A method and system for constructing a bearing fault diagnosis model

Abstract

The present invention provides a method and system for constructing a bearing fault diagnosis model, which relates to the field of data processing technology. For bearing operation scenarios, a monitoring mode is determined and a directed graph based on fault causality is mined. A feature equivalence space is constructed by introducing a loss function. The monitoring data flow structure is determined based on the monitoring mode. Based on the directed graph, single-source directional training and composite cross-training are performed under data-driven conditions to generate a fault decision maker. A bearing fault diagnosis model is generated in combination with the feature equivalence space. The method and system are used to solve the technical problem of how to construct an efficient and accurate bearing fault diagnosis model in the prior art to improve the flexible decision-making ability and adaptability in complex fault scenarios, thereby realizing efficient modeling in complex fault environments and effectively improving the adaptability of the model and the fault diagnosis accuracy.

Description

Technical Field

[0001] The present invention relates to the technical field of data processing, and in particular to a method and system for constructing a bearing fault diagnosis model. Background Art

[0002] Bearings are key components in industrial equipment and play a vital role in many mechanical systems. Their failures usually lead to equipment downtime, reduced production efficiency, and even major safety accidents. Therefore, effective fault diagnosis helps reduce operational risks.

[0003] Most existing fault diagnosis models rely on traditional empirical models and rule bases, lack sufficient flexibility and adaptability, and are unable to cope with complex bearing fault types. At the same time, their diagnostic effects are often limited when faced with complex environments or multiple concurrent faults.

[0004] Therefore, how to build an efficient and accurate bearing fault diagnosis model to improve the flexible decision-making ability and adaptability in complex fault scenarios has become a technical problem to be solved. Summary of the invention

[0005] The present application provides a method and system for constructing a bearing fault diagnosis model, which is used to solve the technical problem existing in the prior art of how to construct an efficient and accurate bearing fault diagnosis model to improve the flexible decision-making ability and adaptability in complex fault scenarios.

[0006] In view of the above problems, the present application provides a method and system for constructing a bearing fault diagnosis model.

[0007] In the first aspect, the present application provides a method for constructing a bearing fault diagnosis model, the method comprising: determining a monitoring mode for a bearing operation scenario, wherein the monitoring mode includes a sensor device group and a sensor coordination mode; interactive bearing fault records, mining a directed graph based on fault causality, wherein the directed graph is marked with single fault labels and compound fault labels based on the monitoring mode, and labels are set based on whether the decision sensor data types of the bearing fault are homologous; traversing the directed graph, introducing a loss function to construct a feature equivalence space, wherein the loss function characterizes the redundant loss amount of features before and after equivalence; determining a monitoring data flow structure based on the monitoring mode, and performing data-driven single-source directed training and compound cross-training based on the directed graph to generate a fault decision maker; integrating the feature equivalence space and the fault decision maker to generate a bearing fault diagnosis model.

[0008] In a second aspect, the present application provides a system for constructing a bearing fault diagnosis model, the system comprising: a mode determination unit, the mode determination unit being used to determine a monitoring mode for a bearing operation scenario, wherein the monitoring mode includes a sensor device group and a sensor coordination mode; a graph mining unit, the graph mining unit being used to interact with bearing fault records and mine a directed graph based on fault causality, wherein the directed graph is marked with a single fault label and a compound fault label based on the monitoring mode, and the label is set based on whether the decision sensor data type of the bearing fault is homologous; an equivalence construction unit, the equivalence construction unit being used to traverse the directed graph and introduce a loss function to construct a feature equivalence space, wherein the loss function characterizes the redundant loss amount of the features before and after equivalence; a training unit, the training unit being used to determine a monitoring data flow structure based on the monitoring mode, and taking the directed graph as a benchmark, performing data-driven single-source directional training and compound cross-training to generate a fault decision maker; a model generation unit, the model generation unit being used to integrate the feature equivalence space and the fault decision maker to generate a bearing fault diagnosis model.

[0009] One or more technical solutions provided in this application have at least the following technical effects or advantages:

[0010] The embodiment of the present application provides a method for constructing a bearing fault diagnosis model. For a bearing operation scenario, a monitoring mode is determined, wherein the monitoring mode includes a sensor device group and a sensor coordination mode; interactive bearing fault records are recorded, and a directed graph based on fault causality is mined, wherein the directed graph is marked with a single fault label and a compound fault label based on the monitoring mode, the directed graph is traversed, and a loss function is introduced to construct a feature equivalence space, wherein the loss function characterizes the redundant loss amount of the features before and after equivalence; the monitoring data stream structure is determined based on the monitoring mode, and single-source directional training and compound cross-training are performed under data-driven conditions based on the directed graph to generate a fault decision maker; the feature equivalence space and the fault decision maker are integrated to generate a bearing fault diagnosis model, which is used to solve the technical problem of how to construct an efficient and accurate bearing fault diagnosis model in the prior art to improve the flexible decision-making ability and adaptability in complex fault scenarios, thereby realizing efficient modeling in complex fault environments and effectively improving the adaptability and fault diagnosis accuracy of the model. BRIEF DESCRIPTION OF THE DRAWINGS

[0011] Figure 1 A flow chart of a method for constructing a bearing fault diagnosis model is provided for this application;

[0012] Figure 2 A schematic diagram of the process of constructing a feature equivalent space in a method for constructing a bearing fault diagnosis model is provided for this application;

[0013] Figure 3 A schematic diagram of the system structure for building a bearing fault diagnosis model is provided for this application.

[0014] Explanation of the accompanying drawings: pattern determination unit 11, graph mining unit 12, equivalence construction unit 13, training unit 14, model generation unit 15. DETAILED DESCRIPTION

[0015] The present application provides a method and system for constructing a bearing fault diagnosis model. For bearing operation scenarios, the monitoring mode is determined and a directed graph based on fault causality is mined. A feature equivalence space is constructed by introducing a loss function. The monitoring data stream structure is determined based on the monitoring mode. Based on the directed graph, data-driven single-source directed training and composite cross-training are performed to generate a fault decision maker. A bearing fault diagnosis model is generated in combination with the feature equivalence space. The application is used to solve the technical problem of how to construct an efficient and accurate bearing fault diagnosis model in the prior art to improve the flexible decision-making ability and adaptability in complex fault scenarios.

[0016] Embodiment 1: Figure 1 As shown, the present application provides a method for constructing a bearing fault diagnosis model, the method comprising:

[0017] S1: Determine a monitoring mode for a bearing operation scenario, wherein the monitoring mode includes a sensor device group and a sensor coordination mode;

[0018] Specifically, in the process of building a bearing fault diagnosis model, targeted processing of the monitoring mode under the bearing operation scenario is the basis for ensuring the accuracy and reliability of fault diagnosis. Among them, the sensor equipment group refers to various sensors deployed to monitor the bearing status, which are used to collect real-time working data of the bearing, such as vibration sensors, temperature sensors, sound sensors and surface image sensors, which are used to monitor different working conditions of the bearing.

[0019] For example, vibration sensors can detect the movement state of bearings in real time and capture abnormal vibration signals, while temperature sensors can monitor overheating that may occur during the operation of bearings. By deploying multiple types of sensors, it is possible to fully obtain key data during the operation of bearings, ensuring more comprehensive and accurate monitoring of bearing faults.

[0020] The sensor coordination method refers to how to effectively coordinate and process the data output by the sensors when multiple sensors work together to form unified and complete diagnostic information.

[0021] Specifically, different data fusion strategies can be adopted according to different sensor data streams and fault diagnosis requirements. For example, in a single fault detection scenario, a single sensor data may be used for independent judgment; while in a compound fault detection scenario, collaborative decision-making of multi-sensor data may be used to obtain more reliable fault diagnosis results. At the same time, some faults can be judged by a single sensor source data; some faults need to be judged in coordination with multiple sensor data to ensure the accuracy of judgment, such as identifying overload or wear through joint analysis of temperature and vibration signals. Specifically, it is necessary to determine the appropriate monitoring mode based on the working characteristics of the bearing, possible fault modes and the characteristics of the sensor to ensure that the monitoring capabilities of each sensor can complement each other and improve the accuracy and robustness of the overall monitoring system.

[0022] S2: Interact with bearing fault records, and mine directed graphs based on fault causality, wherein the directed graphs are marked with single fault labels and compound fault labels based on the monitoring mode, and labels are set based on whether the decision sensing data types of bearing faults are of the same source;

[0023] In the embodiment of the present application, the potential causal relationship of the bearing fault is extracted from the bearing fault record to provide theoretical support for subsequent fault diagnosis. By mining the directed graph based on fault causality, different fault types and their relationships can be effectively captured to establish a complete fault evolution chain.

[0024] Specifically, the bearing fault record refers to a data set including labels of different fault types, corresponding sensor data, environmental conditions when the fault occurred, specific fault characteristics and other information, which is a record of historical fault diagnosis.

[0025] Then, based on the bearing fault records, the directed graph is constructed by identifying the causal chain between faults. In this process, each fault node represents a specific fault type or fault feature, fault indicator and other elements, and the directed connection between nodes represents the causal relationship between faults. For example, excessive temperature may cause bearing wear, which in turn causes abnormal vibration, forming a causal chain from abnormal temperature to abnormal vibration. In this way, the potential connection between different faults can be visualized, and a clear decision-making framework can be provided for subsequent fault diagnosis.

[0026] At the same time, by distinguishing different fault labels, fault judgment is divided, that is, according to the monitoring data stream and its characteristics, the singleness or combination of the fault type is judged. For example, if a fault is caused by the abnormality of only a single sensor, such as the signal of the vibration sensor exceeds the normal range, the fault occurs and the judgment probability meets the standard, and it is marked as a single fault label; if a fault occurs due to the joint abnormality of multiple sensor data, such as the simultaneous abnormality of temperature and vibration signals, the fault is marked as a composite fault label. Such label setting is helpful for subsequent fault classification and diagnosis.

[0027] At the same time, there are differences in the specific fault decisions under single fault labels and compound fault labels. Due to the different feature extraction methods of different sensor data types, such as temperature data, image data, etc., for compound fault labels, the corresponding faults need to consider the coordination status between the sensor data, while the judgment of single fault labels is relatively simple.

[0028] The directed graph constructed through the above steps can not only reveal the causal relationship between different bearing faults, but also provide clear guidance for the subsequent fault diagnosis process through the precise setting of fault labels, thereby improving the accuracy and reliability of fault diagnosis.

[0029] The directed graph is identified with a single fault label and a compound fault label based on the monitoring mode. Step S2 of the present application includes:

[0030] Identify the monitoring mode, and make a single diagnosis based on a single sensor data type decision, and a composite diagnosis based on the collaborative decision of multiple sensor data types; for a single diagnosis, traverse the directed graph, and establish a mapping mark between monitoring features and fault features as the single fault label; for a composite diagnosis, traverse the directed graph, and establish a mapping mark between monitoring features and fault features as the composite fault label.

[0031] Under the constraint of fault diagnosis accuracy, the type of sensor data that needs to be monitored is divided into two types: single diagnosis and composite diagnosis. First, single diagnosis refers to when the diagnosis of the fault depends only on the data of a single sensor. By analyzing the data characteristics collected by a single sensor, the corresponding fault can be directly judged. For example, when a vibration sensor detects abnormal vibration, the system only relies on the vibration signal to determine whether a fault has occurred. In this case, the diagnosis of the fault is based on the data stream of only one type of sensor, so the process is classified as single diagnosis.

[0032] The composite diagnosis is applicable to the fault condition that the accuracy of fault diagnosis can only be guaranteed when the occurrence of the fault involves the joint data of multiple sensors. For example, the accuracy of fault judgment based on vibration signal alone is only 50%, and when combined with temperature signal, the accuracy can reach 90%. For the judgment of this fault, the temperature data and vibration data are used for joint judgment.

[0033] The directed graph is traversed to determine the fault graph part of the single diagnosis and the fault graph part of the composite diagnosis, and a mapping mark based on the monitoring feature and the fault feature is established in the graph. The mapping mark represents the association between a certain monitoring feature (such as the amplitude change of the vibration signal) and a specific fault type (such as damage to the outer ring of the bearing). Exemplarily, when the amplitude of the vibration signal exceeds a predetermined threshold, the signal is mapped to the feature of the bearing fault, thereby generating a single fault label. This label depends only on the vibration data and reflects the type of fault that the sensor can capture.

[0034] Similarly, for the fault types of composite diagnosis, composite fault labels are identified in the directed graph. That is, the monitoring features of multiple data types are mapped and identified with the fault features. For example, the vibration signal and the temperature signal are combined for joint judgment to ensure the judgment accuracy. The composite fault features are combined and mapped, such as the damage that may occur to the bearing under overload conditions. Composite fault labels are generated and mapped and identified.

[0035] In summary, single diagnosis and composite diagnosis are respectively applicable to different fault diagnosis situations. By performing feature mapping and label generation in directed graphs, the accuracy and distinguishability of fault diagnosis can be improved to guide the fault diagnosis decision-making process.

[0036] S3: traversing the directed graph, introducing a loss function to construct a feature equivalence space, wherein the loss function represents the redundant loss of features before and after equivalence;

[0037] In the embodiment of the present application, a loss function is introduced to measure the feature redundancy of the monitoring data, and then a feature equivalence space is constructed to ensure the effectiveness of the features selected for the final fault diagnosis decision, reduce the amount of data and improve the decision accuracy.

[0038] First, each node in the directed graph represents a specific monitoring feature or fault feature, and the directed connections between nodes represent the interdependence between the features.

[0039] On this basis, the purpose of introducing the loss function is to quantify the amount of redundancy loss between features. The loss function characterizes the corresponding feature loss before and after the feature redundancy removal, that is, the degree of information redundancy caused by redundant features when selecting certain features.

[0040] Exemplarily, for the extracted features of the monitoring data, a low-dimensional feature vector or a certain perspective vector of the feature may be required to complete the diagnosis. The remaining part of the extracted features can be regarded as redundant parts and eliminated. The feature difference between the extracted features and the effective features is taken as the redundant loss amount. The loss function is a relationship for measuring the redundant loss amount.

[0041] Furthermore, since the accuracy of fault diagnosis decision-making is not affected before and after feature de-redundancy, the feature equivalent space is constructed based on the loss function that measures the loss amount before and after feature de-redundancy, and is used to optimize and screen the features extracted by monitoring through the loss function, so that the redundancy of the features is minimized, ensuring that each selected feature can effectively provide unique information. The construction of the feature equivalent space makes the input data of the fault diagnosis model more concise and effective, reduces the negative impact of noise and redundant data that may be brought about during processing on the model performance, and thus improves the accuracy and efficiency of the entire diagnosis process.

[0042] Among them, Figure 2 As shown, the loss function is introduced to construct the feature equivalent space, and step S3 of the present application includes:

[0043] The bearing fault records are traversed, clustering is performed based on the diagnosis results, and N clusters are determined; for the N clusters, the redundancy loss of the monitoring features-fault features is measured based on direct static redundancy and dynamic analytical redundancy, and the loss function is constructed; based on the loss function, the feature equivalence space is constructed.

[0044] In an embodiment of the present application, the bearing fault records are clustered based on the diagnosis result, that is, the bearing fault determined by the decision, that is, the faults with the same diagnosis results are classified into the same category, thereby forming N clusters, each of which represents a specific fault type, and N is a positive integer greater than or equal to 1.

[0045] For the N clusters, redundancy loss is measured for the monitoring features and fault features in each cluster, specifically from the perspective of direct static redundancy and dynamic analytical redundancy.

[0046] Among them, the direct static redundancy is the redundancy that exists in the extracted features of the monitoring data itself. For example, a low-dimensional feature vector or a certain perspective vector of the feature may be required to complete the diagnosis. The dynamic analytical redundancy is to analyze the dynamic relationship between features in the time series and evaluate whether these features always carry similar information as time changes. For example, in the diagnosis of bearing faults, vibration signals and sound signals may show different abnormal characteristics in different time periods. At this time, it is necessary to dynamically analyze the redundancy between these features. For example, if the change of the signal is linear, linear substitution can be performed; or if the signal is stable and unchanged, time series identification is sufficient, and the specific data vector can be removed.

[0047] For each cluster, a loss function is determined by measuring the redundant loss to evaluate the change in redundant loss before and after feature equivalence, and then a feature equivalence space is constructed based on the loss function.

[0048] By constructing loss functions and feature equivalent spaces to achieve feature simplification and optimization, redundant information can be removed and features with high diagnostic value can be retained, thereby providing high-quality data support for subsequent fault decision-making and diagnosis, and improving the performance and accuracy of the bearing fault diagnosis model.

[0049] Among them, the redundant loss measurement of the monitoring feature-fault feature is performed to construct the loss function. Step S3 of this application includes:

[0050] Based on the direct static redundancy and the dynamic analytical redundancy, multi-level redundant nodes are defined, wherein the preset data volume is used as the redundancy; the directed graph is identified, the decision process of the monitoring feature-fault feature is determined, and the multi-level redundant nodes are inserted as decision process nodes; the first cluster is identified, and the first redundant node of the first cluster is located with the decision accuracy as a constraint, wherein the first redundant node is the maximum redundant node level that meets the decision accuracy; the first loss function is determined with the monitoring feature-first redundant node corresponding to the first cluster as a benchmark and the feature redundancy loss as a benchmark.

[0051] In the embodiment of the present application, since the redundancy scales of different fault types are different, by setting multi-level redundant nodes, each cluster is adaptively measured to adaptively quantify the redundancy of the features, thereby further improving the accuracy and efficiency of the diagnosis model.

[0052] First, the preset data volume is used as a setting standard for dividing redundancy levels, the preset data volume is used as a data processing difference between neighboring redundancy levels, and the preset data volume is used as redundancy to define the redundancy level.

[0053] Next, the directed graph is identified and the decision process of monitoring feature-fault feature is determined. Before the monitoring feature is processed redundantly, the equivalent feature that completes the accuracy decision of the fault feature is processed redundantly, and then the multi-level redundant node is inserted into the redundant processing stage.

[0054] Furthermore, the first cluster is identified, and the first redundant node is located with the decision accuracy as a constraint. Specifically, the monitoring features are determined based on the first cluster, processed based on the first redundant node, and the fault determination accuracy is evaluated; the multi-level redundant nodes are processed and fault determined in turn, and the node with the maximum redundancy that meets the predetermined decision accuracy requirement is used as the first redundant node. The preset decision accuracy is the accuracy standard for bearing fault determination.

[0055] Furthermore, taking the monitoring feature corresponding to the first cluster - the first redundant node as a benchmark, the first loss function is generated by measuring the redundant loss of invalid features, that is, the amount of feature loss processed based on the redundancy level corresponding to the first redundant node, to characterize the degree to which the extracted features of the direct monitoring data need to be de-redundantly processed to meet the accuracy of fault decision making.

[0056] Similarly, redundant node analysis is performed on each cluster, and N loss functions are constructed. Based on the N loss functions, the feature equivalent space is constructed. By optimizing the feature space in the decision-making process, it is ensured that the features on which the final decision is based are the most effective and representative. This improves the accuracy and efficiency of the bearing fault diagnosis model.

[0057] S4: Determine the monitoring data flow structure based on the monitoring mode, perform data-driven single-source directional training and composite cross-training based on the directed graph, and generate a fault decision maker;

[0058] S5: Integrate the feature equivalent space and the fault decision maker to generate a bearing fault diagnosis model.

[0059] In the embodiment of the present application, the type of sensor used, the coordination mode of the sensors, and the structural arrangement of the data stream are determined based on the monitoring mode. For example, the monitoring mode may include data streams of multiple sensors such as vibration, temperature, and sound signals, intermittent monitoring or continuous monitoring, etc.

[0060] Among them, single-source directional training refers to training based on a single sensor data stream, such as using only vibration signals or temperature signals for fault diagnosis. During this training process, the system performs pattern recognition and feature extraction through the input of a single data source (such as a vibration signal), and finally generates a directional decision maker for the signal. Directed training is performed for each sensor data stream to determine multiple directional decision makers.

[0061] Furthermore, for the composite fault label class, cross-attention training is performed on the basis of the multiple directional decision makers to meet the single diagnosis and composite diagnosis conditions, and to generate the fault decision maker that meets the convergence accuracy, so that intelligent judgment can be made based on data from different sensors, and the specific fault type or status can be output.

[0062] The feature equivalence space and the fault decision maker are integrated, and the optimized feature space and the trained fault decision maker are fused to form the final bearing fault diagnosis model. The feature equivalence space is processed by redundancy measurement and optimization to ensure that the input feature data is streamlined and screened, thereby maximizing the validity and diagnostic accuracy of the data. During the integration process, the system uses the valid features in the feature equivalence space as input and inputs them into the fault decision maker for final fault determination. The purpose of this integration process is to generate an efficient and accurate bearing fault diagnosis model that can process multi-source sensor data in real time in practical applications and make fast and accurate fault diagnosis.

[0063] Wherein, the data-driven single-source directional training, step S4 of the present application includes:

[0064] The monitoring data stream structure at least includes a temperature signal, a sound signal, a vibration signal and a surface image; based on the monitoring data stream structure, the bearing fault records are divided to determine the directional training data; based on the directional training data, the directed graph is traversed and the single fault label is identified, data integration is performed, and pre-training data is determined; based on the pre-training data, a directional decision maker is supervised for training, wherein the directional decision maker corresponds one-to-one to the monitoring data stream structure.

[0065] In the process of constructing the bearing fault diagnosis model, the monitoring data stream structure at least includes temperature signals, sound signals, vibration signals and surface images. Specifically, they represent different monitoring dimensions: the temperature signal reflects the thermal state of the bearing, the sound signal can reveal abnormal sounds during mechanical operation, the vibration signal is an important feature commonly used in bearing fault diagnosis, and the surface image can provide more intuitive fault symptoms, such as cracks or wear, etc., together forming a multi-dimensional fault monitoring system that can comprehensively reflect the operating status of the bearing.

[0066] Furthermore, based on the monitoring data stream structure, the bearing fault records are divided to determine targeted training data, that is, corresponding training data are divided for different monitoring dimensions to ensure that targeted analysis can be performed for different signal sources.

[0067] Identify the first monitoring dimension, such as the directional training data of vibration sensing, and determine the corresponding map part by identifying the single fault label, using the map part as the reference logic, for example, the vibration signal may be directly associated with the fault type of "bearing outer ring damage", and the temperature signal may be associated with "overheating fault". Sort the directional training data of the first monitoring dimension, determine the pre-training data, perform sample-driven training based on the pre-training data until convergence, and obtain the directional decision maker.

[0068] Similarly, for each monitoring dimension, matching directional training data is determined, and graph recognition and data sorting are performed based on a single fault label. Through sample training, multiple directional decision makers are determined. The directional decision makers correspond to the monitoring dimensions, that is, the monitoring data stream structure. Each type of monitoring signal (such as vibration signal, temperature signal, etc.) will correspond to an independent decision maker, and specialized training and decision-making will be performed according to the characteristics of different data streams, thereby improving the accuracy and efficiency of the overall diagnosis model.

[0069] Among them, the composite cross training is performed, and step S4 of the present application includes:

[0070] The composite fault label is identified, and based on the composite fault label, data integration is performed on the bearing fault record to determine composite training data; based on the composite training data, cross-attention training is performed on the directional decision maker to generate the fault decision maker.

[0071] In the embodiment of the present application, in the process of constructing the bearing fault diagnosis model, the sensor source is modeled specifically for the single fault label. At the same time, the composite fault label needs to consider the correlation and synergy between them on the basis of single source processing. The composite fault label refers to the joint determination of the fault through multiple monitoring signals (such as vibration, temperature, sound, etc.) to meet the fault diagnosis accuracy.

[0072] The composite fault label is identified, and composite training data is further determined based on the bearing fault record. Composite training data refers to a training data set generated by the interaction between multiple sensor data sources. These data sets contain not only the characteristics of each individual sensor, but also the correlation and interaction characteristics between them. Specifically, the system collects all monitoring data related to the composite fault label, that is, the bearing fault record, and divides it into training samples according to the definition of the composite fault label. For example, assuming that the vibration and temperature signals exhibit specific synergistic characteristics when a bearing fault occurs, these synergistic data points are aggregated into a composite training data set. By cross-analyzing the relationship between different data sources, it helps the model learn more complex fault characteristics.

[0073] Preferably, data from different sensors are combined into a multi-dimensional feature vector through data integration, while retaining the information of each signal source.

[0074] The constructed directional decision maker is further cross-attention trained based on the composite training data, mainly for processing the interdependence between multiple input data sources. The purpose of cross-attention training is to enable the decision maker to pay attention to data from multiple sensors at the same time, and to perform incremental learning on the correlation between these data, thereby improving the accuracy of composite fault diagnosis. For example, when performing cross-attention training, vibration signals and temperature signals may be weighted through different attention mechanisms, and the importance of each signal is automatically learned through the cross-attention mechanism, and a more accurate fault diagnosis is made based on the relationship between these signals. The trained directional decision maker is used as the fault decision maker, which can effectively identify and classify complex bearing faults, thereby improving the accuracy and robustness of the overall fault diagnosis model.

[0075] After generating the bearing fault diagnosis model, the application further includes:

[0076] Carry out bearing operation monitoring to determine multi-source sensor data; combine the feature equivalence space of the bearing fault diagnosis model, perform data feature extraction and adaptive redundant processing, and determine the effective feature set; transfer the effective feature set to the fault decision maker, determine single faults and compound faults, and integrate and output the fault diagnosis list.

[0077] In the embodiment of the present application, through the monitoring mode, multiple sensors are controlled to monitor the bearing comprehensively at the same time, and sensor information from different dimensions is obtained. The multi-source sensor data includes but is not limited to vibration signals, temperature signals, sound signals and surface images. For example, the vibration sensor can monitor the operating vibration of the bearing in real time, the temperature sensor is used to capture whether the bearing is overheated, the sound sensor can detect whether the bearing produces abnormal noise, and the surface image uses image recognition technology to detect wear, cracks and other damage on the bearing surface.

[0078] Next, the collected raw monitoring data is analyzed and processed based on the feature equivalent space, and fault diagnosis is optimized by reducing data redundancy and extracting important information. Specifically, based on the loss function, feature extraction and loss function matching are performed on the multi-source sensor data, and adaptive redundancy processing is performed to extract key features from the multi-source monitoring data, and determine that the effective feature set is a feature.

[0079] Subsequently, the effective feature set is transferred to the fault decision maker. Single fault diagnosis is performed first, and compound fault diagnosis is performed on this basis. The fault decision maker can distinguish single faults from compound faults based on multiple features in the effective feature set, thereby improving the accuracy and coverage of fault diagnosis.

[0080] The diagnostic results are integrated and a fault diagnosis sheet is output, including but not limited to the classification of fault types, determination of fault locations, and evaluation of fault severity. The diagnosis sheet will provide a detailed fault report for subsequent maintenance personnel or automation systems, helping to take appropriate repair measures in a timely manner.

[0081] The present application provides a method for constructing a bearing fault diagnosis model, which has the following technical effects:

[0082] 1. Based on the accuracy of fault judgment, the fault types are divided, single fault labels and compound fault labels are identified, and the early processing process is optimized with results as the guide to carry out targeted modeling. On this basis, single-source directional training and compound cross-training are used as training methods, so that the model can accurately identify single faults and compound fault modes, improve the adaptability and accuracy of the diagnosis system in complex fault environments, and can handle the interdependence between multi-source data.

[0083] 2. By introducing an adaptive loss function and performing scale measurement for feature redundancy processing, adaptive redundancy processing is performed based on the feature equivalence space in the fault diagnosis model to screen out the most discriminative features for fault identification and eliminate redundant information, thereby reducing data redundancy, reducing computational complexity, and improving the efficiency and accuracy of fault diagnosis. At the same time, the removal of redundant features also enhances the generalization ability of the model, enabling the system to better adapt to different working environments and fault modes.

[0084] Embodiment 2: Based on the same inventive concept as the construction method of a bearing fault diagnosis model in the above embodiment, Figure 3 As shown, the present application provides a system for constructing a bearing fault diagnosis model, the system comprising:

[0085] A mode determination unit 11, the mode determination unit 11 is used to determine a monitoring mode for a bearing operation scenario, wherein the monitoring mode includes a sensor device group and a sensor coordination mode;

[0086] A graph mining unit 12, the graph mining unit 12 is used to interact with bearing fault records and mine a directed graph based on fault causality, wherein the directed graph is marked with a single fault label and a composite fault label based on the monitoring mode, and the label is set based on whether the decision sensor data type of the bearing fault is homologous;

[0087] An equivalent construction unit 13, the equivalent construction unit 13 is used to traverse the directed graph and introduce a loss function to construct a feature equivalent space, wherein the loss function represents the redundant loss amount of the features before and after the equivalence;

[0088] A training unit 14, the training unit 14 is used to determine the monitoring data flow structure based on the monitoring mode, and to perform data-driven single-source directional training and composite cross-training based on the directed graph to generate a fault decision maker;

[0089] The model generating unit 15 is used to integrate the feature equivalent space and the fault decision maker to generate a bearing fault diagnosis model.

[0090] Furthermore, the pattern determination unit 11 is also used to perform the following steps: identifying the monitoring mode, deciding on a single sensor data type as a single diagnosis, and collaboratively deciding on multiple sensor data types as a composite diagnosis; for a single diagnosis, traversing the directed graph, establishing a mapping mark between monitoring features and fault features as the single fault label; for a composite diagnosis, traversing the directed graph, establishing a mapping mark between monitoring features and fault features as the composite fault label.

[0091] Furthermore, the equivalent construction unit 13 is also used to perform the following steps: traverse the bearing fault records, cluster based on the diagnosis results, and determine N cluster clusters; for the N cluster clusters, based on direct static redundancy and dynamic analytical redundancy, perform redundant loss measurement on the monitoring features-fault features to construct the loss function; based on the loss function, construct the feature equivalent space.

[0092] Furthermore, the equivalent construction unit 13 is also used to perform the following steps: based on the direct static redundancy and the dynamic analytical redundancy, define multi-level redundant nodes, wherein the preset data volume is used as the redundancy; identify the directed graph, determine the decision process of the monitoring feature-fault feature, and insert the multi-level redundant nodes as decision process nodes; identify the first clustering cluster, and locate the first redundant node of the first clustering cluster with the decision accuracy as a constraint, wherein the first redundant node is the maximum redundant node level that meets the decision accuracy; determine the first loss function with the monitoring feature-first redundant node corresponding to the first clustering cluster as the benchmark and the feature redundancy loss as the benchmark.

[0093] Furthermore, the training unit 14 is also used to perform the following steps: the monitoring data stream structure includes at least a temperature signal, a sound signal, a vibration signal and a surface image; based on the monitoring data stream structure, the bearing fault records are divided to determine the directional training data; based on the directional training data, the directed graph is traversed and the single fault label is identified, data integration is performed, and pre-training data is determined; based on the pre-training data, a directional decision maker is supervised for training, wherein the directional decision maker corresponds one-to-one to the monitoring data stream structure.

[0094] Furthermore, the training unit 14 is also used to perform the following steps: identifying the composite fault label, and based on the composite fault label, performing data integration on the bearing fault record to determine composite training data; based on the composite training data, performing cross-attention training on the directional decision maker to generate the fault decision maker.

[0095] Furthermore, the system is also used to perform the following steps: perform bearing operation monitoring to determine multi-source sensor data; combine the feature equivalence space of the bearing fault diagnosis model to perform data feature extraction and adaptive redundant processing to determine a valid feature set; transfer the valid feature set to the fault decision maker to determine single faults and compound faults, and integrate and output a fault diagnosis sheet.

[0096] Through the above-mentioned detailed description of the method for constructing a bearing fault diagnosis model in this specification, those skilled in the art can clearly know the method and system for constructing a bearing fault diagnosis model in this embodiment. For the device disclosed in the embodiment, since it corresponds to the method disclosed in the embodiment, the description is relatively simple, and the relevant parts can be referred to the method part description.

[0097] The above description of the disclosed embodiments enables those skilled in the art to implement or use the present application. Various modifications to these embodiments will be apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the present application. Therefore, the present application will not be limited to the embodiments shown herein, but will conform to the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. A method for constructing a bearing fault diagnosis model, characterized in that: The methods include: Determine a monitoring mode for the bearing operation scenario, wherein the monitoring mode includes a sensor device group and a sensor coordination mode; Interactive bearing fault records, mining directed graphs based on fault causality, wherein the directed graphs are marked with single fault labels and compound fault labels based on the monitoring mode, and labels are set based on whether the decision sensing data types of bearing faults are homologous; Traversing the directed graph, introducing a loss function to construct a feature equivalence space, wherein the loss function represents the redundant loss of features before and after equivalence; Determine the monitoring data flow structure based on the monitoring mode, perform data-driven single-source directional training and composite cross-training based on the directed graph, and generate a fault decision maker; Integrating the feature equivalent space and the fault decision maker to generate a bearing fault diagnosis model; The introduction of the loss function to construct the feature equivalent space includes: Traversing the bearing fault records, clustering based on the diagnosis results, and determining N clusters; For the N clusters, based on direct static redundancy and dynamic analytical redundancy, redundancy loss measurement is performed on monitoring features-fault features to construct the loss function; Based on the loss function, the feature equivalence space is constructed.

2. A method for constructing a bearing fault diagnosis model according to claim 1, characterized in that: The directed graph identifies single fault labels and compound fault labels based on the monitoring mode, including: Identify the monitoring mode, and use a single sensor data type decision as a single diagnosis, and use multiple sensor data types collaborative decision as a composite diagnosis; For a single diagnosis, traverse the directed graph and establish a mapping mark between monitoring features and fault features as the single fault label; For composite diagnosis, the directed graph is traversed to establish a mapping mark between monitoring features and fault features as the composite fault label.

3. The method for constructing a bearing fault diagnosis model according to claim 1, characterized in that: The redundant loss measurement of the monitoring feature-fault feature is performed to construct the loss function, including: Based on the direct static redundancy and the dynamic analytical redundancy, multiple levels of redundant nodes are defined, wherein a preset amount of data is used as the redundancy; Identify the directed graph, determine the decision process of the monitoring feature-fault feature, and insert the multi-level redundant nodes as decision process nodes; Identify a first cluster, and locate a first redundant node of the first cluster with decision accuracy as a constraint, wherein the first redundant node is a maximum redundant node level that satisfies the decision accuracy; The first loss function is determined based on the monitoring feature corresponding to the first cluster - the first redundant node and the feature redundancy loss.

4. The method for constructing a bearing fault diagnosis model according to claim 1, characterized in that: The data-driven single-source directional training includes: The monitoring data stream structure at least includes a temperature signal, a sound signal, a vibration signal and a surface image; Based on the monitoring data stream structure, the bearing fault records are divided to determine the directional training data; Based on the directed training data, traverse the directed graph and identify the single fault label, perform data integration, and determine pre-training data; Based on the pre-training data, supervised training of a directional decision maker is performed, wherein the directional decision maker corresponds one-to-one to the monitoring data stream structure.

5. A method for constructing a bearing fault diagnosis model according to claim 4, characterized in that: Perform compound cross-training exercises, including: Identify the composite fault label, and based on the composite fault label, perform data integration on the bearing fault record to determine composite training data; Based on the composite training data, cross-attention training is performed on the directional decision maker to generate the fault decision maker.

6. The method for constructing a bearing fault diagnosis model according to claim 1, characterized in that: After generating the bearing fault diagnosis model, it also includes: Conduct bearing operation monitoring and determine multi-source sensor data; Combined with the feature equivalent space of the bearing fault diagnosis model, data feature extraction and adaptive redundancy processing are performed to determine the effective feature set; The effective feature set is transferred to the fault decision maker to determine single faults and compound faults, and integrated to output a fault diagnosis list.

7. A system for constructing a bearing fault diagnosis model, characterized in that: A method for constructing a bearing fault diagnosis model according to any one of claims 1 to 6, the system comprising: A mode determination unit, the mode determination unit is used to determine a monitoring mode for a bearing operation scenario, wherein the monitoring mode includes a sensor device group and a sensor coordination mode; A graph mining unit, the graph mining unit is used to interact with bearing fault records and mine a directed graph based on fault causality, wherein the directed graph is marked with a single fault label and a compound fault label based on the monitoring mode, and the label is set based on whether the decision sensor data type of the bearing fault is homologous; An equivalent construction unit, the equivalent construction unit is used to traverse the directed graph and introduce a loss function to construct a feature equivalent space, wherein the loss function represents the redundant loss amount of the features before and after the equivalence; A training unit, the training unit is used to determine the monitoring data flow structure based on the monitoring mode, and to perform data-driven single-source directional training and composite cross-training based on the directed graph to generate a fault decision maker; A model generation unit is used to integrate the feature equivalent space and the fault decision maker to generate a bearing fault diagnosis model.

Images