Model training method, life prediction method, and life prediction device

Abstract

The embodiment of the present application discloses a model training method, a life prediction method and a life prediction device, wherein the life prediction method comprises: collecting multi-dimensional health index correlation data of a target prediction object, and inputting the multi-dimensional health index correlation data into a health index prediction correlation model to calculate the current comprehensive health index of the target prediction object through the health index prediction correlation model; constructing a current comprehensive health index sequence according to the current comprehensive health index of the target prediction object; inputting the current comprehensive health index sequence of the target prediction object into a residual life prediction model to predict the residual service life of the target prediction object through the residual life prediction model. The technical scheme of the embodiment of the present application can improve the model prediction accuracy of the health index prediction correlation model and the residual life prediction model, and further improve the accuracy of predicting the residual service life of an object based on the health index prediction correlation model and the residual life prediction model.

Description

Technical Field

[0001] The embodiments of the present invention relate to the field of fault prediction and health management technology, and in particular to a model training method, a life prediction method, an apparatus, an electronic device, a storage medium, and a program. Background Technology

[0002] Components, parts, and complete sets of equipment that have been in continuous service for a long time will experience a gradual decline in performance over time. When the operating condition of such objects approaches the critical failure threshold, it can easily lead to significant safety hazards and cause significant economic losses. To further improve the operational reliability and intrinsic safety of the system, it is necessary to rely on Remaining Useful Life (RUL) prediction technology, combined with the characteristic parameters exhibited by a certain type of object during the performance degradation process, to make forward-looking predictions and quantitative assessments of its failure time.

[0003] RUL technology analyzes real-time operational data and historical operating condition traceability information of objects to accurately predict the time when their health status reaches the critical failure point. This provides core support for the scientific formulation of operation and maintenance strategies and the optimization and upgrading of operation management models, effectively reducing the incidence of sudden object failures and ensuring the continuous, efficient and stable operation of the system.

[0004] In developing this invention, the inventors discovered that existing Remaining Lifespan (RUL) technology requires predicting the health indicators of an object using a health indicator prediction association model, and then using the RUL model to predict the remaining lifespan based on the health indicator prediction results. However, because the health indicator prediction association model does not consider the multi-dimensional distribution characteristics of health status, and the RUL model does not consider the evolutionary laws of object degradation, the prediction accuracy of both the health indicator prediction association model and the RUL model is low, resulting in low accuracy of the remaining lifespan predicted based on these models. Summary of the Invention

[0005] This invention provides a model training method, a lifespan prediction method, an apparatus, an electronic device, a storage medium, and a program, which can improve the model prediction accuracy of health indicator prediction association models and remaining lifespan prediction models, thereby improving the accuracy of predicting the remaining lifespan of an object based on health indicator prediction association models and remaining lifespan prediction models.

[0006] According to one aspect of the present invention, a model training method is provided, comprising:

[0007] Collect multidimensional health index-related sample data of the target prediction object;

[0008] The multidimensional health index associated sample data of the target prediction object is input into the health index prediction association model, so as to calculate the predicted comprehensive health index of the target prediction object through the health index prediction association model;

[0009] The loss value of the predicted comprehensive health index of the target prediction object is calculated by the composite loss function of the health index prediction association model; wherein, the composite loss function of the health index prediction association model includes a reconstruction loss parameter and a compaction loss parameter;

[0010] If it is determined that the loss value of the predicted comprehensive health indicator does not meet the termination condition of the indicator prediction model training, the operation of collecting multidimensional health indicator-related sample data of the target prediction object is returned until it is determined that the loss value of the predicted comprehensive health indicator meets the termination condition of the indicator prediction model training.

[0011] According to another aspect of the present invention, a model training method is provided, comprising:

[0012] Obtain a comprehensive health index sequence sample of the target prediction object and the actual remaining lifespan corresponding to the comprehensive health index sequence sample;

[0013] A life expectancy prediction sample is constructed based on the comprehensive health index sequence sample and the actual remaining life expectancy.

[0014] The comprehensive health index sequence sample in the lifespan prediction sample is input into the remaining lifespan prediction model so as to predict the predicted remaining lifespan corresponding to the comprehensive health index sequence sample through the remaining lifespan prediction model.

[0015] The loss value between the predicted remaining useful life and the actual remaining useful life is calculated using the physical monotonicity composite loss function of the remaining useful life prediction model; wherein, the physical monotonicity composite loss function includes prediction error parameters and physical monotonicity constraint parameters;

[0016] If it is determined that the loss value between the predicted remaining lifespan and the actual remaining lifespan does not meet the lifespan prediction model training termination condition, the operation of obtaining the comprehensive health index sequence sample of the target prediction object and the actual remaining lifespan corresponding to the comprehensive health index sequence sample is returned until it is determined that the loss value between the predicted remaining lifespan and the actual remaining lifespan meets the lifespan prediction model training termination condition.

[0017] The trained remaining life expectancy prediction model is used to calculate the current comprehensive health index of the target prediction object based on the health index prediction association model trained according to the model training method described in the first aspect, and to predict the remaining life expectancy of the target prediction object.

[0018] According to another aspect of the present invention, a lifespan prediction method is provided, comprising:

[0019] Collect multidimensional health indicator correlation data of the target prediction object, and input the multidimensional health indicator correlation data into the health indicator prediction correlation model, so as to calculate the current comprehensive health indicator of the target prediction object through the health indicator prediction correlation model;

[0020] Construct a current comprehensive health index sequence based on the current comprehensive health index of the target prediction object;

[0021] The current comprehensive health index sequence of the target prediction object is input into the remaining life expectancy prediction model so as to predict the remaining life expectancy of the target prediction object through the remaining life expectancy prediction model;

[0022] The health indicator prediction association model is trained using the model training method described in the first aspect, and the remaining life expectancy prediction model is trained using the model training method described in the second aspect.

[0023] According to another aspect of the present invention, a model training apparatus is provided, comprising:

[0024] The multidimensional health indicator correlation sample data acquisition module is used to collect multidimensional health indicator correlation sample data of the target prediction object;

[0025] The predictive comprehensive health index calculation module is used to input the multidimensional health index association sample data of the target prediction object into the health index prediction association model, so as to calculate the predicted comprehensive health index of the target prediction object through the health index prediction association model.

[0026] The comprehensive health index loss value calculation module is used to calculate the loss value of the predicted comprehensive health index of the target prediction object through the composite loss function of the health index prediction association model; wherein, the composite loss function of the health index prediction association model includes a reconstruction loss parameter and a compaction loss parameter;

[0027] The first iterative training module is used to return to the operation of collecting multidimensional health index-related sample data of the target prediction object when it is determined that the loss value of the predicted comprehensive health index does not meet the termination condition of the indicator prediction model training, until it is determined that the loss value of the predicted comprehensive health index meets the termination condition of the indicator prediction model training.

[0028] According to another aspect of the present invention, a model training apparatus is provided, comprising:

[0029] The indicator sequence lifespan data acquisition module is used to acquire a comprehensive health indicator sequence sample of the target prediction object and the actual remaining lifespan corresponding to the comprehensive health indicator sequence sample.

[0030] The lifespan prediction sample construction module is used to construct lifespan prediction samples based on the comprehensive health index sequence samples and the actual remaining lifespan.

[0031] The remaining useful life prediction module is used to input the comprehensive health index sequence sample in the life prediction sample into the remaining useful life prediction model, so as to predict the predicted remaining useful life corresponding to the comprehensive health index sequence sample through the remaining useful life prediction model.

[0032] The predicted lifetime loss value calculation module is used to calculate the loss value between the predicted remaining lifetime and the actual remaining lifetime through the physical monotonic composite loss function of the remaining lifetime prediction model; wherein, the physical monotonic composite loss function includes prediction error parameters and physical monotonicity constraint parameters;

[0033] The second iterative training module is used to return to the operation of obtaining the comprehensive health index sequence sample of the target prediction object and the actual remaining lifespan corresponding to the comprehensive health index sequence sample when it is determined that the loss value between the predicted remaining lifespan and the actual remaining lifespan does not meet the lifespan prediction model training termination condition.

[0034] The trained remaining life expectancy prediction model is used to calculate the current comprehensive health index of the target prediction object based on the health index prediction association model trained according to the model training method described in the first aspect, and to predict the remaining life expectancy of the target prediction object.

[0035] According to another aspect of the present invention, a lifespan prediction device is provided, comprising:

[0036] The current comprehensive health index calculation module is used to collect multidimensional health index correlation data of the target prediction object, and input the multidimensional health index correlation data into the health index prediction correlation model, so as to calculate the current comprehensive health index of the target prediction object through the health index prediction correlation model;

[0037] The current comprehensive health index sequence construction module is used to construct a current comprehensive health index sequence based on the current comprehensive health index of the target prediction object;

[0038] The remaining useful life prediction module is used to input the current comprehensive health index sequence of the target prediction object into the remaining useful life prediction model, so as to predict the remaining useful life of the target prediction object through the remaining useful life prediction model;

[0039] The health indicator prediction association model is trained using the model training method described in the first aspect, and the remaining life expectancy prediction model is trained using the model training method described in the second aspect.

[0040] According to another aspect of the present invention, an electronic device is provided, the electronic device comprising:

[0041] At least one processor;

[0042] and a memory communicatively connected to the at least one processor; wherein,

[0043] The memory stores a computer program that can be executed by the at least one processor, which enables the at least one processor to perform the model training method or lifetime prediction method according to any embodiment of the present invention.

[0044] According to another aspect of the present invention, a computer-readable storage medium is provided, the computer-readable storage medium storing computer instructions for causing a processor to execute and implement the model training method or lifetime prediction method according to any embodiment of the present invention.

[0045] According to another aspect of the present invention, a computer program product is also provided, comprising a computer program that, when executed by a processor, implements the model training method or lifetime prediction method described in any embodiment of the present invention.

[0046] This invention, in its embodiments, inputs multidimensional health indicator-related sample data of the target prediction object into a health indicator prediction association model. The model calculates the predicted comprehensive health indicator of the target prediction object and uses a composite loss function to calculate the loss value of the predicted comprehensive health indicator. The iterative training process of the health indicator prediction association model is then achieved based on the matching between the loss value of the predicted comprehensive health indicator and the training termination condition of the indicator prediction model. Since the composite loss function of the health indicator prediction association model includes a reconstruction loss parameter and a compaction loss parameter, and the reconstruction loss parameter measures the model's ability to reconstruct the original signal, ensuring that the feature vector contains sufficient physical information, while the compaction loss parameter measures the degree to which the sample's feature vector deviates from the preset health center, the composite loss function can balance the preservation of the signal's physical information with the distribution pattern of the features. Training the health indicator prediction association model based on the composite loss function can improve the accuracy of its health indicator prediction. Simultaneously, this embodiment of the invention also obtains comprehensive health index sequence samples of the target prediction object and the actual remaining lifespan corresponding to the comprehensive health index sequence samples to construct lifespan prediction samples. The comprehensive health index sequence samples in the lifespan prediction samples are then input into the remaining lifespan prediction model to predict the predicted remaining lifespan corresponding to the comprehensive health index sequence samples. Furthermore, the loss value between the predicted remaining lifespan and the actual remaining lifespan is calculated using the physical monotonicity composite loss function of the remaining lifespan prediction model. This allows for iterative training of the remaining lifespan prediction model based on the matching between the calculated loss value and the training termination condition of the remaining lifespan prediction model. Since the physical monotonicity composite loss function of the remaining lifespan prediction model includes prediction error parameters and physical monotonicity constraint parameters, and the prediction error parameters ensure that the predicted value is close to the true value, while the physical monotonicity constraint parameters penalize predictions that violate physical degradation laws, an accurate remaining lifespan prediction model that conforms to the physical monotonicity trend can be trained based on the physical monotonicity composite loss function. After model training is completed, multidimensional health index correlation data of the target prediction object can be collected and input into the health index prediction correlation model to calculate the current comprehensive health index of the target prediction object. Furthermore, a current comprehensive health index sequence is constructed based on the current comprehensive health index of the target prediction object. This sequence is then input into the remaining life expectancy prediction model to predict the remaining life expectancy of the target prediction object. Therefore, the above technical solution can improve the prediction accuracy of both the health index prediction association model and the remaining life expectancy prediction model during model training, thereby improving the accuracy of predicting the remaining life expectancy of a specific object based on these two models.

[0047] It should be understood that the description in this section is not intended to identify key or essential features of the embodiments of the present invention, nor is it intended to limit the scope of the invention. Other features of the invention will become readily apparent from the following description. Attached Figure Description

[0048] To more clearly illustrate the technical solutions in the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0049] Figure 1 This is a flowchart of a model training method provided in Embodiment 1 of the present invention;

[0050] Figure 2 This is a schematic diagram of the structure of a health indicator prediction association model provided in Embodiment 1 of the present invention;

[0051] Figure 3 This is a flowchart of a model training method provided in Embodiment 2 of the present invention;

[0052] Figure 4 This is a flowchart of a lifetime prediction method provided in Embodiment 3 of the present invention;

[0053] Figure 5 This is a schematic diagram of a model training device provided in Embodiment 4 of the present invention;

[0054] Figure 6 This is a schematic diagram of a model training device provided in Embodiment 5 of the present invention;

[0055] Figure 7 This is a schematic diagram of a model training device provided in Embodiment Six of the present invention;

[0056] Figure 8 This is a schematic diagram of the structure of an electronic device provided in Embodiment 7 of the present invention. Detailed Implementation

[0057] To enable those skilled in the art to better understand the present invention, the technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort should fall within the scope of protection of the present invention.

[0058] It should be noted that the terms "first," "second," etc., in the specification, claims, and accompanying drawings of this invention are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. It should be understood that such data can be interchanged where appropriate so that the embodiments of the invention described herein can be implemented in orders other than those illustrated or described herein. Furthermore, the terms "comprising" and "having," and any variations thereof, are intended to cover non-exclusive inclusion; for example, a process, method, system, product, or apparatus that comprises a series of steps or units is not necessarily limited to those steps or units explicitly listed, but may include other steps or units not explicitly listed or inherent to such processes, methods, products, or apparatus.

[0059] Example 1

[0060] Figure 1 This is a flowchart of a model training method provided in Embodiment 1 of the present invention. This embodiment is applicable to the situation of training a health indicator prediction correlation model based on a composite loss function. This method can be executed by a model training device, which can be implemented in software and / or hardware, and is generally integrated into an electronic device. This electronic device can be a terminal device or a server device, as long as it can execute the model training method. The present invention does not limit the specific type of electronic device. Correspondingly, as... Figure 1 As shown, the method includes the following operations:

[0061] S110. Collect multidimensional health indicator correlation sample data of the target prediction object.

[0062] The target prediction object can be an object with a certain service life, such as a part, device, component, or equipment. For example, the target prediction object can be a bearing, relay, machining tool, sensor, pressure-bearing component, or engine, as long as it has characteristics of degradation or wear based on usage time, resulting in a predictable service life. This embodiment of the invention does not limit the specific type of the target prediction object. The multidimensional health indicator-related sample data can be collected from the target prediction object and can reflect the health status of the target prediction object. This embodiment of the invention does not limit the specific data types and content included in the multidimensional health indicator-related sample data.

[0063] In a specific example, the machining tool can be considered as the target for prediction. Machining tools are used to machine various parts, such as gears, bearings, or other types of components. The machining accuracy and surface quality of these parts directly affect the performance and reliability of the final product. In gear machining, the machining tool is the key cutting tool, and its performance and lifespan have a decisive impact on machining efficiency, machining quality, and production costs. Machining tools are subject to various forms of damage during cutting, including wear, chipping, and plastic deformation. Tool wear is a complex and gradual process; as cutting time increases, the cutting edge gradually dulls, leading to increased cutting force, higher cutting temperature, and decreased surface quality, and may even cause equipment vibration and workpiece scrap. Therefore, accurately understanding the wear state of the tool and predicting its remaining life is crucial for achieving intelligent, efficient, and cost-optimized machining.

[0064] For example, when the machining tool is used as the target prediction object, multi-source data on tool wear, including but not limited to sensor parameters, equipment operating status parameters, machining process parameters, and tool status data, can be collected in real time during a specific machining scenario (such as gear machining). These data include, but are not limited to, vibration acceleration (m / s²) of key machine tool components measured by vibration sensors to reflect the dynamic stability of the machining process; they can also include parameters from a tool temperature sensor, where increased temperature indicates accelerated wear and affects friction / adhesion. Equipment operating status parameters can include, but are not limited to, the cutting speed, spindle speed, feed rate, and depth of cut of the machining tool, as well as the spindle motor current, power, and torque. Machining process parameters can include, but are not limited to, those recorded by the operating interface or PLC (Programmable Logic Controller) system, including set depth of cut, feed rate, and tool compensation values. Tool status data can include, but is not limited to, tool status labels such as good, usable, critical, and scrap, and can also include related labels such as total machining time and total number of parts.

[0065] After collecting multidimensional health indicator-related sample data of the target prediction object, a series of preprocessing operations can be performed to improve data quality and model training efficiency. For example, data preprocessing operations may include, but are not limited to, data cleaning and denoising, data synchronization and alignment, data normalization, and data segmentation. Data cleaning and denoising can identify and remove outliers, spike noise, outliers, and missing data from the multidimensional health indicator-related sample data, using methods such as median filtering, Kalman filtering, and wavelet denoising. Data synchronization and alignment ensures that all sensor data are accurately aligned on the time axis for subsequent data fusion. Data normalization unifies data with different dimensions and amplitude ranges to the same scale, for example, using Min-Max normalization (deviation standardization or range method) or Z-score normalization (standard score normalization or standard deviation standardization) to eliminate the influence of different features on model training. Data segmentation divides the continuously collected data into multiple data segments based on the periodicity of the target prediction object or a fixed time window, with each data segment corresponding to a window.

[0066] S120. Input the multidimensional health index association sample data of the target prediction object into the health index prediction association model, so as to calculate the predicted comprehensive health index of the target prediction object through the health index prediction association model.

[0067] Among them, the predicted comprehensive health index can be a comprehensive health index calculated by the health index prediction association model based on multidimensional health index association sample data for the target prediction object.

[0068] In an embodiment of the present invention, optionally, a deep neural network model containing an encoder and a decoder can be constructed as a health indicator prediction association model to be trained. Figure 2 This is a schematic diagram of the structure of a health indicator prediction association model provided in Embodiment 1 of the present invention. In a specific example, such as... Figure 2 As shown, the encoder of the health indicator prediction association model can employ a structure combining a convolutional neural network and a multi-head attention mechanism to map the preprocessed high-dimensional original input data X into a low-dimensional latent variable feature vector Z. The decoder of the health indicator prediction association model can use a deconvolutional structure symmetrical to the encoder to reconstruct the latent variable features Z into data approximating the original input. It should be noted that, Figure 2 This is merely a schematic diagram and does not represent that the health indicator prediction correlation model has only one convolutional layer, one deconvolutional layer, and one multi-head attention module.

[0069] Before training the health indicator prediction association model, a vector C with the same dimension as the latent variable feature Z can be defined as the "health hypersphere center" of the feature space to complete the hypersphere center initialization. During the hypersphere center initialization process, a small amount of state data of the target prediction object in a healthy state (such as machining tool data in a healthy state) can be selected and input into the initialized health indicator prediction association model. The mean of its output features is calculated and assigned to the hypersphere center C. In the subsequent training process of the health indicator prediction association model, the hypersphere center C remains fixed (not trainable) and serves as the geometric reference point that all healthy sample features must converge to.

[0070] During each training iteration of the health indicator prediction association model, multidimensional health indicator association sample data of the target prediction object can be input into the model. For example, gear machining data with machining tools in a healthy state (such as new or break-in period) can be used as training samples to calculate the predicted comprehensive health indicator of the target prediction object through the output data of the relevant layers of the health indicator prediction association model. The health indicator prediction association model can predict the predicted value corresponding to the output sample data based on the multidimensional health indicator association sample data, but it cannot directly predict the predicted comprehensive health indicator of the target prediction object. The predicted comprehensive health indicator of the prediction object can be indirectly calculated based on the features output by the internal network structure of the health indicator prediction association model. Optionally, the health indicator prediction association model can be trained by minimizing the model's loss function using the backpropagation algorithm.

[0071] S130. Calculate the loss value of the predicted comprehensive health index of the target prediction object through the composite loss function of the health index prediction association model; wherein, the composite loss function of the health index prediction association model includes a reconstruction loss parameter and a compaction loss parameter.

[0072] The reconstruction loss parameter measures the ability of the health indicator prediction association model to reconstruct the original signal, ensuring that the feature vector contains sufficient physical information. The compaction loss parameter measures the degree to which the feature vector of the current multidimensional health indicator association sample data deviates from the preset health center.

[0073] In an optional embodiment of the present invention, in order to balance the preservation of the physical information of the signal and the distribution pattern of the features, the composite loss function of the health index prediction correlation model can be:

[0074]

[0075]

[0076]

[0077] in, This represents the composite loss function. The reconstruction loss parameter is calculated using the mean squared error method; The compaction loss parameter is defined as follows: it forces the feature vectors of the multidimensional health indicator associated sample data corresponding to all healthy samples to be compressed as much as possible into a minimal hypersphere centered at C; λ1 represents the balance coefficient, used to adjust the weights of the two loss terms to ensure that the health indicator prediction association model satisfies the distribution constraints while learning effective features; N is the number of the first training samples. This represents the i-th data sample. This represents the predicted value of the target object relative to the original data sample. C represents the latent variable feature vector of the current data sample, and C represents the preset health center vector.

[0078] Therefore, the health indicator prediction association model provided in this embodiment of the invention is different from the conventional autoencoder that only targets reconstruction error. It also introduces "hypersphere compact constraint" to force the feature distribution under the health state to exhibit high clustering characteristics in the latent variable space. It can extract health indicators that are highly sensitive to the small impact on the target prediction object from multi-source heterogeneous data.

[0079] S140. If it is determined that the loss value of the predicted comprehensive health indicator does not meet the termination condition of the indicator prediction model training, return to the operation of collecting multidimensional health indicator related sample data of the target prediction object until it is determined that the loss value of the predicted comprehensive health indicator meets the termination condition of the indicator prediction model training.

[0080] The termination condition for training the indicator prediction model can be a condition that constrains the termination of training of the health indicator prediction association model. For example, the termination condition for training the indicator prediction model can be that the calculated loss value of the comprehensive health indicator reaches a minimum or tends to stabilize.

[0081] In each training round, after calculating the loss value of the predicted comprehensive health index of the target object using the composite loss function of the health index prediction association model, it can be determined whether the calculated loss value of the predicted comprehensive health index meets the training termination condition of the indicator prediction model, such as whether the loss value of the predicted comprehensive health index has reached a minimum or tended to a stable state. If it is determined that the loss value of the predicted comprehensive health index does not meet the training termination condition of the indicator prediction model, it indicates that the health index prediction association model has not yet been trained. At this time, an iterative training process can be executed, and multi-dimensional health index association sample data of the target object can be collected again for the next round of model training until it is determined that the loss value of the predicted comprehensive health index meets the training termination condition of the indicator prediction model, and the training of the health index prediction association model is completed.

[0082] Therefore, the above technical solution constructs a deep learning network with geometrical distribution constraints as a health indicator prediction association model. This model not only effectively learns how to reconstruct health signals, but more importantly, it maps all data features belonging to the "healthy" pattern to the vicinity of the health center. In this case, the feature space of the health indicator prediction association model forms a compact "health domain," which is highly sensitive to extracting health indicators with minimal impact on the target prediction object from multi-source heterogeneous data, thus improving the accuracy of health indicator prediction based on the model.

[0083] This invention, in its embodiments, inputs multidimensional health indicator-related sample data of the target prediction object into a health indicator prediction association model. The model calculates the predicted comprehensive health indicator of the target prediction object and uses a composite loss function to calculate the loss value of the predicted comprehensive health indicator. The iterative training process of the health indicator prediction association model is then achieved based on the matching between the loss value of the predicted comprehensive health indicator and the training termination condition of the indicator prediction model. Since the composite loss function of the health indicator prediction association model includes a reconstruction loss parameter and a compaction loss parameter, and the reconstruction loss parameter measures the model's ability to reconstruct the original signal, ensuring that the feature vector contains sufficient physical information, while the compaction loss parameter measures the degree to which the sample's feature vector deviates from the preset health center, the composite loss function can balance the preservation of the signal's physical information with the distribution pattern of the features. Training the health indicator prediction association model based on the composite loss function can improve the accuracy of health indicator prediction.

[0084] Example 2

[0085] Figure 3 This is a flowchart of a model training method provided in Embodiment 2 of the present invention. This embodiment is applicable to the case of training a remaining lifetime prediction model based on a physical monotonicity composite loss function. The method can be executed by a model training device, which can be implemented in software and / or hardware, and is generally integrated into an electronic device. This electronic device can be a terminal device or a server device, as long as it can execute the model training method. The embodiments of the present invention do not limit the specific type of electronic device. Correspondingly, as... Figure 3 As shown, the method includes the following operations:

[0086] S310. Obtain a comprehensive health index sequence sample of the target prediction object and the actual remaining lifespan corresponding to the comprehensive health index sequence sample.

[0087] Among them, the comprehensive health index sequence sample can be a sequence sample constructed based on the comprehensive health index of the target prediction object at different time points.

[0088] The trained health index prediction association model can predict the comprehensive health index of the target prediction object. Based on the comprehensive health index of the target prediction object, a regression model can be used to construct the mapping relationship between the comprehensive health index and the remaining lifespan of the target prediction object, thereby training a remaining lifespan prediction model suitable for the target prediction object.

[0089] When training a remaining life prediction model applicable to the target prediction object, one can first obtain a series of comprehensive health index samples of the target prediction object and the actual remaining life corresponding to each comprehensive health index series sample. In a specific example, assuming that machining tools are the target prediction object, during the comprehensive health index degradation trend modeling and data preparation stage, series sample data of comprehensive health indices of multiple similar machining tools from brand new to scrapped state can be collected, along with their corresponding actual service life or scrapping time points. Optionally, the actual remaining life of the machining tool can be measured by the remaining machining time or the number of remaining machined parts. For each tool, its comprehensive health index HI series sample can be represented as... ,in This indicates the operating time or machining volume corresponding to the machining tool. Simultaneously, it collects the actual service life of each machining tool, i.e., the total operating time or total machining volume from the start of use to its scrapping, denoted as... Then, for a sample of the comprehensive health index of a machining tool... Its corresponding actual remaining lifespan It can be calculated as .

[0090] S320. Construct a lifespan prediction sample based on the comprehensive health index sequence sample and the actual remaining lifespan.

[0091] To train a life expectancy prediction model, the comprehensive health index (HI) sequence samples need to be converted into input-output pairs, serving as life expectancy prediction samples for the model. Optionally, a sliding window method can be used, dividing the HI sequence samples into fixed-length time windows. Each window serves as the input to the life expectancy prediction model, with the actual life expectancy at the end of the window being the input. This serves as the output label for the remaining lifetime prediction model. That is, a lifetime prediction sample can be represented as... , where w is the window length.

[0092] S330. Input the comprehensive health index sequence sample in the lifespan prediction sample into the remaining lifespan prediction model, so as to predict the predicted remaining lifespan corresponding to the comprehensive health index sequence sample through the remaining lifespan prediction model.

[0093] Among them, predicting the remaining useful life is the remaining useful life predicted by the remaining useful life prediction model based on the input comprehensive health index sequence sample for the target prediction object.

[0094] In this embodiment of the invention, optionally, a Long Short-Term Memory (LSTM) network or other suitable network model can be selected as the remaining lifespan prediction model to construct a regression mapping relationship between the comprehensive health index sequence samples in the lifespan prediction sample and the actual remaining lifespan label, and to use its gating mechanism to capture long-term degradation patterns in the comprehensive health index sequence sample sequence. Specifically, the network structure of the remaining lifespan prediction model can be a multi-layer LSTM network, whose input layer receives fragments of the comprehensive health index sequence samples, the hidden layer contains LSTM units, and the output layer can be a fully connected layer used to map the final hidden state of the LSTM to the predicted remaining lifespan value at the current time.

[0095] S340. Calculate the loss value between the predicted remaining useful life and the actual remaining useful life using the physical monotonic composite loss function of the remaining useful life prediction model; wherein the physical monotonic composite loss function includes prediction error parameters and physical monotonicity constraint parameters.

[0096] Traditional lifetime prediction models typically use only Mean Squared Error (MSE) as the loss function to constrain the model training process. This loss function focuses solely on numerical accuracy, ignoring the physical laws governing the degradation of the target object. To address this issue, this invention constructs a physically monotonic composite loss function as the training objective for the lifetime prediction model. The physically monotonic composite loss function includes prediction error parameters and physically monotonic constraint parameters. The prediction error parameters ensure that the predicted values ​​of the lifetime prediction model closely approximate the true values. The physically monotonic constraint parameters penalize predictions that violate the laws of physical degradation.

[0097] In an optional embodiment of the present invention, the physical monotonicity composite loss function of the remaining lifetime prediction model can be:

[0098]

[0099]

[0100]

[0101] in, This represents the physical monotonicity composite loss function. The prediction error parameter represents the physical monotonicity composite loss function. λ² represents the physical monotonicity constraint parameter of the physical monotonicity composite loss function, which uses the ReLU function to penalize predictions that violate the physical degradation law, i.e., penalizes the case where the predicted lifetime at the next time step is actually greater than that at the previous time step; λ² represents the physical constraint weight, and M is the number of the second training samples. This represents the predicted remaining lifespan corresponding to the comprehensive health index sequence sample in the k-th lifespan prediction sample. ReLU() represents the actual remaining lifetime corresponding to the k-th lifetime prediction sample. This represents the predicted remaining useful life at time t+1. This represents the predicted remaining useful life at time t.

[0102] In the above physical monotonic composite loss function, if If the physical laws are violated, the physical monotonicity composite loss function will produce a positive penalty; otherwise, the penalty is 0.

[0103] S350. If it is determined that the loss value between the predicted remaining lifespan and the actual remaining lifespan does not meet the lifespan prediction model training termination condition, return to the operation of obtaining the comprehensive health index sequence sample of the target prediction object and the actual remaining lifespan corresponding to the comprehensive health index sequence sample, until it is determined that the loss value between the predicted remaining lifespan and the actual remaining lifespan meets the lifespan prediction model training termination condition.

[0104] The trained remaining life expectancy prediction model is used to calculate the current comprehensive health index of the target prediction object by the health index prediction association model trained by the model training method described in Embodiment 1 of the present invention, and to predict the remaining life expectancy of the target prediction object.

[0105] In each training epoch, after calculating the loss value between the predicted remaining useful life and the actual remaining useful life using the physical monotonicity composite loss function through the remaining useful life prediction model, it can be determined whether the calculated loss value meets the training termination condition of the remaining useful life prediction model, such as whether the calculated loss value has reached a minimum or tended to a stable state. If the calculated loss value does not meet the training termination condition, it indicates that the remaining useful life prediction model has not yet been trained. At this time, an iterative training process can be executed, re-acquiring comprehensive health index sequence samples of the target prediction object and related data such as the actual remaining useful life corresponding to the comprehensive health index sequence samples for the next round of model training, until it is determined that the loss value calculated by the physical monotonicity composite loss function meets the training termination condition of the remaining useful life prediction model, and the training of the remaining useful life prediction model is completed. Optionally, during the iterative training of the remaining useful life prediction model, the loss value of the physical monotonicity composite loss function of the remaining useful life prediction model can be minimized through the backpropagation algorithm and optimizer, thereby training a remaining useful life prediction model that is both accurate and conforms to the physical monotonicity trend.

[0106] This invention constructs a lifespan prediction sample by acquiring a comprehensive health index sequence sample of the target prediction object and the corresponding actual remaining lifespan. The comprehensive health index sequence sample from the lifespan prediction sample is then input into a remaining lifespan prediction model. The model predicts the predicted remaining lifespan corresponding to the comprehensive health index sequence sample. The loss value between the predicted remaining lifespan and the actual remaining lifespan is calculated using the physical monotonicity composite loss function of the remaining lifespan prediction model. The iterative training process of the remaining lifespan prediction model is then achieved based on the matching between the calculated loss value and the training termination condition of the remaining lifespan prediction model. Since the physical monotonicity composite loss function of the remaining lifespan prediction model includes a prediction error parameter and a physical monotonicity constraint parameter, and the prediction error parameter ensures that the predicted value is close to the true value, while the physical monotonicity constraint parameter penalizes predictions that violate physical degradation laws, a remaining lifespan prediction model that is both accurate and conforms to the physical monotonicity trend can be trained based on the physical monotonicity composite loss function, thus improving the accuracy of the remaining lifespan prediction model in predicting remaining lifespan.

[0107] Example 3

[0108] Figure 4This is a flowchart of a lifespan prediction method provided in Embodiment 3 of the present invention. This embodiment is applicable to situations where the remaining lifespan of a target object is predicted based on a trained health indicator prediction association model and a remaining lifespan prediction model. This method can be executed by a lifespan prediction device, which can be implemented in software and / or hardware, and is generally integrated into an electronic device. This electronic device can be a terminal device or a server device, as long as it can execute the lifespan prediction method. The present invention does not limit the specific type of electronic device. Correspondingly, as... Figure 4 As shown, the method includes the following operations:

[0109] S410. Collect multidimensional health indicator correlation data of the target prediction object, and input the multidimensional health indicator correlation data into the health indicator prediction correlation model to calculate the current comprehensive health indicator of the target prediction object through the health indicator prediction correlation model.

[0110] Among them, the multidimensional health indicator association data can be the association data of health indicators collected in real time from multiple dimensions of the target prediction object, which can reflect the current health status of the target prediction object in real time. The embodiments of the present invention do not limit the specific data types and data content included in the multidimensional health indicator association data.

[0111] For example, when the machining tool is used as the target prediction object, the multidimensional health index associated data may include, but is not limited to, multi-source data on tool wear, including but not limited to, direct physical quantities, indirect characterization quantities and process context, collected in real time during a specific machining scenario (such as gear machining scenario), such as sensor parameters, equipment operating status parameters, machining process parameters and tool status data.

[0112] In this embodiment of the invention, after real-time collection of the current multidimensional health indicator association data of the target prediction object, the multidimensional health indicator association data can be segmented and the segmented multidimensional health indicator association data can be input into the health indicator prediction association model so as to calculate the current comprehensive health indicator of the target prediction object corresponding to each data segment through the health indicator prediction association model.

[0113] In an optional embodiment of the present invention, the step of calculating the current comprehensive health index of the target prediction object through the health index prediction association model may include: extracting the baseline health status features of the baseline health status data sequence of the target prediction object through the health index prediction association model; extracting the current health status features of the corresponding sequence of the multidimensional health index association data through the health index prediction association model; calculating the distance metric between the baseline health status features and the current health status features; and calculating the current comprehensive health index of the target prediction object based on the distance metric between the baseline health status features and the current health status features.

[0114] The baseline health status data sequence can be a sequence composed of multidimensional health indicator correlation data of the target prediction object under a healthy state. The baseline health status features can be features extracted by the health indicator prediction correlation model from the baseline health status data sequence of the target prediction object. The current health status features can be features extracted by the health indicator prediction correlation model from the multidimensional health indicator correlation data of the target prediction object.

[0115] In an optional embodiment of the present invention, the step of extracting the baseline health status features of the baseline health status data sequence of the target prediction object through the health index prediction association model may include: inputting the baseline health status data sequence of the target prediction object into the health index prediction association model; and extracting the encoding feature set output by the target hidden layer of the health index prediction association model as the baseline health status features of the baseline health status data sequence of the target prediction object during the process of the health index prediction association model processing the baseline health status data sequence of the target prediction object.

[0116] The target hidden layer can be multiple selected hidden layers.

[0117] Understandably, when the target object is in a healthy state, the multidimensional health indicator association data collected from it is similar to the multidimensional health indicator association sample data used when training the health indicator prediction association model. Therefore, the hidden layer encoding features of the multidimensional health indicator association data collected from the target object in a healthy state, after being encoded by the health indicator prediction association model, are relatively close to the hidden layer encoding features of the healthy data. However, as the target object begins to deteriorate from a healthy state, the multidimensional health indicator association data collected from it gradually deviates from the data distribution of the healthy state. This causes the distance between the hidden layer encoding features output by the encoder and the hidden layer encoding features of the multidimensional health indicator association data in the healthy state to gradually increase. Therefore, if multidimensional health indicator association data from different health states are input into the health indicator prediction association model, the distance between their hidden layer encoding features can reflect the degree of degradation of the target object.

[0118] For example, taking a machining tool as an example, when the machining tool begins to wear and degrade, its machining data will gradually deviate from the data distribution of the healthy state, causing the distance between the hidden layer coding features output by the encoder and the hidden layer coding features of the healthy data to gradually increase. Therefore, if data of different machining states are input into the health index prediction association model, the distance between its hidden layer coding features can reflect the wear degree of the machining tool.

[0119] Therefore, when calculating the current comprehensive health index of the target prediction object using the health index prediction association model, the hidden layer feature acquisition operation of the health index prediction association model can be performed. This involves dividing the multidimensional health index association data into data segments and inputting each segment into the health index prediction association model, then extracting the relevant features from the data. The outputs of selected hidden layers are used as encoded features. These hidden layers are chosen as the output features of the multi-head attention module. Specifically, a set of representative multidimensional health index correlation data from a novel target prediction object in its health state is first selected as the baseline health state data sequence. The baseline health state data sequence is then divided into multiple baseline data segments, each of which serves as the baseline health state data. The data is then input into a health indicator prediction association model to extract baseline health status data. The corresponding set of m encoded features is denoted as the baseline health status feature. Meanwhile, the multidimensional health indicator correlation data extracted in real time from the online monitoring of the target prediction object is divided into multiple current data segments in the same way as the baseline health status data sequence. This data is then input into the health indicator prediction association model to extract the current data segment. The corresponding set of m encoded features is denoted as the current health status feature. .

[0120] Furthermore, a distance metric operation is performed for each baseline health status data point. The baseline health status characteristics and the corresponding current data segment The current health status characteristics are used to calculate the distance metric between the current health status characteristics and the baseline health status characteristics for each dimension. For example, the Euclidean distance between the current health status characteristics and the baseline health status characteristics can be calculated as the distance metric. The mean of the distance metrics calculated for each dimension is then used as the current comprehensive health index HI of the target prediction object. The calculation formula is as follows:

[0121]

[0122] Where HI represents the current comprehensive health index, m represents the number of hidden layers in the health index prediction correlation model, and j represents the j-th selected hidden layer. , It is the vector of the k-th dimension of the current health status feature output by the j-th hidden layer. It is the vector of the k-th dimension of the baseline health state features output by the j-th hidden layer. It is the dimension of the vector.

[0123] S420. Construct a current comprehensive health index sequence based on the current comprehensive health index of the target prediction object.

[0124] After obtaining the current comprehensive health index of the target prediction object at the current moment, the current comprehensive health index can be added to the real-time comprehensive health index sequence to obtain the current comprehensive health index sequence.

[0125] S430. Input the current comprehensive health index sequence of the target prediction object into the remaining life expectancy prediction model, so as to predict the remaining life expectancy of the target prediction object through the remaining life expectancy prediction model.

[0126] The health indicator prediction association model is trained using the model training method described in Embodiment 1 of the present invention, and the remaining life expectancy prediction model is trained using the model training method described in Embodiment 2 of the present invention.

[0127] Furthermore, the comprehensive health indicator sequence of the current moment and a period preceding it (e.g., a sliding window of data of length w) is fed into the pre-trained remaining lifespan prediction model. The remaining lifespan prediction model will output the predicted remaining lifespan of the target object. .

[0128] Optionally, the predicted remaining useful life of the target object can be continuously monitored. When When the temperature drops to a preset maintenance threshold, the system issues an alert, prompting the operator to prepare to replace the target object. When the target object approaches zero or the current comprehensive health index reaches the set scrapping threshold, the system issues a replacement command to ensure that the object is replaced before it fails, thus avoiding sudden failures and impacting system operation.

[0129] In a specific application scenario, addressing the problems of low accuracy, poor real-time performance, high dependence on expert experience, and difficulty in adapting to complex and changing machining environments in existing methods for predicting the remaining life of machining tools, this invention provides a deep learning-based method for predicting the remaining life of machining tools. This method constructs a deep learning network that fuses multi-source heterogeneous data to achieve unsupervised feature learning of tool wear states during machining. It extracts health indicators that are sensitive to tool wear and exhibit trends, and then accurately predicts the remaining tool life based on these health indicators. This improves the accuracy, real-time performance, and intelligence level of machining tool life prediction, reduces the reliance on human experience in the prediction process, and provides strong support for intelligent maintenance and production optimization in industrial machining.

[0130] The aforementioned technical solution, in the unsupervised feature learning stage of training the health indicator prediction association model, further introduces a "hypersphere compact loss" to build upon the traditional autoencoder's training objective of "minimizing reconstruction error." This introduces a geometric constraint on the feature distribution into the composite loss function of the health indicator prediction association model, forcing the data features of the target prediction object in all health states to be mapped into a minimal hypersphere centered at a fixed center C in the latent variable space. This compact distribution ensures that any tiny degenerate feature will cause its feature vector to rapidly escape from the "health hypersphere," thereby significantly amplifying early fault signals and improving the health indicator prediction association model's extremely high sensitivity to early faults. Meanwhile, to address the issue that existing purely data-driven remaining lifetime prediction models are prone to outputting results that violate physical common sense, such as predicting lifetime to increase over time, a physical monotonic composite loss function was constructed. This loss function introduces a monotonic penalty term into the training of the remaining lifetime prediction model, using the ReLU function to penalize non-physical phenomena such as "predicted lifetime at the next moment being greater than at the current moment." By embedding the physical prior knowledge of "irreversible degradation" into the weight optimization of the neural network, it can ensure that the output remaining lifetime curve exhibits a strictly monotonically decreasing or stable trend, eliminating the "lifetime backflow" oscillation phenomenon commonly found in traditional methods. This makes the remaining lifetime prediction results conform to the physical objective laws of the processing, significantly improving the usability, reliability, and operator trust of the algorithm in actual industrial settings.

[0131] It should be noted that all information (including but not limited to user device information, user personal information, etc.) and data (including but not limited to data used for display, data used for analysis, etc.) involved in this disclosure are information and data authorized by the user or fully authorized by all parties, and the collection, use and processing of the relevant data comply with the relevant laws, regulations and standards of the relevant regions.

[0132] It should be noted that any arrangement or combination of the technical features in the above embodiments also falls within the protection scope of this invention.

[0133] Example 4

[0134] Figure 5 This is a schematic diagram of a model training device provided in Embodiment 4 of the present invention, as shown below. Figure 5 As shown, the device includes: a multidimensional health indicator associated sample data acquisition module 510, a comprehensive health indicator prediction calculation module 520, a comprehensive health indicator loss value calculation module 530, and a first iterative training module 540, wherein:

[0135] The multidimensional health indicator association sample data acquisition module 510 is used to collect multidimensional health indicator association sample data of the target prediction object;

[0136] The predictive comprehensive health index calculation module 520 is used to input the multidimensional health index association sample data of the target prediction object into the health index prediction association model, so as to calculate the predicted comprehensive health index of the target prediction object through the health index prediction association model.

[0137] The comprehensive health index loss value calculation module 530 is used to calculate the loss value of the predicted comprehensive health index of the target prediction object through the composite loss function of the health index prediction association model; wherein, the composite loss function of the health index prediction association model includes a reconstruction loss parameter and a compaction loss parameter;

[0138] The first iterative training module 540 is used to return to the operation of collecting multidimensional health indicator-related sample data of the target prediction object when it is determined that the loss value of the predicted comprehensive health indicator does not meet the termination condition of the indicator prediction model training, until it is determined that the loss value of the predicted comprehensive health indicator meets the termination condition of the indicator prediction model training.

[0139] Optionally, the composite loss function of the health indicator prediction association model is:

[0140] ;

[0141] ;

[0142] ;

[0143] in, This represents the composite loss function. This represents the reconstruction loss parameter. Let λ1 represent the compaction loss parameter, λ1 represent the balance coefficient, and N be the number of the first training samples. This represents the i-th data sample. This represents the predicted value of the target object relative to the original data sample. C represents the latent variable feature vector of the current data sample, and C represents the preset health center vector.

[0144] This invention, in its embodiments, inputs multidimensional health indicator-related sample data of the target prediction object into a health indicator prediction association model. The model calculates the predicted comprehensive health indicator of the target prediction object and uses a composite loss function to calculate the loss value of the predicted comprehensive health indicator. The iterative training process of the health indicator prediction association model is then achieved based on the matching between the loss value of the predicted comprehensive health indicator and the training termination condition of the indicator prediction model. Since the composite loss function of the health indicator prediction association model includes a reconstruction loss parameter and a compaction loss parameter, and the reconstruction loss parameter measures the model's ability to reconstruct the original signal, ensuring that the feature vector contains sufficient physical information, while the compaction loss parameter measures the degree to which the sample's feature vector deviates from the preset health center, the composite loss function can balance the preservation of the signal's physical information with the distribution pattern of the features. Training the health indicator prediction association model based on the composite loss function can improve the accuracy of its health indicator prediction.

[0145] The aforementioned model training apparatus can execute the model training method provided in Embodiment 1 of the present invention, and has the corresponding functional modules and beneficial effects of the method. Technical details not described in detail in this embodiment can be found in the model training method provided in Embodiment 1 of the present invention.

[0146] Example 5

[0147] Figure 6 This is a schematic diagram of a model training device provided in Embodiment 5 of the present invention, as shown below. Figure 6 As shown, the device includes: an index sequence lifetime data acquisition module 610, a lifetime prediction sample construction module 620, a remaining lifetime prediction module 630, a lifetime loss value calculation module 640, and a second iterative training module 650, wherein:

[0148] The indicator sequence lifespan data acquisition module 610 is used to acquire a comprehensive health indicator sequence sample of the target prediction object and the actual remaining lifespan corresponding to the comprehensive health indicator sequence sample.

[0149] Lifespan prediction sample construction module 620 is used to construct lifespan prediction samples based on the comprehensive health index sequence sample and the actual remaining lifespan.

[0150] The remaining useful life prediction module 630 is used to input the comprehensive health index sequence sample in the life prediction sample into the remaining useful life prediction model, so as to predict the predicted remaining useful life corresponding to the comprehensive health index sequence sample through the remaining useful life prediction model.

[0151] The predicted lifetime loss value calculation module 640 is used to calculate the loss value between the predicted remaining lifetime and the actual remaining lifetime through the physical monotonic composite loss function of the remaining lifetime prediction model; wherein, the physical monotonic composite loss function includes prediction error parameters and physical monotonicity constraint parameters;

[0152] The second iterative training module 650 is used to return to the operation of obtaining the comprehensive health index sequence sample of the target prediction object and the actual remaining lifespan corresponding to the comprehensive health index sequence sample when it is determined that the loss value between the predicted remaining lifespan and the actual remaining lifespan does not meet the lifespan prediction model training termination condition.

[0153] The trained remaining life expectancy prediction model is used to calculate the current comprehensive health index of the target prediction object by the health index prediction association model trained by the model training method according to Embodiment 1 of the present invention, and to predict the remaining life expectancy of the target prediction object.

[0154] Optional. The physical monotonicity composite loss function of the remaining lifetime prediction model is:

[0155] ;

[0156] ;

[0157] ;

[0158] in, This represents the physical monotonicity composite loss function. The prediction error parameter represents the physical monotonicity composite loss function. The physical monotonicity constraint parameter represents the physical monotonicity composite loss function, λ2 represents the physical constraint weight, and M is the number of the second training samples. This represents the predicted remaining lifespan corresponding to the comprehensive health index sequence sample in the k-th lifespan prediction sample. ReLU() represents the actual remaining lifetime corresponding to the k-th lifetime prediction sample. This represents the predicted remaining useful life at time t+1. This represents the predicted remaining useful life at time t.

[0159] This invention constructs a lifespan prediction sample by acquiring a comprehensive health index sequence sample of the target prediction object and the corresponding actual remaining lifespan. The comprehensive health index sequence sample from the lifespan prediction sample is then input into a remaining lifespan prediction model. The model predicts the predicted remaining lifespan corresponding to the comprehensive health index sequence sample. The loss value between the predicted remaining lifespan and the actual remaining lifespan is calculated using the physical monotonicity composite loss function of the remaining lifespan prediction model. The iterative training process of the remaining lifespan prediction model is then achieved based on the matching between the calculated loss value and the training termination condition of the remaining lifespan prediction model. Since the physical monotonicity composite loss function of the remaining lifespan prediction model includes a prediction error parameter and a physical monotonicity constraint parameter, and the prediction error parameter ensures that the predicted value is close to the true value, while the physical monotonicity constraint parameter penalizes predictions that violate physical degradation laws, a remaining lifespan prediction model that is both accurate and conforms to the physical monotonicity trend can be trained based on the physical monotonicity composite loss function, thus improving the accuracy of the remaining lifespan prediction model in predicting remaining lifespan.

[0160] The aforementioned model training apparatus can execute the model training method provided in Embodiment 2 of the present invention, and has the corresponding functional modules and beneficial effects of the method. Technical details not described in detail in this embodiment can be found in the model training method provided in Embodiment 2 of the present invention.

[0161] Example 6

[0162] Figure 7 This is a schematic diagram of a lifetime prediction device provided in Embodiment Six of the present invention, as shown below. Figure 7 As shown, the device includes: a current comprehensive health index calculation module 710, a current comprehensive health index sequence construction module 720, and a remaining lifespan prediction module 730, wherein:

[0163] The current comprehensive health index calculation module 710 is used to collect multidimensional health index correlation data of the target prediction object and input the multidimensional health index correlation data into the health index prediction correlation model so as to calculate the current comprehensive health index of the target prediction object through the health index prediction correlation model.

[0164] The current comprehensive health index sequence construction module 720 is used to construct a current comprehensive health index sequence based on the current comprehensive health index of the target prediction object;

[0165] The remaining useful life prediction module 730 is used to input the current comprehensive health index sequence of the target prediction object into the remaining useful life prediction model, so as to predict the remaining useful life of the target prediction object through the remaining useful life prediction model.

[0166] The health indicator prediction association model is trained using the model training method described in Embodiment 1 of the present invention, and the remaining life expectancy prediction model is trained using the model training method described in Embodiment 2 of the present invention.

[0167] Optionally, the current comprehensive health index calculation module 710 is further configured to: extract the baseline health status features of the baseline health status data sequence of the target prediction object through the health index prediction association model; extract the current health status features of the corresponding sequence of the multidimensional health index association data through the health index prediction association model; calculate the distance metric between the baseline health status features and the current health status features; and calculate the current comprehensive health index of the target prediction object based on the distance metric between the baseline health status features and the current health status features.

[0168] Optionally, the current comprehensive health index calculation module 710 is further configured to: input the baseline health status data sequence of the target prediction object into the health index prediction association model; and extract the encoding feature set output by the target hidden layer of the health index prediction association model as the baseline health status feature of the baseline health status data sequence of the target prediction object during the process of the health index prediction association model processing the baseline health status data sequence of the target prediction object.

[0169] Optionally, the current comprehensive health index calculation module 710 is further configured to: input the multidimensional health index associated data into the health index prediction association model; and extract the encoding feature set output by the target hidden layer of the health index prediction association model as the current health status feature of the sequence corresponding to the multidimensional health index associated data during the process of the health index prediction association model processing the multidimensional health index associated data.

[0170] Optionally, the current comprehensive health index calculation module 710 is further configured to: calculate the current comprehensive health index of the target prediction object based on the following formula:

[0171] ;

[0172] Where HI represents the current comprehensive health index, m represents the number of hidden layers in the health index prediction correlation model, and j represents the j-th selected hidden layer. It is the vector of the k-th dimension of the current health status feature output by the j-th hidden layer. It is the vector of the k-th dimension of the baseline health state features output by the j-th hidden layer. It is the dimension of the vector.

[0173] This invention collects multidimensional health indicator correlation data of a target prediction object and inputs this data into a health indicator prediction correlation model to calculate the target prediction object's current comprehensive health indicator. Furthermore, based on the target prediction object's current comprehensive health indicator, a current comprehensive health indicator sequence is constructed and input into a remaining life expectancy prediction model to predict the target prediction object's remaining life expectancy. This improves the accuracy of predicting an object's remaining life expectancy based on the health indicator prediction correlation model and the remaining life expectancy prediction model.

[0174] The above-described model training apparatus can execute the model training method provided in Embodiment 3 of the present invention, and has the corresponding functional modules and beneficial effects of the method. Technical details not described in detail in this embodiment can be found in the model training method provided in Embodiment 3 of the present invention.

[0175] Example 7

[0176] Figure 8 A schematic diagram of an electronic device 10, which can be used to implement embodiments of the present invention, is shown. The electronic device is intended to represent various forms of digital computers, such as laptop computers, desktop computers, workstations, personal digital assistants, servers, blade servers, mainframe computers, and other suitable computers. The electronic device can also represent various forms of mobile devices, such as personal digital processors, cellular phones, smartphones, wearable devices (e.g., helmets, glasses, watches, etc.), and other similar computing devices. The components shown herein, their connections and relationships, and their functions are merely illustrative and are not intended to limit the implementation of the invention described and / or claimed herein.

[0177] like Figure 8As shown, the electronic device 10 includes at least one processor 11 and a memory, such as a read-only memory (ROM) 12 or a random access memory (RAM) 13, communicatively connected to the at least one processor 11. The memory stores computer programs executable by the at least one processor. The processor 11 can perform various appropriate actions and processes based on the computer program stored in the ROM 12 or loaded from storage unit 18 into the RAM 13. The RAM 13 can also store various programs and data required for the operation of the electronic device 10. The processor 11, ROM 12, and RAM 13 are interconnected via a bus 14. An input / output (I / O) interface 15 is also connected to the bus 14.

[0178] Multiple components in electronic device 10 are connected to I / O interface 15, including: input unit 16, such as keyboard, mouse, etc.; output unit 17, such as various types of displays, speakers, etc.; storage unit 18, such as disk, optical disk, etc.; and communication unit 19, such as network card, modem, wireless transceiver, etc. Communication unit 19 allows electronic device 10 to exchange information / data with other devices through computer networks such as the Internet and / or various telecommunications networks.

[0179] Processor 11 can be a variety of general-purpose and / or special-purpose processing components with processing and computing capabilities. Some examples of processor 11 include, but are not limited to, a central processing unit (CPU), a graphics processing unit (GPU), various special-purpose artificial intelligence (AI) computing chips, various processors running machine learning model algorithms, a digital signal processor (DSP), and any suitable processor, controller, microcontroller, etc. Processor 11 performs the various methods and processes described above, such as model training methods or lifetime prediction methods.

[0180] Optionally, a model training method may include: collecting multidimensional health indicator-related sample data of the target prediction object; inputting the multidimensional health indicator-related sample data of the target prediction object into a health indicator prediction association model to calculate the predicted comprehensive health indicator of the target prediction object through the health indicator prediction association model; calculating the loss value of the predicted comprehensive health indicator of the target prediction object through a composite loss function of the health indicator prediction association model; wherein the composite loss function of the health indicator prediction association model includes a reconstruction loss parameter and a compaction loss parameter; if it is determined that the loss value of the predicted comprehensive health indicator does not meet the training termination condition of the indicator prediction model, returning to the operation of collecting multidimensional health indicator-related sample data of the target prediction object until it is determined that the loss value of the predicted comprehensive health indicator meets the training termination condition of the indicator prediction model.

[0181] Optionally, a model training method may include: obtaining a comprehensive health indicator sequence sample of the target prediction object and the actual remaining lifespan corresponding to the comprehensive health indicator sequence sample; constructing a lifespan prediction sample based on the comprehensive health indicator sequence sample and the actual remaining lifespan; inputting the comprehensive health indicator sequence sample from the lifespan prediction sample into a remaining lifespan prediction model to predict the predicted remaining lifespan corresponding to the comprehensive health indicator sequence sample through the remaining lifespan prediction model; calculating the loss value between the predicted remaining lifespan and the actual remaining lifespan through the physical monotonicity composite loss function of the remaining lifespan prediction model; wherein, the physical monotonicity composite loss function includes a prediction error parameter. And physical monotonicity constraint parameters; if it is determined that the loss value between the predicted remaining lifespan and the actual remaining lifespan does not meet the lifespan prediction model training termination condition, the operation of obtaining the comprehensive health index sequence sample of the target prediction object and the actual remaining lifespan corresponding to the comprehensive health index sequence sample is returned until it is determined that the loss value between the predicted remaining lifespan and the actual remaining lifespan meets the lifespan prediction model training termination condition; wherein, the trained remaining lifespan prediction model is used to calculate the current comprehensive health index of the target prediction object by the health index prediction association model trained by the model training method according to Embodiment 1 of the present invention, and predict the remaining lifespan of the target prediction object.

[0182] Optionally, the lifespan prediction method may include: collecting multidimensional health indicator correlation data of the target prediction object, and inputting the multidimensional health indicator correlation data into a health indicator prediction correlation model to calculate the current comprehensive health indicator of the target prediction object through the health indicator prediction correlation model; constructing a current comprehensive health indicator sequence based on the current comprehensive health indicator of the target prediction object; inputting the current comprehensive health indicator sequence of the target prediction object into a remaining lifespan prediction model to predict the remaining lifespan of the target prediction object through the remaining lifespan prediction model; wherein, the health indicator prediction correlation model is trained by the model training method described in Embodiment 1 of the present invention, and the remaining lifespan prediction model is trained by the model training method described in Embodiment 2 of the present invention.

[0183] In some embodiments, the model training method or lifetime prediction method may be implemented as a computer program tangibly contained in a computer-readable storage medium, such as storage unit 18. In some embodiments, part or all of the computer program may be loaded and / or installed on electronic device 10 via ROM 12 and / or communication unit 19. When the computer program is loaded into RAM 13 and executed by processor 11, one or more steps of the model training method or lifetime prediction method described above may be performed. Alternatively, in other embodiments, processor 11 may be configured to perform the model training method or lifetime prediction method by any other suitable means (e.g., by means of firmware).

[0184] Various embodiments of the systems and techniques described above herein can be implemented in digital electronic circuit systems, integrated circuit systems, field-programmable gate arrays (FPGAs), application-specific integrated circuits (ASICs), application-specific standard products (ASSPs), systems-on-a-chip (SoCs), payload-programmable logic devices (CPLDs), computer hardware, firmware, software, and / or combinations thereof. These various embodiments may include implementations in one or more computer programs that can be executed and / or interpreted on a programmable system including at least one programmable processor, which may be a dedicated or general-purpose programmable processor, capable of receiving data and instructions from a storage system, at least one input device, and at least one output device, and transmitting data and instructions to the storage system, the at least one input device, and the at least one output device.

[0185] Computer programs used to implement the methods of the present invention may be written in any combination of one or more programming languages. These computer programs may be provided to a processor of a general-purpose computer, a special-purpose computer, or other programmable data processing device, such that when executed by the processor, the computer programs cause the functions / operations specified in the flowcharts and / or block diagrams to be performed. The computer programs may be executed entirely on a machine, partially on a machine, or as a standalone software package, partially on a machine and partially on a remote machine, or entirely on a remote machine or server.

[0186] In the context of this invention, a computer-readable storage medium can be a tangible medium that may contain or store a computer program for use by or in conjunction with an instruction execution system, apparatus, or device. A computer-readable storage medium may include, but is not limited to, electronic, magnetic, optical, electromagnetic, infrared, or semiconductor systems, apparatus, or devices, or any suitable combination thereof. Alternatively, a computer-readable storage medium may be a machine-readable signal medium. More specific examples of machine-readable storage media include electrical connections based on one or more wires, portable computer disks, hard disks, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM or flash memory), optical fibers, portable compact disk read-only memory (CD-ROM), optical storage devices, magnetic storage devices, or any suitable combination thereof.

[0187] To provide interaction with a user, the systems and techniques described herein can be implemented on an electronic device having: a display device for displaying information to the user (e.g., a CRT (cathode ray tube) or LCD (liquid crystal display) monitor); and a keyboard and pointing device (e.g., a mouse or trackball) through which the user provides input to the electronic device. Other types of devices can also be used to provide interaction with the user; for example, feedback provided to the user can be any form of sensory feedback (e.g., visual feedback, auditory feedback, or tactile feedback); and input from the user can be received in any form (including sound input, voice input, or tactile input).

[0188] The systems and technologies described herein can be implemented in computing systems that include backend components (e.g., as data servers), or middleware components (e.g., application servers), or frontend components (e.g., user computers with graphical user interfaces or web browsers through which users can interact with implementations of the systems and technologies described herein), or any combination of such backend, middleware, or frontend components. The components of the system can be interconnected via digital data communication of any form or medium (e.g., communication networks). Examples of communication networks include local area networks (LANs), wide area networks (WANs), blockchain networks, and the Internet.

[0189] A computing system can include clients and servers. Clients and servers are generally located far apart and typically interact through communication networks. The client-server relationship is created by computer programs running on the respective computers and having a client-server relationship with each other. The server can be a cloud server, also known as a cloud computing server or cloud host, which is a hosting product within the cloud computing service system to address the shortcomings of traditional physical hosts and VPS services, such as high management difficulty and weak business scalability.

[0190] It should be understood that the various forms of processes shown above can be used to reorder, add, or delete steps. For example, the steps described in this disclosure can be executed in parallel, sequentially, or in different orders, as long as the desired result of the technical solution disclosed in this disclosure can be achieved, and this is not limited herein.

[0191] The specific embodiments described above do not constitute a limitation on the scope of protection of this disclosure. Those skilled in the art should understand that various modifications, combinations, sub-combinations, and substitutions can be made according to design requirements and other factors. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of this disclosure should be included within the scope of protection of this disclosure.

Claims

1. A model training method, characterized in that, include: Collect multidimensional health index-related sample data of the target prediction object; The multidimensional health index associated sample data of the target prediction object is input into the health index prediction association model, so as to calculate the predicted comprehensive health index of the target prediction object through the health index prediction association model; The loss value of the predicted comprehensive health index of the target prediction object is calculated by the composite loss function of the health index prediction association model; wherein, the composite loss function of the health index prediction association model includes a reconstruction loss parameter and a compaction loss parameter; the compaction loss parameter is used to measure the degree to which the feature vector of the current multidimensional health index association sample data deviates from the preset health center; If it is determined that the loss value of the predicted comprehensive health indicator does not meet the termination condition of the indicator prediction model training, the operation of collecting multidimensional health indicator-related sample data of the target prediction object is returned until it is determined that the loss value of the predicted comprehensive health indicator meets the termination condition of the indicator prediction model training.

2. The model training method according to claim 1, characterized in that, The composite loss function of the health indicator prediction association model is: ； ； ； in, This represents the composite loss function. This represents the reconstruction loss parameter. Let λ1 represent the compaction loss parameter, λ1 represent the balance coefficient, and N be the number of the first training samples. This represents the i-th data sample. This represents the predicted value of the target object relative to the original data sample. C represents the latent variable feature vector of the current data sample, and C represents the preset health center vector.

3. A model training method, characterized in that, include: Obtain a comprehensive health index sequence sample of the target prediction object and the actual remaining lifespan corresponding to the comprehensive health index sequence sample; A life expectancy prediction sample is constructed based on the comprehensive health index sequence sample and the actual remaining life expectancy. The comprehensive health index sequence sample in the lifespan prediction sample is input into the remaining lifespan prediction model so as to predict the predicted remaining lifespan corresponding to the comprehensive health index sequence sample through the remaining lifespan prediction model. The loss value between the predicted remaining useful life and the actual remaining useful life is calculated using the physical monotonicity composite loss function of the remaining useful life prediction model; wherein, the physical monotonicity composite loss function includes prediction error parameters and physical monotonicity constraint parameters; If it is determined that the loss value between the predicted remaining lifespan and the actual remaining lifespan does not meet the lifespan prediction model training termination condition, the operation of obtaining the comprehensive health index sequence sample of the target prediction object and the actual remaining lifespan corresponding to the comprehensive health index sequence sample is returned until it is determined that the loss value between the predicted remaining lifespan and the actual remaining lifespan meets the lifespan prediction model training termination condition. The trained life expectancy prediction model is used to predict the remaining life expectancy of the target prediction object by calculating the current comprehensive health index of the target prediction object based on the health index prediction association model. The health index prediction association model is trained by any of the model training methods described in claims 1-2.

4. The model training method according to claim 3, characterized in that, The physical monotonicity composite loss function of the remaining lifetime prediction model is: ； ； ； in, This represents the physical monotonicity composite loss function. The prediction error parameter represents the physical monotonicity composite loss function. The physical monotonicity constraint parameter represents the physical monotonicity composite loss function, λ2 represents the physical constraint weight, and M is the number of the second training samples. This represents the predicted remaining lifespan corresponding to the comprehensive health index sequence sample in the k-th lifespan prediction sample. ReLU() represents the actual remaining lifetime corresponding to the k-th lifetime prediction sample. This represents the predicted remaining useful life at time t+1. This represents the predicted remaining useful life at time t.

5. A lifespan prediction method, characterized in that, include: Collect multidimensional health indicator correlation data of the target prediction object, and input the multidimensional health indicator correlation data into the health indicator prediction correlation model, so as to calculate the current comprehensive health indicator of the target prediction object through the health indicator prediction correlation model; Construct a current comprehensive health index sequence based on the current comprehensive health index of the target prediction object; The current comprehensive health index sequence of the target prediction object is input into the remaining life expectancy prediction model so as to predict the remaining life expectancy of the target prediction object through the remaining life expectancy prediction model; The health indicator prediction association model is trained using the model training method described in claim 1 or 2, and the remaining life expectancy prediction model is trained using the model training method described in claim 3 or 4.

6. The lifetime prediction method according to claim 5, characterized in that, The calculation of the current comprehensive health index of the target prediction object through the health index prediction association model includes: The baseline health status features of the baseline health status data sequence of the target prediction object are extracted through the health index prediction association model. The current health status features of the corresponding sequences of the multidimensional health indicator association data are extracted using the health indicator prediction association model. Calculate the distance metric between the baseline health status feature and the current health status feature, and calculate the current comprehensive health index of the target prediction object based on the distance metric between the baseline health status feature and the current health status feature.

7. The lifetime prediction method according to claim 6, characterized in that, The step of extracting the baseline health status features of the baseline health status data sequence of the target prediction object through the health index prediction association model includes: The baseline health status data sequence of the target prediction object is input into the health indicator prediction association model; During the process of the health indicator prediction association model processing the baseline health status data sequence of the target prediction object, the set of encoded features output by the target hidden layer of the health indicator prediction association model is extracted as the baseline health status features of the baseline health status data sequence of the target prediction object.

8. The lifetime prediction method according to claim 6, characterized in that, The step of extracting the current health status features of the corresponding sequences of the multidimensional health indicator correlation data through the health indicator prediction correlation model includes: The multidimensional health indicator correlation data is input into the health indicator prediction correlation model; During the process of the health indicator prediction association model processing the multidimensional health indicator association data, the set of encoded features output by the target hidden layer of the health indicator prediction association model is extracted as the current health status feature of the corresponding sequence of the multidimensional health indicator association data.

9. The lifetime prediction method according to claim 6, characterized in that, The step of calculating the current comprehensive health index of the target prediction object based on the distance metric between the baseline health status characteristics and the current health status characteristics includes: The current comprehensive health index of the target prediction object is calculated based on the following formula: ； Where HI represents the current comprehensive health index, m represents the number of hidden layers in the health index prediction correlation model, and j represents the j-th selected hidden layer. It is the vector of the k-th dimension of the current health status feature output by the j-th hidden layer. It is the vector of the k-th dimension of the baseline health state features output by the j-th hidden layer. It is the dimension of the vector.

10. A lifespan prediction device, characterized in that, include: The current comprehensive health index calculation module is used to collect multidimensional health index correlation data of the target prediction object, and input the multidimensional health index correlation data into the health index prediction correlation model, so as to calculate the current comprehensive health index of the target prediction object through the health index prediction correlation model; The current comprehensive health index sequence construction module is used to construct a current comprehensive health index sequence based on the current comprehensive health index of the target prediction object; The remaining useful life prediction module is used to input the current comprehensive health index sequence of the target prediction object into the remaining useful life prediction model, so as to predict the remaining useful life of the target prediction object through the remaining useful life prediction model; The health indicator prediction association model is trained using the model training method described in claim 1 or 2, and the remaining life expectancy prediction model is trained using the model training method described in claim 3 or 4.

Images