Multi-source data fusion key component fault prediction method and system

Abstract

The invention provides a key component fault prediction method and system based on multi-source data fusion, and the method comprises the steps: S1, collecting key component data of a screen scarifier, constructing a multi-source heterogeneous sensor network, and completing the time-space alignment and quality optimization of multi-dimensional monitoring data; s2, a dynamic topological graph structure is constructed based on the physical connection relation of the components, the fault propagation path and strength are quantified, and multi-level feature representation covering the local state and the overall health degree is formed; s3, designing a hybrid prediction model fusing space-time analysis and memory modeling, and completing accurate description of an equipment degradation trend and early warning of potential faults through a self-adaptive feature integration mechanism; and S4, combining real-time prediction errors and historical operation and maintenance knowledge to dynamically optimize the hybrid prediction model, establishing a data-driven and knowledge-guided dual learning framework, and completing self-adaptive continuous learning along with equipment aging. The method breaks through the limitation of static state of a traditional prediction model, and the accuracy and reliability of a prediction result are remarkably improved.

Description

Technical Field

[0001] The present invention relates to the field of mechanical equipment condition monitoring and maintenance management, and more specifically, to a key component failure prediction method and system based on multi-source data fusion. Background Art

[0002] Currently, fault monitoring of key components of screening machines relies primarily on regular manual inspections and simple threshold alarm systems. The industry generally adopts a simple diagnostic method based on exceeding the limit of vibration amplitude, which determines the status of the equipment by setting fixed thresholds. Some advanced equipment has begun to introduce multi-parameter monitoring such as temperature and pressure, but each parameter is analyzed independently and lacks a comprehensive evaluation mechanism. In terms of prediction technology, it mainly relies on the empirical life curves provided by equipment manufacturers, or simple linear prediction models based on operating time. In recent years, some studies have attempted to apply machine learning algorithms, but due to the harsh working environment and complex working conditions of screening machines, the actual application effect is limited.

[0003] Existing technologies have the following major shortcomings: First, fixed-threshold alarm systems have a high false alarm rate and are unable to identify potential faults at an early stage. Second, there is a lack of effective fusion methods for multi-parameter monitoring data, and the correlations between parameters are not fully utilized. Third, empirical life curves cannot adapt to individual differences and actual operating conditions, resulting in large prediction errors. Finally, existing prediction methods lack self-learning capabilities and are unable to continuously optimize prediction models based on historical data. These issues result in low warning accuracy and poor practicality in actual projects, making it difficult to provide reliable decision support for equipment maintenance. Summary of the Invention

[0004] In response to the technical problems existing in the prior art, the present invention provides a key component fault prediction method and system based on multi-source data fusion. This method significantly improves the accuracy and reliability of the prediction results. The system breaks through the limitations of the static nature of traditional prediction models and achieves continuous improvement in prediction capabilities over time.

[0005] According to a first aspect of the present invention, a method for predicting key component failures by multi-source data fusion is provided, comprising the following steps: S1. Use multi-source sensors to collect data on key components of the screen cleaning machine, build a multi-source heterogeneous sensor network, construct an analysis data set through preprocessing, and complete the spatiotemporal alignment and quality optimization of multi-dimensional monitoring data through signal processing; S2. Build a dynamic topology graph based on the physical connection relationships of components, apply graph neural networks to quantify fault propagation paths and intensity, form a multi-level feature representation covering local status and overall health, and dynamically adjust the fusion weights. S3. Design a hybrid prediction model that integrates spatiotemporal analysis and memory modeling. Through an adaptive feature integration mechanism, it accurately characterizes equipment degradation trends and provides early warning of potential failures. S4. Dynamically optimize the hybrid prediction model by combining real-time prediction errors with historical operation and maintenance knowledge, establish a data-driven and knowledge-guided self-learning system, and complete the continuous optimization of the hybrid prediction model.

[0006] On the basis of the above technical solution, the present invention can also make the following improvements.

[0007] Optionally, the key components of the screening machine include a vibrating screen box, a vibrator, and a conveyor belt drive device; the data of the key components of the screening machine collected using multi-source sensors include: vibration, temperature, and oil data; the preprocessing includes: data cleaning, denoising, and format standardization processing.

[0008] Optionally, the signal processing includes: variational mode decomposition denoising and wavelet transform; completing the spatiotemporal alignment of multi-dimensional monitoring data through signal processing includes: Time synchronization: Using a precise clock protocol, the sampling timestamps of vibration, temperature, and oil sensors are unified to eliminate data time misalignment caused by hardware delays or sampling rate differences. Spatial alignment: Based on the physical location of components, a spatial mapping relationship of sensor data is established to ensure that fault signals can be associated with specific components.

[0009] Optionally, quality optimization of multi-dimensional monitoring data can be achieved through signal processing, including: Condition-adaptive denoising: For vibration signals, variational mode decomposition is used to dynamically adjust modal parameters and separate noise from fault characteristics. For temperature data, sliding window Z-score normalization is used to eliminate environmental interference. Multimodal feature enhancement: Combining time domain, frequency domain, and time-frequency domain features, oil particle data is modeled through Weibull distribution to improve fault sensitivity.

[0010] Optionally, the construction of a dynamic topology graph structure based on the physical connection relationship of components, the use of a graph neural network to quantify the fault propagation path and intensity, the formation of a multi-level feature representation covering local status and overall health, and the dynamic adjustment of fusion weights include: Build a topological graph structure based on the physical connection relationship of the equipment, convert key components into graph nodes, and convert the power transmission paths between components into directed connection edges; Design a multi-dimensional node feature encoding mechanism, where each node feature contains time-frequency domain features, operating condition parameters, and spatial location information data; A multi-head attention mechanism is used to implement feature interaction. Multiple independent attention calculation units are designed, each focusing on a different fault propagation pattern. Each unit calculates the attention weights between nodes using trainable parameters to dynamically capture key fault paths. Fault propagation feature extraction uses a multi-layer processing architecture. High-order node features are first extracted through a three-layer graph convolutional network. Global, path, and timing features are then calculated separately. Global features capture the overall system state; path features analyze historical fault propagation patterns; and timing features track the degradation trajectory of target components. The combination of these three types of features forms a complete representation of the system state. An intelligent feature fusion mechanism is designed to automatically evaluate the contribution of local features and system features through a gated neural network and dynamically adjust the fusion weights.

[0011] Optionally, the design of a hybrid prediction architecture integrating spatiotemporal analysis and memory modeling includes: The dynamic adjacency matrix is ​​used to quantify the fault propagation intensity between components in real time, capturing the spatial diffusion law of faults in the equipment system; an attenuation factor is introduced to strengthen the memory weight of recent fault features; and a triple attention mechanism of nodes, time, and mode is used to adaptively fuse multi-dimensional features. When the fused abnormality probability exceeds the dynamic threshold, an early warning is triggered.

[0012] Optionally, the method of dynamically optimizing a hybrid prediction model by combining real-time prediction errors with historical operation and maintenance knowledge to establish a data-driven and knowledge-guided self-learning system includes: In terms of data-driven development, a dynamic correction mechanism is established to automatically identify changes in operating conditions, new faults, and model degradation, and trigger adaptive optimization strategies for feature weight adjustment, small sample learning, and knowledge distillation. In terms of knowledge guidance, a structured fault knowledge graph was constructed to transform expert experience and maintenance cases into computable and reasonable digital knowledge.

[0013] Optionally, the dynamic optimization of the hybrid prediction model by combining the real-time prediction error with historical operation and maintenance knowledge includes: By monitoring prediction errors in real time, using gradient analysis to dynamically evaluate feature importance, and leveraging small sample learning to quickly adapt to new fault modes, the knowledge graph stores the relationship between fault features and solutions through a semantic network, and uses graph neural networks to achieve knowledge retrieval and reasoning.

[0014] According to a second aspect of the present invention, a key component failure prediction system based on multi-source data fusion is provided, comprising: The data acquisition and processing module is used to collect data from key components of the screen cleaning machine using multi-source sensors, build a multi-source heterogeneous sensor network, construct an analysis data set through preprocessing, and complete the spatiotemporal alignment and quality optimization of multi-dimensional monitoring data through signal processing; A multi-level feature representation module is used to construct a dynamic topology structure based on the physical connection relationships of components. It uses a graph neural network to quantify the fault propagation path and intensity, forming a multi-level feature representation covering local status and overall health, and dynamically adjusts the fusion weights. The hybrid model intelligent prediction module is used to design a hybrid prediction model that integrates spatiotemporal analysis and memory modeling. Through an adaptive feature integration mechanism, it accurately characterizes equipment degradation trends and provides early warning of potential failures. The hybrid prediction model optimization module combines real-time prediction errors with historical operation and maintenance knowledge to dynamically optimize the hybrid prediction model, establishes a data-driven and knowledge-guided self-learning system, and completes the continuous optimization of the hybrid prediction model.

[0015] According to a third aspect of the present invention, an electronic device is provided, comprising a memory and a processor, wherein the processor is configured to implement a key component failure prediction method based on multi-source data fusion when executing a computer program stored in the memory.

[0016] Technical effects and advantages of the present invention: The present invention provides a key component fault prediction method and system based on multi-source data fusion. First, the precise collection and optimization of monitoring data of key components of the screening machine are realized through multi-source heterogeneous sensor networks and signal processing technology. Secondly, the spatiotemporal alignment algorithm is used to unify the multi-dimensional data of vibration, temperature, and oil to the same spatiotemporal benchmark, effectively solving the signal interference problem under complex working conditions. Then, based on the dynamic topological graph structure and graph learning technology, the fault correlation characteristics between components can be accurately captured, and multi-level evaluation from local state to overall health can be realized. Finally, the hybrid prediction model integrating spatiotemporal analysis and memory modeling improves the accuracy of depicting equipment degradation trends and early fault warning capabilities through adaptive feature integration. The dual learning framework of data-driven and knowledge-guided enables the system to have autonomous evolution capabilities, dynamically adapt to equipment aging and changes in working conditions, and continuously optimizes the prediction model through continuous learning of operation and maintenance experience, providing reliable technical support for equipment health management.

[0017] Furthermore, this invention constructs a closed-loop optimization system with self-learning capabilities, enabling the continuous evolution of the prediction model. The system automatically optimizes model parameters by comparing predicted results with actual fault records. Simultaneously, it extracts effective fault feature combinations from newly added repair cases, continuously enriching the knowledge base. This dual optimization mechanism enables the system to adapt to feature changes during equipment aging and promptly identify new fault modes. This system overcomes the static limitations of traditional prediction models, achieving continuous improvement in predictive capabilities over time. BRIEF DESCRIPTION OF THE DRAWINGS

[0018] Figure 1 A flowchart of a key component failure prediction method based on multi-source data fusion provided by an embodiment of the present invention. DETAILED DESCRIPTION

[0019] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of the present invention.

[0020] It is understandable that based on the defects in the background technology, the embodiment of the present invention proposes a key component failure prediction method based on multi-source data fusion, specifically Figure 1 As shown, the method includes the following steps: S1. Use multi-source sensors to collect data on key components of the screen cleaning machine, build a multi-source heterogeneous sensor network, construct an analysis data set through preprocessing, and complete the spatiotemporal alignment and quality optimization of multi-dimensional monitoring data through signal processing; To meet the need for fault prediction of key components of a screen cleaning machine (vibrating screen box, vibrator, and conveyor belt drive device), an embodiment of the present invention constructs a multi-source heterogeneous sensing network for collaborative monitoring of multiple physical quantities.

[0021] Furthermore, the multi-source heterogeneous sensor network was constructed. Specifically, for vibration monitoring, triaxial ICP accelerometers (frequency response 0.5-10kHz, range ±50g) were installed on the screen box side panels, and acoustic emission sensors (resonant frequency 150kHz) were placed on the vibrator bearing seats. Temperature monitoring utilized PT100 embedded thermocouples (vibrator lubrication points) and infrared thermal imagers (conveyor motor windings). Oil condition was monitored in real time using an online particle counter (ISO4406 standard) and a moisture sensor. Furthermore, operating parameters such as hydraulic system pressure (4-20mA pressure transmitter) and vibrating screen inclination (MEMS gyroscope) were collected. All sensor signals were transmitted to edge computing nodes via an industrial CAN bus (ISO11898-2), with μs-level time synchronization achieved using the IEEE1588v2 protocol.

[0022] By constructing a multi-source heterogeneous sensor network to collect screen cleaning machine operating status data and historical maintenance records, and aligning the format of the collected screen cleaning machine operating status data through preprocessing, an analysis data set is constructed. The operating status data includes sensor signals for vibration, temperature, and pressure; the preprocessing includes data cleaning, denoising, and format standardization. Preprocessing first uses the DBSCAN clustering algorithm to automatically identify the equipment's different operating modes (such as high speed and low speed) based on the screen body inclination angle and hydraulic pressure operating condition parameters, providing a basis for operating condition classification for subsequent signal processing. The VMD modal decomposition number of the vibration signal and the wavelet transform parameter processing method of the acoustic emission signal are then dynamically adjusted for different operating conditions. Temperature data is then standardized and the distribution of oil particles is modeled uniformly. Ultimately, a 128-dimensional time domain, frequency domain, and time-frequency domain feature matrix matching the operating conditions is generated, achieving high-quality preprocessing of multi-source monitoring data.

[0023] Specifically, the construction of the analysis data set through preprocessing specifically includes: First, the operating condition is identified. The operating condition feature vector is established based on the screen body inclination angle and hydraulic pressure, and typical operating modes are divided through DBSCAN clustering. The operating condition-adaptive VMD denoising algorithm is used for the vibration signal, and the modal decomposition number K = round(log2(fs / f_gear mesh)) is dynamically determined. The IMF components containing fault characteristics are selected to reconstruct the signal. The acoustic emission signal is extracted through Morlet wavelet transform to extract the impact characteristics. The temperature data is normalized using a sliding window Z-score to eliminate the influence of ambient temperature. The oil particle data is fitted with Weibull distribution to establish a contamination degradation model. Finally, a time-aligned multi-dimensional feature matrix is ​​generated, which includes 128-dimensional features in the time domain (peak factor, pulse index), frequency domain (gear meshing sideband energy ratio), and time-frequency domain (wavelet packet node entropy value), providing high-quality input for subsequent intelligent diagnosis.

[0024] Furthermore, the embodiment of the present invention also realizes the spatiotemporal alignment and quality optimization of multi-dimensional monitoring data of vibration, temperature, and oil through signal processing methods, effectively overcoming the signal interference problem under complex working conditions.

[0025] Specifically, the spatiotemporal alignment of multi-dimensional monitoring data is accomplished through signal processing, including: Time synchronization: Using the IEEE 1588v2 precision clock protocol, the sampling timestamps of vibration, temperature, and oil sensors are unified to eliminate data time misalignment caused by hardware delays or sampling rate differences. Spatial alignment: Based on the physical location of components (e.g., vibration sensors mounted on the sidewall of a gearbox, temperature sensors embedded in a bearing seat), a spatial mapping relationship of sensor data is established to ensure that fault signals can be associated with specific components.

[0026] The quality optimization of multi-dimensional monitoring data through signal processing includes: Condition-adaptive denoising: For vibration signals, variational mode decomposition (VMD) is used to dynamically adjust modal parameters and separate noise from fault characteristics. For temperature data, sliding window Z-score normalization is used to eliminate environmental interference. Multimodal feature enhancement: Combines time domain (such as peak factor), frequency domain (such as gear meshing sideband energy), and time-frequency domain (wavelet packet entropy) features, and uses Weibull distribution to model oil particle data to improve fault sensitivity.

[0027] The aforementioned spatiotemporal alignment and signal optimization address the misjudgments caused by data asynchrony or noise in traditional methods, significantly improving the detection rate of early fault characteristics (such as bearing microcracks and initial gear wear). This solution addresses the signal modulation issues associated with variable speed and high-impact operating conditions in screen cleaning machines, significantly improving the early fault identification rate through spatiotemporal alignment of multi-source data.

[0028] S2. Build a dynamic topology model based on the physical connection relationships of components, apply graph neural networks to quantify fault propagation paths and intensity, form a multi-level feature representation covering local status and overall health, and dynamically adjust the fusion weights. The embodiment of the present invention adopts a graph neural network architecture to convert the physical connection relationship between the key components of the screening machine into a topological graph structure, realizing system-level modeling of fault propagation characteristics.

[0029] Building a dynamic topology model based on the physical connection relationship of components includes: 2.1. Build a topological graph based on the physical connections between the devices. Key components of the vibrating screen box and exciter are converted into graph nodes, and the power transmission paths between components are converted into directed edges. A weight matrix quantifies the connection strength to reflect the fault propagation capacity between different components. Steel components exhibit higher connection strength than rubber damping components.

[0030] This topology supports dynamic adjustment and can adapt to equipment modification needs. The topology structure model is shown as follows:

[0031] : A node set representing each monitoring component (such as vibrating screen, vibrator, etc.).

[0032] : Edge set, representing the power transmission path between components.

[0033] : weight matrix, elements in: : The physical distance between components i and j.

[0034] : Material conductivity coefficient.

[0035] : A very small constant to prevent the denominator from being zero.

[0036] 2.2. Design a multidimensional node feature encoding mechanism; each node feature contains three types of data: time-frequency domain features, operating parameters, and spatial location information. Time-frequency domain features include 60 indicators such as the vibration signal's peak factor and envelope entropy; operating parameters include speed, load, and temperature; and spatial location records the installation coordinates. Standardization and feature selection algorithms ensure the uniformity and validity of input features.

[0037] Each node The eigenvectors of:

[0038]

[0039]

[0040]

[0041]

[0042] 2.3. A multi-head attention mechanism is used to implement feature interaction. Four independent attention calculation units are designed, each focusing on a different fault propagation pattern. Each unit calculates the attention weights between nodes using trainable parameters, dynamically capturing key fault paths. When node features are updated, the combined results of the four outputs are retained, enabling the model to simultaneously identify local anomalies and global propagation patterns. The multi-head attention mechanism is represented as follows: Multi-head attention mechanism (K=4 heads) updates node features:

[0043] The attention coefficient is calculated as:

[0044] : learnable parameters; The characteristic dimensions of each head; : Activation function (such as ReLU).

[0045] : The neighbor set of node i.

[0046] 2.4. Fault propagation feature extraction utilizes a multi-layer processing architecture. High-order node features are first extracted through a three-layer graph convolutional network. Global, path, and time series features are then calculated. Global features capture the overall system state; path features analyze historical fault propagation patterns; and time series features track the degradation trajectory of target components. The combination of these three features forms a complete representation of the system state.

[0047] After convolution of the work layer graph, the system-level fault characteristics are:

[0048] READOUT function includes: global maximum pooling .

[0049] Fault propagation path characteristics:

[0050] Where P represents the three most frequent propagation paths in historical faults.

[0051] 2.5. Design an intelligent feature fusion mechanism; automatically evaluate the contribution of local features and system features through a gated neural network, and dynamically adjust the fusion weights.

[0052] Final eigenvector:

[0053] :

[0054] This mechanism is based on dynamic topological graph structure and graph learning technology, which can accurately capture the fault correlation characteristics between components, realize multi-level evaluation from local status to overall health, and enable the model to flexibly adjust feature utilization strategies according to specific circumstances.

[0055] S3. Design a hybrid prediction model that integrates spatiotemporal analysis and memory modeling. Through an adaptive feature integration mechanism, it accurately characterizes equipment degradation trends and provides early warning of potential failures. In this embodiment, the hybrid prediction model achieves early warning through dual modeling of spatiotemporal graph convolutional network (ST-GCN) and time-aware LSTM. The specific process is as follows: First, the spatiotemporal graph convolution module uses a dynamic adjacency matrix to quantify the fault propagation intensity between components in real time (such as the impact of bearing abnormalities on gearboxes), capturing the spatial diffusion pattern of faults in the equipment system. At the same time, the time-aware LSTM introduces a decay factor to strengthen the memory weight of recent fault features, thereby accurately characterizing the degradation trend; Finally, a triple attention mechanism (node ​​importance, time importance, and modal importance) adaptively fuses multi-dimensional features. When the fused anomaly probability exceeds a dynamic threshold (e.g., vibration, temperature, and oil levels all exceeding limits), an early warning is triggered. For example, initial bearing wear may manifest as subtle anomalies in the vibration frequency sideband energy and temperature gradient. However, through spatiotemporal correlation analysis and historical degradation pattern matching, the system can issue an early warning before the fault develops to a significant stage (e.g., wear <5%).

[0056] The hybrid prediction model that integrates spatiotemporal analysis and memory modeling in the embodiment of the present invention adopts a three-stage architecture of "spatial-temporal graph convolution-memory enhancement-adaptive aggregation", which is specifically implemented as follows: Spatiotemporal graph convolution module: including: dynamic adjacency matrix , A t is the dynamic adjacency matrix, n is the number of components;

[0057] Expressed as the fault propagation delay from component i to j (estimated by cross-correlation analysis), Activation function; MLP stands for Multi-Layer Perceptron, with an input dimension of 2×128+1=2572×128+1=257, an output dimension of 1, and a hidden layer structure of 257→64→1.

[0058] Spatiotemporal graph convolution operation:

[0059] in, It is expressed as a dynamic adjacency matrix, where n is the number of components; Represented as the feature matrix of all components at time t; Represented as a learnable spatial mask matrix; Represented as spatial / temporal convolution weight matrix ; is represented as a bias term. Expressed as a Hadamard product (element-wise multiplication).

[0060] Memory enhancement module: including: improved Time-Aware LSTM unit:

[0061]

[0062] in, Represented as the output features of the spatiotemporal graph convolution module; Represented as the hidden state of the previous time step; Expressed as a bias term; Expressed as a weight matrix ( represents f, i, c, o); It is expressed as the time interval between the current time step and the last abnormal event.

[0063] : time decay factor, Product (element-wise multiplication) The adaptive aggregation module includes three attention mechanisms: Node importance weight:

[0064] Time importance weight:

[0065] Modal Importance Weight (Vibration / Temperature / Oil):

[0066] Where, Expressed as node importance weight (n is the number of components); Expressed as time importance weight (T=60 time steps); Expressed as modal weights (vibration, temperature, oil); Represented as a query vector; Represented as a feature transformation matrix; It is represented as the feature code of each modality.

[0067] Final prediction function:

[0068] Represented as the output layer weight matrix; Function, output failure probability .

[0069] The loss function design is expressed as: Compound loss:

[0070] Expressed as binary cross entropy loss (prediction vs true label); ; ENC / DEC represents encoder / decoder (hidden layer dimension 128→64→128).

[0071] : Contrast loss (temperature coefficient κ=0.5κ=0.5).

[0072] 5. Online update mechanism includes: Model parameter update:

[0073] in, Expressed as learning rate; Expressed as a diagonalized empirical Fisher matrix; Represented as a gradient vector; is expressed as a numerical stability term.

[0074] S4. Dynamically optimize the hybrid prediction model by combining real-time prediction errors with historical operation and maintenance knowledge, establish a data-driven and knowledge-guided self-learning system, and complete the continuous optimization of the hybrid prediction model.

[0075] The dual learning framework proposed in this embodiment of the present invention achieves continuous evolution of the prediction system through two channels: data-driven and knowledge-guided. On the data-driven side, the system establishes a dynamic correction mechanism that automatically identifies three types of issues: operating condition changes, new faults, and model degradation, and triggers adaptive optimization strategies such as feature weight adjustment, small-sample learning, and knowledge distillation. On the knowledge-guided side, a structured fault knowledge graph is constructed to transform expert experience and maintenance cases into computable and reasonable digital knowledge. Specifically, through real-time monitoring of prediction errors, gradient analysis techniques are used to dynamically assess feature importance, and small-sample learning is used to rapidly adapt to new fault modes. Simultaneously, the knowledge graph stores the associations between fault features and solutions through a semantic network, and uses graph neural networks to enable knowledge retrieval and reasoning. These two channels collaborate deeply through three interfaces: feature extraction, decision fusion, and feedback learning, adapting to changes in data distribution while also incorporating expert experience. This architectural design possesses autonomous evolutionary capabilities, continuously optimizing the prediction model through the continuous integration of real-time monitoring data and operational knowledge.

[0076] This embodiment of the present invention builds an intelligent self-learning system with dual evolutionary capabilities, achieving continuous optimization of the prediction model through a closed-loop mechanism of "dynamic feedback and knowledge accumulation." This dual learning framework, driven by data and guided by knowledge, enables the system to evolve autonomously, dynamically adapting to equipment aging and changing operating conditions. By continuously learning from operational experience, the prediction model is continuously optimized, providing reliable technical support for equipment health management.

[0077] According to a second aspect of the present invention, a key component failure prediction system based on multi-source data fusion is provided, comprising: The data acquisition and processing module is used to collect data from key components of the screen cleaning machine using multi-source sensors, build a multi-source heterogeneous sensor network, construct an analysis data set through preprocessing, and complete the spatiotemporal alignment and quality optimization of multi-dimensional monitoring data through signal processing; A multi-level feature representation module is used to construct a dynamic topology structure based on the physical connection relationships of components. It uses a graph neural network to quantify the fault propagation path and intensity, forming a multi-level feature representation covering local status and overall health, and dynamically adjusts the fusion weights. The hybrid model intelligent prediction module is used to design a hybrid prediction model that integrates spatiotemporal analysis and memory modeling. Through an adaptive feature integration mechanism, it accurately characterizes equipment degradation trends and provides early warning of potential failures. The hybrid prediction model optimization module combines real-time prediction errors with historical operation and maintenance knowledge to dynamically optimize the hybrid prediction model, establishes a data-driven and knowledge-guided self-learning system, and completes the continuous optimization of the hybrid prediction model.

[0078] It can be understood that the key component failure prediction system of multi-source data fusion provided by the present invention corresponds to the key component failure prediction system method of multi-source data fusion provided by the aforementioned embodiments. The relevant technical features of the key component failure prediction system of multi-source data fusion can refer to the relevant technical features of the key component failure prediction system method of multi-source data fusion, which will not be repeated here.

[0079] Additionally, an embodiment of the present invention provides an electronic device that may include a processor, a communications interface, a memory, and a communications bus, wherein the processor, the communications interface, and the memory communicate with each other via the communications bus. The processor may invoke logic instructions in the memory to execute steps of a method for predicting key component failures using multi-source data fusion.

[0080] Furthermore, the logical instructions in the aforementioned memory can be implemented as software functional units and, when sold or used as independent products, can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention, or the portion that contributes to the prior art, or a portion of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions for enabling a computer device (which can be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in various embodiments of the present invention. The aforementioned storage media include various media capable of storing program code, such as USB flash drives, mobile hard drives, read-only memories (ROMs), random access memories (RAMs), magnetic disks, or optical disks.

[0081] Through the above description of the embodiments, those skilled in the art will clearly understand that each embodiment can be implemented using software plus a necessary general-purpose hardware platform, or of course, hardware. Based on this understanding, the essence of the above technical solution, or the portion that contributes to the prior art, can be embodied in the form of a software product. This computer software product can be stored in a computer-readable storage medium, such as ROM / RAM, a magnetic disk, or an optical disk, and includes a number of instructions for causing a computer device (such as a personal computer, server, or network device) to execute the methods described in each embodiment or certain portions of the embodiments.

[0082] Finally, it should be noted that the above is only a preferred embodiment of the present invention and is not intended to limit the present invention. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art can still modify the technical solutions described in the aforementioned embodiments or make equivalent substitutions for some of the technical features therein. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principles of the present invention should be included in the scope of protection of the present invention.

Claims

1. A key component failure prediction method based on multi-source data fusion, characterized in that: The following steps are involved: S1. Use multi-source sensors to collect data on key components of the screen cleaning machine, build a multi-source heterogeneous sensor network, construct an analysis data set through preprocessing, and complete the spatiotemporal alignment and quality optimization of multi-dimensional monitoring data through signal processing; S2. Build a dynamic topology graph based on the physical connection relationships of components, apply graph neural networks to quantify fault propagation paths and intensity, form a multi-level feature representation covering local status and overall health, and dynamically adjust the fusion weights. S3. Design a hybrid prediction model that integrates spatiotemporal analysis and memory modeling. Through an adaptive feature integration mechanism, it accurately characterizes equipment degradation trends and provides early warning of potential failures. S4. Dynamically optimize the hybrid prediction model by combining real-time prediction errors with historical operation and maintenance knowledge, establish a data-driven and knowledge-guided self-learning system, and complete the continuous optimization of the hybrid prediction model.

2. The key component failure prediction method based on multi-source data fusion according to claim 1 is characterized in that: The key components of the screen cleaning machine include a vibrating screen box, a vibrator, and a conveyor belt drive device; The data of key components of the cleaning machine collected by multi-source sensors include vibration, temperature and oil data; the preprocessing includes data cleaning, denoising and format standardization.

3. The key component failure prediction method based on multi-source data fusion according to claim 1 is characterized in that: The signal processing includes: variational mode decomposition denoising and wavelet transform; completing the spatiotemporal alignment of multi-dimensional monitoring data through signal processing includes: Time synchronization: Using a precise clock protocol, the sampling timestamps of vibration, temperature, and oil sensors are unified to eliminate data time misalignment caused by hardware delays or sampling rate differences. Spatial alignment: Based on the physical location of components, a spatial mapping relationship of sensor data is established to ensure that fault signals can be associated with specific components.

4. The key component failure prediction method based on multi-source data fusion according to claim 1 is characterized in that: The quality optimization of multi-dimensional monitoring data through signal processing includes: Condition-adaptive denoising: For vibration signals, variational mode decomposition is used to dynamically adjust modal parameters and separate noise and fault characteristics. For temperature data, environmental interference is eliminated by sliding window Z-score normalization; Multimodal feature enhancement: Combining time domain, frequency domain, and time-frequency domain features, oil particle data is modeled through Weibull distribution to improve fault sensitivity.

5. The key component failure prediction method based on multi-source data fusion according to claim 1 is characterized in that: The dynamic topology structure is constructed based on the physical connection relationship of components, and the graph neural network is used to quantify the fault propagation path and intensity, forming a multi-level feature representation covering local status and overall health, and dynamically adjusting the fusion weights. Build a topological graph structure based on the physical connection relationship of the equipment, convert key components into graph nodes, and convert the power transmission paths between components into directed connection edges; Design a multi-dimensional node feature encoding mechanism, where each node feature contains time-frequency domain features, operating condition parameters, and spatial location information data; A multi-head attention mechanism is used to implement feature interaction. Multiple independent attention calculation units are designed, each focusing on a different fault propagation pattern. Each unit calculates the attention weights between nodes using trainable parameters to dynamically capture key fault paths. Fault propagation feature extraction uses a multi-layer processing architecture. High-order node features are first extracted through a three-layer graph convolutional network. Global, path, and time series features are then calculated. Global features are used to capture the overall system state, path features are used to analyze historical fault propagation patterns, and time series features are used to track the degradation trajectory of target components. These three types of features are combined to form a complete representation of the system state. An intelligent feature fusion mechanism is designed to automatically evaluate the contribution of local features and system features through a gated neural network and dynamically adjust the fusion weights.

6. The key component failure prediction method based on multi-source data fusion according to claim 1 is characterized in that: The hybrid prediction architecture designed to integrate spatiotemporal analysis and memory modeling includes: The dynamic adjacency matrix is ​​used to quantify the fault propagation intensity between components in real time, capturing the spatial diffusion law of faults in the equipment system; an attenuation factor is introduced to strengthen the memory weight of recent fault features; and a triple attention mechanism of nodes, time, and mode is used to adaptively fuse multi-dimensional features. When the fused abnormality probability exceeds the dynamic threshold, an early warning is triggered.

7. The key component failure prediction method based on multi-source data fusion according to claim 1 is characterized in that: The method of dynamically optimizing the hybrid prediction model by combining real-time prediction errors with historical operation and maintenance knowledge to establish a data-driven and knowledge-guided self-learning system includes: In terms of data-driven development, a dynamic correction mechanism is established to automatically identify changes in operating conditions, new faults, and model degradation, and trigger adaptive optimization strategies for feature weight adjustment, small sample learning, and knowledge distillation. In terms of knowledge guidance, a structured fault knowledge graph was constructed to transform expert experience and maintenance cases into computable and reasonable digital knowledge.

8. The key component failure prediction method based on multi-source data fusion according to claim 1 is characterized in that: The dynamic optimization hybrid prediction model combining real-time prediction error and historical operation and maintenance knowledge includes: By monitoring prediction errors in real time, using gradient analysis to dynamically evaluate feature importance, and leveraging small sample learning to quickly adapt to new fault modes, the knowledge graph stores the relationship between fault features and solutions through a semantic network, and uses graph neural networks to achieve knowledge retrieval and reasoning.

9. A key component failure prediction system based on multi-source data fusion, characterized in that: include: The data acquisition and processing module is used to collect data from key components of the screen cleaning machine using multi-source sensors, build a multi-source heterogeneous sensor network, construct an analysis data set through preprocessing, and complete the spatiotemporal alignment and quality optimization of multi-dimensional monitoring data through signal processing; A multi-level feature representation module is used to construct a dynamic topology structure based on the physical connection relationships of components. It uses a graph neural network to quantify the fault propagation path and intensity, forming a multi-level feature representation covering local status and overall health, and dynamically adjusts the fusion weights. The hybrid model intelligent prediction module is used to design a hybrid prediction model that integrates spatiotemporal analysis and memory modeling. Through an adaptive feature integration mechanism, it accurately characterizes equipment degradation trends and provides early warning of potential failures. The hybrid prediction model optimization module combines real-time prediction errors with historical operation and maintenance knowledge to dynamically optimize the hybrid prediction model, establishes a data-driven and knowledge-guided self-learning system, and completes the continuous optimization of the hybrid prediction model.

10. An electronic device, characterized in that: The method comprises a memory and a processor, wherein the processor is configured to implement a key component failure prediction method based on multi-source data fusion as claimed in any one of claims 1 to 8 when executing a computer program stored in the memory.

Images