the signal recognition capability in complex noise environments, and significantly reducing the missed detection rate and false alarm rate.

[0017] 2. Intelligent diagnosis and accurate fault tracing: The Gaussian mixture model is combined with the Mahalanobis distance to dynamically update the alarm threshold and consider feature correlation, breaking through the limitations of the traditional Euclidean distance and achieving accurate quantitative detection of multi-dimensional anomalies.

[0018] The three-dimensional digital twin model integrates the fault feature library of key components, maps monitoring data to the physical model in real time, supports multi-level fault tracing such as "abnormal vibration → bearing wear → insufficient lubrication", quickly locates faulty components, and shortens manual troubleshooting time.

[0019] 3. Predictive maintenance strategy optimization: The long short-term memory network is combined with reinforcement learning to predict the remaining life of equipment based on historical data and dynamically generate maintenance strategies. It comprehensively considers maintenance costs, downtime losses and safety risks, realizing the transition from "post-maintenance" to "predictive maintenance", and improving maintenance efficiency and equipment reliability.

[0020] The federated learning architecture supports local training on edge nodes and cloud-based model aggregation, improving model generalization capabilities while protecting data privacy and adapting to differences in working conditions across different devices.

[0021] 4. Efficient system architecture and intelligent interaction: The edge computing layer implements data preprocessing and lightweight model inference, and combines 5G / Ethernet dual-link communication to balance low-latency warning and high-bandwidth data transmission requirements to meet real-time monitoring requirements in scenarios such as high-speed rotating machinery.

[0022] The augmented reality operation and maintenance interface and voice interaction function display the device health heat map in real time and support the automatic generation and synchronization of work orders, improving the efficiency of human-machine collaboration and reducing the complexity of operation and maintenance. BRIEF DESCRIPTION OF THE DRAWINGS

[0023] Figure 1 This is the process of the multi-sensor based mechanical equipment condition monitoring method Figure 1 ; Figure 2 This is the process of the multi-sensor based mechanical equipment condition monitoring method Figure 2 ; Figure 3 It is a data preprocessing and feature fusion module diagram of the present invention; Figure 4 It is the interactive diagram of the digital twin and maintenance decision of the present invention; Figure 5 It is a module diagram of the intelligent diagnosis algorithm of the present invention; Figure 6 It is a system-level interaction and communication link diagram of the present invention. DETAILED DESCRIPTION

[0024] In order to make the technical means, creative features, objectives and effects achieved by the present invention easier to understand, the present invention is further described below in conjunction with specific implementation methods.

[0025] Example 1: refer to Figures 1-6 A multi-sensor-based mechanical equipment condition monitoring method and system involves CNC machine tool applications, is suitable for spindle condition monitoring and predictive maintenance of horizontal machining centers, and solves the problem of early identification of bearing wear under high-load conditions.

[0026] 1. Implementation of Condition Monitoring Methods 1. Multimodal data acquisition of S1: Method and steps: A temperature-vibration integrated sensor is deployed on the main shaft bearing seat to synchronously collect vibration signals and temperature data; a strain sensor and a speed encoder are installed on the transmission shaft, and a current transformer and a pressure sensor are deployed in the auxiliary system; the raw data are denoised by wavelet and then normalized to unify the dimensions.

[0027] System implementation: The sensor network layer adopts a three-layer architecture, and the core component layer sensors achieve hardware synchronization through the IEEE1588 protocol to ensure that the vibration and temperature data timestamps are aligned.

[0028] 2. Dynamic multimodal feature fusion of S2: Method and steps: Build an attention mechanism model to dynamically adjust the modal weights of vibration, temperature, etc. according to the real-time load. Automatically increase the weight of vibration features under high load to highlight high-frequency signals related to bearing faults.

[0029] System implementation: The embedded processor in the edge computing layer performs dynamic weighted calculations and transmits the fused feature vectors to the cloud in real time via the 5G link.

[0030] 3. Intelligent diagnosis and decision-making involved in S3-S5: Method and steps: A Gaussian mixture model is combined with Mahalanobis distance to dynamically update the alarm threshold, and a 3D-CNN outputs the "bearing wear probability." LSTM predicts the remaining life, and reinforcement learning generates maintenance strategies.

[0031] System implementation: A digital twin engine is deployed on the cloud platform to map monitoring data to a three-dimensional model in real time. The human-computer interaction layer displays fault probabilities and maintenance recommendations on the PC.

[0032] 2. System Architecture Implementation Sensor network layer: The core component layer integrates temperature and vibration sensors, the transmission chain layer deploys strain and speed sensors, and the auxiliary system layer collects motor and lubrication data to form a full-area monitoring network.

[0033] Edge computing layer: Preprocesses data and performs lightweight model inference, filters out noise and compresses feature data to reduce cloud transmission pressure.

[0034] Cloud platform layer: The digital twin model locates faulty components, and the knowledge graph associates historical cases to achieve multi-level traceability, such as "abnormal vibration → bearing outer ring wear → grease failure".

[0035] Human-computer interaction layer: The mobile app supports AR scanning, overlaying a digital twin model to display a health heat map, and supports voice generation of "bearing replacement" work orders and synchronizing them to the ERP system.

[0036] Example 2: refer to Figures 1-6 A multi-sensor-based mechanical equipment condition monitoring method and system involves wind turbine applications and is suitable for gearbox fault diagnosis in offshore wind farms, solving the problems of multi-source data fusion and remote operation and maintenance under complex working conditions.

[0037] 1. Implementation of Condition Monitoring Methods 1. Multimodal data acquisition of S1: Method and steps: A temperature and vibration integrated sensor is deployed on the high-speed shaft of the gearbox, a torque and speed sensor is installed on the low-speed shaft, and a particle size and pressure sensor is deployed in the lubrication system. The vibration and temperature signals are jointly processed for noise reduction to retain strongly correlated fault characteristics.

[0038] System implementation: The sensor adopts an explosion-proof design and resists marine electromagnetic interference through differential signal transmission. The edge node locally stores the original data and uploads the feature vector.

[0039] 2. Adaptive Diagnosis and Federated Learning: Method and steps: Improve the DS evidence theory to fuse multi-sensor evidence, train the model on edge nodes under the federated learning architecture and upload the gradient, and aggregate in the cloud to generate a global model.

[0040] System implementation: The edge computing layer uses TensorRT to accelerate time-frequency analysis, the cloud platform uses the FedAvg algorithm to optimize the model, and knowledge distillation improves the fault identification capability of new equipment.

[0041] 3. Predictive maintenance decision-making: Method and steps: An LSTM network analyzes historical vibration and torque data to predict the remaining life of the gearbox. A reinforcement learning strategy comprehensively considers downtime costs and safety risks to generate the optimal maintenance opportunity.

[0042] System implementation: The human-computer interaction layer provides a remote monitoring interface, displays the gearbox health curve in real time, and automatically triggers a "gearbox maintenance" warning with an accompanying maintenance plan.

[0043] 2. System Architecture Implementation Sensor network layer: The three-layer architecture covers the core components of the gearbox, the transmission chain, and the lubrication system. The sensors support a wide operating temperature range of -40°C to 85°C, adapting to the harsh offshore environment.

[0044] Edge computing layer: Heterogeneous computing architecture accelerates feature extraction, 5G slicing network ensures low-latency warning transmission, and Ethernet links support batch upload of historical data.

[0045] Cloud platform layer: The digital twin model simulates the gear meshing state, and the three-dimensional visualization shows the crack propagation process; the knowledge graph associates the causal chain of "vibration peak → gear tooth breakage → bearing overload" and provides maintenance process recommendations.

[0046] Human-computer interaction layer: The PC's large screen displays the status of the wind farm equipment cluster, the mobile terminal supports offline work order viewing, and voice interaction enables quick operations such as "generating a gearbox lubrication work order."

[0047] Example 3: refer to Figures 1-6A multi-sensor-based mechanical equipment status monitoring method and system involves petrochemical centrifugal pump applications, is suitable for monitoring centrifugal pump bearings and impellers in high-temperature and high-pressure environments, and meets the safety monitoring needs of flammable and explosive scenarios.

[0048] 1. Implementation of Condition Monitoring Methods 1. Multimodal data acquisition: Method and steps: An explosion-proof temperature-vibration integrated sensor is deployed on the impeller, current and shaft displacement sensors are installed on the motor end, and pressure and temperature sensors are deployed in the lubrication pipeline. After wavelet denoising and normalization of the data, the vibration-temperature coupling characteristics are retained through cross-correlation filtering.

[0049] System implementation: The sensor network layer adopts an intrinsically safe design, complies with the ExiaIICT6 explosion-proof standard, has a differential signal transmission distance of up to 200 meters, and suppresses pipeline vibration and noise interference.

[0050] 2. Dynamic feature fusion and diagnosis: Method and steps: The attention mechanism adjusts the weights of vibration and current signals according to the pump load. The 3D-CNN analyzes the vibration time-frequency graph and current harmonics and outputs the "impeller wear probability".

[0051] System implementation: The edge computing layer calculates the Mahalanobis distance in real time. When the value exceeds the dynamic threshold, it triggers a local sound and light alarm, and uploads abnormal data to the cloud via Ethernet.

[0052] 3. Maintenance strategy optimization: Method and steps: The reinforcement learning model takes safety production regulations into consideration and prioritizes generating "slowdown" or "emergency shutdown" strategies. LSTM predicts the remaining life of bearings and automatically allocates maintenance resources in conjunction with the work order system.

[0053] System implementation: The cloud platform integrates a safety assessment module, and the human-computer interaction layer displays the risk level. It supports "one-click generation of explosion-proof component replacement work orders" and links with the safety management system.

[0054] 2. System Architecture Implementation Sensor network layer: The core component sensor has a built-in calibration module, which automatically corrects vibration and temperature drift every day to ensure long-term monitoring accuracy.

[0055] Edge computing layer: The lightweight neural network reduces 40% of its parameters to adapt to the limited computing power of the pump control system, and the inference time is less than 20ms to meet real-time requirements.

[0056] Cloud platform layer: The digital twin model simulates the impeller cavitation process, and the three-dimensional thermal map shows the bearing temperature distribution; the knowledge graph associates the "abnormal vibration → seal failure → medium leakage" fault chain and provides emergency treatment guidelines.

[0057] Human-computer interaction layer: The explosion-proof mobile app supports on-site scanning, AR displays the health status of the equipment, and voice commands generate "impeller dynamic balancing verification" work orders and transmit them encrypted to the central control room.

[0058] The basic principles, main features, and advantages of the present invention are shown and described above. Those skilled in the art should understand that the present invention is not limited to the above embodiments. The above embodiments and descriptions are merely illustrative of the principles of the present invention. Various changes and modifications may be made to the present invention without departing from the spirit and scope of the present invention. Such changes and modifications are intended to fall within the scope of the present invention. The scope of protection claimed in the present invention is defined by the appended claims and their equivalents.

Claims

1. A method for monitoring the condition of mechanical equipment based on multiple sensors, characterized in that: The steps include: S1. Multimodal data acquisition and preprocessing: Deploy vibration sensors, temperature sensors, dust concentration sensors, and current transformers to collect vibration signals, temperature field distribution, dust concentration, and electrical parameters of core components, transmission chains, and auxiliary systems. Perform wavelet denoising and normalization on the raw data. The formula is as follows: ， in, It is the normalized data value, which is used to eliminate the dimension difference of different sensor data and facilitate unified analysis. Collect data for raw sensors, is the original data mean, The above formula converts the raw data of different sensors into standard normal distribution data with a mean of 0 and a standard deviation of 1 through normalization, thereby improving the accuracy of subsequent feature fusion. S2. Dynamic multimodal feature fusion: Build a dynamic weighted model of the attention mechanism and adjust the weights of each modal feature according to the real-time load factor L. , the formula is as follows: ， in, For the The feature weight of each modality, It is a real-time load system, calculated by the torque sensor and defined as the ratio of the real-time torque to the rated torque. For the The a priori importance coefficient of each mode, is the total number of sensor modes, is an exponential function used to amplify the weight differences of different modes under different loads. The above formula implements dynamic weighting based on the SoftMax function, which automatically increases the weight of key features such as vibration under high load; S3, Adaptive threshold diagnosis: Use Gaussian mixture model to cluster the fusion features, combined with Mahalanobis distance Dynamic update of alarm thresholds: ， in, is the Mahalanobis distance, which is used to measure the degree of deviation between the current feature vector and the normal operating condition distribution. The larger the value, the higher the abnormal probability. is the fusion feature vector at the current moment, is the cluster center vector of the Gaussian mixture model, which is calculated in real time by the GMM algorithm. is the covariance matrix, The inverse matrix is ​​used to dynamically calculate the difference between the feature vector and the normal distribution through the Mahalanobis distance. Compared with the Euclidean distance, this formula takes into account the correlation between features and can more accurately detect multi-dimensional anomalies. S4. Digital Twin Fault Tracing: Build a 3D digital twin model, integrate the fault signature library of more than 10 key components, and map real-time monitoring data to the model; S5. Predictive maintenance decision-making: Based on the long short-term memory network, the remaining life of the equipment is predicted and maintenance recommendations and spare parts lists are output. The maintenance strategy is dynamically optimized through reinforcement learning. The formula is as follows: ， in, is the state-action value function, is the current device state vector, including multimodal features such as vibration amplitude, temperature, and dust concentration. Its dimension is consistent with the number of sensors. For maintenance actions, is the learning rate, For instant rewards, is the discount factor, Indicates the next state All possible maintenance actions under To perform an action The next state after For the next state All possible moves Maximum value.

2. A method for monitoring the condition of mechanical equipment based on multiple sensors according to claim 1, characterized in that: The sensor deployed in step S1 includes a temperature-vibration integrated sensor, which is specifically designed as follows: Integrate MEMS accelerometers and thin-film thermocouples to replace traditional independent vibration sensors and temperature sensors; The IEEE1588 protocol is used to implement hardware synchronization clock, ensuring that the vibration and temperature data timestamp error is less than 1μs; Differential signal transmission reduces electromagnetic interference and increases transmission distance.

3. The method for monitoring the condition of mechanical equipment based on multiple sensors according to claim 1, characterized in that: In step S1, the vibration signal and the temperature signal are subjected to joint noise reduction processing, specifically including: Construct the signal cross-correlation matrix: , in, Vibration signal With temperature signal The cross-correlation function reflects the time delay between the two signals The degree of linear correlation under the condition, the larger the value, the stronger the correlation. is the numeric expectation operator, is the time domain signal collected by the vibration sensor, is the mean value of the vibration signal, is the mean value of the temperature signal, is the time delay, It is the time domain signal collected by the temperature sensor; Cross-correlation filtering based on wavelet transform: attenuates frequency components with cross-correlation values ​​below the threshold and retains strongly correlated fault feature components.

4. The method for monitoring the condition of mechanical equipment based on multiple sensors according to claim 1, characterized in that: In step S1, the sensors are deployed according to a three-layer network architecture of "core components - transmission chain - auxiliary system", specifically including: Core component layer: Deploy temperature and vibration integrated sensors at key locations such as bearing seats and gearboxes; Transmission chain layer: deploy strain sensors and speed encoders on transmission shafts and couplings; Auxiliary system layer: Current transformers and pressure sensors are deployed in the motor housing and lubrication pipelines.

5. The method for monitoring the condition of mechanical equipment based on multiple sensors according to claim 1, characterized in that: In step 3, the improved DS evidence theory is adopted, which specifically includes: Conflict quantification: The degree of conflict of sensor evidence is calculated using the Bhattacharyya distance, which is: ， in, and The fault propositions for sensor 1 and sensor 2 are The basic probability distribution of The larger the value, the more significant the conflict of evidence; BPA construction: Generate BPA based on fuzzy naive Bayes method, through membership function To quantify the uncertainty of low signal-to-noise data, the formula is: ， in, is a fuzzy membership function used to quantify the uncertainty of low signal-to-noise ratio data. The higher the value, the higher the probability that the data belongs to normal working conditions. is the signal mean, is the standard deviation; Dynamic weighted combination: The weighted average rule is used to fuse evidence. The formula is: , in, For sensors The credibility weight is dynamically adjusted through Bhattacharyya distance and sensor accuracy.

6. The method for monitoring the condition of mechanical equipment based on multiple sensors according to claim 1, characterized in that: In step S3, a three-dimensional convolutional neural network is used, which specifically includes: Data tensor construction: Convert the Mel-spectrogram of the vibration signal and the temperature time series data into a three-dimensional tensor; Cross-modal attention mechanism: The spatiotemporal attention weights are calculated through 3D convolution and Softmax. The formula is: ， Among them, represents the time ,frequency ,sensor The attention coefficient is a local slice of the three-dimensional tensor; Probability output: The fully connected layer outputs the fault probability distribution, such as "gear crack probability 78%", to support graded warnings.

7. The method for monitoring the condition of mechanical equipment based on multiple sensors according to claim 1, characterized in that: In step S3, a federated learning architecture is adopted, which specifically includes: Edge node training: Local device training data fusion model, only uploading gradients To the central server; Global model aggregation: The central server updates the global model parameters using the FedAvg algorithm. The formula is: , in, For the The local model parameters of the edge nodes, including the weights and biases of the neural network, is the total data weight, For the The amount of data per node; Knowledge distillation: Transfer global knowledge through loss function, the formula is: ， in, is the cross entropy loss, which is used to measure the difference between the predicted value and the true label. for Divergence, used to transfer the soft label knowledge of the global model, =0.7 is the weight coefficient to balance the two.

8. A multi-sensor based mechanical equipment condition monitoring system, applicable to the multi-sensor based mechanical equipment condition monitoring method according to any one of claims 1 to 7, characterized in that: include: Sensor network layer, edge computing layer, cloud platform layer and human-computer interaction layer; The sensor network layer includes vibration sensors, temperature sensors, dust concentration sensors, and current transformers, arranged in a three-layer layout of "core components - transmission chain - auxiliary systems" to achieve full coverage of the equipment area; The edge computing layer uses an embedded processor to perform S1-S3 preprocessing, feature fusion, and threshold diagnosis, supports 5G / Ethernet dual communication links, and realizes real-time data interaction between the edge computing layer and the cloud platform layer; The cloud platform layer deploys a digital twin engine and LSTM prediction model, achieves 3D visualization through WebGL, and integrates a reinforcement learning module to optimize maintenance strategies; The human-computer interaction layer provides PC and mobile interfaces, supporting fault warning, historical data query and maintenance work order generation.

9. The multi-sensor based mechanical equipment condition monitoring system according to claim 8, characterized in that: The sensor network layer adopts a hierarchical calibration mechanism: The core component sensor is automatically calibrated every 24 hours, and the zero drift is corrected through the built-in standard vibration source and constant temperature reference point; The transmission chain sensor uses hardware synchronization triggering and achieves nanosecond synchronization through the IEEE1588 precise clock protocol, ensuring that the time stamp error between the transmission chain vibration and speed signals is less than 50ns. The edge computing layer deploys a lightweight neural network inference engine: TensorRT is used to accelerate the time-frequency analysis of vibration signals, shortening the computation time of short-time Fourier transforms. Dynamic sparsification technology is used to prune neural networks, reducing model parameters while maintaining accuracy and improving edge inference speed. The integrated adaptive filter estimates the temperature sensor drift in real time and dynamically adjusts the compensation coefficient to improve temperature measurement accuracy.

10. The multi-sensor based mechanical equipment condition monitoring system according to claim 8, characterized in that: The cloud platform layer integrates a multi-dimensional health assessment model: The digital twin model is used to calculate the "three-dimensional health" of the equipment: mechanical health, thermal health, and energy efficiency health. Mechanical health is calculated based on the permutation entropy of vibration signals, thermal health is based on the temperature field gradient distribution, and energy efficiency health is based on the current-power regression model. The three are weighted using the hierarchical analysis method to generate a comprehensive health index. Using knowledge graph technology to build a fault causal relationship network, linking more than 100 historical fault cases, and supporting multi-level fault tracing from "abnormal vibration → bearing wear → insufficient lubrication"; The human-computer interaction layer provides an augmented reality operation and maintenance interface: Scan the device through the mobile app, overlay the 3D digital twin model in real time, and display the health heat map of each component; Supports voice interaction to generate maintenance work orders: "Create a bearing replacement work order, high priority", and the work order information is automatically synchronized to the ERP system, supporting hierarchical display of fault warnings.

Images