Long-term vibration monitoring method and system based on multi-channel data acquisition

Abstract

The invention discloses a long-term vibration monitoring method and system based on multi-channel data acquisition, and the method comprises the following steps: installing a sensor node integrated with a multi-mode energy collection module on a monitored object, and providing continuous power supply for a multi-channel data acquisition and edge processing unit; according to the invention, through a dynamic adaptive sampling mode and edge end extraction key features, the data load in a distributed monitoring scene is significantly reduced; the multi-mode energy collection module is integrated to collect vibration and environmental mechanical energy, and the energy management unit is combined to dynamically distribute stored energy and supply power, so that long-term autonomous operation of the sensor node is realized; the abnormal state is locally inferred and identified through the edge processing unit, and only key feature data instead of full data is uploaded, so that the real-time response capability to the equipment abnormality is improved; multi-dimensional correlation features are extracted by collecting multi-channel sensing data and combining a deep learning model and a time sequence-space feature fusion algorithm, and the health state of the equipment is comprehensively reflected.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of intelligent sensing systems, in particular to a long-term vibration monitoring method and system based on multi-channel data acquisition. BACKGROUND

[0002] Vibration monitoring, as a core technology for industrial equipment health management and infrastructure safety assessment, is widely used in key fields such as wind power equipment, rail transportation, and nuclear power facilities. It can achieve fault early warning, life prediction, and maintenance decision support by capturing vibration characteristics during equipment operation, which is of great significance to reducing downtime losses and ensuring production safety. With the deep integration of Industry 4.0 and Internet of Things technology, traditional periodic offline monitoring mode has been difficult to meet the needs of real-time and continuous monitoring of large and complex systems such as smart factories and cross-regional pipe networks. Long-term online vibration monitoring has become an inevitable trend in the industry. Traditional vibration monitoring technology mostly uses single-channel sensors, which have limited coverage and single data. Although multi-channel data acquisition technology has been applied, there are still some problems in long-term monitoring: 1. High sampling rate leads to a sharp increase in storage and transmission costs, especially in distributed monitoring scenarios; 2. Sensor nodes need to run for a long time, and traditional power supply methods such as batteries are difficult to meet the demand; 3. Massive data needs to be transmitted to the cloud for processing, resulting in high latency and inability to respond to equipment abnormalities in a timely manner; 4. Single modal data (such as only vibration signals) cannot fully reflect the health status of the equipment; Therefore, a long-term vibration monitoring method and system based on multi-channel data acquisition are proposed. SUMMARY

[0003] The present application aims to provide a long-term vibration monitoring method and system based on multi-channel data acquisition to solve one of the problems raised in the background technology.

[0004] First aspect: To solve the above technical problems, the present application adopts a technical solution: a long-term vibration monitoring method based on multi-channel data acquisition, comprising the following steps: Step 1: Install a sensor node integrated with a multi-modal energy harvesting module on the monitoring object to provide continuous power supply for multi-channel data acquisition and edge processing unit; Step 2: Based on the power supply of the sensor node, the edge processing unit analyzes the vibration signal characteristics in real time and selects the sampling mode according to the vibration signal characteristics to collect multi-channel sensor data; Step 3: Input the collected multi-channel sensor data into a deep learning model to extract multi-dimensional correlation features through time-space feature conversion and fusion algorithm, and construct a device state recognition model; Step four, deploy the constructed equipment state recognition model to the edge processing unit, perform local inference on real-time collected data, identify key feature data of abnormal states, and upload to the cloud platform; Step five, based on the uploaded key feature data, combine the historical database to perform equipment life prediction and state evaluation, and generate evaluation results; The equipment state evaluation sets the state of the equipment as , and predicts the state through a deep learning model, and the model output is the probability distribution of the predicted state : ; Wherein, is the scoring function of the equipment state obtained through the model parameters , is the input sensor data, is all possible states, is the probability of the equipment being in state S when the input data X is given.

[0005] Step six, according to the evaluation results, generate maintenance strategies through digital modeling.

[0006] As a further preferred embodiment of the present technical solution: in step one, the multi-modal energy collection module includes at least two of a friction nano power generation unit, a piezoelectric conversion unit and an electromagnetic induction unit, the multi-modal energy collection module collects at least one energy of the vibration signal of the monitored object and the environmental mechanical energy, and performs dynamic allocation of energy storage and power supply through the energy management unit.

[0007] As a further preferred embodiment of the present technical solution: in step two, the sampling mode includes a low-frequency sampling mode and a high-frequency sampling mode; The edge processing unit selects the sampling mode by judging whether the amplitude, frequency or mutation rate of the vibration signal exceeds the preset threshold value, wherein the sampling frequency of the low-frequency sampling mode is 0.1-10Hz, and the sampling frequency of the high-frequency sampling mode is 50-1000Hz; the preset threshold value is dynamically adjusted according to the type of the monitored object; The sampling rate is adjusted according to the change rate of the vibration signal , and the formula is: ; Wherein, is a control coefficient, is the change rate of the signal at time .

[0008] ​The multi-channel sensing data includes at least two of vibration signals, temperature signals, and acoustic emission signals, and the spatiotemporal alignment of the multi-channel data is achieved through a timestamp synchronization mechanism.

[0009] As a further preferred embodiment of this technical solution: In step three, the deep learning model includes a 1D convolutional layer, a temporal feature encoding layer, a spatial feature fusion layer, and an attention mechanism unit; The temporal-spatial feature conversion and fusion algorithm converts temporal data into a two-dimensional feature map through Gram angle field coding, extracts spatial features by combining convolutional neural networks, and captures temporal correlation features through gated recurrent units.

[0010] As a further preferred embodiment of this technical solution: the convolutional neural network formula is as follows: ; in, It is the first The node feature matrix of the layer It is a normalized adjacency matrix. It is the first The weight matrix of the layer, It is an activation function.

[0011] Modeling complex relationships between data in multimodal data, especially handling the correlation between different sensor nodes, can process data from different sensors and effectively extract spatiotemporal features.

[0012] As a further preferred embodiment of this technical solution: In step four, the key feature data of the abnormal state includes time-domain features, frequency-domain features and multimodal correlation features, wherein the time-domain features include peak value, kurtosis and root mean square value, and the frequency-domain features include feature frequency amplitude and frequency band energy ratio; the local inference latency of the edge processing unit does not exceed 500ms, and the amount of key feature data uploaded to the cloud platform does not exceed 5% of the original collected data.

[0013] As a further preferred embodiment of this technical solution: In step five, the equipment life prediction adopts a time-series prediction model based on long short-term memory network or Transformer, and combines fault sample data in historical database with full life cycle performance degradation curve to generate equipment remaining life assessment results.

[0014] The formula for the Long Short-Term Memory network model is as follows: ; in, It is the input sequence data. It is in a hidden state. It is the hidden state from the previous moment.

[0015] For life prediction, the accuracy and stability of the model can be further improved by combining the LSTM model with a residual network connection. The residual network (ResNet) can help avoid the problem of gradient disappearance.

[0016] As a further preferred embodiment of the technical solution: in step six, the digital modeling adopts digital twin technology to construct a virtual mapping model of the monitored object, simulates the effects of different maintenance schemes through virtual simulation, and generates a maintenance strategy in combination with the device operation cost and downtime loss parameters; The maintenance strategy includes maintenance time, component replacement priority and operation process.

[0017] Second aspect: To solve the above technical problems, another technical solution adopted by the present application is: a long-term vibration monitoring system based on multi-channel data acquisition, comprising an energy supply and sensing node module, a dynamic sampling control module, a model training and feature fusion module, an edge inference and data upload module, a cloud analysis and evaluation module, and a digital maintenance decision module; The energy supply and sensing node module is configured to install a sensor node integrated with a multi-modal energy harvesting module on the monitored object, to provide continuous power supply for the multi-channel data acquisition and edge processing unit; The dynamic sampling control module is configured to analyze vibration signal features in real time based on the powered sensor node by the edge processing unit, and select a sampling mode according to the vibration signal features to collect multi-channel sensing data; The model training and feature fusion module is configured to input the collected multi-channel sensing data into a deep learning model, extract multi-dimensional correlation features through time-space feature conversion and fusion algorithm, and construct a device state recognition model; The edge inference and data upload module is configured to deploy the constructed device state recognition model on the edge processing unit, perform local inference on the real-time collected data, identify key feature data of abnormal states, and upload to the cloud platform; The cloud analysis and evaluation module is configured to perform device life prediction and state evaluation based on the uploaded key feature data in combination with the historical database, and generate an evaluation result; The digital maintenance decision module is configured to generate a maintenance strategy through digital modeling according to the evaluation result.

[0018] As a further preferred embodiment of the technical solution: the multi-modal energy harvesting module in the energy supply and sensing node module includes at least two of a friction nano power generation unit, a piezoelectric conversion unit and an electromagnetic induction unit, and is integrated with an energy management unit; The energy management unit is configured to dynamically allocate the collected monitoring object vibration energy and environmental mechanical energy, and perform adaptive regulation and control of energy storage and power supply.

[0019] As a further preferred embodiment of the technical solution, the dynamic sampling control module comprises a threshold judgment unit and a multi-channel synchronous acquisition unit. The threshold judgment unit is configured to compare the amplitude, frequency or mutation rate of the vibration signal with a preset threshold in real time through the edge processing unit, and select a low-frequency sampling mode or a high-frequency sampling mode according to the comparison result. The multi-channel synchronous acquisition unit is configured to perform time-space alignment acquisition of at least two kinds of data in vibration signals, temperature signals and acoustic emission signals through a timestamp synchronization mechanism.

[0020] Advantages of the present application: 1. The present application effectively solves the problem of rapid increase in storage and transmission costs caused by high sampling rate through dynamic adaptive sampling mode and edge extraction of key features, significantly reducing data load in distributed monitoring scenarios. 2. The present application breaks through the power consumption bottleneck of traditional battery power supply by integrating a multi-modal energy collection module to collect vibration and environmental mechanical energy, and combining an energy management unit to dynamically allocate energy storage and power supply, realizing long-term autonomous operation of the sensor node. 3. The present application avoids the delay problem of massive data transmission to the cloud by local inference of abnormal states through the edge processing unit and uploading only key feature data instead of full data, improving the real-time response capability to equipment abnormalities. 4. The present application breaks through the limitation of single modal data by collecting multi-channel sensing data such as vibration, temperature and acoustic emission, and extracting multi-dimensional correlation features by combining deep learning models and time series-space feature fusion algorithms, fully reflecting the equipment health status and solving the problem of insufficient feature extraction. BRIEF DESCRIPTION OF DRAWINGS

[0021] In order to more clearly illustrate the technical solutions in the embodiments of the present application or the prior art, the drawings needed in the embodiment or prior art description will be briefly introduced below. Obviously, the drawings in the following description are only some embodiments of the present application, and those skilled in the art can obtain other drawings according to these drawings without creative labor.

[0022] Figure 1 A flowchart of a long-term vibration monitoring method based on multi-channel data acquisition according to the present application; Figure 2 A functional module diagram of a long-term vibration monitoring system based on multi-channel data acquisition according to the present application. DETAILED DESCRIPTION

[0023] 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.

[0024] Example 1: Figure 1 This is a flowchart illustrating a long-term vibration monitoring method based on multi-channel data acquisition according to an embodiment of the present invention. It should be noted that if substantially the same result is obtained, the method of this application is not necessarily identical. Figure 1 The illustrated process sequence is limited. For example... Figure 1 As shown: A long-term vibration monitoring method based on multi-channel data acquisition includes the following steps: Step 1: Install sensor nodes with integrated multimodal energy harvesting modules on the monitored objects to provide continuous power to the multi-channel data acquisition and edge processing units; Specifically, firstly, based on the structural characteristics of the monitored object (such as the location of the equipment vibration source and the distribution area of ​​mechanical energy), the installation points of the sensor nodes are determined, and areas with vibration amplitude ≥0.01g and where environmental mechanical energy (such as airflow and friction) can be utilized are given priority (such as gearbox bearing seats and pipe flange connections). Then, the sensor nodes were fixed by a combination of 3M VHB tape and threaded fixing to ensure that the positional offset was less than 0.1mm under a 10g impact and that the sensitive surface of the multimodal energy harvesting module was facing the vibration direction. Next, at least two of the following are integrated into the sensor node: triboelectric nanogenerator, piezoelectric conversion unit, and electromagnetic induction unit (e.g., electromagnetic induction unit is preferentially activated for low-frequency vibration, and piezoelectric conversion unit is activated for high-frequency micro-vibration), and each energy harvesting unit is connected to the energy management unit via wires. Next, configure the operating parameters of the energy management unit, set the energy storage threshold (such as stopping energy storage when the lithium battery voltage is ≥3.6V), power supply priority (the edge processing unit is prioritized over the sensor acquisition module), and enable the dynamic allocation algorithm—when the monitored object is in a steady state (vibration amplitude <0.1g), power is supplied only through the triboelectric nanogenerator; when the vibration amplitude is ≥0.1g, the piezoelectric conversion unit is started simultaneously to supplement the power supply. Finally, a continuous power supply test was conducted to simulate the typical operating conditions of the monitored object (such as no-load and full-load operation) and monitor continuously for 72 hours to verify the power supply stability of the sensor node and edge processing unit (voltage fluctuation ≤ ±5%), ensuring that it can still maintain continuous operation under the lowest energy input scenario (such as vibration amplitude 0.05g, environmental mechanical energy <0.1mW).

[0025] Step two, based on the power supply sensor node, the edge processing unit analyzes the vibration signal characteristics in real time, and selects the sampling mode according to the vibration signal characteristics, and collects multi-channel sensing data; Specifically, first, the edge processing unit initializes and loads the preset parameters, calls the corresponding vibration signal characteristic threshold library (including the baseline values of amplitude, frequency and mutation rate) according to the type of monitoring object (such as wind power gearbox, rail transit bearing), and configures the initial working state of the multi-channel sensor (vibration sensor enabled, temperature and acoustic emission sensor standby); Then, the sensor node starts the initial vibration signal collection under the control of the edge processing unit, and acquires baseline data by using 1Hz low-frequency sampling. The edge processing unit analyzes the real-time collected vibration signal in time domain (peak value, kurtosis) and frequency domain (characteristic frequency component), and identifies whether the current working condition is stable; Next, the edge processing unit dynamically switches the sampling mode according to the analysis result, as follows: When the amplitude of the vibration signal is less than or equal to the preset threshold (such as 0.3g), the frequency fluctuation range is less than or equal to 5%, and there is no mutation (mutation rate <10% / s), maintain 0.1-10Hz low-frequency sampling mode, and only collect vibration signal; When any characteristic exceeds the threshold, immediately switch to 50-1000Hz high-frequency sampling mode, and trigger the temperature sensor (sampling rate 10Hz) and acoustic emission sensor (sampling rate 500Hz) to start synchronously; At the same time, start the multi-channel synchronous collection mechanism, add unified time stamp to the vibration, temperature and acoustic emission signals through the high-precision clock (error <1ms) built-in the edge processing unit, ensure the alignment of different types of data in time dimension, and realize the spatial correlation of data in space dimension through the pre-calibration (installation position coordinate input system) of the sensor node; Finally, the collected multi-channel data is preprocessed in real time to remove environmental interference in the temperature signal (such as using sliding average filtering), and the vibration signal is de-trended. The preprocessed data is temporarily stored in the local cache (capacity ≥1GB) of the edge processing unit, waiting for subsequent feature extraction.

[0026] Step three, input the collected multi-channel sensing data into the deep learning model, extract multi-dimensional associated features through time-space feature conversion and fusion algorithm, and construct the device state recognition model; Specifically, first, the multi-channel sensing data collected in step two is pre-processed, including standardizing (scaling the data to the range of [-1, 1]) and removing outliers (filtering noise points outside the normal fluctuation range based on the 3σ criterion) of the vibration, temperature, and acoustic emission signals respectively, and aligning the multi-channel data by timestamp to form a multi-modal data set arranged in time sequence (such as generating a sample containing 1000 time series data every 10 minutes); Then, the time-space feature conversion is started, and the GASF encoding is used to convert the time domain sequence of the vibration signal into a two-dimensional feature map (size 64x64 pixels), while the time series data of the temperature and acoustic emission signals are converted into corresponding dimension feature maps through the same encoding method, realizing the mapping of time series information to spatial features. Next, the 1D convolution layer of the deep learning model is used to extract local features from the original time series data, as follows: 16 3x1 convolution kernels are used to extract high-frequency impact features for the vibration signal. 8 5x1 convolution kernels are used to extract trend features for the temperature and acoustic emission signals, and output multi-channel local feature vectors. After that, the time series feature encoding layer and the spatial feature fusion layer are entered; the GRU in the time series feature encoding layer captures the long-term dependence relationship of the multi-channel data (such as the lag correlation between vibration amplitude and temperature rise); the spatial feature fusion layer uses CNN to extract spatial features from the two-dimensional feature map generated by GASF encoding (through 2 3x3 convolution kernels), and concatenates the extracted time series features and spatial features into a fusion feature matrix. Then, the attention mechanism unit is used to assign weights to the fusion feature matrix; features sensitive to equipment failure (such as vibration kurtosis and acoustic emission pulse number) are given higher attention weights (weight value ≥ 0.7), and environmental interference features (such as small fluctuations in temperature) are given low weights (weight value ≤ 0.3), focusing on key features. Finally, the labeled device state data (such as normal, slight wear, and serious failure) is used as the label, and the Adam optimizer is used to train the model. Through 5-fold cross-validation, the model parameters (such as the number of convolution kernels and the dimension of the GRU hidden layer) are adjusted until the state recognition accuracy of the model on the validation set is ≥ 95%, and the construction of the device state recognition model is completed.

[0027] Step four, deploy the constructed device state recognition model on the edge processing unit, perform local inference on real-time collected data, identify key feature data of abnormal states, and upload to the cloud platform; Specifically, first, the device state recognition model constructed in step three is subjected to lightweight processing; the model is converted into a format compatible with edge devices (such as TensorFlow Lite), redundant neurons are removed through model pruning (core convolutional layers and attention mechanism units are retained), and the model size is compressed to within 50 MB to adapt to the computing power of edge processing units (such as TISitara AM62x processors); Then, the lightweight model is deployed to the local storage module of the edge processing unit and subjected to deployment verification; 3 sets of test data under typical working conditions (normal state, slight fault, and serious fault) are input, the consistency of the model inference output and the labeled results is verified (accuracy ≥ 95%), and the inference time is tested (to ensure that the single-sample inference time is ≤ 100 ms); Next, the edge processing unit receives the multi-channel data (pre-processed vibration, temperature, and acoustic emission signals) collected in real time in step two, aligns the data according to the model input format (such as generating an input vector containing 3 channels of features every 100 ms), and triggers the local inference process; After that, the model performs state recognition on the input data, as follows: If the output result is "normal state", only the state label is recorded, and no feature uploading is performed; If the recognition result is "abnormal state" (such as bearing peeling or gear crack), the corresponding key feature data is automatically extracted, including time-domain features (peak value, kurtosis, and root mean square value), frequency-domain features (characteristic frequency amplitude, frequency band energy proportion), and multi-modal correlation features (such as the time difference between sudden increase in vibration amplitude and sudden rise in temperature); After that, the extracted key feature data is compressed and encoded; sparse matrix representation is used to retain feature values and corresponding timestamps, and the data volume is compressed to within 5% of the original collected data (single abnormal record ≤ 5 KB); Finally, the edge processing unit uploads the compressed key feature data and abnormal state label to the cloud platform through the 5G / NB-IoT communication module, and stores a backup locally (retaining the last 72 hours of abnormal data), ensuring that the cloud can receive the data again if the upload fails.

[0028] Step five, based on the uploaded key feature data, the device life prediction and state evaluation are performed in combination with the historical database, and the evaluation results are generated; Specifically, first, the cloud platform receives the key feature data uploaded by the edge processing unit, checks the data integrity (such as checking whether the dimension of the feature vector matches the pre-set format and whether the timestamps are continuous), and restores the data using the AES-256 decryption algorithm to ensure that the data has not been tampered with during transmission; Then, the decrypted key feature data is associated and matched with the historical database; the historical database contains the full life cycle data of the same type of equipment (such as vibration features in the past 5 years, fault records, maintenance records), and through data alignment algorithm (based on equipment model, running time), the real-time features (such as current vibration kurtosis value, temperature trend) are compared with the feature curve under the same working condition in history, and the abnormal points deviating from the normal trend are marked; Next, the equipment life prediction process is started; the time series prediction model based on long short-term memory network (LSTM) or Transformer is called, and the sorted key feature time series data (such as vibration root mean square value in the past 24 hours, temperature mean value sequence) is input into the model, the model combines the performance degradation curve in the historical database (such as the change rule of vibration amplitude from normal to failure of bearing), outputs the equipment remaining life prediction value (unit: day), and calculates the prediction error (ensure that the error < 8%); After that, based on the similarity matching of real-time key features and historical fault samples (such as calculating the matching degree of the current feature vector and the "gear tooth breaking" sample by cosine similarity), the type (such as bearing peeling, gear wear) and severity level (mild / moderate / severe) of the current abnormal state are determined, and the reliability of the evaluation result is cross-verified combined with temperature, acoustic emission and other multi-modal features (such as increasing the severity level judgment weight when vibration anomaly is accompanied by temperature rise); Finally, the life prediction value, state evaluation level and key abnormal features (such as "vibration kurtosis value reaches 8.5, which is 3 times higher than the historical normal range") are integrated to generate a structured evaluation report, which includes the current health score of the equipment (0-100 points, above 80 points is normal), the remaining life prediction interval (such as "120 ± 10 days"), the main abnormal feature description and the corresponding historical fault case reference, which provides data support for subsequent maintenance strategy generation.

[0029] Step six, according to the evaluation result, generate maintenance strategy through digital modeling; Specifically, first, the cloud platform receives the evaluation results generated in step five (including equipment health score, remaining life prediction interval, abnormal feature description), and imports the evaluation results into the digital twin system to initialize the virtual mapping model of the monitoring object - the model includes the three-dimensional structure parameters of the equipment (such as size, material, connection relationship), the historical running data interface (associated with the maintenance records in the past 3 years) and the real-time state parameters (such as current vibration amplitude, temperature value), ensuring that the mapping error between the virtual model and the physical equipment is ≤2%; Then, at least three candidate maintenance schemes are generated according to the abnormal type in the evaluation result (such as bearing wear), including "immediate shutdown and replacement", "planned shutdown maintenance", "running with fault and strengthening monitoring", and the corresponding operation parameters (such as replacement time, number of maintenance personnel, spare part model) are configured for each scheme; Then, the implementation effect of each scheme is simulated by virtual simulation. In the simulation process, the equipment operation cost parameters (such as production loss of 1 hour of downtime, spare part procurement cost) and performance recovery indicators (such as vibration amplitude reduction ratio after maintenance, extension of remaining life for days) are introduced, and the comprehensive benefit value of each scheme is calculated (benefit value = (life extension benefit - maintenance cost - downtime loss) / implementation period); Then, the optimal scheme is selected in combination with the remaining life prediction value in the evaluation result; if the remaining life is ≤ 30 days and the abnormality level is “serious”, the scheme of “immediate shutdown and replacement” is preferentially selected; if the remaining life is > 90 days and the abnormality level is “slight”, the scheme of “running with fault and strengthening monitoring” is preferentially selected, and the final scheme is determined by balancing the cost and risk through a multi-objective optimization algorithm (such as genetic algorithm); Finally, a structured maintenance strategy is generated; the strategy includes specific maintenance time (accurate to hours, such as “2025-10-15 8:00-12:00”), component replacement priority (such as “bearing > seal ring > lubrication system”), operation process (including safety specifications, tool list, acceptance criteria), and the strategy is synchronized to the equipment management system, and the operation process is preplayed through the digital twin model to verify the feasibility of the scheme (such as whether the maintenance space is sufficient and whether the spare parts compatibility is matched).

[0030] In this embodiment, specifically: in step one, the multi-modal energy collection module includes at least two of a friction nanometer power generation unit, a piezoelectric conversion unit and an electromagnetic induction unit, the multi-modal energy collection module collects at least one of the vibration signal of the monitoring object and the environmental mechanical energy, and performs dynamic allocation of energy storage and power supply through the energy management unit; The friction nanometer power generation unit works based on the triboelectric effect and electrostatic induction principle, and its core is two layers of different dielectric materials (such as polytetrafluoroethylene and aluminum foil) in contact with each other. When the monitoring object vibrates or the environment has mechanical friction, the two material layers periodically contact and separate, surface charge transfer is formed, an alternating electric field is generated, and then electric energy is output. This unit has high energy conversion efficiency for low-frequency, small-amplitude vibration (such as pipeline slight pulsation, micro-vibration during stable operation of equipment) of 0.5-10Hz, and the output voltage of a single unit can reach 5-30V, which is suitable for supplementing the power supply demand in low-power consumption scenarios; The piezoelectric conversion unit works based on the piezoelectric effect (such as using piezoelectric ceramic PZT or piezoelectric film PVDF). When the vibration mechanical energy of the monitoring object acts on the piezoelectric material, polarization phenomenon occurs inside the material and surface charge is formed, and direct current is led out through the electrode. It is sensitive to high-frequency vibration (such as gear meshing, high-frequency vibration generated by bearing rolling) of 10-50Hz, especially suitable for energy collection under small deformation, with small output current (μA level) but stable voltage, which can quickly respond to the change of instantaneous vibration energy; The electromagnetic induction unit is composed of a permanent magnet and a coil. When the monitoring object vibrates to drive the permanent magnet to move relative to the coil, the coil cuts the magnetic induction lines to generate an induced electromotive force. This unit has strong adaptability to medium-amplitude vibrations of 5-30 Hz (such as motor shell vibrations and periodic shaking of large equipment during operation), and the energy conversion efficiency increases with the increase of vibration amplitude. The output power is relatively high (mW level), which can be used as one of the stable main energy supply units. Specifically, the multi-modal energy harvesting module covers the vibration energy of 0.5-50 Hz generated by the monitoring object through the combination of the above three units (such as simultaneously integrating the friction nanogenerator unit and the electromagnetic induction unit), and simultaneously captures the mechanical energy generated by air flow disturbance and component friction in the environment (such as wind-induced vibration of outdoor equipment and friction mechanical energy of transmission components). The energy management unit realizes dynamic regulation through the following process. The specific process is as follows: First, the output power of each power generation unit is rectified (alternating current is converted into direct current) and stabilized (stabilized to 3.3V or 5V standard voltage), and then connected to an energy storage element (such as a combination of a 1000mAh lithium battery and a 1F super capacitor). When the total output power of each unit is greater than or equal to the total power consumption of the load (the sum of the real-time power consumption of the sensor node and the edge processing unit), the excess power is stored in the energy storage element. When the output power is less than the load power consumption, the energy storage element releases power to supplement the power supply. At the same time, the unit is equipped with an intelligent monitoring chip, which collects the output power of each power generation unit and the remaining power of the energy storage element in real time. The edge processing unit (core control component) is preferentially allocated power supply to ensure its stable operation when the energy fluctuates. The sensor node dynamically adjusts the sampling frequency according to the energy adequacy (reduces the sampling frequency to reduce power consumption when the energy is tight), and finally realizes the closed-loop dynamic balance of "collection-energy storage-power supply", ensuring long-term autonomous operation of the system.

[0031] In this embodiment, specifically, in step two, the sampling mode includes a low-frequency sampling mode and a high-frequency sampling mode. The edge processing unit selects the sampling mode by judging whether the amplitude, frequency or mutation rate of the vibration signal exceeds the preset threshold. The sampling frequency of the low-frequency sampling mode is 0.1-10 Hz, and the sampling frequency of the high-frequency sampling mode is 50-1000 Hz. The preset threshold is dynamically adjusted according to the type of the monitoring object. Specifically, the selection of the sampling mode is based on the real-time comparison of the vibration signal characteristics and the preset threshold, which realizes the precision and energy saving of data acquisition. Specifically as follows: The low-frequency sampling mode (0.1-10Hz) is suitable for monitoring scenarios where the object is in a steady state operation (such as the smooth operation of the device under no load or rated load), at which time the vibration signal fluctuates little and the characteristics are stable, and the key state information can be captured by low-frequency sampling, effectively reducing the generation of invalid data; for example, for the steady state operation stage of a wind power gearbox, a sampling frequency of 1Hz can be used to continuously monitor the basic vibration trend, meeting the low-power consumption demand of long-term monitoring; The high-frequency sampling mode (50-1000Hz) is aimed at scenarios where the vibration signal appears abnormal fluctuations (such as early wear, impact or resonance of device components), and fine features (such as high-frequency impact pulses generated by bearing spalling, abnormal shift of gear meshing frequency) need to be captured by high-frequency sampling; for example, when a rail transit bearing has a slight crack, the vibration signal in the 200-500Hz frequency band will have a characteristic frequency amplitude surge, at which time switching to 500Hz high-frequency sampling can accurately capture the abnormal characteristics; The edge processing unit determines whether to switch modes through the following three dimensions: Amplitude judgment: when the peak value or root mean square value of the vibration signal exceeds the preset threshold value (such as the amplitude threshold value of the wind power gearbox is set to 0.5g, and the rail transit bearing is set to 0.8g), high-frequency sampling is triggered; Frequency judgment: when the dominant frequency of the vibration signal exceeds the normal working frequency range (such as the normal running frequency of the motor is 50Hz, and if more than 20% of the high-frequency component above 150Hz is detected), it is determined to be abnormal and the mode is switched; Mutation rate judgment: calculate the change rate of vibration amplitude per unit time (such as 1 second), when the change rate exceeds the preset value (such as 50% / s for wind power equipment and 30% / s for precision machine tools), it is determined to be a sudden abnormality and high-frequency sampling is immediately started; Among them, the dynamic adjustment of the preset threshold value is strongly related to the type of the monitored object, and the specific is as follows: For high-vibration-intensity equipment (such as a crusher), the threshold value is generally high (amplitude threshold value 1.0g, frequency threshold value 200Hz); For precision equipment (such as the main shaft of a numerical control machine tool), the threshold value is low (amplitude threshold value 0.3g, frequency threshold value 100Hz), and the threshold value can be remotely updated through a cloud platform to adapt to changes in the equipment running stage (such as loosening the threshold value during the running-in period and tightening the threshold value during the aging period); In addition, the time stamp synchronization mechanism is used to ensure the continuity of multi-channel data during mode switching, avoiding data discontinuity caused by sudden changes in sampling frequency, and ultimately achieving a dynamic balance between "steady low-power sampling" and "accurate capture of abnormalities".

[0032] The multi-channel sensing data includes at least two of vibration signals, temperature signals, and acoustic emission signals, and the time and space alignment of multi-channel data is performed through a time stamp synchronization mechanism; Among them, the vibration signal mainly reflects the dynamic response of the equipment structure (such as the increase of vibration amplitude caused by bearing wear and the characteristic frequency generated by abnormal gear meshing), which is the core basis for judging mechanical failure; The temperature signal is used to capture the thermal state change of the equipment during operation (such as the temperature rise of the winding caused by motor overload and the temperature rise caused by lubrication failure), which can assist in verifying whether the vibration anomaly is related to overheating; The acoustic emission signal realizes the detection of early micro-failure (such as the acoustic emission pulse when the metal material plastically deforms) by collecting stress waves released by internal defects of the material (such as crack propagation and fatigue damage); Selecting at least two signals (such as vibration + temperature, vibration + acoustic emission) can form a cross-verification of "mechanical dynamics-thermal state" and "macro-vibration-micro-defect", avoiding the limitations of a single signal (such as distinguishing mechanical failure from environmental interference only by vibration signal, and excluding the misjudgment of environmental temperature fluctuations by combining temperature signal); Specifically, the time stamp is generated by the high-precision real-time clock (RTC, error ≤1ms) built in the edge processing unit, and each sensor module will obtain the time stamp of this clock (accurate to millisecond level) in real time when collecting data and attach it to the corresponding data frame; For example, the vibration amplitude collected by the vibration sensor at t=10:00:00.000 and the temperature value collected by the temperature sensor at the same time are marked with "10:00:00.000" timestamp, ensuring the strict alignment of different types of data in time dimension; When installing the sensor node, record the installation position of each sensor (accurate to centimeter level) through the pre-set spatial coordinate system (such as establishing a three-dimensional coordinate system with the rotating shaft of the equipment as the origin), for example, the vibration sensor is installed on the bearing seat (coordinates X1, Y1, Z1), and the temperature sensor is installed on the adjacent shell (coordinates X2, Y2, Z2), and the position information is pre-stored in the edge processing unit; During data processing, the system automatically associates the sensor data at different positions with the corresponding monitoring area (such as bearing seat vibration and shell temperature, both indicating the running state of the bearing assembly) according to the sensor coordinates and the equipment structure model, realizing the correlation and alignment in spatial dimension; Through space-time alignment, multi-channel data can form a three-dimensional correlation matrix of "time-space-physical quantity" (such as vibration amplitude and temperature value at t time, coordinates (X, Y, Z)), which provides accurate multi-modal input for subsequent time-space feature fusion algorithm, ensuring that the correlation features (such as the time difference between sudden increase of vibration and sudden rise of temperature at a certain time, and the amplitude correlation between acoustic emission signal and vibration signal in the same area) truly reflect the equipment state, and improve the accuracy of the state recognition model.

[0033] In this embodiment, specifically: in step three, the deep learning model includes a 1D convolution layer, a time series feature encoding layer, a spatial feature fusion layer, and an attention mechanism unit; Among them, the 1D convolutional layer is used as the input layer of the model to extract the local features of multi-channel time series data. According to the time series characteristics of vibration, temperature and acoustic emission signals, 1D convolution kernels with different parameters are configured: 16 3x1 convolution kernels (step 1) are used for vibration signals (containing high-frequency impact features) to focus on the local fluctuations of adjacent 3 time points; 8 5x1 convolution kernels (step 1) are used for temperature and acoustic emission signals (containing trend features) to capture the change trend of a longer time window; through convolution operation, multi-channel feature maps are generated (each channel corresponds to a local feature), realizing the dimension reduction and preliminary feature extraction of the original time series data, and the output dimension is (time step, channel number); The time series feature encoding layer is constructed based on the gated recurrent unit (GRU) to capture the long-term time series association of multi-channel data; the GRU unit effectively handles the dependence relationship across time steps such as "impact pulse and subsequent decay" and "slow temperature rise and vibration mutation" in vibration signals through update gate (controls the proportion of historical information retention) and reset gate (controls the input weight of new information); 2 layers of GRU are set in the layer (64 hidden units per layer), the input is the local feature output by the 1D convolutional layer, and the output is a feature vector (dimension 64) containing time series dynamic rules, focusing on mining time correlation features such as "temperature starts to rise 10ms after vibration amplitude suddenly increases"; The spatial feature fusion layer is constructed using a 2D convolutional neural network (CNN) to extract spatial correlation features of multi-modal data; the input is a two-dimensional feature map converted by Gram angle field encoding (see the algorithm below), and the convolution operation is performed through 2 3x3 convolution kernels (step 2), combined with a max pooling layer (2x2) to compress the spatial dimension, focusing on capturing the correlation of different modal features in the spatial distribution (such as the spatial overlap of "high-frequency region" in the vibration feature map and "high-temperature region" in the temperature feature map), and the output dimension is (feature map size, spatial feature number); The attention mechanism unit is used to weight the time series features and spatial features, focusing on key features sensitive to device status; the attention weights of the time series feature vector and the spatial feature vector are calculated through a fully connected layer (the sum of the weights is 1), and high weights (≥0.6) are given to strong correlation features such as "bearing fault corresponding to vibration kurtosis value" and "acoustic emission spatial aggregation features generated by crack propagation", and low weights (≤0.2) are given to weak correlation features such as "environmental temperature fluctuations" and "low-frequency background vibration", and finally a multi-dimensional correlation feature vector (dimension 128) is output.

[0034] The time series-space feature conversion and fusion algorithm converts time series data into a two-dimensional feature map through Gram angle field encoding, and extracts spatial features combined with convolutional neural network, and captures time correlation features through gated recurrent unit; Specifically, first, Gram angle field (GASF / GADF) encoding is performed, which is as follows: The one-dimensional time series data is converted into a two-dimensional feature map, and the amplitude and phase information of the time series data is preserved. Taking the vibration signal as an example, first, the time series data is normalized to [-1, 1], and the data value is mapped to an angle (θ = arcsin(x)) through the inverse sine function, and then the angle sum (GASF: cos(θ i +θ j ) or the angle difference (GADF: sin(θ i -θ j )) of any two time points is calculated, to generate a 64x64 two-dimensional matrix (both horizontal and vertical axes are time steps, and the matrix value is the angle operation result). This process converts the "time sequence relationship" into a "spatial position relationship", so that the periodic and trend characteristics (such as an impact pulse occurring every 100 ms) hidden in the time series are clearly presented as spatial textures in the two-dimensional map, which facilitates CNN extraction. Then, spatial feature extraction is performed, specifically as follows: The Gram angle field feature maps (a total of 3) of the vibration, temperature, and acoustic emission signals are spliced as an input of 3x64x64, and through 2D convolution operation of the spatial feature fusion layer, cross-modal spatial correlation features such as "overlap of vibration high-frequency texture area and high-temperature area" and "matching of spatial distribution of acoustic emission pulse and impact position of vibration" are extracted to make up for the lack of spatial information of a single modality. Finally, time series-spatial feature fusion is performed, specifically as follows: The 64-dimensional time series vector output by the time series feature encoding layer is spliced with the 64-dimensional spatial vector output by the spatial feature fusion layer into a 128-dimensional feature, which is input into the attention mechanism unit for weighted fusion. The final output multi-dimensional correlation feature includes not only the time series law such as "amplitude change of the vibration signal at t1-t5", but also the spatial correlation such as "spatial overlap of the vibration high-frequency area and the acoustic emission pulse area at t3", which provides comprehensive feature input for the device state recognition model and significantly improves the recognition accuracy of early faults and compound faults.

[0035] In this embodiment, specifically, in step four, the key feature data of the abnormal state includes time domain features, frequency domain features, and multi-modal correlation features, wherein the time domain features include peak value, kurtosis, and root mean square value, and the frequency domain features include characteristic frequency amplitude and frequency band energy proportion. Specifically, the time domain features reflect the statistical characteristics of the signal in the time dimension, focusing on the instantaneous and overall energy changes, and specifically include: Peak: refers to the maximum instantaneous amplitude of the vibration signal under abnormal conditions (unit: g or mm / s), which directly reflects the impact intensity of the equipment; for example, when metal spalling occurs between the bearing roller and the inner ring, a sudden impact will occur, causing the vibration peak to increase by 3-5 times compared to the normal state. Through the peak value, the "impact type anomaly" (such as part collision, foreign matter intrusion) can be quickly located. Kurtosis: describes the steepness of the probability density distribution of the signal (dimensionless), and the calculation formula is the ratio of the fourth-order central moment to the square of the second-order central moment; under normal conditions, the vibration signal is close to normal distribution, and the kurtosis is about 3; when early micro-cracks or wear particles occur in the equipment, a small amount of pulse components will be mixed into the signal, causing the kurtosis to increase significantly (usually > 5), which is a sensitive indicator for identifying "early micro-faults" (such as initial pitting on the gear tooth surface). Root Mean Square (RMS): calculates the average of the squares of all sampling points of the signal and then takes the square root (unit: g or mm / s), which reflects the overall energy level of the signal; when the equipment wear intensifies or the gap increases, the vibration energy will continue to accumulate, and the root mean square value will monotonously increase with the development of the fault (for example, during the wear process of the motor bearing, the root mean square value gradually increases from 0.1g to more than 0.5g), which can be used to evaluate the severity of the anomaly; Frequency domain features are analyzed through the frequency dimension to locate the characteristic frequency corresponding to the anomaly, focus on fault source identification, and specifically include: Characteristic frequency amplitude: refers to the vibration amplitude at the characteristic frequency (such as bearing inner ring fault frequency, gear meshing frequency) corresponding to a specific fault under abnormal conditions (unit: g or mm / s); for example, the characteristic frequency of the bearing inner ring fault can be calculated by the formula (related to the speed and bearing parameters), and when the amplitude at this frequency increases by more than 20dB compared to the normal state, it can be determined that the inner ring is abnormal; the characteristic frequency amplitude is directly related to the fault type, and is the core basis for "locating specific fault components"; Frequency band energy ratio: the ratio of the vibration energy of a certain frequency band (such as 100-500Hz) to the total energy of the signal (dimensionless); when the equipment is running normally, the energy distribution is relatively stable; when anomalies such as gear tooth breakage and bearing cage failure occur, the energy of a specific frequency band will increase significantly (such as gear failure often causes the energy ratio of the meshing frequency and its harmonic frequency band to increase from 10% to more than 30%), and through this indicator, the abnormal shift of the frequency distribution can be identified.

[0036] The local inference delay of the edge processing unit does not exceed 500ms, and the amount of key feature data uploaded to the cloud platform does not exceed 5% of the original collected data amount; Specifically, the edge processing unit (such as using TISitara AM62x processor with a main frequency of 1.4GHz) controls the delay through hardware adaptation and model optimization, as follows: Firstly, the device state recognition model constructed in step three is pruned (redundant neurons are removed, and core 1D convolution layers, GRUs, and attention mechanism units are retained), and is converted into TensorFlowLite format, the model volume is compressed to within 50MB, and the calculation load during inference is reduced; Then, the local inference only performs feature matching and state determination on the real-time collected multi-channel data (pre-processed vibration, temperature, and acoustic emission signals), skips the complex iterative calculation in the model training stage, and the single-sample inference process (data input → feature extraction → state output) takes 100-300 ms; Finally, the edge processing unit allocates independent calculation cores (such as 2 CPU cores) for the inference task, avoids resource occupation by data acquisition and communication tasks, and ensures that the inference delay is still stably controlled within 500 ms under extreme working conditions (such as high-frequency sampling), meeting the real-time response requirements of device abnormalities; The key feature data volume is reduced through the process of "feature extraction + compression encoding" as follows: First, taking the high-frequency sampling mode (1000Hz) as an example, a single-channel vibration signal generates 1000 sampling points per second (16-bit precision, i.e. 2 bytes per point), and the original data volume of 3 channels (vibration + temperature + acoustic emission) per second is 1000x2x3=6000 bytes (about 5.86KB), and the original data volume per hour is about 21MB; Then, the edge processing unit only extracts core features (such as peak value, kurtosis in time domain, characteristic frequency amplitude in frequency domain, and correlation time difference of multi-modal) under abnormal state, and the feature vector generated for each abnormal event contains 20-30 key parameters (8 bytes per parameter), and the single-event feature data volume is about 0.2KB; Finally, the sparse matrix representation method is used to record the feature values and corresponding time stamps (only the features at abnormal moments are stored, and the normal state is not transmitted), and the Huffman coding is used to further compress the redundant information, so that the ratio of the uploaded key feature data volume to the original data volume collected at the same period is stably controlled at 3%-5%, significantly reducing the cloud transmission and storage pressure; Through edge "light inference + feature extraction", data dimensionality reduction is realized while ensuring real-time, providing support for efficient operation of long-term vibration monitoring.

[0037] In this embodiment, specifically: in step five, the device life prediction adopts a time series prediction model based on long short-term memory network or Transformer, combines the fault sample data in the historical database and the performance degradation curve in the whole life cycle, and generates a device remaining life evaluation result; Among them, the long short-term memory network (LSTM) model effectively captures the nonlinear variation of key features such as vibration signals and temperature over time (such as the slow upward trend of the root mean square value of vibration during bearing wear) through the synergistic effect of input gate, forget gate and output gate; The model sets 3 layers of hidden layers (128 neurons per layer), the input is the time series sequence of key features in the past 24 hours (such as 10 minutes per data point, a total of 144 time steps), including vibration peak value, kurtosis, temperature mean value, etc.; The output is the remaining life prediction value (unit: day); Its core advantage is to handle the "long-term dependence" in long time series data (such as the correlation between the mutation of vibration features after 1000 hours of device operation and early minor wear), avoiding the gradient disappearance problem of traditional time series models; The Transformer model can simultaneously focus on different time steps of key features (such as a sudden increase in vibration kurtosis at a certain time and a slow increase in temperature three months ago) through self-attention mechanism, which is more suitable for capturing performance degradation correlations across long time scales; The model includes 6 encoder layers (each layer contains a multi-head self-attention module and a feedforward network), the input is the time series embedding vector of key features (combined with timestamp information), which focuses on features that have a significant impact on life (such as the cumulative change in acoustic emission pulse number) through attention weight distribution, and the output is the probability distribution of remaining life (such as "the probability of remaining life 120 days is 75%"), which improves the robustness of life prediction under complex working conditions; The historical database stores the full life cycle data of similar devices, including: Fault sample data: records the complete process data of past devices from normal operation to failure (such as vibration, temperature features and corresponding actual life of 50 bearings of the same type), labels fault types (such as roller wear, inner ring crack) and failure time; Normal operation data: collect key feature baseline values of the device under different working conditions (such as no load, full load) (such as the fluctuation range of the root mean square value of vibration during normal operation), as the reference baseline of the model; During model training, the historical fault sample data is used to optimize parameters (such as the forget gate weight of LSTM), so that the model learns the mapping rule of "feature change trend → remaining life"; During prediction, real-time key features are compared with historical data under the same working conditions to calibrate prediction bias (such as when the similarity between real-time vibration features and a certain fault sample in history is more than 80%, the remaining life curve of the sample is referred to adjust the prediction value); The performance degradation curve is a curve of the key feature of the same type of equipment changing with the running time (such as "the rising curve of the kurtosis value of vibration with the number of running days" and "the change trend of the mean temperature with the life decay"), which is generated by fitting the degradation data of multiple devices in the historical database (using the least square method to fit an S-shaped curve or an exponential curve); in prediction, the model maps the real-time key feature (such as the current kurtosis value of vibration) to the corresponding position on the degradation curve, and calculates the remaining life by combining the slope of the curve (reflecting the degradation speed) - for example, if the current feature is located in the "accelerated degradation stage" of the curve (slope> 0.05 / day), the remaining life is shorter; if it is in the "slow degradation stage" (slope<0.01 / day), the remaining life is longer. The equipment remaining life evaluation result includes: specific remaining life value, key factors affecting life and degradation trend curve.

[0038] In this embodiment, specifically: in step six, digital modeling uses digital twinning technology to construct a virtual mapping model of the monitored object, simulates the effects of different maintenance schemes through virtual simulation, and generates a maintenance strategy combining device operating cost and downtime loss parameters; Specifically, the virtual mapping model constructed by digital modeling through digital twinning technology takes the physical entity of the monitored object as the prototype, and realizes accurate mapping of "physical-virtual" through multi-dimensional data fusion; wherein, based on three-dimensional scanning technology, the device structure parameters are obtained, a three-dimensional geometric model (error≤0.5mm) with a 1:1 scale of the physical device is constructed, and the spatial position and form of the key components are accurately restored; the material properties (such as elastic modulus, friction coefficient) and mechanical properties (such as vibration transfer function, thermal conductivity coefficient) are integrated, and the physical model is constructed through finite element analysis algorithm, which can reproduce the dynamic response of the device during operation; real-time key feature data (vibration, temperature, acoustic emission) and historical operation data (such as load change, maintenance record) are accessed, and the operation behavior law of the device is modeled through algorithm, so that the virtual model can be dynamically updated with the state change of the physical device; Based on the virtual mapping model, at least three candidate maintenance schemes are generated for the abnormal state in the evaluation result, including "immediate shutdown and replacement of components", "planned shutdown maintenance (such as after 30 days)", "running with faults and strengthening monitoring", etc., each scheme clearly specifies the operation object, time length and required resources; the implementation effects of each scheme are simulated through virtual simulation, and the performance recovery indicators (such as the recovery degree of vibration and temperature after maintenance), the life impact indicators (such as the extension effect of remaining life), and the risk probability indicators (such as the probability of failure risk) are specifically evaluated; On this basis, the equipment operation cost and downtime loss parameters are introduced to quantitatively evaluate the simulation results, the operation cost includes spare parts procurement cost, maintenance personnel labor cost, energy consumption cost, etc., and the downtime loss is calculated according to the production capacity to calculate the downtime production value loss and the secondary loss of fault expansion; Taking "maximizing comprehensive benefits" (comprehensive benefits = life extension benefits - maintenance cost - downtime loss) as the target, the candidate schemes are sorted by genetic algorithm, and the final output maintenance strategy clearly includes maintenance time (accurate to hours), component replacement priority (such as "bearing No. 1 > sealing ring > lubrication system"), operation process (including safety specifications), and the operation process is preplayed through the virtual model to verify the feasibility of maintenance space, spare parts compatibility, etc., to ensure that the generated strategy not only guarantees the performance recovery of the equipment, but also minimizes the comprehensive cost, realizes "precise maintenance, cost reduction and efficiency improvement"; By digital modeling, the "equipment state evaluation results", "virtual simulation data" and "cost parameters" are integrated to generate a maintenance strategy that not only guarantees the performance recovery of the equipment, but also minimizes the comprehensive cost, achieving the goal of "precise maintenance, cost reduction and efficiency improvement".

[0039] The maintenance strategy includes maintenance time, component replacement priority and operation process; Among them, the maintenance time needs to be combined with the remaining life prediction results of the equipment and the production plan, and determined by virtual simulation of the impact of downtime at different time points on production through digital twin technology, accurate to the hour level, for example, according to the evaluation results and production scheduling, the maintenance time is set to "2025 October 15 8:00-12:00", to balance the equipment state and production efficiency; The component replacement priority is determined according to the criticality of the component to the overall operation of the equipment, the severity of the fault and the performance recovery effect after replacement, for example, for the composite anomaly of bearing wear and sealing ring aging, because the bearing directly affects the stability of the transmission and the fault worsens faster, the priority is set to "bearing > sealing ring > lubrication system"; The operation process covers safety specifications, specific operation steps and acceptance standards, for example, for bearing replacement, the process includes "30 minutes after shutdown and hanging warning signs → disassembly of protective cover and end cover → use special tools to remove old bearings → clean the bearing seat and apply lubricating grease → install new bearings (ensure that the fit clearance meets the standard) → reset end cover and protective cover → idle running for 30 minutes (monitor vibration and temperature) → load running verification", to ensure that the maintenance operation is standardized and can effectively restore the performance of the equipment.

[0040] In summary, the long-term vibration monitoring method based on multi-channel data acquisition provided by the embodiment of the application breaks through the limitation of traditional battery power supply by integrating a multi-modal energy collection module and an energy management unit to realize long-term self-powered supply of a sensor node; the spatiotemporal alignment of multi-channel data is realized by dynamically switching a low-frequency / high-frequency sampling mode by an edge processing unit in combination with a timestamp synchronization mechanism, so as to reduce the amount of invalid data while ensuring comprehensive monitoring; a high-precision equipment state recognition model is constructed by a deep learning model and a time-space feature fusion algorithm, so as to effectively extract multi-dimensional associated features; the transmission cost and delay are reduced by local reasoning of an edge lightweight model and uploading of key features; equipment life prediction and state evaluation are realized in combination with a historical database and a time series prediction model, and a precise maintenance strategy including maintenance time, component priority and operation process is generated based on digital twin technology, so as to finally realize the goals of low-power consumption, efficient data processing, real-time response and comprehensive state evaluation in long-term vibration monitoring, and provide reliable technical support for industrial equipment health management.

[0041] Embodiment two Figure 2 is a functional module schematic diagram of a long-term vibration monitoring system based on multi-channel data acquisition according to an embodiment of the application, as shown in Figure 2 a long-term vibration monitoring system based on multi-channel data acquisition, comprising an energy supply and sensor node module, a dynamic sampling control module, a model training and feature fusion module, an edge reasoning and data uploading module, a cloud analysis and evaluation module and a digital maintenance decision module; The energy supply and sensor node module is configured to install a sensor node integrated with a multi-modal energy collection module on a monitoring object to provide continuous power supply for multi-channel data acquisition and an edge processing unit; The dynamic sampling control module is configured to analyze vibration signal features in real time by the edge processing unit based on the powered sensor node, and select a sampling mode according to the vibration signal features to collect multi-channel sensing data; The model training and feature fusion module is configured to input the collected multi-channel sensing data into a deep learning model, extract multi-dimensional associated features by a time-space feature conversion and fusion algorithm, and construct an equipment state recognition model; The edge reasoning and data uploading module is configured to deploy the constructed equipment state recognition model on the edge processing unit, perform local reasoning on the real-time collected data, identify key feature data of abnormal states, and upload the key feature data to a cloud platform; The cloud analysis and evaluation module is configured to perform equipment life prediction and state evaluation based on the uploaded key feature data in combination with a historical database, and generate an evaluation result; The digital maintenance decision module is configured to generate a maintenance strategy by digital modeling according to the evaluation result.

[0042] In the embodiment, specifically, the energy supply and the multi-modal energy collection module in the sensor node module comprises at least two of a friction nanogenerator unit, a piezoelectric conversion unit and an electromagnetic induction unit, and an energy management unit is integrated; The energy management unit is configured to dynamically allocate the collected vibration energy of the monitoring object and the environmental mechanical energy, and to perform adaptive regulation and control of energy storage and power supply.

[0043] In the embodiment, specifically, the dynamic sampling control module comprises a threshold judgment unit and a multi-channel synchronous acquisition unit. The threshold judgment unit is configured to compare the amplitude, frequency or mutation rate of the vibration signal with a preset threshold in real time through the edge processing unit, and to select a low-frequency sampling mode or a high-frequency sampling mode according to the comparison result. The multi-channel synchronous acquisition unit is configured to perform time-space alignment acquisition of at least two kinds of data in the vibration signal, the temperature signal and the acoustic emission signal through a timestamp synchronization mechanism.

[0044] In summary, the long-term vibration monitoring system based on multi-channel data acquisition provided by the embodiment of the present application realizes continuous power supply through the energy supply and the sensor node module; the dynamic sampling control module dynamically adjusts the sampling mode and completes the time-space alignment of data; the model training and feature fusion module constructs a device state recognition model using deep learning and feature fusion algorithm; the edge inference and data upload module realizes local rapid inference and efficient upload of key features; the cloud analysis and evaluation module completes life prediction and state evaluation combined with historical data; and the digital maintenance decision module generates a maintenance strategy, forming a complete monitoring system from long-term power supply, efficient acquisition and accurate identification to scientific decision-making, meeting the low-power, comprehensive and real-time needs of long-term vibration monitoring.

[0045] For other details of the technical solutions of each module in the above-mentioned embodiment of the long-term vibration monitoring system based on multi-channel data acquisition, refer to the description in the embodiment of the long-term vibration monitoring method based on multi-channel data acquisition, which will not be repeated here.

[0046] It should be noted that each embodiment in the present specification adopts a progressive description manner, and each embodiment focuses on the difference from other embodiments, and the same and similar parts between embodiments can be referred to each other. For system embodiments, since they are basically similar to method embodiments, the description is relatively simple, and the relevant parts can be referred to the part of the method embodiment.

[0047] The above only describes the preferred embodiments of the present application and should not be used to limit the present application, and any modification, equivalent replacement, improvement, etc. made within the spirit and principles of the present application should be included in the protection scope of the present application.

Claims

1. A long-term vibration monitoring method based on multi-channel data acquisition, characterized in that, The method comprises the following steps: acquiring multi-channel data, powering an edge processing unit according to a sensor node installed with an integrated multi-modal energy harvesting module on a monitored object; based on the powered sensor node, analyzing vibration signal features in real time through the edge processing unit, selecting a sampling mode according to the vibration signal features, and collecting multi-channel sensing data; constructing a device state recognition model, inputting the collected multi-channel sensing data into a deep learning model, and extracting multi-dimensional correlation features through a time-space feature conversion and fusion algorithm; deploying the device state recognition model on the edge processing unit, performing local inference on real-time collected data, identifying key feature data of abnormal states, and uploading the key feature data to a cloud platform; based on the uploaded key feature data, performing device life prediction and state evaluation in combination with a historical database, and generating an evaluation result.

2. The long-term vibration monitoring method based on multi-channel data acquisition according to claim 1, characterized in that, The multi-modal energy harvesting module comprises at least two of a friction nanogenerator unit, a piezoelectric conversion unit, and an electromagnetic induction unit, and the multi-modal energy harvesting module collects at least one of vibration signals of the monitored object and environmental mechanical energy; The vibration signal feature extraction formula : ； wherein, is a wavelet packet transform coefficient, is a wavelet function, is a scale factor, is a translation factor, is an input signal, is a signal length, t is a time index.

3. The long-term vibration monitoring method based on multi-channel data acquisition according to claim 1, characterized in that, The sampling mode comprises a low-frequency sampling mode and a high-frequency sampling mode; The edge processing unit selects the sampling mode by judging whether the amplitude, frequency, or mutation rate of the vibration signal exceeds a preset threshold value, wherein the sampling frequency of the low-frequency sampling mode is 0.1-10 Hz, the sampling frequency of the high-frequency sampling mode is 50-1000 Hz, and the preset threshold value is dynamically adjusted according to the type of the monitored object; The sampling mode selection formula is: ； wherein, is a control coefficient, is a rate of change of the signal at time t, is a sampling rate; The multi-channel sensing data comprises at least two of vibration signals, temperature signals, and acoustic emission signals.

4. The long-term vibration monitoring method based on multi-channel data acquisition according to claim 1, characterized in that, The deep learning model comprises a 1D convolution layer, a time-series feature encoding layer, a spatial feature fusion layer, and an attention mechanism unit; The time-space feature conversion and fusion algorithm converts time-series data into a two-dimensional feature map through Gram angle field coding, extracts spatial features in combination with a convolutional neural network, and captures time-series correlation features through a gated recurrent unit.

5. The long-term vibration monitoring method based on multi-channel data acquisition according to claim 1, characterized in that, The key feature data of the abnormal state comprises time-domain features, frequency-domain features, and multi-modal correlation features, wherein the time-domain features comprise peak value, kurtosis, and root mean square value, the frequency-domain features comprise characteristic frequency amplitude and frequency band energy proportion; the local inference delay of the edge processing unit is not more than 500 ms, and the amount of key feature data uploaded to the cloud platform is not more than 5% of the amount of original collected data.

6. The long-term vibration monitoring method based on multi-channel data acquisition according to claim 1, characterized in that, The device life prediction adopts a time-series prediction model based on a long short-term memory network or a Transformer, combines fault sample data and full-life-cycle performance degradation curves in the historical database, and generates a device remaining life evaluation result.

7. The long-term vibration monitoring method based on multi-channel data acquisition according to claim 1, characterized in that, The method further comprises generating a maintenance strategy through digital modeling according to the evaluation result, wherein the digital modeling adopts digital twin technology to construct a virtual mapping model of the monitored object, simulates the effects of different maintenance schemes through virtual simulation, and generates a maintenance strategy in combination with device operation cost and downtime loss parameters; The maintenance strategy comprises maintenance time, component replacement priority, and operation process.

8. A long-term vibration monitoring system based on multi-channel data acquisition, applied to the long-term vibration monitoring method based on multi-channel data acquisition according to any one of claims 1-7, characterized in that, The energy supply and sensing node module, the dynamic sampling control module, the model training and feature fusion module, the edge inference and data uploading module, the cloud analysis and evaluation module, and the digital maintenance decision module are included. The energy supply and sensing node module is configured to install a sensor node integrated with a multi-modal energy collection module on a monitored object to provide continuous power supply for a multi-channel data acquisition and edge processing unit. The dynamic sampling control module is configured to analyze vibration signal features in real time by the edge processing unit based on the powered sensor node, and select a sampling mode according to the vibration signal features to collect multi-channel sensing data. The model training and feature fusion module is configured to input the collected multi-channel sensing data into a deep learning model, extract multi-dimensional correlation features through a time-space feature conversion and fusion algorithm, and construct a device state recognition model. The edge inference and data uploading module is configured to deploy the constructed device state recognition model on the edge processing unit, perform local inference on the real-time collected data, identify key feature data of abnormal states, and upload the key feature data to a cloud platform. The cloud analysis and evaluation module is configured to perform device life prediction and state evaluation based on the uploaded key feature data in combination with a historical database, and generate an evaluation result. The digital maintenance decision module is configured to generate a maintenance strategy through digital modeling according to the evaluation result.

9. A long-term vibration monitoring system based on multi-channel data acquisition according to claim 8, characterized in that, The multi-modal energy collection module in the energy supply and sensing node module includes at least two of a friction nanogenerator unit, a piezoelectric conversion unit, and an electromagnetic induction unit, and is integrated with an energy management unit. The energy management unit is configured to dynamically allocate the collected vibration energy of the monitored object and environmental mechanical energy, and perform adaptive regulation and control of energy storage and power supply.

10. The long-term vibration monitoring system based on multi-channel data acquisition of claim 8, wherein, The dynamic sampling control module includes a threshold judgment unit and a multi-channel synchronous acquisition unit. The threshold judgment unit is configured to compare the amplitude, frequency, or mutation rate of the vibration signal with a preset threshold in real time through the edge processing unit, and select a low-frequency sampling mode or a high-frequency sampling mode according to the comparison result. The multi-channel synchronous acquisition unit is configured to perform time-space alignment acquisition of at least two of vibration signals, temperature signals, and acoustic emission signals through a timestamp synchronization mechanism.

Images