Electronic device health status prediction method, device, electronic device and storage medium

Abstract

The present invention relates to the field of computer technology, and provides a method, device, electronic device and storage medium for predicting the health status of an electronic device, the method comprising: obtaining the current operating status data of a target electronic device; inputting the operating status data into a health status prediction model, and obtaining the current health status level output by the health status prediction model based on the model to which the target electronic device belongs; wherein the health status prediction model is constructed based on a multi-domain learning mechanism, each domain corresponds to a model of the same type of electronic device, and the health status prediction model is trained based on operating status data samples of the same type of electronic devices of different models, and health status level labels corresponding to the operating status data samples. The present invention adopts a multi-domain learning mechanism, and can output the current health status level suitable for the target electronic device of the model according to the model to which the target electronic device belongs, so that the prediction of the health status is more accurate.

Description

Technical Field

[0001] The present invention relates to the field of computer technology, and in particular to a method, device, electronic device and storage medium for predicting the health status of an electronic device. Background Art

[0002] With the rapid development of technologies such as big data, cloud computing, and 5G, the amount of data has exploded. To this end, cloud service providers have established huge data centers to provide users with high-quality services, and the stable operation of data centers has become the key to influencing user experience. Data centers need to operate for a long time. For the various electronic devices that make up the data center, their physical characteristics determine that the longer they are used, the higher the probability of failure and the worse the health status. For example: In data centers, hard disk failures are the most common. If the hard disk fails, it will cause the task or system running on the hard disk to crash, resulting in service interruption, and may also cause the loss of user stored data.

[0003] At present, the fault protection mechanisms for electronic devices include passive protection and active protection. Passive protection is to take remedial measures after an electronic device fails, and active protection usually predicts the health status of the electronic device through a machine learning model. However, for an electronic device, there are products from different manufacturers, and the same manufacturer also has products of different models. The existing relevant machine learning models do not take into account the differences between these products, that is, they do not consider the impact of the operating status data of electronic devices of different models on the model parameters, resulting in inaccurate predictions of the health status of electronic devices by existing machine learning models. Summary of the invention

[0004] The present invention provides a method, device, electronic device and storage medium for predicting the health status of an electronic device, so as to solve the technical problem in the existing related technology that a machine learning model cannot accurately predict the health status of an electronic device.

[0005] The present invention provides a method for predicting the health status of an electronic device, comprising the following steps.

[0006] Obtain the current operating status data of the target electronic device.

[0007] The operating status data is input into a health status prediction model to obtain a current health status level output by the health status prediction model based on the model of the target electronic component.

[0008] Among them, the health status prediction model is constructed based on a multi-domain learning mechanism, each domain corresponds to a model of the same type of electronic device, and the health status prediction model is trained based on operating status data samples of the same type of electronic devices of different models, and the health status level labels corresponding to the operating status data samples.

[0009] According to a method for predicting the health status of an electronic device provided by the present invention, the health status prediction model includes: a shared fully connected neural network and N specific domain fully connected neural networks, wherein N is greater than or equal to 2, and one specific domain fully connected neural network corresponds to one model of the electronic device.

[0010] The weight of each corresponding model of the specific domain fully connected neural network is obtained by bitwise multiplication of the shared parameters of the shared fully connected neural network and the specific parameters of the specific domain fully connected neural network, and the weight is used to weight the input feature vector of the corresponding model of the input specific domain fully connected neural network.

[0011] The shared parameters are used to learn the common features of the operating status data samples of the same type of electronic devices of different models during the training phase, and the specific parameters are used to learn the specific features of the operating status data samples of the same type of electronic devices of a specific model during the training phase.

[0012] According to a method for predicting the health status of an electronic device provided by the present invention, the shared fully connected neural network and the specific domain fully connected neural network both include multiple feature extraction layers, and each of the feature extraction layers includes: a normalization module and a feature extraction module.

[0013] The normalization module is used to normalize the input feature vector according to the mean and variance corresponding to the model of the target electronic device, and the mean and variance are statistically obtained during the inference and prediction process of the health status prediction model.

[0014] The feature extraction module is used to extract features from the normalized input feature vector.

[0015] According to a method for predicting the health status of an electronic device provided by the present invention, the normalization module is used to perform normalization according to the following formula.

[0016] .

[0017] in, z represents the input feature vector before normalization, z' represents the normalized input feature vector, E p Indicates the model of the target electronic device p The corresponding input feature vector is the mean, Var p Indicates the model of the target electronic device p The variance of the corresponding input eigenvector, γ represents the shared range vector of the normalization module, γ p Indicates the corresponding model of the normalized module p A specific range vector ofβ express γ The shared bias vector of β p express γ p A specific bias vector of ε is a constant.

[0018] According to a method for predicting the health status of an electronic device provided by the present invention, the operating status data is input into a health status prediction model to obtain the current health status level output by the health status prediction model based on the model to which the target electronic device belongs, including the following steps.

[0019] The operating status data is input into a shared fully connected neural network of a health status prediction model and a specific domain fully connected neural network corresponding to the model of the target electronic device, and a first eigenvector output by the shared fully connected neural network and a second eigenvector output by the specific domain fully connected neural network are obtained respectively.

[0020] The first feature vector and the second feature vector are bitwise multiplied to obtain a result feature vector representing the current health status level.

[0021] According to a method for predicting the health status of an electronic device provided by the present invention, the operating status data is input into a shared fully connected neural network of a health status prediction model, and a specific domain fully connected neural network corresponding to the model of the target electronic device, and a first eigenvector output by the shared fully connected neural network and a second eigenvector output by the specific domain fully connected neural network are respectively obtained, including the following steps.

[0022] The running status data is input into the shared fully connected neural network, and each layer output of the shared fully connected neural network is as follows.

[0023] .

[0024] Determine the output of the last layer out is the first eigenvector, where W represents the shared parameter, which is the parameter of the feature extraction module in the shared fully connected neural network, in represents the output of the normalization module of the current layer in the shared fully connected neural network, b express W The shared bias vector of Represents the activation function.

[0025] The operating status data is input into the specific domain fully connected neural network, and each layer output of the specific domain fully connected neural network is as follows.

[0026] .

[0027] Determine the output of the last layer out p is the second eigenvector, where in p Indicates the model p The output of the normalization module of the current layer in the corresponding domain-specific fully connected neural network.

[0028] .

[0029] W p Indicates the model p The corresponding specific parameter, the specific parameter is the model p The parameters of the feature extraction module in the corresponding domain-specific fully connected neural network, b p express W p A specific bias vector of Indicates the model p The weight of express A specific bias vector of Represents a bitwise multiplication operation of vectors.

[0030] According to a method for predicting the health status of an electronic device provided by the present invention, before the operating status data is input into a health status prediction model, the following steps are also included.

[0031] The model of the target electronic device is obtained, and the model is encoded to obtain a model identification code of the target electronic device.

[0032] The operating status data is input into a specific domain fully connected neural network corresponding to the model of the target electronic device, including: identifying the corresponding specific domain fully connected neural network based on the model identification code, and inputting the operating status data into the specific domain fully connected neural network corresponding to the model of the target electronic device.

[0033] According to a method for predicting the health status of an electronic device provided by the present invention, the health status prediction model training process is as follows.

[0034] The running status data sample is input into the shared fully connected neural network of the initial model to obtain a first prediction result output by the shared fully connected neural network.

[0035] The running status data sample corresponding to the model identification code is input into the specific domain fully connected neural network corresponding to the model identification code in the initial model to obtain a second prediction result output by the specific domain fully connected neural network.

[0036] The first prediction result and the second prediction result are multiplied bit by bit to obtain a final prediction result.

[0037] The final prediction result and the health status level label are substituted into the loss function. If the loss function has not converged, back propagation is performed to update the shared parameters, specific parameters, shared range vectors and specific range vectors. If the loss function converges, the model training is completed.

[0038] According to a method for predicting the health status of an electronic device provided by the present invention, the health status prediction model training process is as follows.

[0039] The running status data sample is input into the shared fully connected neural network of the initial model to obtain a first prediction result output by the shared fully connected neural network.

[0040] The running status data sample corresponding to the model identification code is input into the specific domain fully connected neural network corresponding to the model identification code in the initial model to obtain a second prediction result output by the specific domain fully connected neural network.

[0041] The first prediction result and the second prediction result are bit-wise multiplied to obtain a third prediction result.

[0042] The intermediate feature vector output by the intermediate layer of the fully connected neural network of the specific domain corresponding to the model identification code is input into the auxiliary neural network corresponding to the model identification code to obtain a fourth prediction result output by the auxiliary neural network, and the fourth prediction result has the same dimension as the third prediction result.

[0043] A final prediction result is determined based on the third prediction result and the fourth prediction result.

[0044] The final prediction result and the health status level label are substituted into the loss function. If the loss function has not converged, back propagation is performed to update the shared parameters, specific parameters, shared range vectors, specific range vectors and model parameters of the auxiliary neural network. If the loss function converges, the model training is completed.

[0045] According to a method for predicting the health status of an electronic device provided by the present invention, a final prediction result is determined based on the third prediction result and the fourth prediction result, including: determining the final prediction result according to the following formula.

[0046] .

[0047] in, represents the final prediction result, represents the third prediction result, represents the fourth prediction result, softmaxRepresents the normalization function.

[0048] According to a method for predicting the health status of an electronic device provided by the present invention, the health status level label corresponding to the operating status data sample is set as follows.

[0049] Taking the time when the faulty electronic device corresponding to the operating status data sample fails as the starting time, at least two target time points are determined in the forward direction to divide the time in the forward direction of the starting time into at least three time periods.

[0050] The health status level label corresponding to the running status data sample is set according to the time period in which the running status data sample is collected.

[0051] According to a method for predicting the health status of an electronic device provided by the present invention, the moment when a faulty electronic device corresponding to the operating status data sample fails is taken as the starting moment, and at least two target time points are determined in advance, including: for each attribute item in the operating status data sample, a regression decision tree model is used, the time before the starting moment is taken as the fitting target, and the mean absolute error is used as the loss function of the regression decision tree model to fit at least two separated time points that separate the time before the starting moment; and the separated time point fitted by the regression decision tree model corresponding to the attribute item with the smallest sum of the mean absolute errors is selected as the target time point.

[0052] According to the present invention, a device for predicting the health status of an electronic device includes the following units.

[0053] The data acquisition unit is used to acquire the current operating status data of the target electronic device.

[0054] The model execution unit is used to input the operating status data into the health status prediction model to obtain the current health status level output by the health status prediction model based on the model of the target electronic device.

[0055] Among them, the health status prediction model is constructed based on a multi-domain learning mechanism, each domain corresponds to a model of the same type of electronic device, and the health status prediction model is trained based on operating status data samples of the same type of electronic devices of different models, and the health status level labels corresponding to the operating status data samples.

[0056] The present invention also provides an electronic device, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein when the processor executes the computer program, a method for predicting the health status of an electronic device as described above is implemented.

[0057] The present invention also provides a non-transitory computer-readable storage medium having a computer program stored thereon, and when the computer program is executed by a processor, the method for predicting the health status of an electronic device as described in any one of the above is implemented.

[0058] The electronic device health status prediction method, device, electronic device and storage medium provided by the present invention obtain the current health status level output by the health status prediction model based on the model of the target electronic device by inputting the operating status data into the health status prediction model, and the health status prediction model is constructed based on a multi-domain learning mechanism, each domain corresponds to a model of the same type of electronic device, and the health status prediction model is trained based on operating status data samples of the same type of electronic devices of different models, and the health status level labels corresponding to the operating status data samples. Due to the use of the multi-domain learning mechanism, when the health status prediction model is trained, its model parameters can learn common knowledge or features from the operating status data samples of electronic devices of different models, as well as learn specific features in the operating status data samples of electronic devices of different models. The finally trained model has better generalization ability and feature expression ability for electronic devices of different models. Therefore, when predicting, it can output the current health status level suitable for the target electronic device of this model according to the model to which the target electronic device belongs, so that the prediction of the health status is more accurate, that is, the health status of the same type of electronic devices of different models can be accurately predicted. BRIEF DESCRIPTION OF THE DRAWINGS

[0059] In order to more clearly illustrate the technical solutions in the present invention or the prior art, the following briefly introduces the drawings required for use in the embodiments or the description of the prior art. Obviously, the drawings described below are some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative work.

[0060] Figure 1 It is a flow chart of the method for predicting the health status of an electronic device provided by the present invention.

[0061] Figure 2 It is a schematic diagram of attribute items of hard disk operation status data in the electronic device health status prediction method provided by the present invention.

[0062] Figure 3 This is one of the structural intentions of the health status prediction model in the electronic device health status prediction method provided by the present invention.

[0063] Figure 4 This is the second structural intention of the health status prediction model in the electronic device health status prediction method provided by the present invention.

[0064] Figure 5It is a running status data collection table of a faulty hard disk in the electronic device health status prediction method provided by the present invention.

[0065] Figure 6 It is a schematic diagram of using a regression decision tree model for sample balancing in the electronic device health status prediction method provided by the present invention.

[0066] Figure 7 It is a structural schematic diagram of the electronic device health status prediction device provided by the present invention.

[0067] Figure 8 It is a structural schematic diagram of the electronic device provided by the present invention. DETAILED DESCRIPTION

[0068] In order to make the purpose, technical solution and advantages of the present invention clearer, the technical solution of the present invention will be clearly and completely described below in conjunction with the drawings of the present invention. Obviously, the described embodiments are part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0069] The electronic device health status prediction method of the embodiment of the present invention is as follows: Figure 1 As shown, it includes step S110 and step S120.

[0070] Step S110: Acquire the current operating status data of the target electronic device. The target electronic device is the electronic device whose health status is currently to be predicted. Its operating status data is various attribute items (also called indicator items) reflecting whether the target electronic device is operating normally. The vector composed of various attribute items is the operating status data. Specifically, taking a hard disk as an example, Figure 2 As shown, the SMART (Self-Monitoring, Analysis and Report Technology) data of the hard disk is usually used as the operating status data. When predicting the health status of the target hard disk, a group of current SMART data of the target hard disk is collected. The vector composed of the group of SMART data is the current operating status data of the target hard disk.

[0071] Step S120: Input the operating status data into the health status prediction model to obtain the current health status level output by the health status prediction model based on the model of the target electronic device. The health status prediction model is constructed based on a multi-domain learning mechanism, each domain corresponds to a model of the same type of electronic device, and the health status prediction model is trained based on operating status data samples of the same type of electronic devices of different models and the health status level labels corresponding to the operating status data samples.

[0072] Specifically, the multi-domain learning mechanism is an important concept in the field of machine learning. It involves learning knowledge from multiple different data sources or fields and being able to transfer this knowledge to a new or different field. This learning method is particularly suitable for multi-source and heterogeneous data learning, using data from different fields to improve the generalization ability of the model so that it can perform better in new or unknown fields. The same type of electronic devices means that the electronic devices are of the same type, for example: the type is hard disk, memory or GPU, etc., each has a different model. For hard disks, the health status prediction model is a health status prediction model for predicting the health status of hard disks. The model of each hard disk can be regarded as a domain in multi-domain learning. The SMART data of a hard disk of a model is the data of a domain. Due to the use of the multi-domain learning mechanism, the health status prediction model can learn common knowledge or features from the running status data samples of hard disks of different models, and the model has better generalization ability. Therefore, when predicting, the current health status level suitable for the model of hard disk can be output according to the model of the target hard disk, so that the prediction of the health status is more accurate.

[0073] Among them, the health status level can be divided according to the actual situation. For example, the simplest way of division is to divide it into two states: non-fault and fault. It can also be divided into three states: non-fault, warning and fault, and each state corresponds to a number of operation status data samples.

[0074] In the electronic device health status prediction method of the present embodiment, the operating status data is input into the health status prediction model to obtain the current health status level output by the health status prediction model based on the model to which the target electronic device belongs, and the health status prediction model is constructed based on a multi-domain learning mechanism, each domain corresponds to a model of the same type of electronic device, and the health status prediction model is trained based on operating status data samples of the same type of electronic devices of different models, and the health status level labels corresponding to the operating status data samples. Due to the use of the multi-domain learning mechanism, when the health status prediction model is trained, its model parameters can learn common knowledge or features from the operating status data samples of electronic devices of different models, as well as learn specific features in the operating status data samples of electronic devices of different models. The finally trained model has better generalization ability and feature expression ability for electronic devices of different models. Therefore, when predicting, it can output the current health status level suitable for the target electronic device of this model according to the model to which the target electronic device belongs, so that the prediction of the health status is more accurate, that is, the health status of the same type of electronic devices of different models can be accurately predicted.

[0075] In some embodiments, Figure 3As shown, the health status prediction model includes: a shared fully connected neural network and N specific domain fully connected neural networks, wherein N is greater than or equal to 2, and a specific domain fully connected neural network corresponds to a model of the electronic device. Figure 3 In the example, taking N=4, it includes a shared fully connected neural network and four specific domain fully connected neural networks. The four specific domain fully connected neural networks correspond to the models A, B, C and D of the electronic devices respectively, forming a star-structured health status prediction model.

[0076] The weight of each model corresponding to the specific domain fully connected neural network is obtained by bitwise multiplication of the shared parameters of the shared fully connected neural network and the specific parameters of the specific domain fully connected neural network, and the weight is used to weight the input feature vector of the corresponding model of the input specific domain fully connected neural network. The shared parameters are used to learn the common features of the operation status data samples of the same type of electronic devices of different models in the training stage, and the specific parameters are used to learn the specific features of the operation status data samples of the same type of electronic devices of specific models in the training stage.

[0077] Specifically, the operating status data of the same type of electronic devices of different models have the same attribute items, for example: Figure 2 In the example, the SMART data of all hard disks include multiple attribute items, each of which has a unique attribute ID. However, due to different models, the attribute values ​​corresponding to these attribute items have different value ranges or units, that is, specific features. As the use time increases, the performance of electronic devices will decay until failure. Therefore, although the attribute values ​​are different, the trend of these attribute values ​​changing over time is the same, that is, shared features. In this embodiment, the shared features between the running status data samples of electronic devices of different models are learned during the training process by sharing the shared parameters of the fully connected neural network, and the specific features of the running status data samples of the corresponding models of electronic devices are learned during the training process by using the specific parameters of each specific domain fully connected neural network, and the shared parameters are multiplied bit by bit by the specific parameters to obtain the weight of the model corresponding to the specific parameters, and the weight is used to weight the input feature vector of the corresponding model of the input specific domain fully connected neural network, so that the result feature vector representing the current health status level output by the specific domain fully connected neural network corresponding to the model contains both the specific features corresponding to the model and the shared features, and has better and more accurate feature expression capabilities, so that the predicted current health status level is more accurate.

[0078] like Figure 4 As shown, in some embodiments, the shared fully connected neural network and the specific domain fully connected neural network each include multiple feature extraction layers, and each of the feature extraction layers includes: a normalization module and a feature extraction module.

[0079] The normalization module is used to normalize the input feature vector according to the mean and variance corresponding to the model of the target electronic device, and the mean and variance are statistically obtained during the inference and prediction process of the health status prediction model. When the health status prediction model is applied, the mean and variance of the input feature vectors of electronic devices of the same model will be statistically analyzed. It should be noted that: for the normalization module of the first layer, its input feature vector is the operating status data of the target electronic device.

[0080] The feature extraction module is used to extract features from the normalized input feature vector.

[0081] In this embodiment, the mean and variance of each field data are different, that is, the mean and variance of each model are different, and the same mean and variance cannot be simply used for normalization. Therefore, for different models of electronic devices, only the respective data are used to calculate the mean and variance, so that for each model, the normalized input feature vector is more accurate.

[0082] Specifically, the normalization module is used to perform normalization according to the following formula:

[0083] .

[0084] in, z represents the input feature vector before normalization, z' represents the normalized input feature vector, E p Indicates the model of the target electronic device p The corresponding input feature vector is the mean, Var p Indicates the model of the target electronic device p The variance of the corresponding input eigenvector, γ represents the shared range vector (shared scale parameter) of the normalization module, γ p Indicates the corresponding model of the normalized module p A specific range vector (specific scale parameter), β express γ The shared bias vector (shared bias vector), β p express γ p A specific bias vector (specific bias vector), ε is a constant, for example: 0.0001.

[0085] In some embodiments, step S120 specifically includes the following steps.

[0086] The operating status data is input into the shared fully connected neural network of the health status prediction model and the specific domain fully connected neural network corresponding to the model of the target electronic device, and the first feature vector output by the shared fully connected neural network and the second feature vector output by the specific domain fully connected neural network are obtained respectively. Specifically, Figure 3 As shown, if the model of the target electronic device is A, the corresponding operating status data will only be input into the shared fully connected neural network and the specific domain fully connected neural network A.

[0087] The first eigenvector and the second eigenvector are bitwise multiplied to obtain a result feature vector representing the current health status level. The feature vectors output by the shared fully connected neural network and the specific domain fully connected neural network are bitwise multiplied to obtain a result feature vector representing the current health status level, which includes both the specific features corresponding to the model and the shared features, and has better and more accurate feature expression capabilities, so that the predicted current health status level is more accurate.

[0088] In some embodiments, the step of inputting the operating status data into a shared fully connected neural network of a health status prediction model and a specific domain fully connected neural network corresponding to the model of the target electronic device, and obtaining a first eigenvector output by the shared fully connected neural network and a second eigenvector output by the specific domain fully connected neural network, respectively, specifically includes the following steps.

[0089] The running status data is input into the shared fully connected neural network, and each layer output of the shared fully connected neural network is as follows.

[0090] .

[0091] Determine the output of the last layer out is the first eigenvector, where W represents the shared parameter, which is the parameter of the feature extraction module in the shared fully connected neural network, in Represents the output of the normalization module of the current layer in the shared fully connected neural network. The feature vectors output by the previous layer must be normalized by the normalization module of the current layer. b express W The shared bias vector of Represents the activation function, which can be Relu, elu, gelu, sigmoid and other activation functions.

[0092] The operating status data is input into the specific domain fully connected neural network, and each layer output of the specific domain fully connected neural network is as follows.

[0093] .

[0094] Determine the output of the last layer out p is the second eigenvector, where in p Indicates the model p The output of the normalization module of the current layer in the corresponding domain-specific fully connected neural network.

[0095] .

[0096] W p Indicates the model p The corresponding specific parameter, the specific parameter is the model p The parameters of the feature extraction module in the corresponding domain-specific fully connected neural network, b p express W p A specific bias vector of Indicates the model p The weight of express A specific bias vector of Represents a bitwise multiplication operation of vectors.

[0097] In some embodiments, before step S110, the method further includes: obtaining the model of the target electronic device, encoding the model to obtain the model identification code of the target electronic device. Specifically, one-hot encoding may be used, for example: Figure 3 There are four models, and the corresponding codes for each model are: 0001, 0010, 0100 and 1000.

[0098] The step of inputting the operating status data into a specific domain fully connected neural network corresponding to the model of the target electronic device includes: identifying the corresponding specific domain fully connected neural network based on the model identification code, and inputting the operating status data into the specific domain fully connected neural network corresponding to the model of the target electronic device.

[0099] In practical applications, the model identification code can also be used as an attribute item of the operating status data and input into the health status prediction model together. Each specific domain fully connected neural network in the health status prediction model receives the corresponding operating status data according to the model identification code. For example: if the model of the current target electronic device is A, then the specific domain fully connected neural network A receives the corresponding operating status data and performs inference prediction after judging according to the model identification code, while other specific domain fully connected neural networks do not receive the operating status data and do not perform inference prediction.

[0100] In some embodiments, the health status prediction model training process is as follows.

[0101] The operating status data sample is input into the shared fully connected neural network of the initial model to obtain a first prediction result output by the shared fully connected neural network. The initial model is the initial state of the health status prediction model, and the parameters in the model are all initial states.

[0102] The running status data sample corresponding to the model identification code is input into the specific domain fully connected neural network corresponding to the model identification code in the initial model to obtain a second prediction result output by the specific domain fully connected neural network.

[0103] The first prediction result and the second prediction result are multiplied bit by bit to obtain a final prediction result.

[0104] Substitute the final prediction result and the health status level label into the loss function. If the loss function does not converge, perform back propagation to update the shared parameters, specific parameters, shared range vectors and specific range vectors. If the loss function converges, the model training is completed. Specifically, when back propagating to update parameters, specific parameters and specific range vectors are only updated according to the gradient of the corresponding model's corresponding running status data samples, and shared parameters and shared range vectors are updated according to the gradient of all samples to learn shared knowledge between different fields (i.e., different models).

[0105] In some embodiments, in order to better learn the specific features of the operating status data samples corresponding to a specific field (i.e., a specific model), an auxiliary neural network corresponding to each model is also introduced during the training phase. Its purpose is to directly model the differences between different types of electronic devices, focus on using features that are strongly related to the field, directly affect the output, and better learn field-specific knowledge. For example: Figure 3 In the network structure, auxiliary neural network A, auxiliary neural network B, auxiliary neural network C and auxiliary neural network D are introduced during training. The specific training steps of this embodiment are as follows.

[0106] The running status data sample is input into the shared fully connected neural network of the initial model to obtain a first prediction result output by the shared fully connected neural network.

[0107] The running status data sample corresponding to the model identification code is input into the specific domain fully connected neural network corresponding to the model identification code in the initial model to obtain a second prediction result output by the specific domain fully connected neural network.

[0108] The first prediction result and the second prediction result are bit-wise multiplied to obtain a third prediction result.

[0109] The intermediate feature vector output by the intermediate layer of the fully connected neural network of the specific domain corresponding to the model identification code is input into the auxiliary neural network corresponding to the model identification code to obtain a fourth prediction result output by the auxiliary neural network, and the fourth prediction result has the same dimension as the third prediction result. Specifically, the auxiliary neural network can also be a multi-layer fully connected neural network. During the training process, the model parameters are updated only according to the gradient of the running status data samples corresponding to the corresponding model, and have a strong correlation with the specific field.

[0110] A final prediction result is determined based on the third prediction result and the fourth prediction result.

[0111] The final prediction result and the health status level label are substituted into the loss function. If the loss function has not converged, back propagation is performed to update the shared parameters, specific parameters, shared range vectors, specific range vectors and model parameters of the auxiliary neural network. If the loss function converges, the model training is completed.

[0112] In some embodiments, determining a final prediction result based on the third prediction result and the fourth prediction result includes: determining the final prediction result according to the following formula.

[0113] .

[0114] in, represents the final prediction result, represents the third prediction result, represents the fourth prediction result, softmax Represents the normalization function.

[0115] In the health status prediction model training of the above embodiment, a cross entropy loss function (Cross EntropyLoss) is used, and the function expression is: ,in, Indicates the health status level label, Represents the model prediction value.

[0116] Since there are many attribute items of the operating status data during the operation of electronic devices, for example, Figure 2In the SMART data of the hard disk, there are many attribute items, and not all attribute items are strongly correlated with the failure of the electronic device. If the operating status data containing all attribute items are used for the training and reasoning of the health status prediction model, unnecessary computing overhead will be increased. Therefore, in some embodiments, before step S110 and before model training, it also includes: obtaining all historical operating status data of the faulty electronic device; based on the historical operating status data, analyzing the N attribute items that cause the most failures, and using the vector composed of the N attribute items as the operating status data for the health status prediction model, where N is greater than or equal to 1, and can be set according to different types of electronic devices, for example: the hard disk can select 10 attribute items to form the operating status data, and the memory can select 12 attribute items to form the operating status data. Since the N attribute items that cause the most failures have a high correlation with the failure of the electronic device, it can not only accurately predict the health status of the electronic device, but also reduce the number of attribute items used in the prediction, thereby reducing the amount of calculation and improving the efficiency of model prediction.

[0117] In practical applications, for faulty electronic devices, the amount of data of the operating status data samples (negative samples) corresponding to the fault state (e.g., 10 data) is much smaller than the amount of data of the operating status data samples (positive samples) corresponding to the non-fault state (e.g., 100,000 data), that is, the samples are seriously unbalanced. When training the model, it is easier to learn the knowledge of the positive samples. The trained model is more accurate in predicting the non-fault state, but less accurate in predicting the fault state. Therefore, in order to alleviate the problem of serious sample imbalance, in some embodiments, the health status level label corresponding to the operating status data sample is set as follows.

[0118] Taking the time when the faulty electronic device corresponding to the operating status data sample fails as the starting time, at least two target time points are determined in advance to divide the time before the starting time into at least three time periods; the health status level label corresponding to the operating status data sample is set according to the time period in which the collection time of the operating status data sample is located. In this embodiment, the health status level label is set in combination with the physical decay characteristics of the electronic device, that is, the length of time before the electronic device fails. The longer the time, the better the health of the hard disk. On the contrary, the hard disk is in a sub-healthy or even fault warning stage. For example: taking days as the time unit, the operating status data samples 10 days before the starting time of the electronic device failure have a health status level label of failure, the operating status data samples 11 to 40 days before the starting time of the electronic device failure have a health status level label of failure warning, the operating status data samples 41 to 80 days before the starting time of the electronic device failure have a health status level label of good, and the operating status data samples 81 days before the starting time of the electronic device failure and earlier have a health status level label of very good. In this way, the 100,000 operating status data samples corresponding to non-fault states will be relatively evenly distributed within the health status level ranges of failure warning, good and very good, thereby achieving sample balance. The health status prediction model trained based on balanced samples can more accurately predict the health status of electronic devices.

[0119] In some embodiments, taking the time when the faulty electronic device corresponding to the operating status data sample fails as the starting time, the step of determining at least two target time points in advance includes the following steps.

[0120] For each attribute item in the running status data sample, a regression decision tree model is used, the time before the start time is used as the fitting target, and the mean absolute error is used as the loss function of the regression decision tree model to fit at least two separation time points that separate the time before the start time. Specifically, Figure 5 and 6 As shown, taking the hard disk as an example, each attribute item fi, i=1,2,…,m in the hard disk SMART data corresponds to an Figure 6 The regression decision tree model shown, in which different hard disk IDs can be considered as different hard disk models.

[0121] The separation time point fitted by the regression decision tree model corresponding to the attribute item with the smallest sum of mean square error (MSE) is selected as the target time point. Figure 6As shown, for each attribute item, the mean absolute error of the collection time (day) of the running status data sample in each leaf node is calculated. The attribute item fj corresponding to the regression decision tree model with the smallest sum of the mean absolute errors of all leaf nodes is selected. Then the separation time points X1, X2 and X3 fitted by the regression decision tree model of the attribute item fj are selected as the target time points to divide into four time periods, time periods 1-4, where X1<X2<X3, for example: X1 is the 10th day before the starting time, X2 is the 40th day before the starting time, and X2 is the 80th day before the starting time. The running status data samples in time period 1 have a health status level label of fault, the running status data samples in time period 2 have a health status level label of fault warning, the running status data samples in time period 3 have a health status level label of good, and the running status data samples in time period 4 have a health status level label of very good.

[0122] In this embodiment, the health status level labels of the operating status data samples of the electronic device are defined by a regression decision tree model, and sample balance is achieved. These samples are used to train the health status prediction model in the above embodiment. The trained health status prediction model can more accurately predict the current health status of the electronic device.

[0123] The electronic device health status prediction device provided by the present invention is described below. The electronic device health status prediction device described below and the electronic device health status prediction method described above can be referenced to each other.

[0124] The electronic device health status prediction device according to the embodiment of the present invention is as follows: Figure 7 As shown, it includes the following modules.

[0125] The data acquisition unit 710 is used to acquire the current operating status data of the target electronic device.

[0126] The model execution unit 720 is used to input the operating status data into the health status prediction model to obtain the current health status level output by the health status prediction model based on the model of the target electronic device.

[0127] Among them, the health status prediction model is constructed based on a multi-domain learning mechanism, each domain corresponds to a model of the same type of electronic device, and the health status prediction model is trained based on operating status data samples of the same type of electronic devices of different models, and the health status level labels corresponding to the operating status data samples.

[0128] In the electronic device health status prediction device of this embodiment, by inputting the operating status data into the health status prediction model, the current health status level output by the health status prediction model based on the model to which the target electronic device belongs is obtained, and the health status prediction model is constructed based on a multi-domain learning mechanism, each domain corresponds to a model of the same type of electronic device, and the health status prediction model is trained based on operating status data samples of the same type of electronic devices of different models, and the health status level labels corresponding to the operating status data samples. Due to the use of the multi-domain learning mechanism, when the health status prediction model is trained, its model parameters can learn common knowledge or features from the operating status data samples of electronic devices of different models, as well as learn specific features in the operating status data samples of electronic devices of different models. The final trained model has better generalization ability and feature expression ability for electronic devices of different models. Therefore, when predicting, it can output the current health status level suitable for the target electronic device of this model according to the model to which the target electronic device belongs, so that the prediction of the health status is more accurate, that is, the health status of the same type of electronic devices of different models can be accurately predicted.

[0129] In some embodiments, the health status prediction model includes: a shared fully connected neural network and N specific domain fully connected neural networks, where N is greater than or equal to 2, and one specific domain fully connected neural network corresponds to one model of the electronic device.

[0130] The weight of each corresponding model of the specific domain fully connected neural network is obtained by bitwise multiplication of the shared parameters of the shared fully connected neural network and the specific parameters of the specific domain fully connected neural network, and the weight is used to weight the input feature vector of the corresponding model of the input specific domain fully connected neural network.

[0131] The shared parameters are used to learn the common features of the operating status data samples of the same type of electronic devices of different models during the training phase, and the specific parameters are used to learn the specific features of the operating status data samples of the same type of electronic devices of a specific model during the training phase.

[0132] In some embodiments, the shared fully connected neural network and the domain-specific fully connected neural network each include a plurality of feature extraction layers, and each of the feature extraction layers includes: a normalization module and a feature extraction module.

[0133] The normalization module is used to normalize the input feature vector according to the mean and variance corresponding to the model of the target electronic device, and the mean and variance are statistically obtained during the inference and prediction process of the health status prediction model.

[0134] The feature extraction module is used to extract features from the normalized input feature vector.

[0135] In some embodiments, the normalization module is used to perform normalization according to the following formula.

[0136] .

[0137] in, z represents the input feature vector before normalization, z' represents the normalized input feature vector, E p Indicates the model of the target electronic device p The corresponding input feature vector is the mean, Var p Indicates the model of the target electronic device p The variance of the corresponding input eigenvector, γ represents the shared range vector of the normalization module, γ p Indicates the corresponding model of the normalized module p A specific range vector of β express γ The shared bias vector of β p express γ p A specific bias vector of ε is a constant.

[0138] In some embodiments, the model execution unit 720 is specifically used to input the operating status data into a shared fully connected neural network of a health status prediction model, and a specific domain fully connected neural network corresponding to the model of the target electronic device, to respectively obtain a first eigenvector output by the shared fully connected neural network and a second eigenvector output by the specific domain fully connected neural network; and multiply the first eigenvector and the second eigenvector bit by bit to obtain a result eigenvector representing the current health status level.

[0139] In some embodiments, the model execution unit 720 is specifically used to input the operating status data into the shared fully connected neural network, and each layer output of the shared fully connected neural network is as follows.

[0140] .

[0141] Determine the output of the last layer out is the first eigenvector, where W represents the shared parameter, which is the parameter of the feature extraction module in the shared fully connected neural network, in represents the output of the normalization module of the current layer in the shared fully connected neural network, b express WThe shared bias vector of Represents the activation function.

[0142] The operating status data is input into the specific domain fully connected neural network, and each layer output of the specific domain fully connected neural network is as follows.

[0143] .

[0144] Determine the output of the last layer out p is the second eigenvector, where in p Indicates the model p The output of the normalization module of the current layer in the corresponding domain-specific fully connected neural network.

[0145] .

[0146] W p Indicates the model p The corresponding specific parameter, the specific parameter is the model p The parameters of the feature extraction module in the corresponding domain-specific fully connected neural network, b p express W p A specific bias vector of Indicates the model p The weight of express A specific bias vector of Represents a bitwise multiplication operation of vectors.

[0147] In some embodiments, the electronic device health status prediction apparatus further includes: an encoding module for obtaining the model of the target electronic device before inputting the operating status data into the health status prediction model, and encoding the model to obtain a model identification code of the target electronic device.

[0148] The model execution unit 720 is specifically used to identify the corresponding domain-specific fully connected neural network based on the model identification code, and input the operating status data into the domain-specific fully connected neural network corresponding to the model of the target electronic device.

[0149] In some embodiments, the health status prediction model training process is as follows.

[0150] The running status data sample is input into the shared fully connected neural network of the initial model to obtain a first prediction result output by the shared fully connected neural network.

[0151] The running status data sample corresponding to the model identification code is input into the specific domain fully connected neural network corresponding to the model identification code in the initial model to obtain a second prediction result output by the specific domain fully connected neural network.

[0152] The first prediction result and the second prediction result are multiplied bit by bit to obtain a final prediction result.

[0153] The final prediction result and the health status level label are substituted into the loss function. If the loss function has not converged, back propagation is performed to update the shared parameters, specific parameters, shared range vectors and specific range vectors. If the loss function converges, the model training is completed.

[0154] In some embodiments, the health status prediction model training process is as follows.

[0155] The running status data sample is input into the shared fully connected neural network of the initial model to obtain a first prediction result output by the shared fully connected neural network.

[0156] The running status data sample corresponding to the model identification code is input into the specific domain fully connected neural network corresponding to the model identification code in the initial model to obtain a second prediction result output by the specific domain fully connected neural network.

[0157] The first prediction result and the second prediction result are bit-wise multiplied to obtain a third prediction result.

[0158] The intermediate feature vector output by the intermediate layer of the fully connected neural network of the specific domain corresponding to the model identification code is input into the auxiliary neural network corresponding to the model identification code to obtain a fourth prediction result output by the auxiliary neural network, and the fourth prediction result has the same dimension as the third prediction result.

[0159] A final prediction result is determined based on the third prediction result and the fourth prediction result.

[0160] The final prediction result and the health status level label are substituted into the loss function. If the loss function has not converged, back propagation is performed to update the shared parameters, specific parameters, shared range vectors, specific range vectors and model parameters of the auxiliary neural network. If the loss function converges, the model training is completed.

[0161] In some embodiments, determining the final prediction result based on the third prediction result and the fourth prediction result includes: determining the final prediction result according to the following formula.

[0162] .

[0163] in, represents the final prediction result, represents the third prediction result, represents the fourth prediction result, softmax Represents the normalization function.

[0164] In some embodiments, the health status level label corresponding to the operating status data sample is set as follows.

[0165] Taking the time when the faulty electronic device corresponding to the operating status data sample fails as the starting time, at least two target time points are determined in the forward direction to divide the time in the forward direction of the starting time into at least three time periods.

[0166] The health status level label corresponding to the running status data sample is set according to the time period in which the running status data sample is collected.

[0167] In some embodiments, the moment when the faulty electronic device corresponding to the operating status data sample fails is taken as the starting moment, and at least two target time points are determined in advance, including: for each attribute item in the operating status data sample, a regression decision tree model is used, the time before the starting moment is taken as the fitting target, and the mean absolute error is used as the loss function of the regression decision tree model to fit at least two separated time points that separate the time before the starting moment; the separated time point fitted by the regression decision tree model corresponding to the attribute item with the smallest sum of mean absolute errors is selected as the target time point.

[0168] Figure 8 An example of a physical structure diagram of an electronic device is shown in FIG. Figure 8 As shown, the electronic device may include: a processor 810, a communication interface 820, a memory 830 and a communication bus 840, wherein the processor 810, the communication interface 820 and the memory 830 communicate with each other through the communication bus 840. The processor 810 may call the logic instructions in the memory 830 to execute the electronic device health state prediction method, which includes the following steps.

[0169] Obtain the current operating status data of the target electronic device.

[0170] The operating status data is input into a health status prediction model to obtain a current health status level output by the health status prediction model based on the model of the target electronic component.

[0171] Among them, the health status prediction model is constructed based on a multi-domain learning mechanism, each domain corresponds to a model of the same type of electronic device, and the health status prediction model is trained based on operating status data samples of the same type of electronic devices of different models, and the health status level labels corresponding to the operating status data samples.

[0172] In addition, the logic instructions in the above-mentioned memory 830 can be implemented in the form of a software functional unit and can be stored in a computer-readable storage medium when it is sold or used as an independent product. Based on this understanding, the technical solution of the present invention is essentially or the part that contributes to the prior art or the part of the technical solution can be embodied in the form of a software product, and the computer software product is stored in a storage medium, including a number of instructions to enable a computer device (which can be a personal computer, a server, or a network device, etc.) to perform all or part of the steps of the method described in each embodiment of the present invention. The aforementioned storage medium includes: U disk, mobile hard disk, read-only memory (ROM, Read-Only Memory), random access memory (RAM, Random Access Memory), disk or optical disk and other media that can store program codes.

[0173] On the other hand, the present invention also provides a computer program product, which includes a computer program. The computer program can be stored on a non-transitory computer-readable storage medium. When the computer program is executed by a processor, the computer can execute the electronic device health status prediction method provided by the above-mentioned methods, which includes the following steps.

[0174] Obtain the current operating status data of the target electronic device.

[0175] The operating status data is input into a health status prediction model to obtain a current health status level output by the health status prediction model based on the model of the target electronic component.

[0176] Among them, the health status prediction model is constructed based on a multi-domain learning mechanism, each domain corresponds to a model of the same type of electronic device, and the health status prediction model is trained based on operating status data samples of the same type of electronic devices of different models, and the health status level labels corresponding to the operating status data samples.

[0177] On the other hand, the present invention also provides a non-transitory computer-readable storage medium having a computer program stored thereon. When the computer program is executed by a processor, the method for predicting the health status of an electronic device provided by the above-mentioned methods is implemented, and the method includes the following steps.

[0178] Obtain the current operating status data of the target electronic device.

[0179] The operating status data is input into a health status prediction model to obtain a current health status level output by the health status prediction model based on the model of the target electronic component.

[0180] Among them, the health status prediction model is constructed based on a multi-domain learning mechanism, each domain corresponds to a model of the same type of electronic device, and the health status prediction model is trained based on operating status data samples of the same type of electronic devices of different models, and the health status level labels corresponding to the operating status data samples.

[0181] The device embodiments described above are merely illustrative, wherein the units described as separate components may or may not be physically separated, and the components displayed as units may or may not be physical units, that is, they may be located in one place, or they may be distributed on multiple network units. Some or all of the modules may be selected according to actual needs to achieve the purpose of the scheme of this embodiment. Ordinary technicians in this field can understand and implement it without paying creative labor.

[0182] Through the description of the above implementation methods, those skilled in the art can clearly understand that each implementation method can be implemented by means of software plus a necessary general hardware platform, and of course, can also be implemented by hardware. Based on this understanding, the above technical solution is essentially or the part that contributes to the prior art can be embodied in the form of a software product, and the computer software product can be stored in a computer-readable storage medium, such as ROM / RAM, a disk, an optical disk, etc., including a number of instructions for a computer device (which can be a personal computer, a server, or a network device, etc.) to execute the methods described in each embodiment or some parts of the embodiments.

[0183] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit it. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the aforementioned embodiments, or make equivalent replacements for some of the technical features therein. However, these modifications or replacements do not deviate the essence of the corresponding technical solutions from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method for predicting the health status of an electronic device, characterized in that: include: Acquire current operating status data of the target electronic device; Inputting the operating status data into a health status prediction model to obtain a current health status level output by the health status prediction model based on the model of the target electronic device; The health status prediction model is constructed based on a multi-domain learning mechanism, each domain corresponds to a model of the same type of electronic device, and the health status prediction model is trained based on operating status data samples of the same type of electronic devices of different models and health status level labels corresponding to the operating status data samples; The health status prediction model includes: a shared fully connected neural network and N specific domain fully connected neural networks, wherein N is greater than or equal to 2, and one specific domain fully connected neural network corresponds to one model of the electronic device; The weight of each model corresponding to the specific domain fully connected neural network is obtained by bitwise multiplication of the shared parameters of the shared fully connected neural network and the specific parameters of the specific domain fully connected neural network, and the weight is used to weight the input feature vector of the corresponding model of the input specific domain fully connected neural network; The shared parameters are used to learn the common features of the operating status data samples of the same type of electronic devices of different models during the training phase, and the specific parameters are used to learn the specific features of the operating status data samples of the same type of electronic devices of a specific model during the training phase.

2. The method for predicting the health status of an electronic device according to claim 1, characterized in that: The shared fully connected neural network and the domain-specific fully connected neural network each include a plurality of feature extraction layers, each of the feature extraction layers includes: a normalization module and a feature extraction module; The normalization module is used to normalize the input feature vector according to the mean and variance corresponding to the model of the target electronic device, wherein the mean and variance are statistically obtained during the inference and prediction process of the health status prediction model; The feature extraction module is used to extract features from the normalized input feature vector.

3. The method for predicting the health status of an electronic device according to claim 2, characterized in that: The normalization module is used to perform normalization according to the following formula: ； in, z represents the input feature vector before normalization, z' represents the normalized input feature vector, E p Indicates the model of the target electronic device p The corresponding input feature vector is the mean, Var p Indicates the model of the target electronic device p The variance of the corresponding input eigenvector, γ represents the shared range vector of the normalization module, γ p Indicates the corresponding model of the normalized module p A specific range vector of β express γ The shared bias vector of β p express γ p A specific bias vector of ε is a constant.

4. The method for predicting the health status of an electronic device according to claim 2, characterized in that: Inputting the operating status data into a health status prediction model to obtain a current health status level output by the health status prediction model based on the model of the target electronic device includes: Inputting the operating status data into a shared fully connected neural network of a health status prediction model and a domain-specific fully connected neural network corresponding to the model of the target electronic device, respectively obtaining a first eigenvector output by the shared fully connected neural network and a second eigenvector output by the domain-specific fully connected neural network; The first feature vector and the second feature vector are bitwise multiplied to obtain a result feature vector representing the current health status level.

5. The method for predicting the health status of an electronic device according to claim 4, characterized in that: The operating status data is input into a shared fully connected neural network of a health status prediction model and a domain-specific fully connected neural network corresponding to the model of the target electronic device, and a first feature vector output by the shared fully connected neural network and a second feature vector output by the domain-specific fully connected neural network are obtained respectively, including: The running status data is input into the shared fully connected neural network, and each layer output of the shared fully connected neural network is as follows: ； Determine the output of the last layer out is the first eigenvector, where W represents the shared parameter, which is the parameter of the feature extraction module in the shared fully connected neural network, in represents the output of the normalization module of the current layer in the shared fully connected neural network, b express W The shared bias vector of represents the activation function; The running status data is input into the specific domain fully connected neural network, and each layer output of the specific domain fully connected neural network is as follows: ； Determine the output of the last layer out p is the second eigenvector, where in p Indicates the model p The output of the normalization module of the current layer in the corresponding domain-specific fully connected neural network; ； W p Indicates the model p The corresponding specific parameter, the specific parameter is the model p The parameters of the feature extraction module in the corresponding domain-specific fully connected neural network, b p express W p A specific bias vector of Indicates the model p The weight of express A specific bias vector of Represents a bitwise multiplication operation of vectors.

6. The method for predicting the health status of an electronic device according to claim 4, characterized in that: Before inputting the operating status data into the health status prediction model, the method further includes: Acquire the model of the target electronic device, and encode the model to obtain a model identification code of the target electronic device; The operating status data is input into a specific domain fully connected neural network corresponding to the model of the target electronic device, including: identifying the corresponding specific domain fully connected neural network based on the model identification code, and inputting the operating status data into the specific domain fully connected neural network corresponding to the model of the target electronic device.

7. The method for predicting the health status of an electronic device according to claim 3, characterized in that: The health status prediction model training process is as follows: Inputting the running status data sample into a shared fully connected neural network of an initial model to obtain a first prediction result output by the shared fully connected neural network; Inputting the running status data sample corresponding to the model identification code into the specific domain fully connected neural network corresponding to the model identification code in the initial model to obtain a second prediction result output by the specific domain fully connected neural network; Multiplying the first prediction result and the second prediction result bit by bit to obtain a final prediction result; The final prediction result and the health status level label are substituted into the loss function. If the loss function has not converged, back propagation is performed to update the shared parameters, specific parameters, shared range vectors and specific range vectors. If the loss function converges, the model training is completed.

8. The method for predicting the health status of an electronic device according to claim 3, characterized in that: The health status prediction model training process is as follows: Inputting the running status data sample into a shared fully connected neural network of an initial model to obtain a first prediction result output by the shared fully connected neural network; Inputting the running status data sample corresponding to the model identification code into the specific domain fully connected neural network corresponding to the model identification code in the initial model to obtain a second prediction result output by the specific domain fully connected neural network; Multiplying the first prediction result and the second prediction result bit by bit to obtain a third prediction result; Inputting the intermediate feature vector output by the intermediate layer of the specific domain fully connected neural network corresponding to the model identification code into the auxiliary neural network corresponding to the model identification code, to obtain a fourth prediction result output by the auxiliary neural network, wherein the fourth prediction result has the same dimension as the third prediction result; Determine a final prediction result based on the third prediction result and the fourth prediction result; The final prediction result and the health status level label are substituted into the loss function. If the loss function has not converged, back propagation is performed to update the shared parameters, specific parameters, shared range vectors, specific range vectors and model parameters of the auxiliary neural network. If the loss function converges, the model training is completed.

9. The method for predicting the health status of an electronic device according to claim 8, characterized in that: Determining a final prediction result based on the third prediction result and the fourth prediction result includes: determining the final prediction result according to the following formula: ； in, represents the final prediction result, represents the third prediction result, represents the fourth prediction result, softmax Represents the normalization function.

10. The method for predicting the health status of an electronic device according to any one of claims 1 to 9, characterized in that: The health status level label corresponding to the running status data sample is set as follows: Taking the time when the faulty electronic device corresponding to the running status data sample fails as the starting time, at least two target time points are determined forward to divide the time forward of the starting time into at least three time periods; The health status level label corresponding to the running status data sample is set according to the time period in which the running status data sample is collected.

11. The method for predicting the health status of an electronic device according to claim 10, characterized in that: Taking the time when the faulty electronic device corresponding to the running status data sample fails as the starting time, at least two target time points are determined in the forward direction, including: For each attribute item in the running status data sample, a regression decision tree model is used, the time before the starting moment is used as the fitting target, and the mean absolute error is used as the loss function of the regression decision tree model to fit at least two separation time points that separate the time before the starting moment; The separation time point fitted by the regression decision tree model corresponding to the attribute item with the smallest sum of mean absolute errors is selected as the target time point.

12. An electronic device health status prediction device, characterized in that: include: A data acquisition unit, used to acquire current operating status data of the target electronic device; A model execution unit, configured to input the operating status data into a health status prediction model to obtain a current health status level output by the health status prediction model based on the model of the target electronic device; The health status prediction model is constructed based on a multi-domain learning mechanism, each domain corresponds to a model of the same type of electronic device, and the health status prediction model is trained based on operating status data samples of the same type of electronic devices of different models and health status level labels corresponding to the operating status data samples; The health status prediction model includes: a shared fully connected neural network and N specific domain fully connected neural networks, wherein N is greater than or equal to 2, and one specific domain fully connected neural network corresponds to one model of the electronic device; The weight of each model corresponding to the specific domain fully connected neural network is obtained by bitwise multiplication of the shared parameters of the shared fully connected neural network and the specific parameters of the specific domain fully connected neural network, and the weight is used to weight the input feature vector of the corresponding model of the input specific domain fully connected neural network; The shared parameters are used to learn the common features of the operating status data samples of the same type of electronic devices of different models during the training phase, and the specific parameters are used to learn the specific features of the operating status data samples of the same type of electronic devices of a specific model during the training phase.

13. An electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein: When the processor executes the computer program, the method for predicting the health status of an electronic device according to any one of claims 1 to 11 is implemented.

14. A non-transitory computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the method for predicting the health status of an electronic device according to any one of claims 1 to 11 is implemented.

Images