Equipment fault prediction method and system based on deep learning

Abstract

The invention provides an equipment fault prediction method and system based on deep learning, and relates to the technical field of computers, and the method comprises the steps: obtaining first information, second information and third information; extracting historical dynamic operation characteristics of the equipment according to the third information to obtain an operation state characteristic matrix; performing graph construction processing according to the second information, and respectively constructing to obtain a structure graph, a wiring characteristic graph and a control logic graph; according to the operation state characteristic matrix, the structure diagram, the wiring characteristic diagram and the control logic diagram, performing fusion processing to obtain a comprehensive diagram structure; according to the comprehensive graph structure, using a deep learning algorithm to construct and obtain a fault prediction model; and inputting the first information into the fault prediction model to obtain a prediction result. According to the method, the time domain, frequency domain and time-frequency domain characteristics are extracted from the time sequence of the historical operation data in a segmented manner, a multi-dimensional comprehensive graph structure is constructed in combination with the equipment structure, the signal wiring characteristics and the control logic, and the internal characteristics of the equipment and the incidence relation thereof are fully excavated.

Description

Technical Field

[0001] The present invention relates to the field of computer technology, and in particular, to a device fault prediction method and system based on deep learning. Background Art

[0002] With the wide application of industrial equipment, the complexity of its operating state and the variability of the working environment make the fault prediction of equipment a key link to ensure the reliability and stability of the equipment. At present, equipment maintenance mainly relies on two technical routes: regular maintenance and experience-based passive maintenance. Regular maintenance is a time-driven maintenance method, that is, the equipment is inspected and maintained at a predetermined time interval, but this method ignores the actual operating state of the equipment and may lead to problems of over-maintenance or under-maintenance. And experience-based passive maintenance usually diagnoses and repairs faults depending on the experience of engineering and technical personnel after the equipment shows obvious faults or even stops running. This way not only increases the fault downtime of the equipment, but also may cause potential faults not to be discovered in time, thus leading to more serious equipment damage.

[0003] The prior art has proposed some statistical analysis methods based on a single operating state parameter to improve the fault prediction ability of equipment. For example, the abnormality of mechanical components is detected by monitoring the change of vibration signals, or the poor heat dissipation condition is analyzed through temperature signals. However, these methods usually only consider a single type of operating data, ignoring the complex correlation between the internal structure, wiring characteristics and control logic of the equipment, and cannot accurately describe the overall health state of the equipment. In addition, the prior art lacks effective multi-modal data fusion means and is difficult to fully explore the potential relationship between multi-dimensional data, resulting in insufficient accuracy and comprehensiveness of fault prediction.

[0004] Based on the above disadvantages of the prior art, there is an urgent need for a device fault prediction method and system based on deep learning. Summary of the Invention

[0005] The purpose of the present invention is to provide a device fault prediction method and system based on deep learning to improve the above problems. To achieve the above purpose, the technical solutions adopted by the present invention are as follows:

[0006] In the first aspect, the present application provides a device fault prediction method based on deep learning, including:

[0007] Obtain first information, second information and third information, where the first information is the operation state data collected in real time, the second information includes the structure data, wiring data and control logic data of the device, and the third information is the historical data of the device, and the historical data includes historical operation state data, historical fault records, maintenance records and working conditions;

[0008] Extract the historical dynamic operation characteristics of the device according to the third information, extract the time-domain, frequency-domain, and time-frequency domain features through time series segmentation, and normalize the features to obtain the operation state feature matrix;

[0009] Perform graph construction processing according to the second information, and respectively construct a structure diagram, a wiring characteristic diagram, and a control logic diagram;

[0010] Perform fusion processing according to the operation state feature matrix, the structure diagram, the wiring characteristic diagram, and the control logic diagram. By mapping the matrix elements to the nodes in the graph respectively, and combining the physical connection relationship, signal transmission characteristics, and control logic dependency relationship between components, obtain the integrated graph structure;

[0011] Construct a fault prediction model according to the integrated graph structure using a deep learning algorithm;

[0012] Input the first information into the fault prediction model to perform fault prediction on the current operation state of the device to obtain a prediction result, where the prediction result includes the occurrence probability of device faults, the specific fault location nodes, and the change trend of the device health state.

[0013] In a second aspect, the present application also provides a device fault prediction system based on deep learning, including:

[0014] An acquisition module for acquiring the first information, the second information, and the third information. The first information is the operation state data collected in real time, the second information includes the structure data, wiring data, and control logic data of the device, and the third information is the historical data of the device, and the historical data includes historical operation state data, historical fault records, maintenance records, and working conditions;

[0015] An extraction module for extracting the historical dynamic operation characteristics of the device according to the third information, extracting the time-domain, frequency-domain, and time-frequency domain features through time series segmentation, and normalizing the features to obtain the operation state feature matrix;

[0016] A construction module for performing graph construction processing according to the second information, and respectively constructing a structure diagram, a wiring characteristic diagram, and a control logic diagram;

[0017] A fusion module for performing fusion processing according to the operation state feature matrix, the structure diagram, the wiring characteristic diagram, and the control logic diagram. By mapping the matrix elements to the nodes in the graph respectively, and combining the physical connection relationship, signal transmission characteristics, and control logic dependency relationship between components, obtain the integrated graph structure;

[0018] A modeling module for constructing a fault prediction model according to the integrated graph structure using a deep learning algorithm;

[0019] A prediction module for inputting the first information into the fault prediction model to perform fault prediction on the current operating state of the device and obtain a prediction result, where the prediction result includes the occurrence probability of device faults, specific fault location nodes, and the change trend of the device health state.

[0020] The beneficial effects of the present invention are as follows:

[0021] By segmenting the time series of historical operation data to extract time-domain, frequency-domain, and time-frequency-domain features, and combining the device structure, signal wiring characteristics, and control logic, the present invention constructs a multi-dimensional comprehensive graph structure to fully explore the internal characteristics of the device and their correlation relationships; uses a deep learning algorithm to establish a fault prediction model with the comprehensive graph as the input, realizes accurate modeling of the device operating state, and significantly improves the accuracy and robustness of device fault prediction. Description of the Drawings

[0022] To more clearly illustrate the technical solutions of the embodiments of the present invention, the following will briefly introduce the drawings required in the embodiments. It should be understood that the following drawings only show some embodiments of the present invention and should not be regarded as limiting the scope. For those of ordinary skill in the art, other related drawings can be obtained based on these drawings without creative efforts.

[0023] Figure 1 It is a schematic flowchart of a device fault prediction method based on deep learning described in the embodiments of the present invention;

[0024] Figure 2 It is a schematic structural diagram of a device fault prediction system based on deep learning described in the embodiments of the present invention;

[0025] Figure 3 It is a schematic structural diagram of a device fault prediction device based on deep learning described in the embodiments of the present invention.

[0026] Reference numerals in the figure: 800, a device fault prediction device based on deep learning; 801, a processor; 802, a memory; 803, a multimedia component; 804, an I / O interface; 805, a communication component; 901, an acquisition module; 902, an extraction module; 903, a construction module; 904, a fusion module; 905, a modeling module; 906, a prediction module. Detailed Embodiments

[0027] To make the objectives, technical solutions and advantages of the embodiments of the present invention clearer, the technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are some, but not all, of the embodiments of the present invention. Components of the embodiments of the present invention usually described and illustrated in the accompanying drawings herein can be arranged and designed in a variety of different configurations. Therefore, the detailed description of the embodiments of the present invention provided in the accompanying drawings is not intended to limit the scope of the claimed invention, but merely represents selected embodiments of the present invention. All other embodiments obtained by those of ordinary skill in the art based on the embodiments of the present invention without creative efforts shall fall within the scope of protection of the present invention.

[0028] It should be noted that similar reference numerals and letters denote similar items in the following drawings. Therefore, once an item is defined in one drawing, it does not need to be further defined and explained in subsequent drawings. At the same time, in the description of the present invention, the terms "first", "second", etc. are only used for descriptive distinction and cannot be construed as indicating or implying relative importance.

[0029] Embodiment 1:

[0030] This embodiment provides a device fault prediction method based on deep learning.

[0031] See Figure 1 , which shows that this method includes steps S100 to S600.

[0032] Step S100: Obtain the first information, the second information and the third information. The first information is the operation status data collected in real time. The second information includes the structure data, wiring data and control logic data of the device. The third information is the historical data of the device, and the historical data includes historical operation status data, historical fault records, maintenance records and working conditions;

[0033] It is understandable that the first information, namely the operation status data collected in real time, covers the key parameters (such as temperature, vibration, current, voltage, and rotational speed) during the operation of the device. These data directly reflect the current operation status of the device and are the necessary basis for real-time monitoring of the device's health status. The second information, including the structural data, wiring data, and control logic data of the device, is a comprehensive description of the physical and logical characteristics of the device, which can reveal the physical connection relationships between internal components of the device, the electrical signal transmission paths, and the logical dependency relationships of the control system, providing structural and logical support for subsequent graph construction and model training. The third information - historical data, including historical operation status data, historical fault records, maintenance records, and working conditions, can provide the model with the status characteristics and fault patterns during long-term operation. In particular, historical fault records and maintenance records can help the model extract patterns from known fault modes and predict abnormal situations. Different from the traditional method that only relies on real-time data or single-dimensional historical data, this step significantly improves the relevance and integrity of the data by combining operation data with the physical and logical structures of the device.

[0034] Step S200: Extract the historical dynamic operation characteristics of the device according to the third information, extract time-domain, frequency-domain, and time-frequency-domain characteristics through time series segmentation, and normalize the characteristics to obtain an operation status feature matrix;

[0035] It should be noted that during the time-domain feature extraction process, statistical features (such as mean, standard deviation, maximum value, and minimum value) are used to describe the overall fluctuation situation and trend of the device operation parameters. For example, the average value of the temperature can reflect the long-term heat dissipation performance of the device, and the standard deviation of the vibration signal can quantify the stability of mechanical operation. By extracting frequency-domain features, the energy distribution of the device signal at different frequencies can be captured, which is crucial for specific frequency harmonics (such as the resonance frequencies of mechanical bearings or gears) in the vibration signal. At the same time, the proportion of the energy of the frequency band with abnormal harmonic content can intuitively reflect the potential abnormalities of the device within certain frequency ranges. Time-frequency-domain feature extraction further combines the time and frequency information of the signal through wavelet transform or short-time Fourier transform to capture short-time abnormal signals or local fluctuation characteristics in the dynamic changes of the device. For example, wavelet transform can reveal the instantaneous change characteristics of high-frequency vibration energy during the startup, shutdown, or load change process of the device, and these information are particularly important in the early stage of faults. Since the numerical ranges and dimensions of different types of data are different, directly inputting them into the model may lead to unreasonable influences of specific features on the model output. Therefore, by normalizing various feature data to the same range, not only can the convergence speed of model training be improved, but also the adaptability of the model to different data sources can be enhanced.

[0036] Step S300: Perform graph construction processing based on the second information, and respectively construct a structure diagram, a wiring characteristic diagram, and a control logic diagram;

[0037] It can be understood that the structure diagram can completely describe the physical topology of the device, providing an accurate calculation basis for the physical propagation path of faults (such as vibration transmission through mechanical connections); the wiring characteristic diagram reflects the signal transmission process of the electrical system and can provide direct fault location support for electrical abnormalities (such as short circuits or overloads); the control logic diagram explicates the logical behavior of the device and the causal relationship of faults, thus supporting the fault propagation analysis of complex logic systems.

[0038] Step S400: Perform fusion processing based on the operation state feature matrix, the structure diagram, the wiring characteristic diagram, and the control logic diagram. By respectively mapping the matrix elements to the nodes in the graph and combining the physical connection relationship, signal transmission characteristics, and control logic dependency relationship between components, an integrated graph structure is obtained;

[0039] It should be noted that in this step, by integrating the operation state, physical connection, signal transmission, and logical dependency into a unified integrated graph structure, the operation state of the device and its internal relevance are completely described, overcoming the limitations of single data source analysis.

[0040] Step S500: Based on the integrated graph structure, use a deep learning algorithm to construct a fault prediction model;

[0041] It should be noted that in this step, the multi-dimensional information in the integrated graph is fully mined through the deep learning algorithm to achieve accurate prediction of device faults.

[0042] Step S600: Input the first information into the fault prediction model to perform fault prediction on the current operation state of the device to obtain a prediction result, where the prediction result includes the occurrence probability of device faults, the specific fault location nodes, and the change trend of the device health state.

[0043] It can be understood that in this step, a comprehensive analysis of the current operation state of the device is performed, and a prediction result with practical value is given.

[0044] Furthermore, step S200 includes steps S210 to S230.

[0045] Step S210: Perform time series segmentation processing based on the third information. Among them, the temperature data captures the overall slow-changing trend through a long time window, the vibration data captures the transient fluctuation characteristics through a short time window, and the current data reflects the load fluctuation situation through the standard deviation and change amplitude to obtain time-domain feature data;

[0046] It can be understood that this step adopts a differential segmentation strategy to capture the characteristics of various types of data at different time scales. Temperature data is segmented using a long-time window, which can reveal the long-term trend of the device's thermal performance. For example, problems such as poor heat dissipation or heat accumulation can be identified by calculating the average value and change rate within the window. Vibration data, due to its rich transient information and high-frequency characteristics, is segmented through a short-time window to capture the subtle fluctuations during the operation of mechanical components and extract features such as standard deviation, maximum value, and amplitude change, which helps to detect early mechanical vibration anomalies. The segmentation process of current data focuses on the dynamic fluctuations of the load. By calculating the standard deviation and change amplitude within the window, the stability and abnormal fluctuation degree of the electrical load can be quantified, reflecting problems such as motor overload or short-term current instability. The results of the above segmentation process form a multi-dimensional time-domain feature matrix, covering the dynamic changes of the device's operating state at different time scales. This feature extraction process significantly enhances the pertinence and descriptive ability of the data, providing high-quality input data for subsequent frequency-domain and time-frequency-domain analyses. Finally, through the hierarchical extraction of time-domain features, slow thermal trends, transient mechanical anomalies, and electrical fluctuation characteristics are captured, laying a solid foundation for the accurate diagnosis and prediction of device failures, while improving the efficiency and expression ability of feature input.

[0047] Step S220: Perform frequency-domain analysis based on the time-domain feature data. Extract the main frequency, the energy ratio of the abnormal frequency band of harmonic content, and the total spectral energy through fast Fourier transform to obtain frequency-domain feature data.

[0048] The core of this step is to reveal the frequency components and their distribution characteristics hidden in the time domain, thus providing support for identifying the periodic anomalies and vibration characteristics of the device. In frequency-domain analysis, the extraction of the main frequency is used to locate the frequency component with the highest energy ratio in the signal, such as the natural frequency during the operation of mechanical components or the rotational speed frequency of rotating equipment. These main-frequency characteristics can directly reflect the core operating state of the device. In addition, by calculating the energy ratio of the abnormal frequency band of harmonic content, the model can capture specific frequency bands related to faults during the device operation. For example, gear wear may cause abnormal enhancement of specific high-frequency signals, or the failure of rolling bearings will lead to harmonic distortion. At the same time, the extraction of the total spectral energy is used to quantify the overall vibration or energy level of the signal to detect whether the energy change during the device operation exceeds the normal range, such as excessive energy caused by overload or abnormal vibration. The generation of frequency-domain feature data significantly improves the level of signal analysis, providing richer and more direct operating state information than the time domain.

[0049] Step S230: Perform time-frequency analysis based on the frequency-domain feature data and the time-domain feature data. Decompose the signal through wavelet transform to capture the frequency energy distribution, instantaneous frequency change, and local fluctuation characteristics in a specific time period to obtain time-frequency-domain feature data.

[0050] Specifically, wavelet transform decomposes a signal into a series of sub-signals with different frequencies and time windows to identify the frequency energy distribution within a specific time period. For example, during the operation of mechanical equipment, an increase in high-frequency energy at a certain stage may indicate transient abnormalities in bearings or gears. In addition, capturing the instantaneous frequency changes enables the model to perceive dynamic adjustments or sudden abnormalities in the equipment operating state, such as the frequency drift of current or vibration signals during motor startup or sudden load changes. The extraction of local fluctuation characteristics can reveal short-term disturbances in the signal, such as shock vibrations or non-stationary fluctuation behaviors caused by mechanical failures. The time-frequency domain feature data obtained through time-frequency analysis can comprehensively describe the coupling characteristics of equipment operating signals in the time and frequency dimensions, providing high-dimensional and high-resolution input data for subsequent model training.

[0051] Furthermore, step S300 includes steps S310 to S330.

[0052] Step S310: Based on the equipment component decomposition information and component function descriptions in the structure data, perform hierarchical modeling on the key components of the equipment. By defining components as nodes, physical connections as edges, and constructing the equipment physical topology structure, a structure diagram is obtained;

[0053] First, hierarchical modeling decomposes the equipment into several key components according to functional and structural levels by analyzing the structure data of the equipment. For example, in a rotating machinery system, components such as gears, bearings, and drive shafts are regarded as independent functional nodes, while the equipment housing and brackets exist as auxiliary nodes. Each node is assigned corresponding functional attributes (such as load-bearing capacity, material characteristics, operating type, etc.), and these attributes can reflect the operating characteristics and physical states of the components. Secondly, the physical connections between components are modeled as edges, and the weights of the edges represent the connection strength, mechanical coupling characteristics, or stress distribution. For example, the connection weight between a gear and a bearing is calculated based on their transmission force characteristics, while the connection weight between a bracket and a shaft is calculated based on the force distribution. By corresponding the attributes of nodes and edges to the physical topology of the equipment, the structure diagram can comprehensively describe the physical relevance and interaction between equipment components.

[0054] Step S320: Based on the signal nodes and electrical parameters in the wiring data, model the transmission path of the equipment electrical signals. By defining signal input / output ports as nodes, signal connections as edges, and annotating the electrical characteristic parameters of each node, a wiring characteristic diagram is constructed;

[0055] Specifically, first, signal ports in the device (such as the output port of a sensor, the input port of a controller, and the drive port of an actuator) are defined as nodes. Each node is assigned electrical characteristic parameters related to its physical function. For example, a sensor node includes a voltage range and a signal output frequency; a controller node includes an input impedance and a response time; an actuator node includes a drive current and a rated power, etc. These characteristic parameters provide a detailed description of the electrical characteristics of the device for the model. Next, the signal transmission paths in the device wiring are modeled as edges between nodes, and the weights of the edges are used to reflect the physical characteristics of signal transmission. For example, the weight of an edge represents the attenuation degree, transmission delay, or interference intensity of the signal. Through these modelings, the wiring characteristic diagram can truly reproduce the transmission relationship and its dynamic characteristics of internal signals in the device among various nodes.

[0056] Step S330: According to the logical trigger conditions and action timings in the control logic data, model the control process and logical relationships of the device. By defining logical trigger points as nodes and dependencies as edges, and extracting the action trigger time intervals and periodic characteristics, a control logic diagram is constructed.

[0057] In the specific implementation process, first, logical trigger points during device operation (such as sensor signals, switch commands, controller outputs, etc.) are defined as nodes. For example, the over-limit signal of a temperature sensor, the trigger state of a pressure sensor, the start signal of a cooling system, etc. are modeled as nodes. Each node is assigned attribute parameters related to its logical function, such as a trigger threshold, a signal type (such as a digital signal or an analog signal), and a response priority. These attribute parameters can intuitively reflect the role of the node in the device control logic. Secondly, the dependencies during device operation (such as the logical trigger chain between an input signal and an output action) are modeled as edges, and the weights of the edges describe the strength of the logical dependency, the probability of signal response, or the delay. For example, when the trigger signal of a temperature sensor directly controls the start of a cooling system, the weight of the edge of this logical trigger chain can reflect the reliability or time delay of this response path.

[0058] To further enhance the timing characteristics of the model, this step also extracts the action trigger time intervals and periodic characteristics of the logical trigger points. For example, for the start signal of the cooling system, its trigger period is extracted to determine whether the cooling system starts and stops frequently; for the trigger logic of the alarm system, its response time interval is analyzed to evaluate the real-time performance of fault detection. These timing characteristics are encoded as additional attributes of nodes or edges in the control logic diagram to describe the dynamic characteristics of the device operation logic.

[0059] Furthermore, step S400 includes steps S410 to S430.

[0060] Step S410: Node feature mapping is performed according to the operating state feature matrix and the structure diagram, by matching the temperature, vibration and speed parameters of the matrix with the component nodes in the structure diagram, and establishing an association between the operating state and the component attributes in combination with the physical connection relationship between the equipment components, so as to obtain a structure diagram with operating state characteristics;

[0061] It should be noted that the various parameters in the operating status feature matrix (such as temperature, vibration, and speed) are first matched one by one to the corresponding component nodes in the structure diagram. Specifically, the temperature characteristics are mapped to nodes such as bearings and gears, the vibration characteristics are mapped to mechanical moving parts nodes, and the speed characteristics are matched with motor or drive shaft nodes. This mapping relationship is based on the physical topology and operating characteristics of the equipment, and is determined by the attribute definition of the node in the structure diagram and the source of the operating status data. After each node is assigned dynamic characteristics, its attributes include not only the original physical characteristics (such as component functions, materials, etc.), but also combine real-time operating status data (such as the current temperature value or vibration amplitude of the node).

[0062] After completing the node feature mapping, the physical connection relationship between the equipment components is further combined to establish the association between the operating status and the properties between the components. For example, the temperature anomaly of a node may affect the status of adjacent nodes through the physical connection relationship. For example, the high bearing temperature may cause the lubrication failure of the adjacent gear, thereby triggering a chain reaction. This association is modeled through the edge weights in the structure diagram, which represent the physical properties such as heat transfer efficiency and mechanical coupling strength between components.

[0063] Step S420, performing graph topology characteristic fusion according to the structure graph and the wiring characteristic graph with operation status characteristics, by modeling the association between the signal nodes and the physical component nodes in the wiring characteristic graph, using a weighted adjacency matrix to represent the electrical characteristics between the nodes and superimposing them on the connection relationship of the physical structure graph, and generating a fused topology graph;

[0064] Specifically, first, based on the signal nodes (such as sensors, controllers, actuators, etc.) in the wiring characteristic diagram and the physical component nodes (such as bearings, gears, motors, etc.) in the structure diagram, an association model is established. For example, the temperature data collected by the sensor node is directly associated with the operating state of the bearing node in the structure diagram, and the drive signal output by the controller is associated with the dynamic characteristics of the motor node. This association is determined by the design logic and wiring data of the device, forming a mapping relationship between the signal node and the physical node. Next, a weighted adjacency matrix is used to quantitatively model the connection relationship between the signal node and the physical node. The elements of the weighted adjacency matrix represent the electrical characteristics of the signal transmission path, such as resistance, impedance, signal delay, attenuation coefficient, etc. By superimposing these weight parameters of the wiring characteristic diagram onto the physical connection relationship of the structure diagram, the edge weights of the fusion topology diagram not only include the physical characteristics between components but also incorporate the dynamic characteristics of electrical signals. For example, the edge weight can simultaneously reflect the mechanical transmission relationship between the gear and the bearing and the influence range of the control signal between these components. This fusion mechanism enables the model to simultaneously describe the complex interactions between the mechanical system and the electrical system.

[0065] Step S430: Perform control dependency modeling according to the fusion topology diagram and the control logic diagram. By mapping the trigger conditions of the control logic between the physical components and the signal nodes, establish a logic-driven dependency relationship to obtain the integrated graph structure.

[0066] In the specific implementation process, first, according to the triggering conditions in the control logic diagram (such as sensor thresholds, action logics, and triggering sequences), the logical trigger points are mapped to the physical component nodes and signal nodes in the fusion topology diagram. For example, when the reading of the temperature sensor exceeds the set threshold, the start signal of the cooling system is triggered, and this logical trigger condition is mapped to the dependency relationship between the sensor node (signal node) and the cooling system node (physical component node). Through this mapping process, the causal relationship of logical control is injected into the physical and signal networks of the device. Next, by modeling the control dependencies, the association between the logical trigger points and the physical nodes is quantified. Specifically, the edge weights of the dependencies represent the priorities, response times, or triggering probabilities of logical control. For example, in complex multi-level logical control, when multiple triggering conditions are satisfied simultaneously, the triggering condition with a higher priority will have a stronger impact on the operating states of the relevant nodes; and the quantification of the response time can help the model capture the real-time characteristics of logical control and reflect the potential impact of control delays on the operating states of the device. In addition, this modeling process also integrates periodic logical characteristics (such as the periodic triggering of control signals) or dynamic logical characteristics (such as condition changes caused by environmental changes), further enhancing the model's adaptability to complex logical systems. The finally generated comprehensive graph structure can not only describe the physical structure and electrical signal characteristics of the device, but also intuitively present how the control logic drives the operating dynamics of the device and their interdependencies.

[0067] Further, step S500 includes steps S510 to S530.

[0068] Step S510: Perform feature hierarchical aggregation processing according to the comprehensive graph structure, and obtain the hierarchical feature representation of the graph by classifying and modeling and interacting with the features of the operating state nodes, structure nodes, signal transmission nodes, and logical nodes and hierarchically aggregating them.

[0069] In specific implementation, first, the nodes in the comprehensive graph are classified and modeled. According to the types of nodes (operation status nodes, structural nodes, signal transmission nodes, and logic nodes), their unique features are extracted respectively. For example, the operation status nodes mainly extract dynamic operation features to describe the changes in the real-time status of the device; the structural nodes extract physical connection relationships and mechanical coupling characteristics to describe the physical topology inside the device; the signal transmission nodes focus on signal transmission characteristics and extract features such as signal strength, delay, or interference; the logic nodes extract control logic dependency relationships and trigger timing characteristics. The purpose of classification modeling is to perform targeted modeling on each type of node feature and retain the unique attributes of each type of node. Subsequently, through interactive modeling, the relevance between different nodes is modeled. For example, the features of the operation status nodes may be associated with their adjacent structural nodes (such as the temperature change of a bearing may affect the status of its adjacent gear); the features of the signal transmission nodes may be affected by the operation status of physical components (such as the signal output of a sensor may be affected by the vibration of a bearing); the trigger relationship of the logic nodes may be coupled with the operation characteristics of the signal nodes and structural nodes (such as an abnormality at the logic trigger point may be caused by a control signal or a physical fault). This interactive modeling realizes the interactive integration of different types of node features through the edge relationships and node neighbor characteristics in the comprehensive graph. Finally, through hierarchical aggregation, the local features of the nodes and edges are integrated layer by layer into a global feature representation. The aggregation operation at each layer is based on the graph neural network (GNN) framework, combining node features and neighbor features to generate a hierarchical embedding representation of the comprehensive graph layer by layer. For example, the first layer captures the direct associations between local nodes, the second layer extends to the interactions of multi-hop neighbors, and gradually extends to the global scope. Hierarchical aggregation can effectively integrate the local characteristics of nodes and global structure information to generate multi-level graph feature representations.

[0070] Step S520: Extract global graph structure features based on the hierarchical feature representation, dynamically assign weights to the importance between nodes using the attention mechanism, and establish the relationship between logical trigger and operation anomaly using the association between the control logic node and the operation status node to obtain the global graph features;

[0071] This process dynamically assigns weights to the nodes and their association relationships in the comprehensive graph using the attention mechanism, and at the same time combines the characteristics of the control logic node and the operation status node, focusing on depicting the potential relevance between logical trigger and operation anomaly, forming a global modeling of the device operation status. The global graph features not only include the static structure and signal transmission relationship of the device, but also depict the dynamic changes in the operation status and its interaction characteristics with logical control, and can comprehensively reflect the overall operation status of the device and its potential anomaly patterns.

[0072] Step S530: Perform joint modeling processing on the time-dynamic characteristics of the device according to the global graph features and the third information to construct a fault prediction model.

[0073] It is understandable that the fault prediction model can not only identify potential anomalies in the current operating state, but also predict the change trend of the future operating state of the device and its possible fault risks.

[0074] Furthermore, step S600 includes step S610 to step S630.

[0075] Step S610: According to the first information and the fault prediction model, perform fault correlation calculation processing on the real-time operating state characteristics. By inputting the real-time operating state characteristics into the classification module of the fault prediction model, calculate the matching probability between the current operating state of the device and each known fault category to obtain the fault correlation probability.

[0076] It should be noted that first, input the real-time operating state characteristics in the first information into the fault prediction model. The preset classification module in the prediction model performs multi-dimensional feature interaction analysis on the real-time operating state data by combining global graph features (such as device physical topology, signal transmission characteristics, control logic relationships, etc.) and time dynamic characteristics. The classification module of the model infers the real-time state characteristics through a deep learning framework, preferably such as a Softmax classification layer, and outputs the matching probability between the current operating state and each known fault category. For example, the vibration characteristics of a certain device have a high degree of match with the bearing wear pattern in the historical records, and at the same time, the temperature characteristics are also related to the heat dissipation fault mode. The model will calculate the correlation probabilities with bearing wear and heat dissipation faults respectively. During the fault correlation calculation process, the model dynamically adjusts the contribution of the operating state characteristics to different fault modes through an attention mechanism or node importance weighting in the graph neural network. For example, for a device with obvious real-time temperature anomalies, the model will give priority to focusing on the fault modes related to temperature rise (such as heat dissipation problems or mechanical friction overheating), and assign lower weights to other non-related features, thereby improving the matching accuracy. In addition, the model also uses the results of time dynamic characteristic modeling to combine the current characteristics with the historical trend of the device to determine whether the anomaly is a short-term fluctuation or a continuation of a long-term trend. Through this step, the real-time operating state characteristics are efficiently interpreted and matched with known fault modes, significantly improving the real-time performance and accuracy of device fault analysis, and laying a solid foundation for the intelligent operation monitoring of complex devices.

[0077] Step S620: According to the fault correlation probability and the comprehensive graph structure, perform positioning processing on the specific fault location of the device, map the fault correlation probability to the node space of the comprehensive graph, and obtain the specific fault location node in the device.

[0078] In the specific implementation process, first, using the node information (such as the operating status node, signal node, and control logic node) in the comprehensive graph structure and the association relationship of the edges, the fault correlation probability is assigned to the nodes that may be related to the fault. For example, when the correlation probability of a certain type of fault is relatively high (such as the bearing wear fault probability is 0.85), the model will, according to the adjacency relationship in the comprehensive graph, first locate the bearing node and at the same time analyze whether the operating status characteristics of this node (such as temperature rise, vibration intensification) are consistent with the fault type. If the consistency is relatively high, then initially locate this node as the fault location. Subsequently, through the edge weights and association relationships in the comprehensive graph structure, the state of adjacent nodes is analyzed for propagation. Specifically, the model uses a graph neural network (GNN) or an algorithm based on weighted random walk to evaluate the risk that the fault may spread to adjacent nodes. For example, the abnormal temperature of the bearing node may be transmitted to the adjacent gear node through mechanical connection, and the model will evaluate the fault probability of the adjacent node according to the edge weight between the bearing and the gear (such as mechanical coupling strength or heat transfer efficiency). This propagation analysis ensures that the fault location is not limited to a single node but can also identify potential cascading fault locations. Through the above mapping and propagation analysis, the model finally generates a set of fault location nodes and assigns corresponding fault correlation scores to each node, intuitively reflecting the fault possibility of each node. Through this step, the fault correlation probability is transformed from an abstract numerical value into a specific spatial position, providing highly practical information for equipment operation monitoring and fault repair.

[0079] Step S630: According to the fault location nodes and the comprehensive graph structure, evaluate and process the current operating state of the equipment, and obtain the health state index of the current equipment by analyzing the node characteristics and adjacency relationships of the fault location.

[0080] Through the comprehensive analysis of the fault location nodes and their association relationships in this step, the complex operating state data is transformed into an intuitive health state index, providing key support for the intelligent operation monitoring and maintenance optimization of the equipment.

[0081] Furthermore, step S630 includes steps S631 to S633.

[0082] Step S631: According to the fault location nodes and the comprehensive graph structure, perform abnormal scoring processing on the state characteristics of the fault nodes, and calculate the state abnormal score of the nodes by analyzing the deviation degree between the real-time characteristics of the nodes and the historical normal characteristic distribution.

[0083] In this step, by introducing statistical methods to accurately model the deviation degree of the characteristics, a mathematical basis is provided for identifying the health state of key nodes. Specifically, a method based on Mahalanobis distance is used to compare the real-time characteristics of the nodes with the mean and covariance matrix of the historical normal operating state characteristics, and calculate its deviation degree. The formula is:

[0084]

[0085] Among them, D M represents the degree of deviation; x represents the real-time feature vector; μ represents the historical feature mean; Σ represents the covariance matrix of historical features.

[0086] Compared with the Euclidean distance, the Mahalanobis distance can consider the correlation and scale differences between features, so it is more applicable in multi-dimensional feature analysis. The greater the degree of deviation, the more significant the difference between the current operating state of the node and the historical normal state, which also means a higher probability of abnormality.

[0087] Step S632: According to the state abnormality score, perform a failure risk propagation analysis and processing on the nodes adjacent to the faulty node, and calculate the diffusion influence of the abnormality score among adjacent nodes through a preset graph propagation model to obtain the adjacent node failure risk value;

[0088] The purpose of this process is to evaluate the potential impact of the abnormal state of the faulty node on adjacent nodes, help identify the areas that may be affected by the failure, and provide a quantitative basis for the diffusion analysis of the failure. Specifically, a graph propagation algorithm based on weighted random walk is adopted. Using the abnormality score of the node as the initial influence value and combining the weights of the adjacency relationship, calculate the failure risk of each adjacent node:

[0089] R v = ∑ u∈N(v) w uv ·S u ;

[0090] Among them, R v is the risk value of the current node v; S u is the state abnormality score of the adjacent node u; w uv represents the weight of the edge between node v and node u; N(v) represents the set of adjacent nodes of node v.

[0091] In this step, by combining the state abnormality score of the node and the weight of the adjacency relationship, it can accurately quantify the propagation path and scope of the failure impact, revealing the chain effect of equipment failures; at the same time, the calculation of the failure risk of adjacent nodes enhances the model's perception ability of failure propagation, and can identify potential failure areas and key nodes with greater influence in the equipment; further, the introduction of the weighted random walk algorithm considers the connection strength between nodes, ensuring the accuracy and rationality of failure propagation, and is particularly suitable for analyzing multi-level and cross-domain failure impacts in complex equipment systems.

[0092] Step S633: Based on the status anomaly score and the adjacent node failure risk value, comprehensively evaluate the health status of the device. Use a preset health scoring function to perform weighted accumulation on the status anomaly score of the faulty node and the adjacent node failure risk value to obtain the health status indicator of the device.

[0093] The calculation formula for the health status indicator is:

[0094] H = ∑ v∈V α v ·S v + ∑ u∈V β u ·R u ;

[0095] Wherein, H represents the overall health score of the device; S v is the status anomaly score of the current node v; R u is the risk propagation value of the adjacent node u; α v , β u are weight coefficients; V is the set of all nodes in the graph.

[0096] This evaluation method can combine real-time data with historical health data to improve the decision-making quality of fault diagnosis and preventive maintenance.

[0097] Embodiment 2:

[0098] As Figure 2 shown, this embodiment provides a device fault prediction system based on deep learning. The system includes:

[0099] An acquisition module 901, configured to acquire the first information, the second information, and the third information. The first information is the operation status data collected in real time. The second information includes the structure data, wiring data, and control logic data of the device. The third information is the historical data of the device, and the historical data includes historical operation status data, historical fault records, maintenance records, and working conditions;

[0100] An extraction module 902, configured to extract the historical dynamic operation characteristics of the device according to the third information, extract time domain, frequency domain, and time-frequency domain features through time series segmentation, and perform normalization processing on the features to obtain an operation status feature matrix;

[0101] A construction module 903, configured to perform graph construction processing according to the second information, and respectively construct a structure graph, a wiring characteristic graph, and a control logic graph;

[0102] The fusion module 904 is used to perform fusion processing based on the operation status feature matrix, structure diagram, wiring characteristic diagram, and control logic diagram. By mapping matrix elements to nodes in the diagrams respectively, and combining the physical connection relationships, signal transmission characteristics, and control logic dependencies between components, a comprehensive diagram structure is obtained;

[0103] The modeling module 905 is used to construct a fault prediction model based on the comprehensive diagram structure using deep learning algorithms;

[0104] The prediction module 906 is used to input the first information into the fault prediction model to perform fault prediction on the current operation status of the device, and the prediction result includes the occurrence probability of device faults, the specific fault location nodes, and the change trend of the device health status.

[0105] In a specific embodiment of the present invention, the extraction module includes:

[0106] The first extraction unit is used to perform time series segmentation processing according to the third information. Among them, temperature data captures the overall slow-changing trend through a long-time window, vibration data captures the transient fluctuation characteristics through a short-time window, and current data reflects the load fluctuation situation through standard deviation and change amplitude, so as to obtain time-domain feature data;

[0107] The second extraction unit is used to perform frequency-domain analysis according to the time-domain feature data. By performing fast Fourier transform to extract the main frequency, the energy ratio of the abnormal frequency band of the harmonic content, and the total energy of the spectrum, frequency-domain feature data is obtained;

[0108] The third extraction unit is used to perform time-frequency analysis according to the frequency-domain feature data and the time-domain feature data. By decomposing the signal through wavelet transform to capture the frequency energy distribution, instantaneous frequency change, and local fluctuation characteristics in a specific time period, time-frequency domain feature data is obtained.

[0109] In a specific embodiment of the present invention, the construction module includes:

[0110] The first construction unit is used to perform hierarchical modeling on the key components of the device according to the device component decomposition information and component function description in the structure data. By defining components as nodes and physical connections as edges, and constructing the physical topology structure of the device, a structure diagram is obtained;

[0111] The second construction unit is used to model the transmission path of the device electrical signal according to the signal nodes and electrical parameters in the wiring data. By defining the signal input / output ports as nodes and signal connections as edges, and annotating the electrical characteristic parameters of each node, a wiring characteristic diagram is constructed;

[0112] A third construction unit is used to model the control process and logical relationship of a device according to the logical trigger conditions and action timings in the control logic data. By defining the logical trigger points as nodes and the dependency relationships as edges, and extracting the action trigger time intervals and periodic characteristics, a control logic diagram is constructed.

[0113] Embodiment 3:

[0114] Corresponding to the above method embodiment, in this embodiment, a device fault prediction device based on deep learning is also provided. A device fault prediction device based on deep learning described below can be mutually corresponding and referred to with a device fault prediction method based on deep learning described above.

[0115] Figure 3 It is a block diagram of a device fault prediction device 800 based on deep learning shown according to an exemplary embodiment. As Figure 3 shown, the device fault prediction device 800 based on deep learning may include: a processor 801, a memory 802. The device fault prediction device 800 based on deep learning may further include one or more of a multimedia component 803, an I / O interface 804, and a communication component 805.

[0116] Among them, the processor 801 is used to control the overall operation of the device fault prediction device 800 based on deep learning to complete all or part of the steps in the above-mentioned device fault prediction method based on deep learning. The memory 802 is used to store various types of data to support the operation of the device fault prediction device 800 based on deep learning. These data may include, for example, instructions for any application or method operating on the device fault prediction device 800 based on deep learning, as well as application-related data, such as contact data, sent and received messages, pictures, audio, video, and so on. The memory 802 can be implemented by any type of volatile or non-volatile storage device or a combination thereof, such as Static Random Access Memory (SRAM), Electrically Erasable Programmable Read-Only Memory (EEPROM), Erasable Programmable Read-Only Memory (EPROM), Programmable Read-Only Memory (PROM), Read-Only Memory (ROM), magnetic memory, flash memory, magnetic disk or optical disk. The multimedia component 803 may include a screen and an audio component. Among them, the screen can be, for example, a touch screen, and the audio component is used to output and / or input audio signals. For example, the audio component may include a microphone, and the microphone is used to receive external audio signals. The received audio signal can be further stored in the memory 802 or sent through the communication component 805. The audio component also includes at least one speaker for outputting audio signals. The I / O interface 804 provides an interface between the processor 801 and other interface modules, and the above-mentioned other interface modules can be a keyboard, a mouse, buttons, etc. These buttons can be virtual buttons or physical buttons. The communication component 805 is used for wired or wireless communication between the device fault prediction device 800 based on deep learning and other devices. Wireless communication, such as Wi-Fi, Bluetooth, Near Field Communication (NFC), 2G, 3G or 4G, or a combination of one or more of them. Therefore, the corresponding communication component 805 may include: a Wi-Fi module, a Bluetooth module, and an NFC module.

[0117] In one exemplary embodiment, a device failure prediction device 800 based on deep learning may be implemented by one or more application specific integrated circuits (ASICs), digital signal processors (DSPs), digital signal processing devices (DSPDs), programmable logic devices (PLDs), field programmable gate arrays (FPGAs), controllers, microcontrollers, microprocessors or other electronic components, and is used to execute the above-mentioned device failure prediction method based on deep learning.

[0118] In another exemplary embodiment, a computer-readable storage medium including program instructions is further provided. When the program instructions are executed by a processor, the steps of the above-mentioned device failure prediction method based on deep learning are implemented. For example, the computer-readable storage medium may be the above-mentioned memory 802 including program instructions, and the above-mentioned program instructions may be executed by the processor 801 of a device failure prediction device 800 based on deep learning to complete the above-mentioned device failure prediction method based on deep learning.

[0119] As described above, the above are only specific embodiments of the present invention, but the protection scope of the present invention is not limited thereto. Any person skilled in the art within the technical scope disclosed by the present invention can easily think of changes or substitutions, which should all be covered within the protection scope of the present invention.

Claims

1. A device fault prediction method based on deep learning, characterized in that Including: Obtain the first information, the second information, and the third information. The first information is the operation status data collected in real time. The second information includes the structure data, wiring data, and control logic data of the device. The third information is the historical data of the device, and the historical data includes historical operation status data, historical fault records, maintenance records, and working conditions; Extract the historical dynamic operation characteristics of the device according to the third information. Extract time-domain, frequency-domain, and time-frequency-domain features through time series segmentation, and perform normalization processing on the features to obtain an operation status feature matrix; Perform graph construction processing according to the second information, and respectively construct a structure graph, a wiring characteristic graph, and a control logic graph; Perform fusion processing according to the operation status feature matrix, the structure graph, the wiring characteristic graph, and the control logic graph. By mapping matrix elements to nodes in the graph respectively, and combining the physical connection relationship, signal transmission characteristics, and control logic dependency relationship between components, obtain a comprehensive graph structure; Construct a fault prediction model according to the comprehensive graph structure using a deep learning algorithm; Input the first information into the fault prediction model to perform fault prediction on the current operation status of the device to obtain a prediction result. The prediction result includes the occurrence probability of device faults, the specific fault location nodes, and the change trend of the device health status.

2. The device fault prediction method based on deep learning according to claim 1, characterized in that, Extract the historical dynamic operation characteristics of the device according to the third information. Extract time-domain, frequency-domain, and time-frequency-domain features through time series segmentation, and perform normalization processing on the features to obtain an operation status feature matrix, including: Perform time series segmentation processing according to the third information. Among them, the temperature data captures the overall slow-changing trend through a long time window, the vibration data captures the transient fluctuation characteristics through a short time window, and the current data reflects the load fluctuation situation through the standard deviation and the change amplitude to obtain time-domain feature data; Perform frequency-domain analysis according to the time-domain feature data. Extract the main frequency, the energy ratio of the abnormal frequency band of the harmonic content, and the total spectrum energy through fast Fourier transform to obtain frequency-domain feature data; Perform time-frequency analysis according to the frequency-domain feature data and the time-domain feature data. Decompose the signal through wavelet transform to capture the frequency energy distribution, instantaneous frequency change, and local fluctuation characteristics in a specific time period to obtain time-frequency-domain feature data.

3. The device fault prediction method based on deep learning according to claim 1, wherein Perform graph construction processing according to the second information, and respectively construct a structure graph, a wiring characteristic graph, and a control logic graph, including: According to the device component decomposition information and component function description in the structure data, perform hierarchical modeling on the key components of the device. By defining components as nodes and physical connections as edges, and constructing the physical topology structure of the device, obtain a structure graph; According to the signal nodes and electrical parameters in the wiring data, model the transmission path of the device electrical signal. By defining the signal input / output ports as nodes and signal connections as edges, and annotating the electrical characteristic parameters of each node, construct a wiring characteristic graph; Model the control flow and logical relationships of the device according to the logical trigger conditions and action timings in the control logic data. By defining the logical trigger points as nodes and the dependency relationships as edges, and extracting the action trigger time intervals and periodic characteristics, a control logic diagram is constructed.

4. A device fault prediction method based on deep learning according to claim 1, characterized in that, Perform fusion processing based on the operating state feature matrix, the structure diagram, the wiring characteristic diagram, and the control logic diagram. By mapping the matrix elements to the nodes in the diagram respectively, and combining the physical connection relationships, signal transmission characteristics, and control logic dependency relationships between components, a comprehensive diagram structure is obtained, including: Perform node feature mapping according to the operating state feature matrix and the structure diagram. By matching the temperature, vibration, and rotation speed parameters of the matrix with the component nodes in the structure diagram, and establishing the association between the operating state and component attributes in combination with the physical connection relationships between device components, a structure diagram with operating state characteristics is obtained. Perform graph topology feature fusion according to the structure diagram with operating state characteristics and the wiring characteristic diagram. By modeling the association between the signal nodes and physical component nodes in the wiring characteristic diagram, representing the electrical characteristics between nodes using a weighted adjacency matrix and superimposing it into the connection relationships of the physical structure diagram, a fused topology graph is generated. Perform control dependency relationship modeling according to the fused topology graph and the control logic diagram. By mapping the trigger conditions of the control logic between physical components and signal nodes, a logic-driven dependency relationship is established to obtain a comprehensive diagram structure.

5. A device fault prediction method based on deep learning according to claim 1, characterized in that, Based on the comprehensive diagram structure, use a deep learning algorithm to construct a fault prediction model, including: Perform feature hierarchical aggregation processing according to the comprehensive diagram structure. By classifying and modeling and interacting with the features of the operating state nodes, structure nodes, signal transmission nodes, and logical nodes and hierarchically aggregating them, a hierarchical feature representation of the graph is obtained. Extract the global graph structure features according to the hierarchical feature representation. Use the attention mechanism to dynamically assign weights to the importance between nodes, and use the association between the control logic nodes and the operating state nodes to establish the relationship between logical triggers and operating anomalies to obtain the global graph features. Perform joint modeling processing on the time dynamic characteristics of the device according to the global graph features and the third information to construct a fault prediction model.

6. The device fault prediction method based on deep learning according to claim 1, characterized in that, Input the first information into the fault prediction model to perform fault prediction on the current operating state of the device to obtain a prediction result. The prediction result includes the occurrence probability of the device fault, the specific fault location node, and the change trend of the device health state, including: Perform fault correlation calculation processing on the real-time operating state features according to the first information and the fault prediction model. By inputting the real-time operating state features into the classification module of the fault prediction model, calculate the matching probability between the current operating state of the device and each known fault category to obtain the fault correlation probability. Locate the specific fault location of the device according to the fault correlation probability and the comprehensive diagram structure. Map the fault correlation probability to the node space of the comprehensive diagram to obtain the specific fault location node in the device. Based on the fault location node and the comprehensive graph structure, evaluate and process the current operating state of the device, and obtain the health status index of the current device by analyzing the node characteristics and adjacency relationships of the fault location.

7. The method for predicting device faults based on deep learning according to claim 6, wherein, Based on the fault location node and the comprehensive graph structure, evaluate and process the current operating state of the device, and obtain the health status index of the current device by analyzing the node characteristics and adjacency relationships of the fault location, including: Based on the fault location node and the comprehensive graph structure, perform an abnormal scoring process on the state characteristics of the fault node, and calculate the state abnormal score of the node by analyzing the deviation degree between the real-time characteristics of the node and the historical normal characteristic distribution. Based on the state abnormal score, perform a fault risk propagation analysis process on the nodes adjacent to the fault node, and calculate the diffusion influence of the abnormal score among the adjacent nodes through a preset graph propagation model to obtain the adjacent node fault risk value. Based on the state abnormal score and the adjacent node fault risk value, perform a comprehensive evaluation process on the health state of the device, and use a preset health scoring function to perform weighted accumulation on the state abnormal score of the fault node and the adjacent node fault risk value to obtain the health status index of the device.

8. A device fault prediction system based on deep learning, characterized in that, Including: An acquisition module for acquiring the first information, the second information, and the third information. The first information is the operation state data collected in real time. The second information includes the structure data, wiring data, and control logic data of the device. The third information is the historical data of the device, and the historical data includes historical operation state data, historical fault records, maintenance records, and working conditions. An extraction module for extracting the historical dynamic operation characteristics of the device according to the third information, extracting time domain, frequency domain, and time-frequency domain characteristics through time series segmentation, and performing normalization processing on the characteristics to obtain an operation state feature matrix. A construction module for performing graph construction processing according to the second information, and respectively constructing a structure graph, a wiring characteristic graph, and a control logic graph. A fusion module for performing fusion processing according to the operation state feature matrix, the structure graph, the wiring characteristic graph, and the control logic graph. By mapping the matrix elements to the nodes in the graph respectively, and combining the physical connection relationship, signal transmission characteristics, and control logic dependency relationship between components, a comprehensive graph structure is obtained. A modeling module for constructing a fault prediction model using a deep learning algorithm according to the comprehensive graph structure. A prediction module for inputting the first information into the fault prediction model to perform fault prediction on the current operating state of the device to obtain a prediction result, where the prediction result includes the occurrence probability of device faults, the specific fault location node, and the change trend of the device health state.

9. The device fault prediction system based on deep learning according to claim 8, characterized in that, The extraction module includes: A first extraction unit for performing time series segmentation processing according to the third information, where the temperature data captures the overall slow-changing trend through a long time window, the vibration data captures the transient fluctuation characteristics through a short time window, and the current data reflects the load fluctuation situation through the standard deviation and the change amplitude, to obtain time domain feature data. A second extraction unit, configured to perform frequency-domain analysis based on the time-domain feature data, and extract the main frequency, the energy ratio of the abnormal frequency band of the harmonic content, and the total spectrum energy through fast Fourier transform, so as to obtain frequency-domain feature data; A third extraction unit, configured to perform time-frequency analysis based on the frequency-domain feature data and the time-domain feature data, decompose the signal through wavelet transform, capture the frequency energy distribution, the instantaneous frequency change, and the local fluctuation characteristics in a specific time period, so as to obtain time-frequency domain feature data.

10. The device fault prediction system based on deep learning according to claim 8, characterized in that, The construction module includes: A first construction unit, configured to perform hierarchical modeling on the key components of the device according to the device component decomposition information and component function description in the structure data, construct the physical topology structure of the device by defining components as nodes and physical connections as edges, and obtain a structure diagram; A second construction unit, configured to model the transmission path of the electrical signals of the device according to the signal nodes and electrical parameters in the wiring data, construct a wiring characteristic diagram by defining signal input / output ports as nodes and signal connections as edges, and annotating the electrical characteristic parameters of each node; A third construction unit, configured to model the control process and logical relationship of the device according to the logical trigger conditions and action time sequences in the control logic data, construct a control logic diagram by defining logical trigger points as nodes and dependency relationships as edges, and extracting the action trigger time interval and cycle characteristics.

Images