Mechanical part life prediction method and device, computer equipment and storage medium

Abstract

The invention relates to a mechanical part life prediction method and device, computer equipment and a storage medium, and relates to the technical field of vehicle data processing. The method comprises the steps of determining health index features of a to-be-predicted component based on feature signals of the to-be-predicted component in a running state and recession knowledge corresponding to the to-be-predicted component; the recession knowledge is used for representing a corresponding relation between the characteristic signal of the to-be-predicted component in the operation process and the residual life of the component; performing recession feature iteration prediction based on health index moment features corresponding to at least two moments in the health index features to obtain health index moment features corresponding to prediction moments; determining a predicted health index value corresponding to the prediction moment according to the health index moment characteristics corresponding to the prediction moment; and under the condition that the predicted health index value represents that the to-be-predicted part is invalid, determining the predicted residual life of the to-be-predicted part according to the prediction moment. The method can improve the accuracy of predicting the service life of the mechanical part.

Description

Technical Field

[0001] The present application relates to the technical field of vehicle data processing, and in particular to a method, device, computer equipment and storage medium for predicting the life of a mechanical component. Background Art

[0002] The working environment of each mechanical component in the motor of a new energy vehicle is complex, and the performance of the mechanical components will be affected by many factors such as temperature, humidity, load, and speed. During the use of the car, it is necessary to predict the life of each mechanical component in the motor in real time to ensure vehicle safety. The above factors that affect the performance of mechanical components may affect the life of mechanical components, but it is currently difficult to accurately capture and quantify the impact of the above factors on the life of mechanical components. In addition, the life of mechanical components of different models and materials will also be different.

[0003] In the related art, the life of mechanical parts is mainly predicted based on statistical methods and empirical formulas, but the related art requires a large amount of real historical data as support. In addition, the related art usually only provides the total life of a certain type of mechanical part from the factory to the damage of the mechanical part to predict the remaining life of the mechanical part. It is impossible to predict the remaining life of any mechanical part based on the current performance data of the mechanical part, resulting in insufficient accuracy in predicting the remaining life of each independent mechanical part. Summary of the invention

[0004] Based on this, it is necessary to provide a mechanical component life prediction method, device, computer equipment, computer-readable storage medium and computer program product that can improve prediction accuracy in response to the above technical problems.

[0005] In a first aspect, the present application provides a method for predicting the life of a mechanical component. The method comprises:

[0006] Determine the health index characteristics of the component to be predicted based on the characteristic signal of the component to be predicted in the operating state and the decay knowledge corresponding to the component to be predicted; the decay knowledge is used to represent the corresponding relationship between the characteristic signal of the component to be predicted in the operating process and the remaining life of the component, and the health index characteristics include health index time characteristics corresponding to at least two moments in the operating state;

[0007] Performing iterative prediction of the decline characteristics based on the health indicator moment characteristics corresponding to the at least two moments in the health indicator characteristics to obtain the health indicator moment characteristics corresponding to the predicted moment; in each iterative prediction of the decline characteristics, determining the decline fluctuation characteristics at the current moment based on the decline fluctuation characteristics at the previous moment, and determining the decline trend characteristics at the current moment based on the decline fluctuation characteristics at the current moment, the decline trend characteristics at the previous moment and the health indicator moment characteristics at the current moment;

[0008] Determining a predicted health indicator value corresponding to the predicted time according to the health indicator moment characteristic corresponding to the predicted time;

[0009] In a case where the predicted health index value indicates that the component to be predicted has failed, the predicted remaining life of the component to be predicted is determined according to the predicted time.

[0010] In one embodiment, the iterative prediction of the decay characteristics based on the health indicator time characteristics corresponding to the at least two moments in the health indicator characteristics to obtain the health indicator time characteristics corresponding to the predicted moment includes:

[0011] For a first moment of the at least two moments, determining a health indicator moment characteristic corresponding to the first moment from the health indicator characteristics, and obtaining a decline trend characteristic corresponding to the first moment according to the health indicator moment characteristic corresponding to the first moment and the decline fluctuation characteristic of the first moment;

[0012] At each current moment starting from the second moment of the at least two moments, determining the decay fluctuation feature of the current moment according to the decay fluctuation feature corresponding to the previous moment of the current moment;

[0013] Obtaining the recession trend characteristics corresponding to the current moment according to the recession fluctuation characteristics at the current moment and the health indicator moment characteristics corresponding to the current moment;

[0014] In the case that the current moment does not belong to the at least two moments, the current moment is used as the predicted moment, and based on the decay trend characteristics corresponding to the current moment, the health indicator moment characteristics of the predicted moment are obtained.

[0015] In one embodiment, for each current moment starting from the second moment of the at least two moments, determining the decay fluctuation characteristics of the current moment according to the decay fluctuation characteristics corresponding to the previous moment of the current moment includes:

[0016] For each current moment starting from the second moment of the at least two moments, determining a first previous moment of the current moment and a second previous moment before the first previous moment;

[0017] determining the moment fluctuation feature of the current moment according to the decay trend feature of the first previous moment and the decay trend feature of the second previous moment;

[0018] The decay fluctuation characteristic at the current moment is obtained according to the moment fluctuation characteristic and the decay fluctuation characteristic at the first previous moment.

[0019] In one embodiment, obtaining the decay trend feature corresponding to the current moment according to the decay fluctuation feature at the current moment and the health indicator moment feature corresponding to the current moment includes:

[0020] Based on the attention mechanism, obtaining an intermediate attention feature according to the decay fluctuation feature at the current moment and the decay trend feature at the first previous moment;

[0021] Obtaining an intermediate fusion feature according to the intermediate attention feature and the decay trend feature of the first prior moment;

[0022] According to the intermediate fusion feature and the health indicator moment feature corresponding to the current moment, the decline trend feature corresponding to the current moment is obtained.

[0023] In one embodiment, the method further comprises:

[0024] In the case where the current moment belongs to the at least two moments, determining the health indicator moment feature corresponding to the current moment from the health indicator features;

[0025] In a case where the current moment does not belong to the at least two moments, the health indicator moment feature corresponding to the current moment is obtained according to the decay trend feature of the first previous moment.

[0026] In one embodiment, determining the health indicator characteristics of the component to be predicted based on the characteristic signal of the component to be predicted in the operating state and the decay knowledge corresponding to the component to be predicted includes:

[0027] Acquire characteristic signals of the component to be predicted in the running state, and obtain target characteristic signals corresponding to at least two moments from the characteristic signals;

[0028] Obtaining an initial feature according to the target feature signals corresponding to each of the at least two moments;

[0029] The health indicator characteristics of the component to be predicted are obtained based on the initial characteristics through the trained decay knowledge adaptation model; the trained decay knowledge adaptation model is obtained based on the decay knowledge training.

[0030] In one embodiment, the method further comprises:

[0031] Acquire a historical feature signal corresponding to a mechanical component associated with the component to be predicted, obtain target historical feature signals corresponding to at least two historical moments from the historical feature signal, and obtain historical initial features according to the target historical feature signals corresponding to the at least two historical moments;

[0032] The historical health index feature and the historical intermediate feature of the mechanical component are obtained based on the historical initial feature by using the decay knowledge adaptation model to be trained; the historical intermediate feature is obtained by encoding the decay knowledge adaptation model to be trained based on the historical initial feature, and the historical health index feature is obtained by decoding the decay knowledge adaptation model to be trained based on the historical intermediate feature;

[0033] Obtaining a health indicator loss according to the historical health indicator feature, the historical intermediate feature, the historical initial feature, and a decay constraint condition determined based on the decay knowledge;

[0034] The decay knowledge adaptation model to be trained is updated according to the health indicator loss and then continued to be trained until a training end condition is met, thereby obtaining the trained decay knowledge adaptation model.

[0035] In one embodiment, the health indicator loss is obtained according to the historical health indicator feature, the historical intermediate feature, the historical initial feature and the decay constraint condition determined based on the decay knowledge, including:

[0036] For each of the at least two historical moments, determine the historical health indicator moment feature of the historical moment from the historical health indicator features, determine the intermediate feature of the historical moment from the historical intermediate features, and determine the initial feature of the historical moment from the historical initial features;

[0037] Obtaining, according to the initial features of the at least two historical moments in the historical initial features, a first feature statistical parameter of the historical initial features, and determining a first loss based on the first feature statistical parameter;

[0038] Obtaining a second loss according to a feature difference between the historical health indicator moment feature at the historical moment and the initial feature at the historical moment;

[0039] Obtaining a third loss according to the intermediate feature of the historical moment, the intermediate feature of the moment adjacent to the historical moment, and a decay constraint parameter determined based on the decay knowledge;

[0040] Based on the first loss, the second loss and the third loss, obtaining the target loss for the historical moment;

[0041] A health indicator loss is obtained according to the target losses of each of the at least two historical moments.

[0042] In one embodiment, the health indicator feature of the component to be predicted is obtained based on the initial feature by using the trained decay knowledge adaptation model, including:

[0043] Inputting the initial features into the trained decay knowledge adaptation model to obtain initial health indicator features corresponding to the initial features;

[0044] According to the second characteristic statistical parameter corresponding to the trained decay knowledge adaptation model, the initial health indicator characteristic is mapped into the health indicator characteristic of the component to be predicted;

[0045] The second characteristic statistical parameter is obtained based on the historical intermediate characteristic statistics corresponding to each of the historical characteristic signals.

[0046] In one embodiment, obtaining the initial feature according to the target feature signals corresponding to the at least two moments respectively includes:

[0047] According to the transformation method corresponding to the first dimension, the target feature signals corresponding to the at least two moments are transformed to obtain the initial features in the first dimension;

[0048] The method further comprises:

[0049] According to the predicted remaining life, obtaining a first predicted life of the component to be predicted in the first dimension;

[0050] Acquire a second predicted life of the component to be predicted in at least one second dimension, where the second dimension is different from the first dimension;

[0051] According to the first predicted life and the second predicted life, a comprehensive predicted remaining life of the component to be predicted is obtained.

[0052] In one embodiment, determining the predicted health indicator value corresponding to the predicted moment according to the health indicator moment characteristic corresponding to the predicted moment includes:

[0053] The health indicator moment feature corresponding to the predicted moment is mapped into a predicted health indicator value corresponding to the predicted moment in numerical form by a preset feature mapping method.

[0054] In one of the embodiments, when the predicted health index value indicates failure of the component to be predicted, determining the predicted remaining life of the component to be predicted according to the prediction time includes:

[0055] In the case where the predicted health index value represents the failure of the component to be predicted, determining the time span between the prediction moment and the acquisition moment of the characteristic signal;

[0056] The predicted remaining life of the component to be predicted at the acquisition moment is determined according to the time span.

[0057] In a second aspect, the present application also provides a mechanical component life prediction device. The device comprises:

[0058] A first feature determination module is used to determine the health index feature of the component to be predicted based on the feature signal of the component to be predicted in the operating state and the decay knowledge corresponding to the component to be predicted; the decay knowledge is used to represent the corresponding relationship between the feature signal of the component to be predicted in the operating process and the remaining life of the component, and the health index feature includes the health index moment features corresponding to at least two moments in the operating state;

[0059] A second feature prediction module is used to perform iterative prediction of the decline feature based on the health indicator moment features corresponding to the at least two moments in the health indicator features, and obtain the health indicator moment features corresponding to the prediction moment; in each iterative prediction of the decline feature, the decline fluctuation feature at the current moment is determined based on the decline fluctuation feature at the previous moment, and the decline trend feature at the current moment is determined based on the decline fluctuation feature at the current moment, the decline trend feature at the previous moment and the health indicator moment feature at the current moment;

[0060] An indicator value determination module, used to determine the predicted health indicator value corresponding to the predicted moment according to the health indicator moment characteristic corresponding to the predicted moment;

[0061] The life determination module is used to determine the predicted remaining life of the component to be predicted according to the prediction time when the predicted health index value indicates that the component to be predicted has failed.

[0062] In a third aspect, the present application further provides a computer device, wherein the computer device comprises a memory and a processor, wherein the memory stores a computer program, and when the processor executes the computer program, the steps of the method described in the first aspect are implemented.

[0063] In a fourth aspect, the present application further provides a computer-readable storage medium having a computer program stored thereon, and when the computer program is executed by a processor, the steps of the method described in the first aspect are implemented.

[0064] In a fifth aspect, the present application further provides a computer program product, wherein the computer program product comprises a computer program, and when the computer program is executed by a processor, the steps of the method described in the first aspect are implemented.

[0065] The above-mentioned mechanical component life prediction method, device, computer equipment, storage medium and computer program product determine the health index characteristics of the component to be predicted through the characteristic signal of the component to be predicted in the operating state and the decay knowledge corresponding to the component to be predicted, and obtain the health index time characteristics corresponding to at least two moments in the operating state. The health index characteristics contain the decay knowledge of the component to be predicted, which can represent the correspondence between the characteristic signal and the remaining life of the component. Based on this, the decay characteristics can be iteratively predicted based on the health index characteristics corresponding to at least two moments, and the health index time characteristics corresponding to the prediction moment can be obtained. In the iterative prediction of the decay characteristics, the decay fluctuation characteristics of the current moment can be determined based on the decay fluctuation characteristics of the previous moment, and the decay trend characteristics of the current moment can be determined based on the decay fluctuation characteristics of the current moment, the decay trend characteristics of the previous moment and the health index time characteristics of the current moment, so as to mine the decay trend and decay fluctuation corresponding to each moment in the operating state of the component to be predicted to the prediction moment, so as to improve the accuracy of determining the health index time characteristics corresponding to the prediction moment.

[0066] This method determines the health index characteristics of the predicted component at each future prediction moment based on the characteristic signal and decay knowledge of the predicted component in the running state through iterative prediction, and determines the predicted health index value corresponding to the health index characteristics at each future prediction moment. According to the corresponding relationship between the characteristic signal and the remaining life of the component, the decay trend and decay fluctuation of the predicted component at each future moment can be determined, and the predicted health index value at the prediction moment can be accurately obtained. Afterwards, by using the predicted health index value at the prediction moment, it is determined whether the predicted component fails at the prediction moment, and the predicted remaining life of the predicted component can be determined, thereby improving the accuracy of mechanical component life prediction. BRIEF DESCRIPTION OF THE DRAWINGS

[0067] In order to more clearly illustrate the technical solutions in the embodiments of the present application or the related technologies, the drawings required for use in the embodiments or the related technical descriptions are briefly introduced below. Obviously, the drawings described below are only some embodiments of the present application. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative work.

[0068] Figure 1 A diagram showing an application environment of a method for predicting the life of a mechanical component in an embodiment;

[0069] Figure 2 A schematic diagram of a flow chart of a method for predicting the life of a mechanical component in one embodiment;

[0070] Figure 3 A schematic diagram of a flow chart of a step of determining a health indicator moment characteristic corresponding to a prediction moment in an embodiment;

[0071] Figure 4 A schematic diagram of a flow chart of a step of determining a decay fluctuation characteristic at a current moment in an embodiment;

[0072] Figure 5 A schematic diagram of a flow chart of a step of determining a characteristic decline trend at a current moment in an embodiment;

[0073] Figure 6 A schematic diagram of a flow chart of a health indicator characteristic step of a component to be predicted in an embodiment;

[0074] Figure 7 A flowchart of the steps of determining a trained decay knowledge adaptation model in one embodiment;

[0075] Figure 8 A schematic diagram of a process for determining a health indicator loss step in one embodiment;

[0076] Fig. 9 A schematic diagram of the structure of a decay knowledge adaptation model in one embodiment;

[0077] Fig.10 A schematic diagram of the structure of a dual-line gated fusion learning model in an embodiment;

[0078] Fig.11 is a structural block diagram of a mechanical component life prediction device in one embodiment;

[0079] Fig.12 FIG. 4 is a diagram showing the internal structure of a computer device in one embodiment. DETAILED DESCRIPTION

[0080] In order to make the purpose, technical solution and advantages of the present application more clearly understood, the present application is further described in detail below in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present application and are not used to limit the present application.

[0081] The mechanical component life prediction method provided in the embodiment of the present application can be applied to Figure 1In the application environment shown. Among them, the terminal 102 communicates with the server 104 through the network. The data storage system can store the data that the server 104 needs to process. The data storage system can be integrated on the server 104, or it can be placed on the cloud or other network servers. The terminal 102 can send the acquired characteristic signals of each component to be predicted in the running state to the server 104. The server 104 can determine the health indicator characteristics of the component to be predicted in the running state based on the characteristic signals and decay knowledge corresponding to the component to be predicted. The health indicator characteristics include health indicator moment characteristics corresponding to at least two moments. The server 104 can perform iterative prediction of decay characteristics for the health indicator moment characteristics of at least two moments, obtain the health indicator moment characteristics corresponding to the future prediction moment, and determine the predicted health indicator value at the prediction moment. During the iterative prediction of the decay characteristics, the server can continuously predict the predicted health indicator values ​​at each prediction moment in the future. If the predicted health indicator value at a certain target prediction moment in the future indicates that the component to be predicted has failed, the predicted remaining life of the component to be predicted can be determined according to the target prediction moment.

[0082] It should be understood that when the server 104 determines the health indicator moment characteristics corresponding to the predicted moment, it is obtained by predicting the decay fluctuations and decay trends corresponding to the previous moment and the current moment, respectively, and the server can continuously predict the health indicator moment characteristics corresponding to each future moment through iterative prediction. In one example, the server determines the decay fluctuation characteristics of the current moment through the decay fluctuation characteristics of the previous moment, and determines the decay trend characteristics of the current moment through the decay fluctuation characteristics of the current moment, the decay trend characteristics of the previous moment, and the health indicator moment characteristics, and then uses the health indicator moment characteristics corresponding to at least two moments to determine the health indicator moment characteristics corresponding to the future predicted moment.

[0083] The terminal 102 may be a new energy vehicle, and obtains characteristic signals of each mechanical component from the sensor of the new energy vehicle, and synchronously sends the characteristic signals to the server 104, which uniformly predicts the corresponding life of the mechanical components in each operating state. In addition, in addition to being a new energy vehicle, the terminal 102 may also be, but is not limited to, various personal computers, laptops, smart phones, tablet computers, and Internet of Things devices, which may be smart speakers, smart TVs, smart air conditioners, smart car-mounted devices, etc. The server 104 may be implemented as an independent server or a server cluster consisting of multiple servers.

[0084] In an exemplary embodiment, Figure 2 As shown in the figure, a method for predicting the life of mechanical parts is provided. Figure 1 The server 104 in the example is used as an example to illustrate the method, which includes the following steps S202 to S208. Among them:

[0085] Step S202: determining the health indicator characteristics of the component to be predicted based on the characteristic signal of the component to be predicted in the operating state and the decay knowledge corresponding to the component to be predicted.

[0086] Among them, the decay knowledge is used to represent the correspondence between the characteristic signal of the component to be predicted during operation and the remaining life of the component, and the health indicator characteristics include the health indicator moment characteristics corresponding to at least two moments in the operating state. The operating state is used to represent the state corresponding to the component to be predicted when the device or terminal to which the component to be predicted belongs is in operation. The characteristic signal is used to represent the physical signal generated by the component to be predicted during operation. For example, the characteristic signal can be a vibration signal of the component to be predicted in the operating state, a current signal and / or a voltage signal inside the component to be predicted. It should be understood that the decay knowledge corresponding to each type of component to be predicted is different depending on the model or material of the component to be predicted. Components of the same model or material can use the same decay knowledge.

[0087] Specifically, the server can obtain the characteristic signal of the component to be predicted obtained by the terminal sampling in the running state. In one example, the characteristic signal can be the vibration signal of the component to be predicted. The server can determine the health indicator characteristics corresponding to the time period based on the vibration signal in a certain time period under the running state and the decay knowledge corresponding to the component to be predicted. For this time period, the server can determine the health indicator moment characteristics corresponding to at least two moments in the time period. After arranging multiple health indicator moment characteristics in chronological order, the health indicator characteristics corresponding to the component to be predicted can be obtained.

[0088] Step S204, performing iterative prediction of decay characteristics based on health indicator time characteristics corresponding to at least two time points in the health indicator characteristics, and obtaining health indicator time characteristics corresponding to the predicted time point.

[0089] Among them, the iterative prediction of decay characteristics is used to mine and predict the decay characteristics implicit in the health indicator moment characteristics, and based on the decay characteristics predicted in each iteration, predict the decay characteristics corresponding to the next iteration, so as to obtain the health indicator moment characteristics corresponding to multiple consecutive moments. Decay characteristics may include decay fluctuation characteristics and decay trend characteristics. Decay fluctuation characteristics are local and unstable fluctuation characteristics of the component to be predicted in its life cycle. For example, the decay fluctuation characteristics may change dramatically in multiple consecutive moments; the decay trend characteristics are the overall degradation trend characteristics of the component to be predicted in its life cycle. For example, the decay trend characteristics may represent a gentle decay process in multiple moments corresponding to the life cycle. The prediction moment may be any moment after at least two moments.

[0090] In each iterative prediction of the recession characteristics, the recession fluctuation characteristics at the current moment are determined based on the recession fluctuation characteristics at the previous moment, and the recession trend characteristics at the current moment are determined based on the recession fluctuation characteristics at the current moment, the recession trend characteristics at the previous moment and the health indicator moment characteristics at the current moment.

[0091] In one example, each iterative prediction of the decay feature may include the decay fluctuation feature and the decay trend feature corresponding to the two moments of the previous moment and the current moment, respectively, wherein the decay fluctuation feature of the previous moment and the decay trend feature of the previous moment are known quantities, and the decay fluctuation feature of the current moment and the decay trend feature of the current moment are unknown quantities. The server can predict the decay fluctuation feature of the next moment, that is, the current moment, based on the decay fluctuation feature of the previous moment. The server can process the decay trend feature of the previous moment, the decay fluctuation feature of the current moment, and the health indicator moment feature of the current moment, and merge the fluctuation features, trend features and health indicator moment features involved in the above different moments to obtain the decay trend feature of the current moment. In general, the decay fluctuation feature, the decay trend feature of the previous moment, and the health indicator moment feature of the current moment are predicted to obtain the decay trend feature of the current moment. Based on this, the server can predict the health indicator moment feature corresponding to the predicted moment through the health indicator moment features in at least two moments.

[0092] Specifically, the server can perform iterative prediction of the decline characteristics of the health indicator moment characteristics corresponding to at least two moments in the health indicator characteristics in chronological order, and each iterative prediction of the decline characteristics can be based on the prediction process between the previous moment and the current moment in the above example to obtain the decline trend characteristics corresponding to the current moment. In the subsequent iteration process, the current moment can be regarded as the previous moment, and the same prediction process is used to predict the decline trend characteristics corresponding to the current moment in the next iteration, thereby obtaining the decline trend characteristics of one or more predicted moments.

[0093] Step S206, determining the predicted health index value corresponding to the predicted time according to the health index time characteristic corresponding to the predicted time.

[0094] Specifically, when a health indicator moment feature corresponding to a predicted moment is obtained through each iteration, the server can determine the predicted health indicator value corresponding to the predicted moment. The server can map the health indicator moment feature according to a predefined first mapping strategy to obtain the corresponding predicted health indicator value.

[0095] Step S208, when the predicted health index value indicates failure of the component to be predicted, determining the predicted remaining life of the component to be predicted according to the prediction time.

[0096] Specifically, the predicted health index value can represent the health state of the component to be predicted at a future time. When the predicted health index value indicates the failure of the component to be predicted, the server can determine that the predicted time corresponding to the predicted health index value is the time when the component to be predicted fails, that is, the time corresponding to the end of its life. The server can determine the predicted remaining life of the component to be predicted based on the time corresponding to the end of its life.

[0097] In the above-mentioned mechanical component life prediction method, the health index characteristics of the component to be predicted are determined by the characteristic signal of the component to be predicted in the operating state and the decay knowledge corresponding to the component to be predicted, and the health index time characteristics corresponding to at least two moments in the operating state are obtained. The health index characteristics contain the decay knowledge of the component to be predicted, which can represent the correspondence between the characteristic signal and the remaining life of the component. Based on this, the decay characteristics can be iteratively predicted based on the health indicator characteristics corresponding to at least two moments, and the health indicator time characteristics corresponding to the prediction moment can be obtained. In the iterative prediction of the decay characteristics, the decay fluctuation characteristics of the current moment can be determined based on the decay fluctuation characteristics of the previous moment, and the decay trend characteristics of the current moment can be determined based on the decay fluctuation characteristics of the current moment, the decay trend characteristics of the previous moment and the health indicator time characteristics of the current moment, so as to mine the decay trend and decay fluctuation corresponding to each moment in the operating state of the component to be predicted to the prediction moment, so as to improve the accuracy of determining the health indicator time characteristics corresponding to the prediction moment.

[0098] This method determines the health index characteristics of the predicted component at each future prediction moment based on the characteristic signal and decay knowledge of the predicted component in the running state through iterative prediction, and determines the predicted health index value corresponding to the health index characteristics at each future prediction moment. According to the corresponding relationship between the characteristic signal and the remaining life of the component, the decay trend and decay fluctuation of the predicted component at each future moment can be determined, and the predicted health index value at the prediction moment can be accurately obtained. Afterwards, by using the predicted health index value at the prediction moment, it is determined whether the predicted component fails at the prediction moment, and the predicted remaining life of the predicted component can be determined, thereby improving the accuracy of mechanical component life prediction.

[0099] In an exemplary embodiment, Figure 3 As shown, the specific implementation process of the step of "iteratively predicting the decay characteristics based on the health indicator moment characteristics corresponding to at least two moments in the health indicator characteristics to obtain the health indicator moment characteristics corresponding to the prediction moment" includes steps S302 to S308. Among them:

[0100] Step S302, for the first moment of at least two moments, determine the health indicator moment feature corresponding to the first moment from the health indicator features, and obtain the decline trend feature corresponding to the first moment based on the health indicator moment feature corresponding to the first moment and the decline fluctuation feature of the first moment.

[0101] The first moment is the first element in the time series composed of at least two moments. Since there is no previous moment before the first moment, the decay trend characteristics and decay fluctuation characteristics of the previous moment corresponding to the first moment can be determined by initialization.

[0102] Specifically, the server can obtain the health indicator moment characteristics corresponding to the first moment from the health indicator characteristics, and obtain the decay fluctuation characteristics of the previous moment corresponding to the first moment and the decay trend characteristics of the previous moment corresponding to the first moment through initialization. The server can determine the decay fluctuation characteristics corresponding to the first moment based on the decay fluctuation characteristics of the previous moment corresponding to the first moment. Based on this, the server can determine the decay trend characteristics corresponding to the first moment based on the decay fluctuation characteristics corresponding to the first moment, the decay trend characteristics of the previous moment corresponding to the first moment, and the health indicator moment characteristics of the first moment.

[0103] Step S304, for each current moment starting from the second moment of the at least two moments, the decay fluctuation characteristics of the current moment are determined according to the decay fluctuation characteristics corresponding to the previous moment of the current moment.

[0104] Specifically, each moment after the first moment has a corresponding previous moment, and the server can perform iterative prediction of the decay feature for each current moment starting from the second moment of at least two moments. For each current moment, the server can determine the decay fluctuation feature corresponding to the previous moment of the current moment, and mine the decay fluctuation feature of the current moment based on the decay fluctuation feature corresponding to the previous moment to obtain the decay fluctuation feature of the current moment.

[0105] Step S306, obtaining the decay trend characteristics corresponding to the current moment according to the decay fluctuation characteristics at the current moment and the health indicator time characteristics corresponding to the current moment.

[0106] Specifically, for each current moment, the server may fuse the decay trend characteristics of the previous moment corresponding to the current moment, the decay fluctuation characteristics of the current moment, and the health indicator moment characteristics corresponding to the current moment to obtain the decay trend characteristics corresponding to the current moment.

[0107] Step S308, when the current moment does not belong to at least two moments, the current moment is used as the predicted moment, and based on the decay trend characteristics corresponding to the current moment, the health indicator moment characteristics of the predicted moment are obtained.

[0108] Specifically, the server can determine whether the current moment belongs to at least two moments previously defined. If the current moment does not belong to at least two moments, the current moment is determined to be a predicted moment, and the server can map the decay trend feature corresponding to the current moment to obtain the health indicator moment feature of the predicted moment. If the current moment belongs to at least two moments, the decay feature iterative prediction is continued until the current moment does not belong to at least two moments.

[0109] In this embodiment, the first moment of at least two moments is initialized, and the first iterative prediction of the decay characteristics is performed based on the result of the initialization. Based on this, the decay characteristics of each current moment after the first moment can be iteratively predicted to obtain the decay trend characteristics corresponding to each current moment. When the current moment after the iteration does not belong to at least two moments, the current moment is determined to be the predicted moment, and the health indicator moment characteristics of the predicted moment are obtained. The decay trend characteristics of subsequent predicted moments can be iteratively predicted through the decay trend characteristics and decay fluctuation characteristics of consecutive moments, thereby obtaining the health indicator moment characteristics of the predicted moments, thereby improving the accuracy of determining the health indicator moment characteristics.

[0110] In an exemplary embodiment, Figure 4 As shown, the specific implementation process of the step of "determining the decay fluctuation characteristics of the current moment according to the decay fluctuation characteristics corresponding to the previous moment of the current moment at each current moment starting from the second moment of at least two moments" includes steps S402 to S406. Among them:

[0111] Step S402: for each current moment starting from the second moment of at least two moments, determine a first previous moment of the current moment and a second previous moment before the first previous moment.

[0112] Specifically, each current moment starting from the second moment of at least two moments is traversed to determine the first previous moment and the second previous moment of the current moment. The second previous moment is before the first previous moment, and the first previous moment is before the current moment. In one example, the server may determine that the current moment is t, the first previous moment is t-1, and the second previous moment is t-2.

[0113] Step S404, determining the moment fluctuation characteristics of the current moment according to the decay trend characteristics of the first previous moment and the decay trend characteristics of the second previous moment.

[0114] The moment fluctuation feature at the current moment is an intermediate variable corresponding to the decay fluctuation feature at the current moment. Before determining the decay fluctuation feature at the current moment, it is necessary to determine the moment fluctuation feature at the current moment.

[0115] Specifically, since the decay feature is predicted iteratively according to the time sequence, the decay trend features and the decay fluctuation features at each moment before the current moment are known quantities. The server can obtain the decay trend features of the first previous moment and the decay trend features of the second previous moment. The server can map the decay trend features of the first previous moment and the decay trend features of the second previous moment through a predefined second mapping strategy to obtain the moment fluctuation features of the current moment.

[0116] In one example, the predefined second mapping strategy may be to perform a difference processing between the decay trend feature of the first previous moment and the decay trend feature of the second previous moment, and determine the difference as the moment fluctuation feature of the current moment. t is the momentary fluctuation characteristic of the current moment, h t-1 is the recession trend characteristic of the first prior moment, h t-2 is the decay trend characteristic of the second previous moment. The specific implementation process of determining the moment fluctuation characteristic of the current moment may be: .

[0117] Step S406, obtaining the decay fluctuation characteristics at the current moment according to the moment fluctuation characteristics and the decay fluctuation characteristics at the first previous moment.

[0118] Specifically, the server may connect the moment fluctuation feature and the decay fluctuation feature of the first previous moment, that is, splice the moment fluctuation feature and the decay fluctuation feature of the first previous moment to obtain a first spliced ​​feature. The server processes the first spliced ​​feature through the update gate and the reset gate in the first GRU (Gated Recurrent Unit) to obtain a first update gate feature and a first reset gate feature. The server may fuse the first update gate feature and the first reset gate feature to obtain the decay fluctuation feature of the current moment. In an example, the process of splicing the moment fluctuation feature and the decay fluctuation feature of the first previous moment to obtain the first spliced ​​feature is: Where N is the decay fluctuation feature. The process of the server determining the first update gate feature and the first reset gate feature through the first GRU is: , .in, is the first update gate feature determined by the update gate, is a first reset gate characteristic determined by the reset gate, and is a constant, is the activation function, is the linear transformation matrix of the first update gate, is the bias of the first update gate, is the linear transformation matrix of the first reset gate, is the bias of the first reset gate, A is used to indicate the first, to distinguish it from the update gate and reset gate that appear later, t is used to indicate the time, subscript z is used to indicate the update gate, and subscript r is used to indicate the reset gate.

[0119] In addition, in one example, the process of fusing the first update gate feature and the first reset gate feature is as follows: the server may fuse the first reset gate feature corresponding to the current moment and the decay fluctuation feature of the first previous moment, and splice the fused result with the moment fluctuation feature of the current moment to obtain a second splicing feature. The server may map the second splicing feature based on a predefined third mapping strategy to obtain a mapped second splicing feature. The server may fuse the first update gate feature corresponding to the current moment and the decay fluctuation feature of the first previous moment to obtain a first fusion result; and fuse the first update gate feature corresponding to the current moment and the mapped second splicing feature to obtain a second fusion result, and add the first fusion result and the second fusion result to obtain the decay fluctuation feature of the current moment. The formula for fusing the first update gate feature and the first reset gate feature of the server is as follows:

[0120] ,

[0121] Among them, the server can determine D t is the momentary fluctuation characteristic of the current moment, N t-1 is the recession fluctuation characteristic of the first prior moment, is the first update gate feature, is the first reset gate feature, To obtain the intermediate eigenvectors required to calculate the decay wave characteristics, is the linear transformation matrix for feature fusion, is the bias, N t It is the recession fluctuation characteristic at the current moment.

[0122] In this embodiment, the moment fluctuation feature of the current moment is determined by the decay trend feature corresponding to the first previous moment and the decay trend feature corresponding to the second previous moment, and the moment fluctuation feature and the decay fluctuation feature of the first previous moment are spliced ​​to obtain a first spliced ​​feature, and the first update gate feature and the first reset gate feature corresponding to the first spliced ​​feature are respectively determined by the update gate and the reset gate in the first GRU, and finally the first update gate feature and the first reset gate feature are fused to obtain the decay fluctuation feature corresponding to the current moment, which can fully explore the moment fluctuation feature of the current moment in the decay trend feature of the first previous moment and the decay trend feature of the second previous moment, and splice the moment fluctuation feature with the decay fluctuation feature of the first previous moment, retain and fuse the decay fluctuation of the previous moment, and thus improve the accuracy of determining the decay fluctuation feature of the current moment.

[0123] In an exemplary embodiment, Figure 5 As shown, the specific implementation process of the step "obtaining the corresponding recession trend characteristics at the current moment according to the recession fluctuation characteristics at the current moment and the corresponding health index time characteristics at the current moment" includes steps S502 to S506. Among them:

[0124] Step S502, based on the attention mechanism, obtains the intermediate attention feature according to the decay fluctuation feature at the current moment and the decay trend feature at the first previous moment.

[0125] Specifically, the server can add an attention mechanism, and process the decay fluctuation characteristics of the current moment through the attention weight in the attention mechanism to obtain the intermediate attention characteristics of the current moment. Among them, the attention weights of the decay fluctuation characteristics corresponding to each current moment in the attention mechanism can be determined based on the decay trend characteristics of each first previous moment. The process of determining the intermediate attention characteristics is as follows:

[0126]

[0127] in, is the intermediate attention feature, N t is the recession fluctuation characteristic at the current moment, is the attention weight.

[0128] In one example, the construction process of the attention mechanism is as follows: for each current moment, the server can determine the scoring function of the attention mechanism based on the decay trend characteristics of the first previous moment and the decay fluctuation characteristics of the current moment. For each current moment, the server can determine the attention weight corresponding to the current moment based on the score value of the current moment in the scoring function and the ratio of the total score values ​​of each current moment, thereby obtaining the attention weight corresponding to each current moment.

[0129] The process of determining the attention weight corresponding to each current moment in the attention mechanism is as follows: ,

[0130] in, is the scoring function, is the attention weight, , , They are all linear transformation matrices that can be learned during model training.

[0131] Step S504, obtaining an intermediate fusion feature according to the intermediate attention feature and the decay trend feature of the first previous moment.

[0132] Specifically, the server may concatenate the intermediate attention feature at the current moment with the decay trend feature at the first previous moment to obtain a third concatenated feature, and map the third concatenated feature based on a predefined fourth mapping strategy to obtain an intermediate fusion feature.

[0133] In one example, the process of determining the intermediate fusion features is as follows:

[0134]

[0135] in, is the intermediate fusion feature, is the intermediate attention feature, h t-1 is the recession trend characteristic of the first prior moment, is the activation function, is the linear transformation matrix that can be learned during model training. The subscript m is used to distinguish the linear transformation matrix and bias corresponding to the intermediate fusion feature.

[0136] Step S506, obtaining the decline trend feature corresponding to the current moment according to the intermediate fusion feature and the health indicator moment feature corresponding to the current moment.

[0137] Specifically, the server may concatenate the intermediate fusion feature with the health indicator moment feature corresponding to the current moment to obtain a fourth concatenated feature. The server processes the fourth concatenated feature through the second update gate and the second reset gate in the second GRU, respectively, to obtain a second update gate feature and a second reset gate feature. The server may fuse the second update gate feature and the second reset gate feature to obtain the decay trend feature at the current moment. In one example, the process of determining the second update gate feature and the second reset gate feature is as follows: , .

[0138] in, is the second update gate feature, is the second reset gate feature, is the momentary feature of the health indicator corresponding to the current moment, is the intermediate fusion feature, and is the linear transformation matrix that can be learned during model training. and For offset, the superscript M is distinguished from the above-mentioned A and is used to indicate the second.

[0139] In this embodiment, the decay fluctuation characteristics of the current moment are processed through the attention mechanism, which can improve the accuracy of the intermediate attention characteristics, and determine the intermediate fusion characteristics based on the intermediate attention, and process the intermediate fusion characteristics through the second GRU to determine the second update gate characteristics and the second reset gate characteristics respectively. Finally, the second update gate characteristics and the second reset gate characteristics are fused to obtain the decay trend characteristics of the current moment. The decay trend characteristics can be fully mined in the intermediate fusion characteristics and the current moment characteristics of health indicators, thereby improving the accuracy of determining the decay trend characteristics of the current moment.

[0140] In an exemplary embodiment, the method also includes: when the current moment belongs to at least two moments, determining the health indicator moment characteristics corresponding to the current moment from the health indicator characteristics; when the current moment does not belong to at least two moments, obtaining the health indicator moment characteristics corresponding to the current moment based on the decay trend characteristics of the first previous moment.

[0141] Specifically, in the process of iterative prediction of the decay feature, if the current moment iterated to belongs to at least two moments, it is determined that the server has not iterated to the predicted moment, and the server can determine the health indicator moment feature corresponding to the current moment in the health indicator feature. In the process of iterative prediction of the decay feature, if the current moment iterated to does not belong to at least two moments, it is determined that the server has iterated to the predicted moment, and the health indicator moment feature corresponding to the current moment does not exist in the health indicator feature. The server can obtain the health indicator moment feature corresponding to the current moment based on the decay trend feature of the first previous moment corresponding to the current moment. In one example, the server can determine the decay trend feature of the first previous moment as the health indicator moment feature corresponding to the current moment.

[0142] In this embodiment, by determining the health indicator moment feature corresponding to the current moment from the health indicator features when the current moment belongs to at least two moments, the health indicator features can be fully feature mined, and when the current moment does not belong to at least two moments, the health indicator moment feature corresponding to the current moment can be obtained based on the decline trend feature of the first previous moment. The health indicator moment feature corresponding to the current moment can be determined based on the decline trend feature mined at the previous moment, thereby improving the accuracy of determining the health indicator moment feature.

[0143] In an exemplary embodiment, Figure 6 As shown, the specific implementation process of the step "determining the health indicator characteristics of the component to be predicted based on the characteristic signal of the component to be predicted in the operating state and the decay knowledge corresponding to the component to be predicted" includes steps S602 to S606. Among them:

[0144] Step S602: Acquire a characteristic signal of the component to be predicted in operation, and obtain target characteristic signals corresponding to at least two moments from the characteristic signal.

[0145] Specifically, the server can obtain the characteristic signal of the component to be predicted within a certain period of time in the running state through sensor sampling. For example, the characteristic signal can be sampled from time t1 to time t n Characteristic signal at the end of sampling For example, the characteristic signal may be a vibration signal a1 with a sequence length of 7 obtained by sampling from time t1. The characteristic signal may be a vibration signal a1 with a sequence length of 7 obtained by sampling from time t2. It can be seen that the sequence lengths of a1 and a2 are the same, and there are n-1 repeated vibration signals. Similarly, the server can obtain the target feature signals corresponding to multiple moments such as a3, a4, and a5.

[0146] Step S604, obtaining initial features according to target feature signals corresponding to at least two moments.

[0147] Specifically, for the target characteristic signal at each moment, the server can perform feature extraction to obtain the distribution of the characteristic signal in different frequency domains or time, thereby determining the time series feature sequence or frequency domain feature sequence, and use it as the initial feature corresponding to the target characteristic signal. For example, feature extraction is performed on a1 to obtain the time domain feature sequence at time t1. At this time, b1 is the initial feature corresponding to time t1. Accordingly, feature extraction is performed on a2 to obtain the initial feature b2 at time t2, and b2 is the target feature signal corresponding to time t2. The type of initial feature can be frequency domain feature or time domain feature.

[0148] Step S606, using the trained decay knowledge adaptation model, the health indicator characteristics of the component to be predicted are obtained based on the initial characteristics.

[0149] Among them, the trained decay knowledge adaptation model is obtained based on decay knowledge training.

[0150] Specifically, the server can input the initial features corresponding to the target feature signals corresponding to at least two moments into the trained decay knowledge adaptation model. The decay knowledge can be integrated into the initial features through the trained decay knowledge adaptation model to obtain the health indicator moment features corresponding to each of the at least two moments. After determining the health indicator moment features corresponding to each moment, the server merges the health indicator moment features to obtain the health indicator features corresponding to at least two moments.

[0151] In this embodiment, by acquiring characteristic signals within a period of time under the running state and sampling the characteristic signals at at least two moments, multiple target characteristic signals with repeated parts are obtained, and each target characteristic signal corresponds to a moment. Based on this, the server can determine the initial features of each target characteristic signal respectively, and add the decay knowledge corresponding to the component to be predicted to the initial features through the trained decay knowledge adaptation model, so as to obtain the health indicator moment features corresponding to each initial feature, which can fully explore the decay knowledge of multiple groups of target characteristic signals separated by one moment, and improve the accuracy of determining the health indicator features.

[0152] In an exemplary embodiment, Figure 7 As shown, the method further includes steps S702 to S708. Among them:

[0153] Step S702, obtaining a historical feature signal corresponding to a mechanical component associated with the component to be predicted, obtaining a target historical feature signal corresponding to at least two historical moments from the historical feature signal, and obtaining a historical initial feature according to the target historical feature signal corresponding to at least two historical moments.

[0154] The mechanical component associated with the component to be predicted may be a mechanical component of the same model as the component to be predicted, or a mechanical component made of the same material as the component to be predicted.

[0155] Specifically, the server can obtain the historical characteristic signal of the operating state of the mechanical component associated with the component to be predicted in a certain historical time period in the past through sensor sampling. For each historical moment in the historical time period, the target historical characteristic signal corresponding to the historical moment is sampled according to a preset sequence length.

[0156] For the target historical feature signal at each moment in the historical time period, the server can perform feature extraction to obtain the distribution of the target historical feature signal in different frequency domains or time, thereby determining the time series feature sequence or frequency domain feature sequence corresponding to the target historical feature signal, and use it as the historical initial feature corresponding to the target feature signal.

[0157] Step S704 , obtaining historical health index features and historical intermediate features of the mechanical component based on historical initial features through the decay knowledge adaptation model to be trained.

[0158] Among them, the historical intermediate features are obtained by encoding the historical initial features based on the decay knowledge adaptation model to be trained, and the historical health indicator features are obtained by decoding the historical intermediate features based on the decay knowledge adaptation model to be trained. The decay knowledge adaptation model includes a set of encoders and a set of decoders. The encoder can encode the historical initial features to obtain the historical intermediate features. The decoder can decode the historical intermediate features to obtain the health indicator features containing decay knowledge.

[0159] Specifically, the server can input the historical initial features into the decay knowledge adaptation model to be trained, and first determine the mean, standard deviation, and sample set that obeys the normal distribution corresponding to the historical initial features through the encoder. The server can multiply the standard deviation corresponding to the historical initial features and the sample set that obeys the normal distribution, and sum the product result with the mean corresponding to the historical initial features to obtain the historical intermediate features. The server inputs the historical intermediate features into the decoder, and processes the historical intermediate features through the predefined fifth mapping strategy to obtain the historical health indicator features corresponding to the historical initial features.

[0160] Step S706, obtaining health indicator loss according to the historical health indicator characteristics, the historical intermediate characteristics, the historical initial characteristics and the decay constraint conditions determined based on the decay knowledge.

[0161] Specifically, the server may determine the first loss in the health indicator loss based on the historical initial features; the server may determine the second loss in the health indicator loss based on the difference between the historical health indicator features and the historical initial features; the server may determine the third loss in the health indicator loss based on the historical intermediate features and the decay constraint determined based on the decay knowledge. The server combines the first loss, the second loss, and the third loss to obtain the health indicator loss.

[0162] Step S708, the decay knowledge adaptation model to be trained is updated according to the health indicator loss and then continued to be trained until the training end condition is met, thereby obtaining a trained decay knowledge adaptation model.

[0163] Specifically, the server can take the loss value that minimizes the loss of health indicators as the target, update the decay knowledge adaptation model to be trained, and continue training through historical initial features after the update, continuously optimizing the loss value until the loss value meets the training end condition, or the number of trainings meets the training end condition, thereby obtaining a trained decay knowledge adaptation model.

[0164] In this embodiment, the target historical feature signal can be obtained through the historical feature signal of the mechanical component associated with the component to be predicted, and the historical initial feature corresponding to each historical moment can be determined. Based on this, the server can update the decay knowledge adaptation model to be trained through the health indicator loss to obtain a trained decay knowledge adaptation model. The health indicator loss includes a first loss, a second loss, and a third loss. The first loss is determined based on the historical initial feature, the second loss is determined based on the difference between the historical health indicator feature and the historical initial feature, and the third loss is determined based on the historical intermediate feature and the decay constraint determined based on the decay knowledge. It can construct a loss function from multiple aspects to improve the accuracy of the health indicator loss, thereby improving the training efficiency and prediction accuracy of the decay knowledge adaptation model trained by the health indicator loss.

[0165] In an exemplary embodiment, Figure 8 As shown, the specific implementation process of the step "obtaining the health indicator loss according to the historical health indicator characteristics, the historical intermediate characteristics, the historical initial characteristics and the decline constraint conditions determined based on the decline knowledge" includes steps S802 to S812. Among them:

[0166] Step S802, for each of at least two historical moments, determine the historical health indicator moment characteristics for the historical moment from the historical health indicator characteristics, determine the intermediate characteristics for the historical moment from the historical intermediate characteristics, and determine the initial characteristics for the historical moment from the historical initial characteristics.

[0167] Specifically, for each of the at least two historical moments, the server may determine the historical health indicator moment feature of each historical moment in the historical health indicator feature. The server may determine the intermediate feature corresponding to each historical moment in the historical intermediate feature. The server may determine the initial feature corresponding to each historical moment in the historical initial feature. In one example, the historical health indicator feature includes the historical health indicator moment features of multiple historical moments, the historical intermediate feature includes the intermediate features of multiple historical moments, and the historical initial feature includes the initial features of multiple historical moments.

[0168] Step S804: obtaining first characteristic statistical parameters of the historical initial characteristics according to respective initial characteristics of at least two historical moments in the historical initial characteristics, and determining the first loss based on the first characteristic statistical parameters.

[0169] Specifically, the server can perform statistics on the initial features corresponding to at least two historical times to obtain the first feature statistical parameters of the historical initial features. The first feature statistical parameters can be the mean and standard deviation of the historical initial features. The server can determine the first loss function based on the mean and standard deviation, and determine the first loss through the first loss function. In one example, the server can determine the KL divergence loss (Kullback-LeiblerDivergence) based on the first feature statistical parameters. The KL divergence loss is used to measure the degree of difference between two probability distributions. When the two distributions are exactly the same, the KL divergence loss is 0; using the KL divergence loss as the optimization target can make the original distribution closer to the normal distribution N (0, 1), and the data of the original distribution is more random and reasonable. The greater the difference between the two distributions, the greater the KL divergence loss. The KL divergence loss function can be shown as follows:

[0170]

[0171] in, and represent mean and standard deviation respectively.

[0172] Step S806, obtaining a second loss according to a feature difference between the historical health indicator moment feature at the historical moment and the initial feature at the historical moment.

[0173] Specifically, the server can determine the historical health indicator moment feature corresponding to the same historical moment, and the initial feature corresponding to the same historical moment, and determine the feature difference between the historical health indicator moment feature and the initial feature at the same historical moment, and the server can calculate the second loss based on the second loss function. In one example, the feature difference can be the square difference between the historical health indicator moment feature and the initial feature. The server can determine the second loss based on the feature difference and the second loss function. The second loss function can be as follows:

[0174]

[0175] in, It is used to represent the historical indicator moment characteristics corresponding to the historical moment i. It is used to represent the initial features corresponding to the historical moment i, and J represents The dimension of express The jth element of represents the weight of error reconstruction.

[0176] Step S808, obtaining a third loss according to the intermediate features of the historical moment, the intermediate features of moments adjacent to the historical moment, and the decay constraint parameters determined based on decay knowledge.

[0177] Specifically, the server may determine the intermediate features corresponding to the same historical moment, and the intermediate features of adjacent moments of the same historical moment. The server may obtain the decay constraint parameter determined based on the decay knowledge. The server may determine the third loss of the historical moment based on the third loss function. In one example, the intermediate features corresponding to the historical moment may be , the intermediate features of adjacent moments corresponding to historical moments can be , the decay constraint parameter can be The third loss function can be shown as follows:

[0178]

[0179] Step S810, obtaining a target loss for a historical moment based on the first loss, the second loss and the third loss.

[0180] Specifically, for each historical moment, the first loss, the second loss, and the third loss corresponding to the historical moment may be added together to obtain the target loss at the historical moment.

[0181] In an example, the loss function for the target loss looks like this:

[0182]

[0183] Step S812, obtaining the health indicator loss according to the respective target losses at at least two historical moments.

[0184] Specifically, the server may determine a target loss at each of at least two historical moments, and add the target losses at each historical moment to obtain a health indicator loss corresponding to the historical initial feature.

[0185] In this embodiment, by respectively calculating the first loss, the second loss, and the third loss of each historical moment, and summing the first loss, the second loss, and the third loss to obtain the target loss at each historical moment, and then summing the target losses of each historical moment, the health indicator loss corresponding to the historical initial characteristics is obtained, and the loss in three dimensions at each moment can be calculated, thereby calculating the total target loss of at least two historical moments, thereby improving the accuracy of determining the health indicator loss.

[0186] In an exemplary embodiment, the specific implementation process of the step of "obtaining the health indicator characteristics of the component to be predicted based on the initial characteristics by adapting the trained decay knowledge model" includes:

[0187] The initial features are input into the trained decay knowledge adaptation model to obtain the initial health indicator features corresponding to the initial features; and the initial health indicator features are mapped into the health indicator features of the component to be predicted according to the second feature statistical parameters corresponding to the trained decay knowledge adaptation model.

[0188] The second characteristic statistical parameter is obtained based on the historical intermediate characteristics corresponding to each historical characteristic signal. The second characteristic statistical parameter can be a set of maximum and minimum values ​​obtained from the statistics of each historical intermediate characteristic. Each element in the historical intermediate characteristic can determine the maximum and minimum values ​​corresponding to the element, thereby obtaining a second characteristic statistical parameter composed of the maximum and minimum values ​​corresponding to multiple elements.

[0189] Specifically, the server can input the initial features into the trained decay knowledge adaptation model, and the trained decay knowledge adaptation model processes the initial features and outputs the initial health indicator features corresponding to the initial features. The server can map the initial health indicator features based on the second feature statistical parameters, and map the initial health indicator features into the health indicator features of the component to be predicted. In an example, the process of mapping the initial health indicator features based on the second feature statistical parameters is as follows:

[0190]

[0191] in, is the health indicator characteristic of the component to be predicted, is the initial health indicator feature corresponding to the initial feature, is the minimum value obtained from statistics of each historical intermediate feature, It is the maximum value obtained from statistics of various historical intermediate features.

[0192] In this embodiment, the initial health indicator feature corresponding to the initial feature can be mapped through the second feature statistical parameter, so as to standardize the initial health indicator and obtain the health indicator feature of the component to be predicted. Since the second feature statistical parameter is obtained by statistics from each historical intermediate feature, the standardization process is determined based on a large number of historical intermediate features, which can improve the accuracy of standardization and improve the accuracy of determining the health indicator feature.

[0193] In an exemplary embodiment, the specific implementation process of the step of "obtaining the initial feature according to the target feature signals corresponding to at least two moments" includes:

[0194] According to the transformation method corresponding to the first dimension, the target feature signals corresponding to at least two moments are transformed to obtain the initial features in the first dimension.

[0195] Among them, the first dimension can be a time domain dimension or a frequency domain dimension, and the server can determine the frequency domain characteristics or time domain characteristics corresponding to the target characteristic signal through the transformation method corresponding to the first dimension.

[0196] Specifically, the server transforms the target feature signals corresponding to at least two moments respectively by the transformation method corresponding to the first dimension, and obtains the initial features of each moment under the first dimension. In one example, when the first dimension is a frequency domain dimension, the server can transform the target feature signals corresponding to at least two moments respectively by the transformation method of the frequency domain dimension, and obtain the frequency domain features of each moment under the first dimension.

[0197] Optionally, the server can transform the target feature signals corresponding to at least two moments respectively by a transformation method corresponding to at least one second dimension, and obtain the initial features of each moment under at least one second dimension. The first dimension and at least one second dimension can be a time domain dimension or a frequency domain dimension. The time domain dimension can be divided into multiple sub-time domain dimensions according to different transformation methods, and the frequency domain dimension can be divided into multiple sub-frequency domain dimensions according to different transformation methods. For example, if different transformation methods are selected, the obtained time domain features and frequency domain features are different. For example, the transformation method can be a wavelet basis function, and the wavelet basis function can be selected as a Haar wavelet basis, a db series wavelet basis, etc.

[0198] In an exemplary embodiment, the method further includes: obtaining a first predicted life of the component to be predicted in a first dimension based on the predicted remaining life; obtaining a second predicted life of the component to be predicted in at least one second dimension, the second dimension and the first dimension being different dimensions; and obtaining a comprehensive predicted remaining life of the component to be predicted based on the first predicted life and the second predicted life.

[0199] Specifically, the server can execute subsequent steps according to the initial features in the first dimension until the step of determining the predicted remaining life of the component to be predicted according to the prediction moment is executed, and the first predicted life in the first dimension can be obtained. Similarly, the server can transform the target feature signals corresponding to at least two moments respectively by a transformation method in at least one second dimension, obtain the initial features in at least one second dimension, and execute subsequent steps until the step of determining the predicted remaining life of the component to be predicted according to the prediction moment is executed, and the second predicted life in at least one second dimension is obtained. Based on this, the server can fuse the first predicted life and at least one second predicted life to determine the comprehensive predicted remaining life of the component to be predicted. The fusion processing process can be to determine the average value of the first predicted life and at least one second predicted life, and determine the average value as the comprehensive predicted remaining life, or the fusion processing process can be to perform weighted summation of the first predicted life and at least one second predicted life to obtain the comprehensive predicted remaining life.

[0200] In this embodiment, the initial features of the target characteristic signal in different dimensions can be obtained by transforming different dimensions, thereby obtaining the predicted life corresponding to different dimensions. After the predicted life corresponding to multiple dimensions is obtained, the comprehensive predicted remaining life of the component to be predicted can be obtained by weighted summation or average value, thereby improving the accuracy of the predicted remaining life.

[0201] In an exemplary embodiment, the specific implementation process of the step of "determining the predicted health indicator value corresponding to the predicted moment according to the health indicator moment characteristic corresponding to the predicted moment" includes:

[0202] The preset feature mapping method is used to map the health indicator moment feature corresponding to the prediction moment into a predicted health indicator value corresponding to the prediction moment in numerical form.

[0203] Specifically, the preset feature mapping method can be as follows:

[0204]

[0205] Among them, g represents the linear activation function, W t and b t denote the weight and bias matrices respectively, represents the predicted health index value at time t, h t is the momentary feature of the health indicator corresponding to the prediction moment.

[0206] In this embodiment, the health indicator moment feature corresponding to the prediction moment is mapped into a predicted health indicator value corresponding to the prediction moment in numerical form through a preset feature mapping method, which can improve the accuracy of the predicted health indicator value.

[0207] In an exemplary embodiment, the specific implementation process of the step of "determining the predicted remaining life of the component to be predicted according to the prediction time when the predicted health indicator value indicates the failure of the component to be predicted" includes:

[0208] When the predicted health index value represents the failure of the component to be predicted, the time span between the prediction moment and the acquisition moment of the characteristic signal is determined; based on the time span, the predicted remaining life of the component to be predicted at the acquisition moment is determined.

[0209] Specifically, the server may obtain a preset failure value, and determine whether the predicted health index value is greater than the preset failure value. If the predicted health index value is greater than the preset failure value, it is determined that the predicted health index value indicates that the component to be predicted has failed. If the predicted health index value is less than the preset failure value, it is determined that the predicted health index value indicates that the component to be predicted has not failed. In the case where the predicted health index value indicates that the component to be predicted has failed, the time span between the prediction moment and the acquisition moment of the characteristic signal is determined, and the time span is determined as the predicted remaining life of the component to be predicted at the acquisition moment.

[0210] In this embodiment, by continuously judging whether each predicted health indicator value is invalid, the moment of the predicted health indicator value corresponding to the failure is determined, and based on the moment and the moment of obtaining the characteristic signal, the time span is determined, and the time span is determined as the predicted remaining life, which can improve the accuracy of the predicted remaining life.

[0211] The following is a detailed description of the specific execution process of the above-mentioned mechanical component life prediction method in conjunction with a specific embodiment. The embodiment of the present application realizes intelligent prediction of the remaining life of new energy vehicle motor bearings by designing an unsupervised knowledge-adaptive dual-line gated learning network, effectively utilizes data in an unsupervised manner, reduces data processing costs, and alleviates the problem of small sample size. In addition, the embodiment of the present application introduces decay constraint loss, so that the obtained health index HI (Health Index) has a more monotonous decay trend, and can also improve the accuracy of the prediction. The health index HI can be the predicted health index value in the above embodiment. Through the dual-line gated fusion learning module, the embodiment of the present application can also realize the deep fusion of the overall decay trend characteristics and the local fluctuation characteristics, so as to more accurately predict the remaining life of the motor bearings.

[0212] In the first step, the server can obtain the vibration signal of the bearing and convert the vibration signal into time domain features or frequency domain features.

[0213] Specifically, for the vibration signal of the bearing, the time-frequency feature sequence can be obtained by wavelet transform. The specific steps are as follows: For example, for a segment from Vibration signal with sampling starting at time and sequence length of 7 By using wavelet transform, we can get the distribution of wavelet coefficients in different frequency domains and time. For example, when the Haar wavelet basis is selected for transformation, the time domain feature sequence can be obtained. and frequency domain feature sequence Among them, the lengths of the frequency domain feature sequence and the time domain feature sequence can be different. The horizontal axis of the time domain feature sequence represents time, and the vertical axis represents the wavelet coefficient. The horizontal axis of the frequency domain feature sequence represents frequency, and the vertical axis represents the wavelet coefficient. Then from Vibration signal with sampling starting at time and sequence length of 7 , similarly, and Then put and They are respectively imported as data sets X into the subsequent unsupervised decay knowledge adaptation model to first obtain the health index HI with decay knowledge added, and then predict the remaining life expectancy.

[0214] Optionally, due to the subsequent steps through and The predicted remaining life obtained is different, so the mean is calculated to estimate the remaining life. In addition, when different wavelet basis functions are selected, the obtained time-frequency feature sequences are also different. For example, the wavelet basis function can be selected as Haar wavelet basis, db series wavelet basis, etc., so the feature sequence can include b, c, d, e.

[0215] In the second step, the server can input the time domain features or frequency domain features into the trained unsupervised decay knowledge adaptation model to obtain the health index HI with decay knowledge added, that is, to obtain the health index features in the above embodiment.

[0216] Among them, the embodiment of the present application designs an unsupervised knowledge adaptive dual-line gated learning network to realize intelligent prediction of the remaining life of motor bearings. The unsupervised knowledge adaptive dual-line gated learning network mainly includes two parts: an unsupervised decay knowledge adaptation model and a cross-domain dual-line gated learning model. The unsupervised decay knowledge adaptation model uses unsupervised learning and combines the knowledge of bearing decay trend to accurately obtain the health index HI of the bearing under complex working conditions. The cross-domain dual-line gated learning model learns the overall degradation trend and local unstable fluctuation information of the bearing from HI through a dual-line memory structure, thereby more accurately predicting the remaining life of the bearing.

[0217] The bearing health index HI has a strong correlation with the decay process. In order to ensure the accuracy and reliability of the remaining life prediction, the decay knowledge involved in the decay process should be considered in the process of establishing the health index HI. Fig. 9 As shown in Figure 2, the unsupervised decay knowledge adaptation model introduces the decay constraint loss into the decoder, so that the obtained health index HI has a more monotonous decay trend. Assuming the dataset , then the loss function of the unsupervised decay knowledge adaptation module can be defined as:

[0218]

[0219] in, and Respectively represent encoder and The output, namely the mean and standard deviation, J represents The dimension of express The jth element of represents the weight of error reconstruction, represents the loss of knowledge of the bearing decay process, is the hidden state of the unsupervised decay knowledge adaptation model, and r is a constraint value.

[0220] In one example, the vibration signal of the bearing contains a lot of decay knowledge and can well reflect the operating status of the bearing. Therefore, in this example, the vibration signal is used to predict the remaining life of the bearing. The data set can be: Specifically, the dataset can be Indicates that b1 and x 1 Correspondingly, b i With x i Based on this, the loss function of the unsupervised decay knowledge adaptation module can be expressed as Fig. 9 The three yellow optimization objectives in . The first loss in the loss function can be based on and OK, D KL KL divergence loss is used to measure the difference between two probability distributions. When the two distributions are exactly the same, the KL divergence loss value is 0. Using KL divergence loss as the optimization target can make the original distribution closer to the normal distribution N (0, 1). KL divergence loss can be expressed as The second loss in the loss function can be the difference between the dataset X and the one predicted by the decoder The third loss in the loss function can be the loss caused by the knowledge of the bearing decay process constraints, such as . Fig. 9 The samples that obey the normal distribution N(0, 1) are specifically samples formed after the frequency domain features are regularized.

[0221] Specifically, the server can input the time domain features or frequency domain features into the trained unsupervised decay knowledge adaptation model, and the process of obtaining the health index HI with decay knowledge added is as follows:

[0222] First, multiple frequency domain features and / or time domain features are calculated from the vibration signal of the bearing to be predicted, and then the time domain features and / or frequency domain features are regularized respectively. The regularization process can be defined as:

[0223]

[0224] in, represents the jth eigenvector, and denote the mean and standard deviation of the j-th eigenvector, respectively. Represents the regularization vector of the feature vector.

[0225] Specifically, the data in the dataset can be . Each element in each data can be regularized by traversing. This regularization process can be used in both model training and model application. The data set can be a training set or a test set. n With f j Correspondingly, b1 and f 1 Corresponding.

[0226] Then, the regularized feature vector is divided into a training set and a test set, and an unsupervised decay knowledge adaptation model is established. The training set is mainly used to train the unsupervised decay knowledge adaptation model, and the test set does not need to participate in the training process of the unsupervised decay knowledge adaptation model. The ratio is 7:3. The unsupervised decay knowledge adaptation model is mainly composed of an encoder and a decoder. The weights of the encoder and the decoder are first randomly initialized, and then the training set is imported for training in an unsupervised manner. The output of the unsupervised decay knowledge adaptation model is the health indicator feature vector of the training set. During the training process, the server can combine the health indicator feature vector of the training set The maximum value in and minimum value , calculate the health index HI of the test set:

[0227]

[0228] Specifically, the data in the test set can be After the test set is input into the unsupervised decay knowledge adaptation model, the test set is output The health indicator feature vector , the server can The maximum value corresponding to each element in and minimum value Calculate and get the health index of the test set .

[0229] In the third step, the server can input the health indicator HI with decay knowledge added into the trained two-line gated fusion learning model to obtain the predicted health indicator value corresponding to the health indicator HI.

[0230] Among them, the health index HI with decay knowledge added contains two features: decay trend feature and decay fluctuation feature. The decay trend feature is a stable and monotonous feature sequence, while the decay fluctuation feature is an unstable feature sequence. In the embodiment of the present application, LSTM (Long Short-Term Memory, long short-term memory network) can be used to capture the time series features through a single-line transmission method. As the time step increases, the information remembered by its cell unit gradually tends to be full. Since the amount of information contained in the steady change is not as complex as the non-steady change, the memory cell is more likely to remember the steady change features under limited information capacity, and LSTM is difficult to process non-stationary time series. For a cell unit of LSTM, it learns the long-term decay trend and short-term local fluctuation information from the health indicator sequence at the current moment, and stores the learned decay trend and fluctuation information in the hidden state. Between two adjacent hidden states, the long-term decay trend changes only slightly, while the short-term fluctuation changes dramatically, so the short-term fluctuation information can be extracted by calculating the difference between two adjacent hidden states. The embodiment of the present application designs a dual-line gated fusion learning model based on this conclusion.

[0231] like Fig.10 As shown in , the dual-line gated fusion learning module contains two GRU units, one of which is used to learn the overall degradation trend from the health indicator feature vector HI, and the other is used to learn unstable fluctuation information, so the two GRUs are called the main line and the secondary line respectively. In the secondary line, the GRU unit can learn the overall decay trend and local decay fluctuation of the input time series in one time step. Since the decay trends of adjacent time steps are close, the local decay fluctuation features can be extracted from the difference in information of two adjacent time steps of the GRU unit. The detailed structure of the dual-line gated fusion learning model is shown in Fig.10 As shown, the model has a total of 4 inputs, including an input sequence , the two hidden states of the main line at the previous two moments and , a hidden state of the secondary line , the process of capturing features is as follows:

[0232] In the secondary line, first calculate and The difference between them can be used to obtain the non-stationary fluctuation information D in the time series. t , and then use a GRU to extract the unstable fluctuation information D t Mining non-stable deep time series features and storing them in The attention mechanism can be and Filter out important knowledge from the fusion generated vector and obtain the short-term memory state .Bundle and The fused vectors are then fed into a fully connected layer to obtain the feature vector. , Contains long-term and short-term state information, and is also the input of the next GRU. The state change of the feature captured by the sideline GRU in this process can be described as:

[0233]

[0234] Among them, D t is the momentary fluctuation characteristic of the current moment, h t-1 is the recession trend characteristic of the first prior moment, h t-2 is the decay trend characteristic of the second prior moment; is the first update gate feature determined by the update gate, is a first reset gate characteristic determined by the reset gate, and is a constant, is the activation function, is the linear transformation matrix of the first update gate, is the bias of the first update gate, is the linear transformation matrix of the first reset gate, is the bias of the first reset gate, A is used to indicate the first, to distinguish it from the update gate and reset gate that appear later, t is used to indicate the time, subscript z is used to indicate the update gate, and subscript r is used to indicate the reset gate. To obtain the intermediate eigenvectors required to calculate the decay wave characteristics, is the linear transformation matrix for feature fusion, is the bias, N t It is the recession fluctuation characteristic at the current moment.

[0235] Subsequently, the updating process of the attention module can be described as:

[0236]

[0237] in, is the intermediate attention feature, N t is the recession fluctuation characteristic at the current moment, is the attention weight; is the scoring function, is the attention weight, , , They are all linear transformation matrices that can be learned during model training. is the intermediate fusion feature, is the intermediate attention feature, h t-1 is the recession trend characteristic of the first prior moment, is the activation function, is the linear transformation matrix that can be learned during model training. The subscript m is used to distinguish the linear transformation matrix and bias corresponding to the intermediate fusion feature.

[0238] Get the sub-line GRU Combined with Input into the main line GRU, the hidden state of the main line GRU at time t can be calculated as:

[0239]

[0240] Among them, among them, is the second update gate feature, is the second reset gate feature, is the momentary feature of the health indicator corresponding to the current moment, is the intermediate fusion feature, and is the linear transformation matrix that can be learned during model training. and For offset, the superscript M is distinguished from the above-mentioned A and is used to indicate the second.

[0241] Finally, the output of the dual-line gating fusion learning module is passed through the fully connected layer. It can be expressed as:

[0242]

[0243] Among them, g represents the linear activation function, W and b represent the weight and bias matrices respectively, Represents the predicted health index HI at the tth moment. The dual-line gated fusion learning module is essentially a regression model. Its main goal is to make the model's predicted value closer to the actual HI sequence. Therefore, the mean square error loss function of this module can be defined as:

[0244]

[0245] in, represents the target output by the dual-line gated fusion learning module, that is, the predicted HI value, and N represents the length of the HI vector. Compared with the traditional LSTM, the dual-line gated fusion learning module mines features through the dual memory structure of the main line GRU and the sub-line GRU. The main line GRU can capture the overall decay trend and global fluctuation information of the time series data, while the sub-line GRU mines the short-term fluctuation characteristics between two adjacent hidden states. Therefore, this module can predict the remaining life of motor bearings under complex working conditions.

[0246] In the fourth step, the server can determine the remaining life of the bearing according to the difference between the prediction time of the predicted health index value and the time of obtaining the vibration signal.

[0247] In the embodiments of the present application, firstly, by predicting the life of the motor bearings, guidance can be provided for enterprises and car owners to better plan maintenance plans. The guidance content may include determining when to replace or repair the bearings to avoid equipment downtime or vehicle inoperability due to bearing failures, thereby ensuring the normal operation of the equipment and the safe driving of the vehicle. Secondly, life prediction helps prevent bearing failures. Through prediction, appropriate maintenance measures can be taken in advance before the bearings have problems, reducing losses and maintenance costs caused by failures. Finally, accurate bearing life prediction also helps to reduce unnecessary downtime losses. For new energy vehicle production lines or vehicles in operation, bearing failures may cause the production line to stop or the vehicle to stop operating, resulting in economic losses. Losses can be reduced by predicting and performing maintenance in advance.

[0248] In a specific embodiment, the embodiment of the present application can construct a health index HI with a decay trend and a unified failure threshold from the collected vibration signal based on the unsupervised decay knowledge adaptation model, and then construct a two-line gated fusion learning model, and predict the remaining life of the bearing based on the health index HI. The detailed process of the remaining life prediction is as follows:

[0249] First, the sampling length T and sampling interval can be obtained through the accelerometer The number of the signal samples can be expressed as n. Then the obtained signal samples are sent to the unsupervised decay knowledge adaptation model to generate the health index HI. The set of its active health index vectors is expressed as , the first m feature vectors in I can be used as training samples It is used to train the dual-line gated fusion learning module, and the feature vectors of the remaining moments can be used as a test set Using training samples To construct the input matrix, the input matrix can be expressed as:

[0250]

[0251] Where k represents the number of input neurons, During model training, the first k row vectors of matrix X are used as input, and the remaining row vectors is regarded as the output target, and the dual-line gated fusion learning module is also regarded as a mapping function y. Therefore, the parameters W and b of the mean-square error (MSE) loss function of the updated network can be defined as:

[0252]

[0253] After training the dual-line gated fusion learning model, the feature vectors of the remaining moments are fed into the trained model to predict the health index HI at subsequent moments. This iterative process can be described as:

[0254]

[0255] The above formula can be understood as mapping function y to X2 to X k+1 Mapping is performed to predict X k+2 , and iterate based on this, continue to predict to the next moment, and get the health index HI corresponding to moment j.

[0256] It can be understood that the prediction process can predict the health index HI corresponding to the k+1 moment based on the health index HI of the previous k moments, and determine the predicted health index value corresponding to the HI at each moment. When the HI value at the final moment predicted by the model exceeds the preset failure threshold, the bearing can be considered to have failed, and the remaining life can be determined by the difference between the failure moment and the moment when the bearing vibration signal is obtained.

[0257] It should be understood that, although the various steps in the flowcharts involved in the above-mentioned embodiments are displayed in sequence according to the indication of the arrows, these steps are not necessarily executed in sequence according to the order indicated by the arrows. Unless there is a clear explanation in this article, the execution of these steps does not have a strict order restriction, and these steps can be executed in other orders. Moreover, at least a part of the steps in the flowcharts involved in the above-mentioned embodiments can include multiple steps or multiple stages, and these steps or stages are not necessarily executed at the same time, but can be executed at different times, and the execution order of these steps or stages is not necessarily to be carried out in sequence, but can be executed in turn or alternately with other steps or at least a part of the steps or stages in other steps.

[0258] Based on the same inventive concept, the embodiment of the present application also provides a mechanical component life prediction device for implementing the mechanical component life prediction method involved above. The implementation scheme for solving the problem provided by the device is similar to the implementation scheme recorded in the above method, so the specific limitations in one or more mechanical component life prediction device embodiments provided below can refer to the limitations of the mechanical component life prediction method above, and will not be repeated here.

[0259] In an exemplary embodiment, Fig.11 As shown, a mechanical component life prediction device 1100 is provided, comprising: a first feature determination module 1101, a second feature prediction module 1102, an index value determination module 1103 and a life determination module 1104, wherein:

[0260] The first feature determination module 1101 is used to determine the health index feature of the component to be predicted based on the feature signal of the component to be predicted in the running state and the decay knowledge corresponding to the component to be predicted; the decay knowledge is used to represent the corresponding relationship between the feature signal of the component to be predicted in the running process and the remaining life of the component, and the health index feature includes the health index moment feature corresponding to at least two moments in the running state;

[0261] The second feature prediction module 1102 is used to perform iterative prediction of the decline feature based on the health indicator moment features corresponding to at least two moments in the health indicator features, and obtain the health indicator moment features corresponding to the prediction moment; in each iterative prediction of the decline feature, the decline fluctuation feature at the current moment is determined based on the decline fluctuation feature at the previous moment, and the decline trend feature at the current moment is determined based on the decline fluctuation feature at the current moment, the decline trend feature at the previous moment and the health indicator moment feature at the current moment;

[0262] The indicator value determination module 1103 is used to determine the predicted health indicator value corresponding to the predicted time according to the health indicator moment characteristic corresponding to the predicted time;

[0263] The life determination module 1104 is used to determine the predicted remaining life of the component to be predicted according to the prediction time when the predicted health index value indicates that the component to be predicted has failed.

[0264] Furthermore, the second feature determination module 1102 is specifically used to: for the first moment of at least two moments, determine the health indicator moment feature corresponding to the first moment from the health indicator features, and obtain the decline trend feature corresponding to the first moment based on the health indicator moment feature corresponding to the first moment and the decline fluctuation feature of the first moment; for each current moment starting from the second moment of at least two moments, determine the decline fluctuation feature of the current moment based on the decline fluctuation feature corresponding to the previous moment of the current moment; obtain the decline trend feature corresponding to the current moment based on the decline fluctuation feature of the current moment and the health indicator moment feature corresponding to the current moment; when the current moment does not belong to at least two moments, take the current moment as the predicted moment, and obtain the health indicator moment feature of the predicted moment based on the decline trend feature corresponding to the current moment.

[0265] Furthermore, the second feature determination module 1102 is specifically used to: determine the first previous moment of the current moment and the second previous moment before the first previous moment for each current moment starting from the second moment of at least two moments; determine the moment fluctuation feature of the current moment based on the decay trend feature of the first previous moment and the decay trend feature of the second previous moment; obtain the decay fluctuation feature of the current moment based on the moment fluctuation feature and the decay fluctuation feature of the first previous moment.

[0266] Furthermore, the second feature determination module 1102 is specifically used for: based on the attention mechanism, obtaining the intermediate attention feature according to the decay fluctuation feature of the current moment and the decay trend feature of the first previous moment; obtaining the intermediate fusion feature according to the intermediate attention feature and the decay trend feature of the first previous moment; obtaining the decay trend feature corresponding to the current moment according to the intermediate fusion feature and the health indicator moment feature corresponding to the current moment.

[0267] Furthermore, the second feature determination module 1102 is specifically used to: when the current moment belongs to at least two moments, determine the health indicator moment feature corresponding to the current moment from the health indicator features; when the current moment does not belong to at least two moments, obtain the health indicator moment feature corresponding to the current moment based on the decay trend feature of the first previous moment.

[0268] Furthermore, the first feature determination module 1101 is specifically used to: obtain a feature signal of the component to be predicted in the operating state, and obtain a target feature signal corresponding to at least two moments from the feature signal; obtain an initial feature based on the target feature signal corresponding to at least two moments; obtain a health indicator feature of the component to be predicted based on the initial feature through a trained decay knowledge adaptation model; the trained decay knowledge adaptation model is obtained based on decay knowledge training.

[0269] Furthermore, the first feature determination module 1101 is specifically used to: obtain historical feature signals corresponding to mechanical parts associated with the parts to be predicted, obtain target historical feature signals corresponding to at least two historical moments from the historical feature signals, and obtain historical initial features according to the target historical feature signals corresponding to at least two historical moments; obtain historical health indicator features and historical intermediate features of the mechanical parts based on the historical initial features through the decay knowledge adaptation model to be trained; the historical intermediate features are encoded by the decay knowledge adaptation model to be trained based on the historical initial features, and the historical health indicator features are decoded by the decay knowledge adaptation model to be trained based on the historical intermediate features; obtain health indicator losses according to the historical health indicator features, the historical intermediate features, the historical initial features and the decay constraints determined based on the decay knowledge; continue training after updating the decay knowledge adaptation model to be trained according to the health indicator losses until the training end conditions are met to obtain a trained decay knowledge adaptation model.

[0270] Furthermore, the first feature determination module 1101 is specifically used to: for each of at least two historical moments, determine the historical health indicator moment feature of the historical moment from the historical health indicator feature, determine the intermediate feature of the historical moment from the historical intermediate feature, and determine the initial feature of the historical moment from the historical initial feature; obtain the first feature statistical parameter of the historical initial feature according to the initial features of at least two historical moments in the historical initial feature, and determine the first loss based on the first feature statistical parameter; obtain the second loss according to the feature difference between the historical health indicator moment feature of the historical moment and the initial feature of the historical moment; obtain the third loss according to the intermediate feature of the historical moment, the intermediate feature of the adjacent moments of the historical moment, and the recession constraint parameter determined based on the recession knowledge; obtain the target loss of the historical moment based on the first loss, the second loss and the third loss; obtain the health indicator loss according to the target loss of at least two historical moments.

[0271] Furthermore, the first feature determination module 1101 is specifically used to: input the initial feature into the trained decay knowledge adaptation model to obtain the initial health indicator feature corresponding to the initial feature; map the initial health indicator feature into the health indicator feature of the component to be predicted according to the second feature statistical parameter corresponding to the trained decay knowledge adaptation model; wherein the second feature statistical parameter is obtained based on the historical intermediate feature statistics corresponding to each historical feature signal.

[0272] Furthermore, the first feature determination module 1101 is specifically used to: transform the target feature signals corresponding to at least two moments respectively according to the transformation method corresponding to the first dimension, so as to obtain the initial features in the first dimension; the life determination module 1104 is specifically used to: obtain the first predicted life of the component to be predicted in the first dimension according to the predicted remaining life; obtain the second predicted life of the component to be predicted in at least one second dimension, the second dimension and the first dimension are different dimensions; obtain the comprehensive predicted remaining life of the component to be predicted according to the first predicted life and the second predicted life.

[0273] Furthermore, the indicator value determination module 1103 is specifically used to: map the health indicator moment feature corresponding to the prediction moment into a predicted health indicator value corresponding to the prediction moment in numerical form by using a preset feature mapping method.

[0274] Furthermore, the life determination module 1104 is specifically used to: when the predicted health index value represents the failure of the component to be predicted, determine the time span between the prediction moment and the acquisition moment of the characteristic signal; and determine the predicted remaining life of the component to be predicted at the acquisition moment based on the time span.

[0275] Each module in the above-mentioned mechanical component life prediction device can be implemented in whole or in part by software, hardware and a combination thereof. Each of the above-mentioned modules can be embedded in or independent of a processor in a computer device in the form of hardware, or can be stored in a memory in a computer device in the form of software, so that the processor can call and execute the corresponding operations of each of the above modules.

[0276] In an exemplary embodiment, a computer device is provided. The computer device may be a server, and its internal structure diagram may be as shown in FIG. Fig.12 As shown. The computer device includes a processor, a memory, an input / output interface (Input / Output, referred to as I / O) and a communication interface. Among them, the processor, the memory and the input / output interface are connected through a system bus, and the communication interface is connected to the system bus through the input / output interface. Among them, the processor of the computer device is used to provide computing and control capabilities. The memory of the computer device includes a non-volatile storage medium and an internal memory. The non-volatile storage medium stores an operating system, a computer program and a database. The internal memory provides an environment for the operation of the operating system and the computer program in the non-volatile storage medium. The database of the computer device is used to store characteristic signals of the components to be predicted in the operating state, the decay knowledge corresponding to the components to be predicted, etc. The input / output interface of the computer device is used to exchange information between the processor and the external device. The communication interface of the computer device is used to communicate with an external terminal through a network connection. When the computer program is executed by the processor, a method for predicting the life of a mechanical component is implemented.

[0277] Those skilled in the art will understand that Fig.12 The structure shown in the figure is only a block diagram of a part of the structure related to the solution of the present application, and does not constitute a limitation on the computer device to which the solution of the present application is applied. The specific computer device may include more or fewer components than those shown in the figure, or combine certain components, or have a different arrangement of components.

[0278] In an exemplary embodiment, a computer device is further provided, including a memory and a processor, wherein a computer program is stored in the memory, and the processor implements the steps in the above-mentioned method embodiments when executing the computer program.

[0279] In one embodiment, a computer-readable storage medium is provided, on which a computer program is stored. When the computer program is executed by a processor, the steps in the above-mentioned method embodiments are implemented.

[0280] In one embodiment, a computer program product is provided, including a computer program, which implements the steps in the above method embodiments when executed by a processor.

[0281] Those skilled in the art can understand that all or part of the processes in the above-mentioned embodiment methods can be completed by instructing the relevant hardware through a computer program, and the computer program can be stored in a non-volatile computer-readable storage medium. When the computer program is executed, it can include the processes of the embodiments of the above-mentioned methods. Among them, any reference to the memory, database or other medium used in the embodiments provided in this application can include at least one of non-volatile and volatile memory. Non-volatile memory can include read-only memory (ROM), magnetic tape, floppy disk, flash memory, optical memory, high-density embedded non-volatile memory, resistive random access memory (ReRAM), magnetoresistive random access memory (MRAM), ferroelectric random access memory (FRAM), phase change memory (PCM), graphene memory, etc. Volatile memory can include random access memory (RAM) or external cache memory, etc. As an illustration and not limitation, RAM can be in various forms, such as static random access memory (SRAM) or dynamic random access memory (DRAM). The database involved in each embodiment provided in this application may include at least one of a relational database and a non-relational database. Non-relational databases may include distributed databases based on blockchains, etc., but are not limited to this. The processor involved in each embodiment provided in this application may be a general-purpose processor, a central processing unit, a graphics processor, a digital signal processor, a programmable logic device, a data processing logic device based on quantum computing, etc., but are not limited to this.

[0282] The technical features of the above embodiments may be combined arbitrarily. To make the description concise, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0283] The above-described embodiments only express several implementation methods of the present application, and the descriptions thereof are relatively specific and detailed, but they cannot be understood as limiting the scope of the present application. It should be pointed out that, for a person of ordinary skill in the art, several variations and improvements can be made without departing from the concept of the present application, and these all belong to the protection scope of the present application. Therefore, the protection scope of the present application shall be subject to the attached claims.

Claims

1. A method for predicting the life of a mechanical component, characterized in that: The method comprises: Determine the health index characteristics of the component to be predicted based on the characteristic signal of the component to be predicted in the operating state and the decay knowledge corresponding to the component to be predicted; the decay knowledge is used to represent the corresponding relationship between the characteristic signal of the component to be predicted in the operating process and the remaining life of the component, and the health index characteristics include health index time characteristics corresponding to at least two moments in the operating state; Performing iterative prediction of the decline characteristics based on the health indicator moment characteristics corresponding to the at least two moments in the health indicator characteristics to obtain the health indicator moment characteristics corresponding to the predicted moment; in each iterative prediction of the decline characteristics, determining the decline fluctuation characteristics at the current moment based on the decline fluctuation characteristics at the previous moment, and determining the decline trend characteristics at the current moment based on the decline fluctuation characteristics at the current moment, the decline trend characteristics at the previous moment and the health indicator moment characteristics at the current moment; Determining a predicted health indicator value corresponding to the predicted time according to the health indicator moment characteristic corresponding to the predicted time; In a case where the predicted health index value indicates that the component to be predicted has failed, a predicted remaining life of the component to be predicted is determined according to the predicted time.

2. The method according to claim 1, characterized in that The iterative prediction of the decay characteristics based on the health indicator time characteristics corresponding to the at least two moments in the health indicator characteristics to obtain the health indicator time characteristics corresponding to the predicted moment includes: For a first moment of the at least two moments, determining a health indicator moment characteristic corresponding to the first moment from the health indicator characteristics, and obtaining a decline trend characteristic corresponding to the first moment according to the health indicator moment characteristic corresponding to the first moment and the decline fluctuation characteristic of the first moment; At each current moment starting from the second moment of the at least two moments, determining the decay fluctuation feature of the current moment according to the decay fluctuation feature corresponding to the previous moment of the current moment; Obtaining the recession trend characteristics corresponding to the current moment according to the recession fluctuation characteristics at the current moment and the health indicator moment characteristics corresponding to the current moment; In the case that the current moment does not belong to the at least two moments, the current moment is used as the predicted moment, and based on the decay trend characteristics corresponding to the current moment, the health indicator moment characteristics of the predicted moment are obtained.

3. The method according to claim 2, characterized in that The step of determining the decay fluctuation feature of the current moment according to the decay fluctuation feature corresponding to the previous moment of the current moment at each current moment starting from the second moment of the at least two moments includes: For each current moment starting from the second of the at least two moments, determining a first previous moment of the current moment and a second previous moment before the first previous moment; determining the moment fluctuation feature of the current moment according to the decay trend feature of the first previous moment and the decay trend feature of the second previous moment; The decay fluctuation characteristic at the current moment is obtained according to the moment fluctuation characteristic and the decay fluctuation characteristic at the first previous moment.

4. The method according to claim 3, characterized in that The step of obtaining the decay trend feature corresponding to the current moment according to the decay fluctuation feature at the current moment and the health indicator moment feature corresponding to the current moment includes: Based on the attention mechanism, obtaining an intermediate attention feature according to the decay fluctuation feature at the current moment and the decay trend feature at the first previous moment; Obtaining an intermediate fusion feature according to the intermediate attention feature and the decay trend feature of the first prior moment; According to the intermediate fusion feature and the health indicator moment feature corresponding to the current moment, the decline trend feature corresponding to the current moment is obtained.

5. The method according to claim 2, characterized in that: The method further comprises: In the case where the current moment belongs to the at least two moments, determining the health indicator moment feature corresponding to the current moment from the health indicator features; In a case where the current moment does not belong to the at least two moments, the health indicator moment feature corresponding to the current moment is obtained according to the decay trend feature of the first previous moment.

6. The method according to claim 1, characterized in that The determining of the health indicator characteristics of the component to be predicted based on the characteristic signal of the component to be predicted in the operating state and the decay knowledge corresponding to the component to be predicted includes: Acquire characteristic signals of the component to be predicted in the running state, and obtain target characteristic signals corresponding to at least two moments from the characteristic signals; Obtaining an initial feature according to the target feature signals corresponding to each of the at least two moments; The health indicator characteristics of the component to be predicted are obtained based on the initial characteristics through the trained decay knowledge adaptation model; the trained decay knowledge adaptation model is obtained based on the decay knowledge training.

7. The method according to claim 6, characterized in that The method further comprises: Acquire a historical feature signal corresponding to a mechanical component associated with the component to be predicted, obtain target historical feature signals corresponding to at least two historical moments from the historical feature signal, and obtain historical initial features according to the target historical feature signals corresponding to the at least two historical moments; The historical health index feature and the historical intermediate feature of the mechanical component are obtained based on the historical initial feature by using the decay knowledge adaptation model to be trained; the historical intermediate feature is obtained by encoding the decay knowledge adaptation model to be trained based on the historical initial feature, and the historical health index feature is obtained by decoding the decay knowledge adaptation model to be trained based on the historical intermediate feature; Obtaining health indicator loss according to the historical health indicator feature, the historical intermediate feature, the historical initial feature, and a decay constraint condition determined based on the decay knowledge; The decay knowledge adaptation model to be trained is updated according to the health indicator loss and then continued to be trained until a training end condition is met, thereby obtaining the trained decay knowledge adaptation model.

8. The method according to claim 7, characterized in that The step of obtaining the health indicator loss according to the historical health indicator feature, the historical intermediate feature, the historical initial feature and the decay constraint condition determined based on the decay knowledge includes: For each of the at least two historical moments, determine the historical health indicator moment feature of the historical moment from the historical health indicator features, determine the intermediate feature of the historical moment from the historical intermediate features, and determine the initial feature of the historical moment from the historical initial features; Obtaining a first feature statistical parameter of the historical initial feature according to the initial features of the at least two historical moments in the historical initial feature, and determining a first loss based on the first feature statistical parameter; Obtaining a second loss according to a feature difference between the historical health indicator moment feature at the historical moment and the initial feature at the historical moment; Obtaining a third loss according to the intermediate feature of the historical moment, the intermediate feature of the moment adjacent to the historical moment, and a decay constraint parameter determined based on the decay knowledge; Based on the first loss, the second loss and the third loss, obtaining the target loss for the historical moment; A health indicator loss is obtained according to the target losses of each of the at least two historical moments.

9. The method according to claim 7, characterized in that: The health indicator feature of the component to be predicted is obtained based on the initial feature by using the trained decay knowledge adaptation model, including: Inputting the initial features into the trained decay knowledge adaptation model to obtain initial health indicator features corresponding to the initial features; According to the second characteristic statistical parameter corresponding to the trained decay knowledge adaptation model, the initial health indicator characteristic is mapped into the health indicator characteristic of the component to be predicted; The second characteristic statistical parameter is obtained based on the historical intermediate characteristic statistics corresponding to each of the historical characteristic signals.

10. The method according to claim 6, characterized in that The obtaining of the initial feature according to the target feature signals corresponding to the at least two moments respectively comprises: According to the transformation method corresponding to the first dimension, the target feature signals corresponding to the at least two moments are transformed to obtain the initial features in the first dimension; The method further comprises: According to the predicted remaining life, obtaining a first predicted life of the component to be predicted in the first dimension; Acquire a second predicted life of the component to be predicted in at least one second dimension, where the second dimension is different from the first dimension; According to the first predicted life and the second predicted life, a comprehensive predicted remaining life of the component to be predicted is obtained.

11. The method according to any one of claims 1 to 10, characterized in that: The determining, according to the health indicator moment characteristic corresponding to the prediction moment, the predicted health indicator value corresponding to the prediction moment comprises: The health indicator moment feature corresponding to the predicted moment is mapped into a predicted health indicator value corresponding to the predicted moment in numerical form by a preset feature mapping method.

12. The method according to any one of claims 1 to 10, characterized in that: When the predicted health index value indicates that the component to be predicted has failed, determining the predicted remaining life of the component to be predicted according to the predicted time includes: In the case where the predicted health index value represents the failure of the component to be predicted, determining the time span between the prediction moment and the acquisition moment of the characteristic signal; The predicted remaining life of the component to be predicted at the acquisition moment is determined according to the time span.

13. A mechanical component life prediction device, characterized in that: The device comprises: A first feature determination module is used to determine the health index feature of the component to be predicted based on the feature signal of the component to be predicted in the operating state and the decay knowledge corresponding to the component to be predicted; the decay knowledge is used to represent the corresponding relationship between the feature signal of the component to be predicted in the operating process and the remaining life of the component, and the health index feature includes the health index moment features corresponding to at least two moments in the operating state; A second feature prediction module is used to perform iterative prediction of the decline feature based on the health indicator moment features corresponding to the at least two moments in the health indicator features, and obtain the health indicator moment features corresponding to the prediction moment; in each iterative prediction of the decline feature, the decline fluctuation feature at the current moment is determined based on the decline fluctuation feature at the previous moment, and the decline trend feature at the current moment is determined based on the decline fluctuation feature at the current moment, the decline trend feature at the previous moment and the health indicator moment feature at the current moment; An indicator value determination module, used to determine the predicted health indicator value corresponding to the predicted moment according to the health indicator moment characteristic corresponding to the predicted moment; The life determination module is used to determine the predicted remaining life of the component to be predicted according to the prediction time when the predicted health index value indicates that the component to be predicted has failed.

14. A computer device comprising a memory and a processor, wherein the memory stores a computer program, wherein: When the processor executes the computer program, the steps of the method according to any one of claims 1 to 12 are implemented.

15. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the steps of the method according to any one of claims 1 to 12 are implemented.

Images