An aeration blower predictive maintenance method, system, device and medium based on multi-modal perception and adaptive learning

Abstract

The application relates to an aeration fan predictive maintenance method, system, equipment and medium based on multi-modal perception and adaptive learning, which comprises the following steps: generating a time series data set through synchronous acquisition and combined noise reduction processing of a sensor group; generating a multi-dimensional feature vector through time-frequency feature map representation and explainability contribution analysis, combined with dynamic weight distribution of environmental factors; based on the multi-dimensional feature vector, implementing real-time anomaly detection through a lightweight model at an edge end, and uploading an abnormal data segment to a cloud end; implementing cross-sensor bidirectional reasoning on the abnormal data segment through an inference model deployed on the cloud end, reconstructing a sensor topology graph, intelligently triggering elastic incremental learning, cooperatively processing equipment degradation trend analysis, multi-source evidence fusion and spatial calibration, and outputting a life prediction result and a fault heat map. The application solves core pain points such as high early fault omission rate and insufficient model robustness, and realizes cost reduction and benefit increase and equipment life extension while maintaining diagnostic accuracy.

Description

Technical Field

[0001] This invention relates to the field of intelligent maintenance technology for industrial equipment, and in particular to a predictive maintenance method, system, equipment, and medium for aeration blowers based on multimodal perception and adaptive learning. Background Technology

[0002] The stable operation of aeration blowers in industrial processes is a key factor in ensuring system energy efficiency and production safety. Existing maintenance technologies face the following bottlenecks that urgently need to be addressed: Current detection methods primarily rely on single-modal analysis of vibration or current signals. This lack of effective integration of the gas leakage characteristic spectrum in the 20-30kHz frequency band of acoustic signals, coupled with the use of a fixed weight allocation strategy in multi-source information fusion, results in significant blind spots in identifying early bearing wear. Furthermore, existing machine learning models, employing static parameter configuration architectures, exhibit significant performance degradation under features distributed shifts during equipment aging and dynamic fluctuations in operating conditions, necessitating manual model retraining every 72 hours on average.

[0003] At the maintenance decision-making level, alarm mechanisms based on fixed thresholds can only identify discrete abnormal states. Due to the lack of continuous modeling capabilities for equipment degradation trends, it is difficult to trigger early warnings in the early stages of faults. Furthermore, the internal decision-making mechanisms of existing predictive models such as deep neural networks exhibit "black box" characteristics. Because they cannot provide interpretive outputs on the physical location of faults, technicians face substantial obstacles when verifying the rationality of maintenance strategies.

[0004] It is worth noting that while existing improved technical solutions attempt to enhance detection accuracy through multimodal combinations of vibration and acoustic signals, they still employ a fixed-weight fusion method in the feature fusion stage, resulting in insufficient effective utilization of gas leakage features. Furthermore, the design architecture of edge computing nodes, which only perform raw data uploads, causes network bandwidth utilization to consistently remain above 90%, and there is significant redundancy in the allocation of cloud computing resources. The systemic existence of these technical deficiencies objectively restricts the evolution and development of predictive maintenance technology systems for industrial equipment. Summary of the Invention

[0005] (a) Technical problems to be solved

[0006] In view of the above-mentioned shortcomings and deficiencies of the prior art, the present invention provides a predictive maintenance method, system, equipment and medium for aeration blowers based on multimodal perception and adaptive learning. It solves the technical problems of the prior art, such as early fault missed due to the lack of dynamic weighting in multimodal feature fusion, parameter drift due to static model's inability to adapt to operating condition fluctuations, lack of trend prediction in threshold alarm mechanism, and excessive bandwidth resource occupation caused by raw data transmission at the edge.

[0007] (II) Technical Solution

[0008] To achieve the above objectives, the main technical solutions adopted by the present invention include:

[0009] In a first aspect, embodiments of the present invention provide a predictive maintenance method for aeration blowers based on multimodal perception and adaptive learning, comprising:

[0010] A normalized time-series dataset is generated by synchronous data acquisition from sensor groups on the aeration blower and feature-guided noise reduction processing.

[0011] The time-series dataset is represented by time-frequency feature maps and analyzed for interpretability contribution. Combined with dynamic weight allocation coupled with environmental factors, a multi-dimensional feature vector is generated.

[0012] Based on multi-dimensional feature vectors, real-time anomaly detection is implemented at the edge using a lightweight compressed sensing model. When multiple periodic features exceed the standard and at least two modalities are jointly abnormal, the abnormal data fragment is uploaded to the cloud.

[0013] By deploying a spatiotemporal graph attention inference model in the cloud, cross-sensor bidirectional inference is performed on abnormal data fragments. The sensor topology is reconstructed based on physical adjacency connections and dynamic statistical correlations, and elastic incremental learning is conditionally triggered. Combined with degradation trajectory temporal evolution analysis, multi-evidence hierarchical fusion, and spatial registration, component-level remaining lifetime prediction results and fault heatmaps are output.

[0014] Optionally, a normalized time-series dataset is generated by synchronously acquiring data from the sensor array on the aeration blower and performing feature-guided combined noise reduction processing, including:

[0015] By deploying a sensor group at the preset detection points of the aeration blower, hardware-level synchronous data acquisition is implemented in combination with a preset clock protocol and a synchronization trigger. The sensor group includes a triaxial vibration sensor, an acoustic emission probe, a Hall current sensor, an infrared temperature sensor, and a pressure sensor.

[0016] At the edge end, the raw vibration signal collected by the triaxial vibration sensor is decomposed into 5 levels of db4 wavelet, retaining the high-frequency detail components of 10-50kHz while suppressing 50Hz power frequency interference, to obtain the vibration time sequence signal.

[0017] A 100kHz high-pass filter is applied to the raw acoustic emission signal collected by the acoustic emission probe at the edge end to extract the acoustic emission event count and acoustic feature parameters, and generate acoustic features.

[0018] At the edge, the raw temperature signal collected by the infrared temperature sensor is subjected to time series smoothing, and the temperature gradient slope is calculated by combining the sliding window to output the temperature change trend characteristics.

[0019] At the edge end, the raw pressure signal collected by the pressure sensor is processed to suppress pulse interference and extract the pressure pulsation period and amplitude characteristics.

[0020] A dynamic Kalman filter state-space model is established on the raw current signal collected by the Hall current sensor at the edge end to separate the fundamental and harmonic components and obtain the current time sequence signal.

[0021] At the edge, the sampling rate of each sensor in the sensor group is dynamically adjusted to the range of 1kHz-20kHz according to the real-time load rate of the device. Z-Score normalization is performed on the multi-modal signals, which include vibration time series signals, acoustic features, temperature change trend features, pressure pulsation period and amplitude features, and current time series signals, to generate a normalized time series dataset.

[0022] Optionally, the time-series dataset is represented by time-frequency feature maps and analyzed for interpretability contributions, combined with dynamic weight allocation coupled with environmental factors, to generate multi-dimensional feature vectors, including:

[0023] The modal signals in the time-series dataset are converted into multiple time-frequency feature maps using time-frequency analysis methods. Specifically, for vibration time-series signals, wavelet packet energy spectrum reconstruction is performed to divide multiple frequency bands aligned with high-frequency detail components, generating a vibration frequency domain energy distribution feature map. For acoustic features, an acoustic emission time-frequency map is constructed using Mel frequency cepstral coefficients. For current time-series signals, the total harmonic distortion rate and harmonic phase shift are calculated based on the separated fundamental and harmonic components to generate a distortion feature map. For temperature change trend features, a temperature gradient feature map is constructed using cubic spline interpolation based on the temperature gradient slope. For pressure pulsation period and amplitude features, a pressure amplitude-frequency joint distribution map is constructed by performing a short-time Fourier transform.

[0024] Gradient-weighted class activation mapping analysis is performed on each time-frequency feature map to calculate the contribution score of each feature map to fault detection. The contribution score represents the importance distribution of the spatiotemporal dimension in the feature map.

[0025] Based on contribution scores and real-time acquired environmental perception data, weights are dynamically allocated through an adaptive weight allocation function driven by learnable parameters.

[0026] The weighted time-frequency feature maps are concatenated along the channel dimension to generate a multi-dimensional feature vector sensitive to operating conditions.

[0027] Optionally, based on multi-dimensional feature vectors, real-time anomaly detection is implemented at the edge using a lightweight compressed sensing model. When multi-period features exceed the limit and at least two modalities are jointly anomaly detected, anomaly data fragments are uploaded to the cloud, including:

[0028] A lightweight compressed sensing model, which is a fusion of convolutional neural networks and long short-term memory networks, is pre-deployed at the edge to perform hierarchical feature abstraction and correlation analysis on the input multidimensional feature vector, and generates anomaly probability prediction values ​​in real time. After compression optimization, the lightweight compressed sensing model meets the storage capacity constraints of edge devices.

[0029] Based on the historical operating data, dynamic detection thresholds for each feature dimension are established. When the predicted abnormal probability value continuously exceeds the dynamic detection threshold for a preset number of cycles, and the feature values ​​of at least two modalities in the multidimensional feature vector simultaneously trigger abnormal conditions, it is determined to be a valid abnormal event.

[0030] Automatically extract abnormal data fragments with complete working conditions from the records before and after the effective timestamp of abnormal events, attach the corresponding feature vectors, and then upload them to the cloud;

[0031] When the network connection is interrupted, a circular data buffer is established in the edge storage unit to temporarily store abnormal data fragments. After the network is restored, the data is retransmitted according to the priority of the event urgency. The buffer capacity is designed to support the data storage needs of the device for the longest continuous operation cycle.

[0032] Optionally, a spatiotemporal graph attention inference model deployed in the cloud is used to perform cross-sensor bidirectional inference on abnormal data fragments. Based on physical adjacency connectivity and dynamic statistical correlation, the sensor topology is reconstructed and conditionally triggered for elastic incremental learning. Combined with degradation trajectory temporal evolution analysis, multi-evidence hierarchical fusion, and spatial registration, the output component-level remaining lifetime prediction results and fault heatmaps include:

[0033] Static connection edges of nodes are established based on the physical adjacency layout of the sensor group. At the same time, statistical correlation is calculated based on the acquired historical pre-training data, and dynamic connection edges that meet the correlation threshold are selected to form a sensor topology graph.

[0034] Abnormal data fragments are input into a spatiotemporal graph attention inference model deployed in the cloud, which is a fusion of convolutional neural networks and graph attention networks. The abnormal data fragments are processed by the branches of the convolutional neural network to extract local spatiotemporal features, and the branches of the graph attention network are used to analyze the cross-node correlation features in the current sensor topology graph. The local spatiotemporal features are then fused into a cross-sensor spatiotemporal correlation feature matrix.

[0035] The statistical correlation is continuously verified based on the sliding window mechanism. When the dynamic connection edge is detected to not meet the correlation threshold or an unidentified fault mode is detected, the statistical correlation is recalculated and the dynamic connection edge is updated based on the latest window data of the abnormal data fragment to reconstruct the sensor topology.

[0036] When the data distribution of the spatiotemporal correlation matrix drifts beyond a preset drift threshold or the abrupt change in the sensor topology exceeds a preset tolerance, an elastic incremental learning mechanism is triggered. The weight solidification method is used to constrain the update range of the preset core parameters, and the active learning process is triggered to manually label and update the preset uncertainty screening conditions. The labeled data is used for incremental training of the spatiotemporal correlation model to update the node association weight parameters of the graph attention network branch.

[0037] Based on the updated spatiotemporal correlation model, a time-series evolution analysis of the spatiotemporal correlation feature matrix is ​​performed using multi-scale degradation indices, including vibration energy growth rate, current harmonic distortion rate growth rate, temperature rise slope, acoustic event density, pressure fluctuation spectrum entropy change rate, and harmonic phase shift dispersion. This analysis identifies the characteristics of component degradation stages, dynamically adjusts the prediction time window length in conjunction with the degradation rate, and outputs component-level remaining life prediction values ​​with confidence intervals.

[0038] A hierarchical fusion is performed on the spatiotemporal correlation feature matrix and the dynamic topology map. Multimodal evidence including acoustic emission event counts and vibration kurtosis is fused through DS evidence theory to generate a fault probability heatmap that matches the three-dimensional structure of the equipment.

[0039] By registering spatial coordinates, the fault probability heatmap is mapped to the physical component location of the equipment's digital twin model, generating 3D highlighted annotations in the augmented reality scene. The spatiotemporal correlation feature matrix, dynamic topology update log, and fusion positioning process data are encrypted and stored to form a traceable and verifiable diagnostic evidence chain.

[0040] Optionally, based on a sliding window mechanism, statistical correlation is continuously verified. When a dynamically connected edge is detected to not meet the correlation threshold or an unidentified fault mode occurs, the statistical correlation is recalculated and the dynamically connected edges are updated according to the latest window data of the abnormal data segment to reconstruct the sensor topology graph, including:

[0041] For abnormal data segments, a sliding window segmentation process is performed in chronological order of collection time to ensure that the last window contains the latest monitoring data, the window length is consistent with the calculation window of the statistical correlation of historical pre-training data, and the overlap rate of adjacent windows is not less than 50%.

[0042] For each data group within a sliding window, extract the feature vectors corresponding to the abnormal data segments and calculate the Pearson correlation coefficient matrix across sensor nodes;

[0043] When the correlation coefficient matrix in three consecutive overlapping windows has a node statistical value lower than a preset threshold, or when a new mutation waveform that does not match the historical fault mode library is detected, the dynamic edge validity verification program is triggered.

[0044] The abnormal data segments are statistically verified by significance hypothesis testing. When the statistical correlation calculation results reach the preset probability inference threshold, the correlation coefficient matrix of the current terminal window is retained as a temporary benchmark matrix.

[0045] Based on the temporary reference matrix, the dynamic connection edge weight values ​​of the corresponding nodes in the sensor topology graph are replaced, and the failed connection edges that do not meet the preset correlation coefficient persistence condition are deleted. The static connection edges established by physical adjacency are kept unaffected by the dynamic edge updates, and the reconstructed sensor topology graph is obtained.

[0046] Optionally, a hierarchical fusion is performed on the spatiotemporal correlation feature matrix and the dynamic topology map. Multimodal evidence, including acoustic emission event counts and vibration kurtosis, is fused using DS evidence theory to generate a fault probability heatmap matching the equipment's three-dimensional structure, including:

[0047] Based on the preset component feature frequency band division rules, the spatiotemporal correlation feature matrix is ​​decomposed in the frequency domain to extract the frequency band energy ratio of each component and generate a preliminary fault area probability distribution map.

[0048] Using the preliminary fault area probability distribution map as basic evidence, and combining the time-frequency domain evidence parameters decomposed from the spatiotemporal correlation feature matrix and the node association weights in the dynamic topology graph, a basic probability allocation function for the preset fault type is constructed.

[0049] Cross-modal joint inference is performed on the basic probability assignment function using DS synthesis rules. When the frequency domain evidence parameters meet the preset discrimination threshold, the confidence assignment value of the corresponding fault proposition is increased to a preset proportion.

[0050] The confidence assignment values ​​of the improved fault propositions are spatially superimposed with the preliminary fault area probability distribution map. A grid vertex probability mapping mechanism is used to perform bilinear interpolation calculation based on the geometric topological relationship of the equipment's 3D model, generating a continuous fault probability heatmap that integrates multi-level evidence.

[0051] Secondly, embodiments of the present invention provide a predictive maintenance system for aeration blowers based on multimodal perception and adaptive learning, comprising:

[0052] The preprocessing module is used to generate a normalized time-series dataset by synchronously collecting data from the sensor group on the aeration blower and performing feature-guided combined noise reduction processing.

[0053] The vector generation module is used to generate multi-dimensional feature vectors by representing time-frequency feature maps and analyzing interpretability contributions of time-series datasets, combined with dynamic weight allocation coupled with environmental factors.

[0054] The edge coarse calculation module is used to perform real-time anomaly detection at the edge using a lightweight compressed sensing model based on multi-dimensional feature vectors. When multi-period features exceed the standard and at least two modalities are jointly abnormal, the abnormal data fragment is uploaded to the cloud.

[0055] The cloud-based fine processing module is used to perform cross-sensor bidirectional reasoning on abnormal data fragments through a spatiotemporal graph attention reasoning model deployed in the cloud. It verifies and reconstructs the sensor topology based on physical adjacency connections and dynamic statistical correlations, and conditionally triggers elastic incremental learning. Combined with degradation trajectory temporal evolution analysis, multi-evidence hierarchical fusion, and spatial registration, it outputs component-level remaining lifetime prediction results and fault heatmaps.

[0056] Thirdly, embodiments of the present invention provide a predictive maintenance device for aeration blowers based on multimodal perception and adaptive learning, comprising: at least one processor; and a memory communicatively connected to the at least one processor; wherein the memory stores instructions executable by the at least one processor, the instructions being executed by the at least one processor to enable the at least one processor to perform the predictive maintenance method for aeration blowers based on multimodal perception and adaptive learning as described above.

[0057] Fourthly, embodiments of the present invention provide a computer-readable storage medium storing computer-executable instructions, which, when executed by a processor, implement the predictive maintenance method for aeration blowers based on multimodal perception and adaptive learning as described above.

[0058] (III) Beneficial Effects

[0059] The beneficial effects of this invention are as follows: Through systematic technological integration, this invention significantly improves the intelligence level and operation and maintenance efficiency of industrial equipment health monitoring. Firstly, based on synchronous acquisition from multiple sources and feature-guided noise reduction processing, this invention constructs a normalized time-series dataset with a high signal-to-noise ratio, effectively eliminating interference from operating condition fluctuations. On this basis, through time-frequency feature map representation and interpretability contribution analysis, combined with a dynamic weight allocation mechanism coupled with the environment, a multi-dimensional feature vector with physical meaning is constructed. This not only amplifies and extracts early, weak fault features but also dynamically optimizes the feature fusion strategy, significantly advancing the detection time of hidden faults such as gas leaks.

[0060] Furthermore, at the edge, a lightweight compressed sensing model is used to perform real-time analysis of multi-dimensional feature vectors. When multimodal correlation anomalies are detected, only key anomaly fragments are uploaded to the cloud. This mechanism significantly reduces network load while maintaining detection sensitivity. In the cloud, a spatiotemporal graph attention inference model is used to achieve bidirectional correlation analysis across sensors, dynamically reconstructing the sensor topology network and triggering incremental learning. By combining degradation trajectory evolution patterns with multi-dimensional evidence fusion, accurate component-level remaining life predictions and visualized fault heatmaps are ultimately output.

[0061] Through the aforementioned edge and cloud collaboration mechanism, not only can abnormal components be quickly located, but the remaining lifespan evolution trend can also be predicted. Combined with visualized heat maps, a closed-loop management system of "precise identification - intelligent diagnosis - on-demand maintenance" can be achieved. This technology system systematically solves the core pain points of traditional solutions, such as high early fault detection rates, insufficient model robustness, lagging alarm mechanisms, and redundant data transmission. While ensuring the accuracy of fault diagnosis, it significantly reduces operation and maintenance costs and extends equipment lifespan, providing an innovative technical path for predictive maintenance of industrial equipment. Attached Figure Description

[0062] Figure 1 A flowchart illustrating the method provided in an embodiment of the present invention;

[0063] Figure 2 This is a schematic diagram illustrating the specific process of step S1 of the method provided in this embodiment of the invention;

[0064] Figure 3 This is a detailed flowchart illustrating step S2 of the method provided in this embodiment of the invention;

[0065] Figure 4 This is a detailed flowchart illustrating step S3 of the method provided in this embodiment of the invention;

[0066] Figure 5 This is a detailed flowchart illustrating step S4 of the method provided in this embodiment of the invention;

[0067] Figure 6 This is a detailed flowchart illustrating step S43 of the method provided in this embodiment of the invention;

[0068] Figure 7 A detailed flowchart illustrating step S46 of the method provided in this embodiment of the invention;

[0069] Figure 8 A schematic diagram of the thermal map and AR maintenance guidance interface of the method provided in the embodiments of the present invention;

[0070] Figure 9 A system architecture diagram provided for embodiments of the present invention;

[0071] Figure 10 This is a schematic diagram of the overall process provided for an embodiment of the present invention. Detailed Implementation

[0072] To better explain and facilitate understanding of the present invention, the present invention will be described in detail below with reference to the accompanying drawings and specific embodiments.

[0073] like Figure 1 As shown in the embodiment of the present invention, a predictive maintenance method for aeration blowers based on multimodal perception and adaptive learning includes: generating a normalized time-series dataset through synchronous acquisition of sensor data from a group on the aeration blower and feature-guided combined noise reduction processing; generating a multidimensional feature vector by characterizing the time-frequency feature map and analyzing the interpretability contribution of the time-series dataset, combined with dynamic weight allocation coupled with environmental factors; implementing real-time anomaly detection at the edge end through a compressed sensing lightweight model; uploading anomaly data fragments to the cloud when multi-period features exceed the standard and at least two modalities are jointly abnormal; performing cross-sensor bidirectional reasoning on the anomaly data fragments through a spatiotemporal graph attention inference model deployed in the cloud, verifying and reconstructing the sensor topology based on physical adjacency connections and dynamic statistical correlations, and conditionally triggering elastic incremental learning; and outputting component-level remaining life prediction results and fault heatmaps by combining degradation trajectory temporal evolution analysis, multi-evidence hierarchical fusion, and spatial registration.

[0074] This invention significantly improves the intelligence level and operation and maintenance efficiency of industrial equipment health monitoring through systematic technology integration. Firstly, based on synchronous acquisition from multiple sources and feature-guided noise reduction processing, a high signal-to-noise ratio normalized time-series dataset is constructed, effectively eliminating interference from operating condition fluctuations. Secondly, through time-frequency feature map representation and interpretability contribution analysis, combined with a dynamic weight allocation mechanism coupled with the environment, a physically meaningful multidimensional feature vector is constructed. This not only amplifies and extracts early, weak fault features but also dynamically optimizes the feature fusion strategy, significantly advancing the detection time of hidden faults such as gas leaks.

[0075] Furthermore, at the edge, a lightweight compressed sensing model is used to perform real-time analysis of multi-dimensional feature vectors. When multimodal correlation anomalies are detected, only key anomaly fragments are uploaded to the cloud. This mechanism significantly reduces network load while maintaining detection sensitivity. In the cloud, a spatiotemporal graph attention inference model is used to achieve bidirectional correlation analysis across sensors, dynamically reconstructing the sensor topology network and triggering incremental learning. By combining degradation trajectory evolution patterns with multi-dimensional evidence fusion, accurate component-level remaining life predictions and visualized fault heatmaps are ultimately output.

[0076] Through the aforementioned edge and cloud collaboration mechanism, not only can abnormal components be quickly located, but the remaining lifespan evolution trend can also be predicted. Combined with visualized heat maps, a closed-loop management system of "precise identification - intelligent diagnosis - on-demand maintenance" can be achieved. This technology system systematically solves the core pain points of traditional solutions, such as high early fault detection rates, insufficient model robustness, lagging alarm mechanisms, and redundant data transmission. While ensuring the accuracy of fault diagnosis, it significantly reduces operation and maintenance costs and extends equipment lifespan, providing an innovative technical path for predictive maintenance of industrial equipment.

[0077] To better understand the above technical solutions, exemplary embodiments of the present invention will be described in more detail below with reference to the accompanying drawings. Although exemplary embodiments of the present invention are shown in the drawings, it should be understood that the present invention can be implemented in various forms and should not be limited to the embodiments set forth herein. Rather, these embodiments are provided so that the present invention can be understood more clearly and thoroughly, and that the scope of the present invention can be fully conveyed to those skilled in the art.

[0078] Specifically, embodiments of the present invention provide a predictive maintenance method for aeration blowers based on multimodal perception and adaptive learning, including:

[0079] S1. A normalized time-series dataset is generated by synchronously collecting data from the sensor group on the aeration blower and performing feature-guided noise reduction processing.

[0080] Furthermore, such as Figure 2 As shown, step S1 includes:

[0081] S11. Hardware-level synchronous data acquisition is implemented by deploying a sensor group at the preset detection points of the aeration blower, combined with a preset clock protocol and a synchronization trigger.

[0082] The following sensor group is deployed in key parts of the aeration blower and hardware synchronization is achieved through the IEEE 1588 Precision Clock Protocol (PTP). The NI 9401 synchronization trigger module equipped with an OCXO temperature-controlled crystal oscillator (clock drift ≤ 0.1ppm) is used to ensure that the timestamp alignment error of the multi-mode signal is ≤ 10μs (designed for an acoustic emission probe with a sampling rate of 80-100kHz, the synchronization error is less than 1 / 10 of the signal period).

[0083] Triaxial vibration sensor: PCB352C33 (16kHz sampling rate, IEPE type, model PCB352C33).

[0084] Acoustic sensor: PAC acoustic emission probe (20-100kHz frequency band, model PAC, sampling rate 80-100kHz).

[0085] Current sensor: LEM LTS6-NP (Hall effect, range 0-200A).

[0086] Temperature sensor: FLIR TG165 (infrared thermal imager, accuracy ±0.5℃).

[0087] Pressure sensor: measuring range 0-1.6MPa, installed on the blower outlet pipe.

[0088] The aforementioned sensors are calibrated annually. A laser Doppler vibration meter is used to compare the vibration sensors to ensure that the sensitivity error is ≤ ±5%. Redundant sensors are designed for cross-verification to achieve self-diagnosis of sensor faults.

[0089] S12. At the edge end, the original vibration signal collected by the triaxial vibration sensor is decomposed into 5-level db4 wavelet, retaining the 10-50kHz high-frequency detail components to extract bearing fault sensitive features, while suppressing 50Hz power frequency interference to obtain the vibration time sequence signal.

[0090] S13. Perform a 100kHz high-pass filter on the raw acoustic emission signal collected by the acoustic emission probe at the edge end, extract the acoustic emission event count and acoustic feature parameters, and generate acoustic features.

[0091] S14. Perform time-series smoothing processing on the raw temperature signal collected by the infrared temperature sensor at the edge end, calculate the temperature gradient slope by combining the sliding window, and output the temperature change trend characteristics.

[0092] S15. At the edge end, the raw pressure signal collected by the pressure sensor is processed to suppress pulse interference and extract the pressure pulsation period and amplitude characteristics.

[0093] S16. At the edge end, a dynamic Kalman filter state space model is established for the raw current signal collected by the Hall current sensor to separate the fundamental and harmonic components. The signal-to-noise ratio is improved by 3dB compared with single wavelet denoising, and the current time sequence signal is obtained.

[0094] Specifically, the established Kalman filter state-space model is as follows:

[0095] x t =Fx t-1 +w t ,z t =Hx t +v t ;

[0096] In the formula, x t =Fx t-1 +w t x t: The system state vector at time t (internal variables that cannot be directly observed, such as current), F: the state transition matrix (describing the evolution of the state from time t-1 to time t, determined by the system's physical characteristics, such as kinematic equations, circuit equations, etc.), x t-1 : The state vector at time t-1 (the state at the previous time step). w t Process noise (Gaussian white noise, representing disturbances or system uncertainties not fully modeled in the model, such as external force interference, circuit thermal noise, etc., with the covariance matrix denoted as Q). Observation equation z t =Hx t +v t , z t H: Observation vector at time t (signal directly measured by the sensor, current sensor data), H: Observation matrix (mapping "unobservable states" to "observable measurements", such as converting velocity state into current observation through the relationship between current and velocity), v t Observation noise (Gaussian white noise, representing sensor measurement errors such as temperature drift and electromagnetic interference, with the covariance matrix denoted as R).

[0097] S17. At the edge end, dynamically adjust the sampling rate of each sensor in the sensor group to the range of 1kHz-20kHz according to the real-time load rate of the device (the vibration sensor is fixed at 128kHz, the acoustic emission probe is 80-100kHz, and the rest are 1-20kHz). Perform Z-Score normalization processing on the multi-mode signal containing vibration time sequence signal, acoustic characteristics, temperature change trend characteristics, pressure pulsation period and amplitude characteristics, and current time sequence signal to generate a normalized time sequence dataset.

[0098] Therefore, at the edge, the sampling rate of each sensor in the sensor group is dynamically adjusted to the range of 1kHz-20kHz according to the real-time load rate of the device (vibration sensor is fixed at 128kHz, acoustic emission probe is 80-100kHz, and the rest are 1-20kHz). Based on the hardware-level synchronization mechanism of IEEE 1588 Precision Clock Protocol (PTP) (equipped with a dedicated clock chip, supporting sub-microsecond synchronization), the timestamp alignment error of multimodal signals is ≤10μs. By using linear interpolation combined with a time-domain compensation algorithm (using linear interpolation based on the slope of adjacent sampling points to correct small synchronization deviations for acoustic emission signals), data at different sampling rates are resampled to a unified time axis. Z-Score normalization processing is performed on multimodal signals including vibration time series signals, acoustic features, temperature change trend features, pressure pulsation period and amplitude features, and current time series signals to generate a normalized time series dataset.

[0099] S2. The time series dataset is represented by time-frequency feature maps and analyzed for interpretability contribution. Combined with dynamic weight allocation coupled with environmental factors, a multi-dimensional feature vector is generated.

[0100] Furthermore, such as Figure 3 As shown, step S2 includes:

[0101] S21. The modal signals in the time-series dataset are converted into multiple time-frequency feature maps using time-frequency analysis methods. Specifically, for the vibration time-series signal, wavelet packet energy spectrum reconstruction is performed to divide multiple frequency bands aligned with high-frequency detail components, generating a vibration frequency domain energy distribution feature map; for acoustic features, an acoustic emission time-frequency map is constructed using Mel frequency cepstral coefficients; for the current time-series signal, the total harmonic distortion rate and harmonic phase offset are calculated based on the separated fundamental and harmonic components to generate a distortion feature map; for temperature change trend features, a temperature gradient feature map is constructed based on the temperature gradient slope using cubic spline interpolation; for pressure pulsation period and amplitude features, a pressure amplitude-frequency joint distribution map is constructed by implementing short-time Fourier transform.

[0102] S22. Perform gradient-weighted class activation mapping analysis on each time-frequency feature map and calculate the contribution score of each feature map to fault detection. The contribution score represents the importance distribution of the spatiotemporal dimension in the feature map.

[0103] In this step, the time-series signal is converted into a time-frequency graph (short-time Fourier transform, window length 512 points, overlap rate 50%), and the feature contribution score f is calculated using the gradient-weighted class activation map Grad-CAM. i (1) For the i-th type of mode time frequency Figure X ∈R T×F (T represents the time dimension, F represents the frequency dimension); the input to the CNN generates a feature map F∈R. H×W×C (H and W are feature map dimensions, and C is the number of channels); (2) Calculate the score S for category c. c =∑ k α k F k +b, where α k is the weight of the k-th feature map (obtained through backpropagation optimization), and b is the bias; therefore, the contribution is: f i =∑ h，w (α k / (H×W)).

[0104] S23. Based on contribution scores and real-time acquired environmental perception data, weights are dynamically allocated using an adaptive weight allocation function driven by learnable parameters. The formula for the weight allocation function is:

[0105]

[0106] Among them, f iα represents the contribution score of the i-th feature class (calculated via Grad-weighted Class Activation Mapping, Grad-CAM), ReLU is the linear rectified function, and α is a learnable parameter dynamically adjusted by the Adam optimizer, with an initial value of 0.5 and a learning rate of 0.001, and adaptive weighting is achieved through Softmax normalization. N is the total number of all feature classes, i is the weight of the i-th feature class currently being calculated, and j is the loop variable for summing the denominator. Under normal operating conditions, the weights for vibration / acoustics / current / temperature are 40% / 30% / 20% / 10%. In high-temperature environments (>40℃), α automatically increases to 0.8, and the temperature weight increases to 35%.

[0107] S24. After assigning weights, the time-frequency feature maps are concatenated along the channel dimension to generate a multi-dimensional feature vector sensitive to operating conditions.

[0108] S3. Based on multi-dimensional feature vectors, real-time anomaly detection is implemented at the edge using a lightweight compressed sensing model. When multiple periodic features exceed the standard and at least two modalities are jointly abnormal, the abnormal data fragment is uploaded to the cloud.

[0109] Furthermore, such as Figure 4 As shown, step S3 includes:

[0110] S31. A lightweight compressed sensing model, which is a fusion of convolutional neural networks and long short-term memory networks, is pre-deployed at the edge to perform hierarchical feature abstraction and correlation analysis on the input multidimensional feature vector, and generates anomaly probability prediction values ​​in real time. After compression optimization, the lightweight compressed sensing model meets the storage capacity constraints of the edge device.

[0111] S32. Based on the historical operating data, establish dynamic detection thresholds for each feature dimension. When the predicted abnormal probability value continuously exceeds the dynamic detection threshold for a preset number of cycles, and the feature values ​​of at least two modalities in the multidimensional feature vector simultaneously trigger abnormal conditions, it is determined to be a valid abnormal event.

[0112] S33. Automatically extract abnormal data fragments with complete working conditions from the records before and after the effective abnormal event timestamps, attach the corresponding feature vectors, and then upload them to the cloud.

[0113] S34. When the network connection is interrupted, a circular data buffer is established in the edge storage unit to temporarily store abnormal data fragments. After the network is restored, the data is retransmitted according to the priority of the event urgency. The buffer capacity is designed to support the data storage requirements of the device for the longest continuous operation cycle.

[0114] In one embodiment, a compressed sensing lightweight model of CNN-LSTM (1D convolutional kernel size = 16, LSTM hidden layer nodes = 64, size < 5MB) optimized by compression is deployed at the edge to perform hierarchical feature abstraction on the temporal feature vector containing key indicators such as vibration kurtosis and acoustic high-frequency energy. Based on this, a dynamic threshold triggering rule is established. When the feature value exceeds three times the standard deviation of historical normal data for three consecutive detection cycles (each cycle is based on a 10-second baseline, calculated based on a 10-minute sliding window with a 10-second window step, adjusted to two cycles under high load conditions), and at least two heterogeneous modes such as vibration and current / temperature are simultaneously abnormal, an abnormal event response mechanism is triggered.

[0115] Subsequently, complete operational data segments, 10 seconds before and after the timestamps that meet the criteria, are automatically extracted, feature vectors are attached, and then uploaded to the cloud. This strategy can reduce network traffic consumption by 80%. For network connection interruption scenarios, a circular buffer is designed to temporarily store abnormal data segments. Its capacity is configured according to the device's maximum continuous operation for 24 hours. Once the network is restored, data is retransmitted based on urgency priority. This mechanism effectively balances edge storage resources and communication reliability requirements while ensuring the integrity of faulty data.

[0116] S4. The spatiotemporal graph attention inference model deployed in the cloud performs cross-sensor bidirectional inference on abnormal data fragments. Based on physical adjacency connection and dynamic statistical correlation, the sensor topology is reconstructed and conditionally triggered for elastic incremental learning. Combined with degradation trajectory temporal evolution analysis, multi-evidence hierarchical fusion and spatial registration, the component-level remaining lifetime prediction results and fault heat map are output.

[0117] This invention deploys a spatiotemporal graph attention inference model (CNN-GAT+Transformer architecture) in the cloud for fault classification and fine-grained RUL prediction. In cross-device migration scenarios, by freezing the parameters of the first three CNN convolutional kernels and only fine-tuning the weights of the last two network layers, the model migration efficiency is improved by 30%. This inference model also integrates a degradation trajectory temporal evolution analysis module, generating component-level remaining lifetime predictions and three-dimensional fault heatmaps based on multi-evidence hierarchical fusion and sensor spatial registration.

[0118] Furthermore, such as Figure 5 As shown, step S4 includes:

[0119] S41. Establish static connection edges for nodes based on the physical adjacency layout of the sensor group. Simultaneously, calculate statistical correlation based on the acquired historical pre-training data, and filter dynamic connection edges that meet the correlation threshold to form a sensor topology graph. The physical adjacency criteria are: the distance between sensors on mechanically related components is <50cm; functionally related sensors (such as pressure and temperature sensors) are not subject to distance restrictions. The correlation threshold is dynamically adjusted according to the fault type: 0.6 for bearing faults and 0.5 for gas leaks.

[0120] S42. Abnormal data fragments are input into a spatiotemporal graph attention inference model deployed in the cloud, which is a fusion of convolutional neural networks and graph attention networks. The convolutional neural network branches process the abnormal data fragments to extract local spatiotemporal features, and the graph attention network branches analyze the cross-node correlation features in the current sensor topology graph. The local spatiotemporal features are then fused into a cross-sensor spatiotemporal correlation feature matrix. It should be noted that the spatiotemporal correlation feature matrix is ​​a multi-dimensional representation matrix formed by fusing the local temporal features extracted by CNN with the cross-sensor topological relationships modeled by GAT. Its dimensions correspond to the number of sensor network nodes × temporal depth × number of feature channels, dynamically representing the evolution characteristics of spatiotemporal coupling relationships during device operation.

[0121] S43. Based on the sliding window mechanism, the statistical correlation is continuously verified. When the dynamic connection edge is detected to not meet the correlation threshold or an unidentified fault mode occurs, the statistical correlation is recalculated and the dynamic connection edge is updated according to the latest window data of the abnormal data fragment to reconstruct the sensor topology.

[0122] Furthermore, such as Figure 6 As shown, step S43 includes:

[0123] S431. Perform sliding window segmentation on abnormal data segments in the order of collection time to ensure that the last window contains the latest monitoring data, the window length is consistent with the calculation window of the statistical correlation of historical pre-training data, and the overlap rate of adjacent windows is not less than 50%.

[0124] S432. For each data group within a sliding window, extract the feature vectors corresponding to the abnormal data segments and calculate the Pearson correlation coefficient matrix across sensor nodes.

[0125] S433. When the correlation coefficient matrix of three consecutive overlapping windows has a node statistical value lower than the preset threshold, or a new mutation waveform that does not match the historical fault mode library is detected, the dynamic edge validity verification program is triggered.

[0126] S434. Perform statistical verification on abnormal data segments through significance hypothesis testing. When the statistical correlation calculation result reaches the preset probability inference threshold, retain the correlation coefficient matrix of the current terminal window as a temporary benchmark matrix.

[0127] S435. Based on the temporary reference matrix, replace the dynamic connection edge weight values ​​of the corresponding nodes in the sensor topology graph, and delete the failed connection edges that do not meet the preset correlation coefficient persistence condition. Keep the static connection edges established by physical adjacency unaffected by the dynamic edge update, and obtain the reconstructed sensor topology graph.

[0128] In one embodiment, the present invention builds a CNN-GAT-Transformer composite architecture (i.e., spatiotemporal graph attention inference model) in the cloud by adding an acoustic-temperature feature branch: the inference module is based on transfer learning and reuses the parameters of the pre-trained ResNet-1D health baseline model. The pre-training data comes from the continuous operation data of 10 wind turbines of the same model for 12 months, covering 8 types of faults such as normal state, bearing wear, and gas leakage, totaling 500,000 sets of data.

[0129] Specifically, the CNN convolutional neural network branch uses a 16-dimensional 1D convolutional kernel (stride 8, padding = same) to extract the local spatiotemporal features of the vibration time-frequency diagram and current waveform, focusing on capturing subtle distortions in the time-domain waveform.

[0130] The GAT graph attention network branch models the sensors as an undirected graph G = (V, E), where nodes V are sensors (vibration, temperature, etc.), and edges E exist under the following dual criteria: physical adjacency (sensor spacing < 50 cm) or statistical correlation (based on a 10-minute sliding window to calculate the Pearson correlation coefficient, requiring three consecutive windows with 50% overlap and a correlation coefficient > 0.6 with a p-value < 0.05). Initial edge weights are the correlation coefficients. The GAT graph attention network branch models coupling relationships through a three-head attention mechanism, and the multi-head outputs are fused using a spatiotemporal correlation feature matrix.

[0131] The dynamic edge update mechanism performs topology reconstruction every hour: the latest abnormal data is divided into 10-minute sliding windows according to the acquisition time sequence (the end window is forced to include the latest monitoring data, and the window overlap rate is ≥50%); then, the correlation coefficient matrix of sensor nodes in the current window is calculated, and reconstruction is triggered when the following situations occur in three consecutive overlapping windows: the correlation coefficient between nodes is lower than the 0.6 threshold, or a new mutation waveform in the high-frequency band (10-50kHz) is detected (not matching the historical fault mode library); after verification through significance hypothesis testing, the correlation coefficient matrix of the end window is set as a temporary benchmark, the dynamic edge weights are updated and the failed connections are deleted, while the static edges established by physical adjacency are retained.

[0132] The Transformer module contains a 6-layer encoder, each layer containing a multi-head attention (number of heads = 8) and a feedforward neural network, with a dropout rate of 0.1. It adaptively enhances the attention weight of the 10-50kHz high-frequency vibration component for bearing fault characteristics, and suppresses environmental noise interference through a time-frequency masking mechanism to improve the decoding sensitivity of fault characteristics.

[0133] The above architecture achieves end-to-end optimization of local feature extraction, cross-sensor correlation modeling, and fault-sensitive signal enhancement through a three-level collaborative processing of CNN-GAT-Transformer, resulting in a significant improvement in edge computing efficiency compared to traditional solutions.

[0134] S44. When the data distribution drift of the spatiotemporal correlation matrix exceeds the preset drift threshold or the abrupt change in the sensor topology exceeds the preset tolerance, the elastic incremental learning mechanism is triggered. The weight solidification method is used to constrain the update range of the preset core parameters, and the active learning process is triggered to manually label and update the preset uncertainty screening conditions. The labeled data is used for incremental training of the spatiotemporal correlation model to update the node association weight parameters of the graph attention network branch.

[0135] In this step, when the KL divergence of the spatiotemporal correlation matrix is ​​detected in real time (threshold 0.3) or new fault types (confidence = 0.5), the learning mechanism is activated when the following occurs: data distribution drift (KL divergence > 0.3) and unidentified fault patterns are detected (classification confidence < 0.5).

[0136] At this point, Elastic Weight Fixation (EWC) is implemented: This constrains the preset core parameters, with a weight protection coefficient λ = 0.8, used to balance the weights of learning new data and retaining historical knowledge. A larger value indicates a stronger constraint on the core parameters, preventing "catastrophic forgetting" caused by over-updating. A penalty term is added to the loss function. Fisher i This is the parameter importance matrix. (Fischer) i (Fisher information matrix elements) are calculated using the gradient variance of historical data to characterize the importance of the i-th parameter to model performance (the larger the value, the more critical the parameter), ensuring that important parameters are better protected during updates. i (Current parameter value) is the real-time updated value of the i-th parameter during the training process on new data. (Initial parameter values) are the baseline values ​​of the i-th parameter during the initial training (or the previous stable state), representing the historical knowledge that needs to be retained. Squared terms It is used to measure the degree of deviation between the current parameters and the initial parameters. The greater the deviation, the stronger the impact of the penalty term on the loss function, forcing the model to update as close as possible to the initial values ​​of the core parameters.

[0137] Furthermore, the active learning process begins: samples with a confidence level of 0.5-0.7 and a KL divergence > 0.4 are automatically selected and prioritized by fault type (bearing > leakage > impeller imbalance > gearbox failure) and pushed to the manual annotation interface. After quality verification, the annotated data is input into the incremental training process to update the node association weight parameters of the GAT layer in a targeted manner.

[0138] S45. Based on the updated spatiotemporal correlation model, a time-series evolution analysis is performed on the spatiotemporal correlation feature matrix using multi-scale degradation indices including vibration energy growth rate, current harmonic distortion rate growth rate, temperature rise slope, acoustic event density, pressure fluctuation spectrum entropy change rate, and harmonic phase shift dispersion. This identifies the characteristics of component degradation stages, dynamically adjusts the prediction time window length in conjunction with the degradation rate, and outputs a component-level remaining life prediction value with a confidence interval.

[0139] In another embodiment, based on the updated spatiotemporal correlation model, a cross-period time-series evolution analysis is performed on the spatiotemporal correlation feature matrix using multi-scale degradation indices including vibration energy growth rate, current harmonic distortion rate growth rate, temperature rise slope, acoustic event density, pressure fluctuation spectral entropy change rate, and harmonic phase shift dispersion. The Transformer module models historical degradation trajectories (input features including the aforementioned degradation indices), setting the input sequence to 7 days of continuous monitoring data (sampling frequency 10 minutes / time), and outputting the component-level remaining life prediction value for the next 7-30 days (confidence interval ±12 hours, prediction error ≤8%) and failure probability. A dynamic window length adjustment mechanism is also built-in: when a sudden increase in vibration energy growth rate (e.g., daily increase > 50%) or other abrupt degradation characteristics are detected, the input sequence length is automatically compressed to 1 day of data, and the prediction frequency is increased to minute-level sampling, achieving rapid response to degradation inflection points. This mechanism adaptively adjusts the time window scale by analyzing degradation rate changes in real time, ensuring the prediction model maintains long-term prediction stability during stable degradation phases and enhances short-term prediction sensitivity during accelerated degradation phases.

[0140] S46. Perform hierarchical fusion of the spatiotemporal correlation feature matrix and the dynamic topology map, and fuse multimodal evidence including acoustic emission event counts and vibration kurtosis through DS evidence theory to generate a fault probability heatmap that matches the three-dimensional structure of the equipment.

[0141] Furthermore, such as Figure 7 As shown, step S46 includes:

[0142] S461. Based on the preset component characteristic frequency band division rules, the spatiotemporal correlation feature matrix is ​​decomposed in the frequency domain, the energy proportion of the characteristic frequency band corresponding to each component is extracted, and a preliminary fault area probability distribution map is generated.

[0143] S462. Using the preliminary fault area probability distribution map as basic evidence, and combining the time-frequency domain evidence parameters decomposed from the spatiotemporal correlation feature matrix with the node association weights in the dynamic topology map, a basic probability allocation function for the preset fault type is constructed.

[0144] S463. Perform cross-modal joint inference on the basic probability allocation function through DS synthesis rules. When the frequency domain evidence parameters meet the preset discrimination threshold, increase the confidence allocation value of the corresponding fault proposition to a preset ratio.

[0145] S464. The improved fault confidence allocation results are spatially superimposed with the preliminary fault area probability distribution map. A grid vertex probability mapping mechanism is used to perform bilinear interpolation calculation based on the geometric topological relationship of the equipment's three-dimensional model to generate a continuous fault probability heatmap that integrates multi-level evidence.

[0146] S47. By registering spatial coordinates, the fault probability heatmap is mapped to the physical component location of the equipment's digital twin model, generating three-dimensional highlighted annotations in the augmented reality scene, and the spatiotemporal correlation feature matrix, dynamic topology update log, and fusion positioning process data are encrypted and stored to form a traceable and verifiable diagnostic evidence chain.

[0147] In another embodiment, a preliminary heatmap is first generated based on Gradient Class Activation Mapping (Grad-CAM), and spatial positioning is then performed using a 3D digital twin model of the equipment (built according to CAD drawings, with fewer than 50,000 triangular faces). Its core process includes the following technical points:

[0148] First, hierarchical localization: The time-frequency feature matrix is ​​decomposed according to a preset component characteristic frequency band division rule (10-50kHz for bearings, 5-15kHz for blades), and the energy proportion of each frequency band is extracted to generate an initial probability distribution map for preliminary localization. Next, in fine-grained localization, a basic probability allocation function is constructed, including acoustic emission event counts (converted to a density factor through exponential smoothing; when the acoustic emission event count > 100 times / second, the density factor is mapped to 0.7-0.9), vibration kurtosis (converted to anomaly coefficient through normalization; when the vibration kurtosis > 5.0, the anomaly coefficient is mapped to 0.6-0.8), and topological node weights. When multimodal evidence meets the joint discrimination threshold (e.g., vibration kurtosis ↑30% and acoustic emission count ↑50%), the bearing fault confidence is increased to 92% using the DS synthesis rule.

[0149] Second, heat map output: The improved fault confidence allocation results are spatially superimposed with the preliminary fault area probability distribution map. A grid vertex probability mapping mechanism is used to perform bilinear interpolation calculation based on the geometric topological relationship of the equipment's three-dimensional model to generate a continuous fault probability heat map that integrates multi-level evidence.

[0150] Third, 3D registration: Utilizing ARKit spatial positioning technology, combined with device QR code markers (5cm×5cm, spacing ≥50cm), under strong light (illuminance >5000 lux) and dust (concentration >50mg / m³) conditions. 3 Under these conditions, through 20 repeated positioning tests, a fault thermal map with an accuracy of ±8mm was achieved and aligned with the actual equipment at the millimeter level under strong light, dust, and low temperature (-10℃) conditions.

[0151] Fourth, AR-enhanced operations and maintenance: such as Figure 8 As shown, the AR visualization interface displays the complete closed-loop process: Step 1: Load the 3D model of the equipment (based on CAD drawings, triangular facets <50,000); Step 2: Render the fault heat map (DS evidence theory fusion, vibration kurtosis ↑30% + acoustic emission event ↑50% → bearing failure probability 92%); Step 3: Highlight abnormal areas (energy concentration in bearing housing 10-50kHz); Step 4: Generate maintenance instructions (ARKit spatial positioning ±8mm accuracy); Specifically extending to each maintenance branch: Replace 6308 bearing (recommended to be performed within 48 hours); RUL prediction 72±12 hours (confidence interval ±12 hours); Lubrication remaining 35% (ΔT / Δt>0.5℃ / h and ΔK<-1.5); Step 5: Record the work order (data immutability 100%), presenting the entire process of equipment from AR visualization detection to generating maintenance instructions, performing maintenance, and recording the work order.

[0152] In one specific embodiment, this embodiment demonstrates the complete implementation process of an intelligent maintenance system for aeration blowers in a wastewater treatment plant:

[0153] Step 1: Deploy a multimodal sensor network in key components of the aeration blower: Install a PCB352C33 vibration sensor (16kHz sampling) and a PAC acoustic emission probe (20-100kHz frequency band) on the bearing housing; configure a LEM LTS6-NP current sensor (0-200A range) and a FLIR TG165 infrared thermal imager (±0.5℃ accuracy) at the motor end; and install a 0-1.6MPa pressure sensor on the blower outlet pipe. Data acquisition uses an NICDAQ-9178 module, and synchronization with an IEEE 1588 clock ensures that the timestamp error of each sensor is ≤0.6ms (meeting the ≤1ms requirement).

[0154] Step 2: Construct a composite feature system: Vibration signals are decomposed using wavelet packets to extract energy in the 10-50kHz frequency band (kurtosis > 5.0 triggers initial judgment); 13-dimensional MFCC coefficients are extracted from acoustic signals and acoustic emission events are statistically analyzed (>100 times / second early warning); harmonic distortion rate (THD) and time series trends are analyzed from current signals. A dynamic weighting mechanism is established: Under normal operating conditions, vibration (40%), acoustic (30%), current (20%), and temperature (10%) are weighted; at high temperatures (>40℃), the temperature weight automatically increases to 35%.

[0155] The initial model was pre-trained on 100,000 sets of data from 3 wind turbines of the same model operating continuously for 6 months (covering 20%-100% load conditions, including 8 types of faults such as normal state, bearing wear, gas leakage, impeller imbalance, and gearbox failure, with balanced samples for each fault type). After being transferred to the target equipment, the parameters were updated every 24 hours through incremental learning, with the EWC protection bearing fault classification layer weight (λ = 0.8).

[0156] Referring to Table 1, compare the impact of different numbers of GAT attention heads (1 / 3 / 5 heads) on fault location accuracy:

[0157] Table 1. Impact of different GAT attention head numbers (1 / 3 / 5 heads) on fault location accuracy.

[0158]

[0159]

[0160] Therefore, the parameter verification experiment shows that, considering both accuracy and edge computing resource limitations, a 3-head attention mechanism is selected based on the edge processor (model: NVIDIA Jetson Xavier, 8GB memory). Dynamic weight α parameter testing: covering a temperature range of 20℃-60℃, with a fixed high-temperature condition (45℃), the temperature weight increases to 25% when α=0.5 and to 35% when α=0.8, the latter advancing the detection time of temperature-related faults by 2 hours. Under 20℃ conditions, when α=0.5, the modal weights stabilize at 40% for vibration and 30% for acoustics, verifying the weight stability under low-temperature conditions.

[0161] Step 3: Achieve closed-loop management through a multi-level response mechanism:

[0162] 1. Edge warning: When the vibration kurtosis is 5.8 and the acoustic high-frequency energy increases by 40%, an warning is triggered and abnormal data is uploaded to the cloud for 10 seconds.

[0163] 2. Cloud analysis: Grad-CAM locates the bearing housing area (10-50kHz energy concentration), and the Transformer model outputs RUL = 72 ± 12 hours.

[0164] 3. AR Interaction: HoloLens2 displays a highlighted heat map of the bearing area and pushes a work order: "Replace bearing 6308, recommended to be executed within 48 hours, current lubrication remaining 35% (calculated based on temperature slope of 0.6℃ / h and vibration kurtosis increment of 1.8)".

[0165] 4. Post-maintenance processing: The bearing replacement data is stored on the blockchain. When migrating across devices, the training time of the target device model is reduced from 48 hours to 6 hours. The migration efficiency has been verified on three different types of wind turbines.

[0166] Therefore, 12 consecutive months of operational data show that: unplanned downtime decreased by 60%, maintenance costs decreased by 32%; and operating conditions (illuminance > 5000 lux / dust > 50 mg / m³) were improved. 3 Fault location time under heavy rain is reduced to 30 minutes (±8mm accuracy), and stable performance is maintained even at -10℃. Long-term operational stability tests show that the model's performance degradation is <5% over the 5-year operating cycle of the equipment, and the sensor time synchronization error remains ≤1ms.

[0167] The following are the remaining test data for the present invention:

[0168] Table 2 Test data for multiple indicators of the present invention

[0169]

[0170] Additionally, embodiments of the present invention provide a predictive maintenance system for aeration blowers based on multimodal perception and adaptive learning, comprising:

[0171] The preprocessing module is used to generate a normalized time-series dataset by synchronously acquiring data from the sensor group on the aeration blower and performing feature-guided noise reduction.

[0172] The vector generation module is used to generate multi-dimensional feature vectors by representing time-frequency feature maps and analyzing interpretability contributions of time-series datasets, combined with dynamic weight allocation coupled with environmental factors.

[0173] The edge coarse calculation module is used to perform real-time anomaly detection at the edge using a lightweight compressed sensing model based on multi-dimensional feature vectors. When multi-period features exceed the standard and at least two modalities are jointly abnormal, the abnormal data fragment is uploaded to the cloud.

[0174] The cloud-based fine processing module is used to perform cross-sensor bidirectional reasoning on abnormal data fragments through a spatiotemporal graph attention reasoning model deployed in the cloud. It verifies and reconstructs the sensor topology based on physical adjacency connections and dynamic statistical correlations, and conditionally triggers elastic incremental learning. Combined with degradation trajectory temporal evolution analysis, multi-evidence hierarchical fusion, and spatial registration, it outputs component-level remaining lifetime prediction results and fault heatmaps.

[0175] Furthermore, this embodiment of the invention also provides a predictive maintenance device for aeration blowers based on multimodal perception and adaptive learning, comprising: at least one processor; and a memory communicatively connected to the at least one processor; wherein the memory stores instructions executable by the at least one processor, the instructions being executed by the at least one processor to enable the at least one processor to perform the predictive maintenance method for aeration blowers based on multimodal perception and adaptive learning as described above.

[0176] Furthermore, embodiments of the present invention provide a computer-readable storage medium storing computer-executable instructions, which, when executed by a processor, implement the predictive maintenance method for aeration blowers based on multimodal perception and adaptive learning as described above.

[0177] In summary, a predictive maintenance system and method for aeration blowers based on multimodal perception and adaptive learning, referencing... Figure 9 The overall architecture of this invention is: multimodal sensing layer → edge computing layer → cloud analytics layer → AR interaction layer. (See also...) Figure 10 The overall process is as follows: First, equipment status signals are simultaneously acquired through multiple source sensors such as vibration and acoustic sensors. Preprocessing (wavelet packet decomposition and harmonic extraction) is performed by the edge computing layer to extract dynamic features such as vibration kurtosis and acoustic MFCC. Second, a weight allocation mechanism is introduced to optimize feature contribution, and acoustic transient impact detection (e.g., acoustic emission event count > 100 times / second) is combined to enhance anomaly sensitivity. Then, a lightweight CNN-LSTM model is deployed at the edge for anomaly detection. Subsequently, a CNN-GAT-Transformer composite model is constructed, where the CNN extracts local spatiotemporal features, and the GAT... The layer models the sensor topology dependency through a three-head attention mechanism, and the spatiotemporal features are fused and input into the Transformer to enhance the fault-sensitive signal. Furthermore, the model performance is monitored in real time. When the KL divergence is detected to be >0.3 or a new fault type is detected, the incremental learning of EWC constraints (weights of the protective bearing fault classification layer) is triggered and actively labeled data is integrated. Under normal operating conditions, the remaining life prediction is directly performed (the measured prediction error is ≤8%). Finally, the DS evidence theory is used to fuse multimodal evidence, and the three-dimensional heat map positioning result is output through the AR interaction layer to achieve the end-to-end connection from data acquisition to fault visualization.

[0178] Therefore, the present invention has the following specific innovative points:

[0179] (1) Enhanced multimodal dynamic features: Grad-CAM feature contribution analysis is introduced into time series data fusion. Combined with acoustic signal MFCC and transient impact detection, in the actual test of the sewage treatment plant, the detection time of gas leakage fault is 48 hours earlier than the traditional scheme, and the missed detection rate is reduced from 20% to less than 3%.

[0180] (2) Adaptive hybrid model architecture: The CNN-GAT hybrid network realizes the modeling of spatiotemporal features and sensor dependence. The dynamic weight allocation mechanism automatically optimizes the multimodal fusion weights, avoids the bias of manual parameter tuning, and compresses the model volume to 1 / 5 of the traditional scheme. The edge inference latency is <50ms. The 3-head attention mechanism of the GAT layer enables the bearing fault location accuracy to reach 95%.

[0181] (3) Explainable prediction and intelligent operation and maintenance: Grad-CAM heat map combined with DS evidence theory realizes multi-component fault decoupling and localization, and AR interactive interface realizes "what you see is what you repair", which improves maintenance efficiency by 40% and shortens fault diagnosis time from 2 hours to 30 minutes.

[0182] (4) Edge-cloud collaboration and incremental learning: Dynamic threshold triggering reduces invalid data transmission by 80%, and the EWC+active learning combination mechanism reduces annotation costs by 50%, adapting to complex industrial working conditions. When the model is migrated across devices, the training time is shortened from 48 hours to 6 hours.

[0183] Since the systems / devices described in the above embodiments of the present invention are systems / devices used to implement the methods of the above embodiments of the present invention, those skilled in the art can understand the specific structure and modifications of the systems / devices based on the methods described in the above embodiments of the present invention, and therefore will not be repeated here. All systems / devices used in the methods of the above embodiments of the present invention fall within the scope of protection of the present invention.

[0184] Those skilled in the art will understand that embodiments of the present invention can be provided as methods, systems, or computer program products. Therefore, the present invention can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present invention can take the form of a computer program product embodied on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0185] This invention is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of the invention. It will be understood that each block of the flowchart illustrations and / or block diagrams, as well as combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions.

[0186] It should be noted that, in this invention, any reference numerals placed between parentheses should not be construed as limiting the scope of the invention. The word "comprising" does not exclude the presence of components or steps not listed in this invention. The word "a" or "an" preceding a component does not exclude the presence of a plurality of such components. This invention can be implemented by means of hardware comprising several different components and by means of a suitably programmed computer. In this invention, where several devices are listed, several of these devices may be embodied by the same hardware. The use of the terms first, second, third, etc., is merely for convenience and does not indicate any order. These terms can be understood as part of the component names.

[0187] Furthermore, it should be noted that in the description of this specification, the terms "one embodiment," "some embodiments," "embodiment," "example," "specific example," or "some examples," etc., refer to specific features, structures, materials, or characteristics described in connection with that embodiment or example, which are included in at least one embodiment or example of the present invention. In this specification, the illustrative expressions of the above terms do not necessarily refer to the same embodiment or example. Moreover, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in one or more embodiments or examples. Furthermore, without contradiction, those skilled in the art can combine and integrate the different embodiments or examples described in this specification, as well as the features of different embodiments or examples.

[0188] Although preferred embodiments of the invention have been described, those skilled in the art, upon learning the basic inventive concept, can make other changes and modifications to these embodiments. Therefore, the invention should be interpreted as including the preferred embodiments as well as all changes and modifications falling within the scope of the invention.

[0189] Obviously, those skilled in the art can make various modifications and variations to this invention without departing from its spirit and scope. Therefore, if these modifications and variations fall within the scope of this invention and its equivalents, then this invention should also include these modifications and variations.

Claims

1. A predictive maintenance method for aeration blowers based on multimodal perception and adaptive learning, characterized in that, include: A normalized time-series dataset is generated by synchronous data acquisition from sensor groups on the aeration blower and feature-guided noise reduction processing. The time-series dataset is represented by time-frequency feature maps and analyzed by gradient-weighted class activation mapping. The contribution score of each feature map to fault detection is calculated. Combined with the dynamic weight allocation coupled with environmental factors, a multi-dimensional feature vector is generated. Based on multidimensional feature vectors, real-time anomaly detection is implemented at the edge through a lightweight compressed sensing model that integrates convolutional neural networks and long short-term memory networks. When the anomaly probability prediction value continuously exceeds the dynamic detection threshold for a preset number of cycles, and the feature values ​​of at least two modalities in the multidimensional feature vector simultaneously trigger the anomaly condition, the anomaly data fragment is uploaded to the cloud. A spatiotemporal graph attention inference model, fused with convolutional neural networks and graph attention networks, deployed in the cloud, performs cross-sensor bidirectional inference on anomalous data fragments. Based on physical adjacency connections and dynamic statistical correlations, it verifies and reconstructs the sensor topology graph and conditionally triggers elastic incremental learning. Combining degradation trajectory temporal evolution analysis, multi-evidence hierarchical fusion, and spatial registration, it outputs component-level remaining lifetime prediction results and fault heatmaps. This includes: establishing static connection edges between nodes based on the physical adjacency layout of the sensor group; calculating statistical correlations based on acquired historical pre-training data; selecting dynamic connection edges that meet the correlation threshold; and forming the sensor topology graph. Anomalous data fragments are then input into the spatiotemporal graph attention inference model to extract local spatiotemporal features and analyze cross-node connections in the current sensor topology graph. The associated features are fused to obtain a spatiotemporal correlation feature matrix across sensors. Statistical correlation is continuously verified based on a sliding window mechanism. When a dynamic connection edge fails to meet the correlation threshold or an unidentified fault mode is detected, the statistical correlation is recalculated and the dynamic connection edge is updated based on the latest window data of the abnormal data fragment to reconstruct the sensor topology. When the data distribution drift of the spatiotemporal correlation feature matrix exceeds a preset drift threshold or the structural abrupt change of the sensor topology exceeds a preset tolerance, an elastic incremental learning mechanism is triggered. A weight solidification method is used to constrain the update range of preset core parameters, and an active learning process is triggered to manually label and update the preset uncertainty screening conditions. The labeled data is used for incremental training of the spatiotemporal correlation model to update the node association weight parameters of the graph attention network branches.

2. The predictive maintenance method for aeration blowers based on multimodal perception and adaptive learning as described in claim 1, characterized in that, By synchronously acquiring data from sensor arrays on the aeration blower and using feature-guided combined noise reduction processing, a normalized time-series dataset is generated, including: By deploying a sensor group at the preset detection points of the aeration blower, hardware-level synchronous data acquisition is implemented in combination with a preset clock protocol and a synchronization trigger. The sensor group includes a triaxial vibration sensor, an acoustic emission probe, a Hall current sensor, an infrared temperature sensor, and a pressure sensor. At the edge end, the raw vibration signal collected by the triaxial vibration sensor is decomposed into 5 levels of db4 wavelet, retaining the high-frequency detail components of 10-50kHz while suppressing 50Hz power frequency interference, to obtain the vibration time sequence signal. A 100kHz high-pass filter is applied to the raw acoustic emission signal collected by the acoustic emission probe at the edge end to extract the acoustic emission event count and acoustic feature parameters, and generate acoustic features. At the edge, the raw temperature signal collected by the infrared temperature sensor is subjected to time series smoothing, and the temperature gradient slope is calculated by combining the sliding window to output the temperature change trend characteristics. At the edge end, the raw pressure signal collected by the pressure sensor is processed to suppress pulse interference and extract the pressure pulsation period and amplitude characteristics. A dynamic Kalman filter state-space model is established on the raw current signal collected by the Hall current sensor at the edge end to separate the fundamental and harmonic components and obtain the current time sequence signal. At the edge, the sampling rate of each sensor in the sensor group is dynamically adjusted to the range of 1kHz-20kHz according to the real-time load rate of the device. Z-Score normalization is performed on the multi-modal signals, which include vibration time series signals, acoustic features, temperature change trend features, pressure pulsation period and amplitude features, and current time series signals, to generate a normalized time series dataset.

3. The predictive maintenance method for aeration blowers based on multimodal perception and adaptive learning as described in claim 2, characterized in that, The time-series dataset is represented by time-frequency feature maps and analyzed using gradient-weighted class activation mapping. The contribution score of each feature map to fault detection is calculated. Combined with dynamic weight allocation coupled with environmental factors, a multi-dimensional feature vector is generated, including: The modal signals in the time-series dataset are converted into multiple time-frequency feature maps using time-frequency analysis methods. Specifically, for vibration time-series signals, wavelet packet energy spectrum reconstruction is performed to divide multiple frequency bands aligned with high-frequency detail components, generating a vibration frequency domain energy distribution feature map. For acoustic features, an acoustic emission time-frequency map is constructed using Mel frequency cepstral coefficients. For current time-series signals, the total harmonic distortion rate and harmonic phase shift are calculated based on the separated fundamental and harmonic components to generate a distortion feature map. For temperature change trend features, a temperature gradient feature map is constructed using cubic spline interpolation based on the temperature gradient slope. For pressure pulsation period and amplitude features, a pressure amplitude-frequency joint distribution map is constructed by performing a short-time Fourier transform. Gradient-weighted class activation mapping analysis is performed on each time-frequency feature map to calculate the contribution score of each feature map to fault detection. The contribution score represents the importance distribution of the spatiotemporal dimension in the feature map. Based on contribution scores and real-time acquired environmental perception data, weights are dynamically allocated through an adaptive weight allocation function driven by learnable parameters. The weighted time-frequency feature maps are concatenated along the channel dimension to generate a multi-dimensional feature vector sensitive to operating conditions.

4. The predictive maintenance method for aeration blowers based on multimodal perception and adaptive learning as described in claim 1, characterized in that, Based on multidimensional feature vectors, a lightweight compressed sensing model, fused with convolutional neural networks and long short-term memory networks, is deployed at the edge to perform real-time anomaly detection. When the predicted anomaly probability continuously exceeds the dynamic detection threshold for a preset number of cycles, and at least two modalities of the feature vector simultaneously trigger anomaly conditions, anomaly data fragments are uploaded to the cloud, including: A lightweight compressed sensing model, which is a fusion of convolutional neural networks and long short-term memory networks, is pre-deployed at the edge to perform hierarchical feature abstraction and correlation analysis on the input multidimensional feature vector, and generates anomaly probability prediction values ​​in real time. After compression optimization, the lightweight compressed sensing model meets the storage capacity constraints of edge devices. Based on the historical operating data, dynamic detection thresholds for each feature dimension are established. When the predicted abnormal probability value continuously exceeds the dynamic detection threshold for a preset number of cycles, and the feature values ​​of at least two modalities in the multidimensional feature vector simultaneously trigger abnormal conditions, it is determined to be a valid abnormal event. Automatically extract abnormal data fragments with complete working conditions from the records before and after the effective timestamp of abnormal events, attach the corresponding feature vectors, and then upload them to the cloud; When the network connection is interrupted, a circular data buffer is established in the edge storage unit to temporarily store abnormal data fragments. After the network is restored, the data is retransmitted according to the priority of the event urgency. The buffer capacity is designed to support the data storage needs of the device for the longest continuous operation cycle.

5. The predictive maintenance method for aeration blowers based on multimodal perception and adaptive learning as described in claim 1, characterized in that, A spatiotemporal graph attention inference model, fused with convolutional neural networks and graph attention networks and deployed in the cloud, performs cross-sensor bidirectional inference on anomalous data fragments. Based on physical adjacency connectivity and dynamic statistical correlation, it verifies and reconstructs the sensor topology and conditionally triggers elastic incremental learning. Combining degradation trajectory temporal evolution analysis, multi-evidence hierarchical fusion, and spatial registration, it outputs component-level remaining lifetime prediction results and fault heatmaps. It also includes: Based on the updated spatiotemporal correlation model, a time-series evolution analysis of the spatiotemporal correlation feature matrix is ​​performed using multi-scale degradation indices, including vibration energy growth rate, current harmonic distortion rate growth rate, temperature rise slope, acoustic event density, pressure fluctuation spectrum entropy change rate, and harmonic phase shift dispersion. This analysis identifies the characteristics of component degradation stages, dynamically adjusts the prediction time window length in conjunction with the degradation rate, and outputs component-level remaining life prediction values ​​with confidence intervals. A hierarchical fusion is performed on the spatiotemporal correlation feature matrix and the dynamic topology map. Multimodal evidence including acoustic emission event counts and vibration kurtosis is fused through DS evidence theory to generate a fault probability heatmap that matches the three-dimensional structure of the equipment. By registering spatial coordinates, the fault probability heatmap is mapped to the physical component location of the equipment's digital twin model, generating 3D highlighted annotations in the augmented reality scene. The spatiotemporal correlation feature matrix, dynamic topology update log, and fusion positioning process data are encrypted and stored to form a traceable and verifiable diagnostic evidence chain.

6. The predictive maintenance method for aeration blowers based on multimodal perception and adaptive learning as described in claim 5, characterized in that, The statistical correlation is continuously verified based on a sliding window mechanism. When a dynamically connected edge is detected that does not meet the correlation threshold or an unidentified fault mode occurs, the statistical correlation is recalculated and the dynamically connected edges are updated based on the latest window data of the abnormal data segment to reconstruct the sensor topology graph, including: For abnormal data segments, a sliding window segmentation process is performed in chronological order of collection time to ensure that the last window contains the latest monitoring data, the window length is consistent with the calculation window for the statistical correlation of historical pre-training data, and the overlap rate of adjacent windows is not less than 50%. For each data group within a sliding window, extract the feature vectors corresponding to the abnormal data segments and calculate the Pearson correlation coefficient matrix across sensor nodes; When the correlation coefficient matrix in three consecutive overlapping windows has a node statistical value lower than a preset threshold, or when a new mutation waveform that does not match the historical fault mode library is detected, the dynamic edge validity verification program is triggered. The abnormal data segments are statistically verified by significance hypothesis testing. When the statistical correlation calculation results reach the preset probability inference threshold, the correlation coefficient matrix of the current terminal window is retained as a temporary benchmark matrix. Based on the temporary reference matrix, the dynamic connection edge weight values ​​of the corresponding nodes in the sensor topology graph are replaced, and the failed connection edges that do not meet the preset correlation coefficient persistence condition are deleted. The static connection edges established by physical adjacency are kept unaffected by the dynamic edge updates, and the reconstructed sensor topology graph is obtained.

7. The predictive maintenance method for aeration blowers based on multimodal perception and adaptive learning as described in claim 5, characterized in that, A hierarchical fusion of the spatiotemporal correlation feature matrix and the dynamic topology map is performed. Multimodal evidence, including acoustic emission event counts and vibration kurtosis, is fused using DS evidence theory to generate a fault probability heatmap matching the equipment's three-dimensional structure, including: Based on the preset component feature frequency band division rules, the spatiotemporal correlation feature matrix is ​​decomposed in the frequency domain to extract the frequency band energy ratio of each component and generate a preliminary fault area probability distribution map. Using the preliminary fault area probability distribution map as basic evidence, and combining the time-frequency domain evidence parameters decomposed from the spatiotemporal correlation feature matrix and the node association weights in the dynamic topology graph, a basic probability allocation function for the preset fault type is constructed. Cross-modal joint inference is performed on the basic probability assignment function using DS synthesis rules. When the frequency domain evidence parameters meet the preset discrimination threshold, the confidence assignment value of the corresponding fault proposition is increased to a preset proportion. The confidence assignment values ​​of the improved fault propositions are spatially superimposed with the preliminary fault area probability distribution map. A grid vertex probability mapping mechanism is used to perform bilinear interpolation calculation based on the geometric topological relationship of the equipment's 3D model, generating a continuous fault probability heatmap that integrates multi-level evidence.

8. A predictive maintenance system for aeration blowers based on multimodal perception and adaptive learning, employing the method described in any one of claims 1-7, characterized in that, include: The preprocessing module is used to generate a normalized time-series dataset by synchronously collecting data from the sensor group on the aeration blower and performing feature-guided combined noise reduction processing. The vector generation module is used to generate multi-dimensional feature vectors by representing time-frequency feature maps and analyzing interpretability contributions of time-series datasets, combined with dynamic weight allocation coupled with environmental factors. The edge coarse calculation module is used to perform real-time anomaly detection at the edge using a lightweight compressed sensing model based on multi-dimensional feature vectors. When multi-period features exceed the standard and at least two modalities are jointly abnormal, the abnormal data fragment is uploaded to the cloud. The cloud-based fine processing module is used to perform cross-sensor bidirectional reasoning on abnormal data fragments through a spatiotemporal graph attention reasoning model deployed in the cloud. It verifies and reconstructs the sensor topology based on physical adjacency connections and dynamic statistical correlations, and conditionally triggers elastic incremental learning. Combined with degradation trajectory temporal evolution analysis, multi-evidence hierarchical fusion, and spatial registration, it outputs component-level remaining lifetime prediction results and fault heatmaps.

9. A predictive maintenance device for aeration blowers based on multimodal perception and adaptive learning, characterized in that, include: At least one processor; and memory that is communicatively connected to at least one processor; The memory stores instructions that can be executed by at least one processor, which enables the at least one processor to perform the predictive maintenance method for aeration blowers based on multimodal perception and adaptive learning as described in any one of claims 1-7.

10. A computer-readable storage medium storing computer-executable instructions thereon, characterized in that, When the executable instructions are executed by the processor, they implement the predictive maintenance method for aeration blowers based on multimodal perception and adaptive learning as described in any one of claims 1-7.

Images