Model training method, maintenance method, computer equipment and medium

Abstract

The invention discloses a model training method. The method comprises the steps of constructing a modeling data set based on operation parameters, historical failure events and historical failure timestamps; and inputting the time sequence features in the training set of the data set into an LSTM time sequence coding layer to output a hidden state sequence. The query mapping layer maps the hidden state sequence into a single query vector. And inputting keys and values obtained by linear transformation of the single query vector and the hidden state sequence into a Transform attention mechanism layer to generate a fusion feature vector. And inputting the fusion feature vector into a regression output layer, mapping the fusion feature vector into a residual service life prediction value of the part in a scalar form, and forming an initial evaluation model. Inputting time sequence features in a verification set of the data set into the model to obtain a prediction result of the verification set; and based on the prediction result and the real remaining service life of the part, optimizing the initial evaluation model to obtain an evaluation model. The evaluation model can effectively predict the remaining service life of the component.

Description

Technical Field

[0001] This application relates to the field of industrial manufacturing technology, and in particular to a model training method, a maintenance method and apparatus thereof, computer equipment and media. Background Technology

[0002] In the electronics manufacturing industry, industrial equipment typically operates continuously, and the performance degradation of its key components directly affects processing accuracy and equipment reliability. A sudden failure of a key component can not only bring the entire production line to a standstill but may also trigger batch-wide quality defects, resulting in significant economic losses.

[0003] To ensure reliable equipment operation, existing technologies generally employ a time- or usage-based periodic maintenance strategy. This involves a unified shutdown inspection of the equipment within a preset maintenance cycle, and determining whether to replace critical components based on manual experience or simple thresholds. While this process can obtain some wear or degradation parameters, it relies solely on offline detection or periodic sampling, which only reflects the current state of components and cannot capture the real-time degradation characteristics of critical components, thus failing to effectively predict the remaining service life of critical components. Summary of the Invention

[0004] In view of this, this application aims to at least partially solve one of the problems in the related art. Therefore, the purpose of this application is to provide a model training method, a maintenance method and apparatus thereof, a computer device and a medium.

[0005] This application provides a model training method for training an evaluation model to assess the remaining service life of components of an evaluation device based on a base model. The base model includes an LSTM temporal encoding layer, a query mapping layer, a Transformer attention mechanism layer, and a regression output layer. The model training method includes: acquiring the operating parameters of the component during operation, the historical failure events of the component, and the historical failure timestamps of the component; constructing a modeling dataset containing time-series features and the actual remaining service life of the component corresponding to the time-series features based on the operating parameters, the historical failure events, and the historical failure timestamps; dividing the modeling dataset into a training set and a validation set; inputting the time-series features from the training set into the LSTM temporal encoding layer, which encodes the input time-series features to output a hidden state sequence representing the time-series dynamic characteristics; and the query mapping layer mapping the hidden state sequence into a single... A query vector is used; the single query vector and the keys and values ​​obtained by linear transformation of the hidden state sequence are input into the Transformer attention mechanism layer, which performs weighted fusion through a self-attention mechanism to generate a fused feature vector; the fused feature vector is input into the regression output layer, which maps the fused feature vector into a scalar form of the predicted remaining useful life of the component to predict the remaining useful life of the component, forming an initial evaluation model; the time series features in the validation set are input into the initial evaluation model to obtain the prediction results of the validation set; based on the prediction results of the validation set and the actual remaining useful life of the component, the running parameters of the initial evaluation model are optimized using a preset loss function to construct the final evaluation model.

[0006] In some implementations, the step of constructing a modeling dataset containing time-series features and the actual remaining service life of the component corresponding to the time-series features, based on the operating parameters, the historical failure events, and the historical failure timestamps, involves: processing the operating parameters of the component during the operation of the equipment to obtain multi-dimensional feature data; obtaining sample data of the remaining service life of the component based on the historical failure events and historical failure timestamps of the component; and associating the multi-dimensional feature data with the sample data to obtain the modeling dataset.

[0007] In some embodiments, the process of obtaining multidimensional feature data based on the operating parameters of the components during the operation of the device includes: preprocessing the operating parameters of the components to obtain processed operating parameters, wherein the preprocessing includes at least one of resampling and alignment, outlier handling, missing value handling, standardization, and Fourier denoising; extracting feature data of multiple dimensions from the processed operating parameters to constitute the multidimensional feature data; wherein the multidimensional feature data includes first-dimensional data corresponding to vibration signals, second-dimensional data corresponding to current signals, and third-dimensional data corresponding to temperature signals.

[0008] In some implementations, obtaining sample data for the remaining service life of the component based on its historical failure events and timestamps includes: constructing a timeline for the component from the start of its operation to its failure based on its historical failure events and timestamps; dividing the timeline within a time interval defined by the timeline according to a preset fixed time interval to obtain multiple observation moments; using the time difference between each observation moment and the corresponding failure timestamp as the remaining service life of the component at that observation moment; and associating the multidimensional feature data with the sample data to obtain the modeling dataset includes: aligning the multidimensional feature data with the remaining service life of the component according to the timestamps of the observation moments; and constructing the modeling dataset based on the aligned multidimensional feature data and the corresponding remaining service life.

[0009] In some implementations, before optimizing the parameters of the initial evaluation model using a preset loss function based on the prediction results of the validation set and the actual remaining service life of the component to construct the final evaluation model, the method further includes: calculating the mean squared error loss function term, the physical impossibility penalty term, and the dynamic weight decay term of the preset loss function based on the sample data; and summing the mean squared error loss function term, the physical impossibility penalty term, and the dynamic weight decay term to obtain the preset loss function.

[0010] In some implementations, the model training method further includes: dividing the modeling dataset into test sets; testing the feasibility of the evaluation model based on the test sets and outputting the feasibility results of the evaluation model; and improving the evaluation model based on the feasibility results to obtain a improved evaluation model.

[0011] In some embodiments, the model training method further includes: acquiring test data of the remaining service life of the component; processing the test data according to a preset data stream window length and a preset step size to obtain model test data; inputting the model test data into the evaluation model to obtain test results; performing drift detection on the test results; and retraining the evaluation model if the drift value of the test results is greater than a preset drift threshold.

[0012] This application also provides a maintenance method, which uses an evaluation model trained using the model training method described in any of the above embodiments. The maintenance method includes: acquiring operational data of a component to be evaluated in a device; evaluating the remaining service life of the component to be evaluated based on the operational data and the evaluation model; plotting a remaining service life curve based on the remaining service life of the component to be evaluated; analyzing and processing the remaining service life of the component to be evaluated to obtain early warning information; controlling the component to be evaluated to undergo shutdown maintenance or replacement based on the remaining service life curve and / or the early warning information; and / or displaying the remaining service life curve and / or the early warning information of the component to be evaluated in real time.

[0013] This application also provides a computer device. The computer device includes a memory and a processor. The memory stores a computer program, which, when executed by the processor, implements the model training method and the maintenance method described in any of the above embodiments.

[0014] This application also provides a computer-readable storage medium. The computer-readable storage medium stores a computer program that, when executed by one or more processors, implements the model training method and the maintenance method described in any of the above embodiments.

[0015] Thus, the evaluation model trained by the model training method of this application can intelligently quantify and effectively predict the remaining service life of equipment components based on equipment operation data, replacing manual experience judgment and achieving an objective, fair, and accurate assessment of the remaining service life. This avoids premature replacement of equipment components before they fail, prevents performance redundancy waste of equipment components, and reduces the production cost of equipment components. At the same time, the model of this application integrates LSTM and Transformer, adopting a cascaded structure. It uses LSTM to extract long-range temporal dependencies and generate global prior anchor points, and then uses a single query vector to significantly reduce computational complexity while retaining sensitivity to key degradation features. This application adopts a cascaded attention coupling scheme, constructing the evaluation model in a two-step manner of first temporal and then global, which can both compress the amount of computation and highlight early weak features, further improving the evaluation efficiency and accuracy.

[0016] Additional aspects and advantages of this application will be set forth in part in the description which follows, and in part will be obvious from the description, or may be learned by practice of this application. Attached Figure Description

[0017] The above and / or additional aspects and advantages of this application will become apparent and readily understood from the following description of the embodiments taken in conjunction with the accompanying drawings, wherein: Figure 1 This is a flowchart illustrating the model training method of some embodiments of this application; Figure 2 This is a schematic diagram of the structure of a model training device according to certain embodiments of this application; Figure 3 This is a flowchart illustrating the model training method of some embodiments of this application; Figure 4 This is a flowchart illustrating the model training method of some embodiments of this application; Figure 5 This is a schematic diagram of Fourier denoising in the model training method of some embodiments of this application; Figure 6 This is a flowchart illustrating the model training method of some embodiments of this application; Figure 7 This is a flowchart illustrating the model training method of some embodiments of this application; Figure 8 This is a schematic diagram of the structure of a model training device according to certain embodiments of this application; Figure 9 This is a flowchart illustrating the model training method of some embodiments of this application; Figure 10 This is a flowchart illustrating the model training method of some embodiments of this application; Figure 11 This is a schematic diagram of the structure of a model training device according to certain embodiments of this application; Figure 12 This is a flowchart illustrating a repair method according to certain embodiments of this application; Figure 13 This is a structural schematic diagram of a maintenance device according to certain embodiments of this application. Detailed Implementation

[0018] The embodiments of this application are described in detail below. Examples of these embodiments are shown in the accompanying drawings, wherein the same or similar reference numerals denote the same or similar elements or elements having the same or similar functions throughout. The embodiments described below with reference to the accompanying drawings are exemplary and are only used to explain this application, and should not be construed as limiting this application.

[0019] In the description of this application, the terms "first" and "second" are used for descriptive purposes only and should not be construed as indicating or implying relative importance, or implicitly specifying the number of indicated technical features. Thus, a feature defined as "first" or "second" may explicitly or implicitly include one or more of the stated features. In the description of this application, "a plurality of" means two or more, unless otherwise explicitly specified.

[0020] In the description of this application, it should be noted that, unless otherwise explicitly specified and limited, the terms "installation" and "connection" should be interpreted broadly, referring to fixed connections, detachable connections, or integral connections; mechanical connections, electrical connections, or connections that allow communication between components; direct connections or indirect connections through an intermediate medium; and connections within two components or interactions between two components. Those skilled in the art can understand the specific meaning of the above terms in this application based on the specific circumstances.

[0021] The following disclosure provides many different implementations or examples for carrying out different structures of this application. To simplify the disclosure, specific examples of components and arrangements are described below. Of course, these are merely examples and are not intended to limit the scope of this application. Furthermore, reference numerals and / or letters may be repeated in different examples; such repetition is for simplification and clarity and does not in itself indicate a relationship between the various implementations and / or arrangements discussed.

[0022] The embodiments of this application are described in detail below. Examples of these embodiments are shown in the accompanying drawings, wherein the same or similar reference numerals denote the same or similar elements or elements having the same or similar functions throughout. The embodiments described below with reference to the accompanying drawings are exemplary and are only used to explain this application, and should not be construed as limiting this application.

[0023] Please see Figure 1 This application provides a model training method for training an evaluation model to assess the remaining useful life of components of a device based on a base model. The base model includes an LSTM temporal encoding layer, a query mapping layer, a Transformer attention mechanism layer, and a regression output layer. The model training method includes: 01: Obtain the operating parameters, historical failure events, and historical failure timestamps of the component during its operation; 02: Based on operating parameters, historical failure events, and historical failure timestamps, construct a modeling dataset that includes time-series features and the actual remaining service life of the components corresponding to those time-series features; 03: Divide the modeling dataset into a training set and a validation set; 04: Input the time series features in the training set into the LSTM temporal coding layer. The LSTM temporal coding layer encodes the input time series features to output a hidden state sequence that represents the dynamic characteristics of the time series. 05: The query mapping layer maps the hidden state sequence to a single query vector; 06: Input the keys and values ​​obtained by linear transformation of the single query vector and the hidden state sequence into the Transformer attention mechanism layer. The Transformer attention mechanism layer performs weighted fusion through self-attention mechanism to generate a fused feature vector. 07: Input the fused feature vector into the regression output layer. The regression output layer maps the fused feature vector into a scalar form of the predicted remaining useful life of the component, so as to predict the remaining useful life of the component and form the initial evaluation model. 08: Input the time series features from the validation set into the initial evaluation model to obtain the prediction results of the validation set; 09: Based on the prediction results of the validation set and the actual remaining service life of the components, the operating parameters of the initial evaluation model are optimized using a preset loss function to construct the final evaluation model.

[0024] Please see Figure 2 This application also provides a model training device 100 for training an evaluation model for assessing the remaining lifespan of components of an evaluation device based on a base model. The base model includes an LSTM temporal encoding layer, a query mapping layer, a Transformer attention mechanism layer, and a regression output layer. The model training device 100 includes an acquisition module 10, a dataset construction module 20, a partitioning module 30, a temporal encoding module 40, a mapping module 50, a weighting module 60, an initial construction module 70, a validation module 80, and an optimization module 90.

[0025] Step 01 can be implemented by the acquisition module 10, step 02 by the dataset construction module 20, step 03 by the partitioning module 30, step 04 by the timing coding module 40, step 05 by the mapping module 50, step 06 by the weighting module 60, step 07 by the initial construction module 70, step 08 by the verification module 80, and step 09 by the optimization module 90. That is, the acquisition module 10 is used to acquire the operating parameters of the component during operation, the component's historical failure events, and the component's historical failure timestamps. The dataset construction module 20 is used to construct a modeling dataset containing time-series features and the actual remaining service life of the components corresponding to those time-series features, based on operating parameters, historical failure events, and historical failure timestamps. The partitioning module 30 is used to partition the modeling dataset into a training set and a validation set. The time-series encoding module 40 is used to input the time-series features from the training set into an LSTM time-series encoding layer, which encodes the input time-series features to output a hidden state sequence representing the time-series dynamic characteristics. The mapping module 50 is used to query the mapping layer to map the hidden state sequence into a single query vector. The weighting module 60 is used to input the single query vector and the keys and values ​​obtained by linear transformation of the hidden state sequence into Transf. The Ormer attention mechanism layer and the Transformer attention mechanism layer perform weighted fusion through self-attention mechanism to generate a fused feature vector. The initial construction module 70 is used to input the fused feature vector into the regression output layer. The regression output layer maps the fused feature vector into a scalar form of the predicted remaining useful life of the component to predict the remaining useful life of the component and form an initial evaluation model. The validation module 80 is used to input the time series features in the validation set into the initial evaluation model to obtain the prediction results of the validation set. The optimization module 90 is used to optimize the running parameters of the initial evaluation model based on the prediction results of the validation set and the actual remaining useful life of the component using a preset loss function to construct the final evaluation model.

[0026] Specifically, the method acquires the operating parameters of the component during operation, the component's historical failure events, and the component's historical failure timestamps. The component can be a core component or other functional part of the equipment, without limitation. The operating parameters of the component during operation can include data such as vibration, temperature, rotational speed, torque, current, power consumption, and load. More specifically, the model training method of this application can acquire the operating parameters of the component during equipment operation using industrial-grade sensors. That is, by deploying industrial-grade sensors on the equipment components and connecting these sensors to the sensor detection interface provided by the equipment manufacturer, real-time data acquisition of the equipment's operating status can be achieved, ensuring the real-time nature and accuracy of the acquired operating parameters and providing a reliable data foundation for subsequent analysis. The component's historical failure events represent the actual time when the component's function is lost. The component's historical failure timestamp refers to the time stamp of the component's failure from the start of operation to the point of failure. More specifically, the model training method of this application can extract the component's historical failure events and historical failure timestamps from a historical operation database.

[0027] Based on operating parameters, historical failure events, and historical failure timestamps, a modeling dataset is constructed that includes time-series features and the actual remaining service life of the components corresponding to those features. In other words, this application can first process and aggregate the operating parameters of components, historical failure events, and historical failure timestamps during the operation of the aforementioned equipment to form the modeling dataset required for constructing the evaluation model.

[0028] Dividing the modeling dataset into training and validation sets means that the modeling dataset obtained in the previous step can be shuffled and divided into two parts. Specifically, one part of the modeling dataset can be used as the data for training the initial evaluation model to obtain the training set, and the other part of the modeling dataset can be used as the data for validating whether the initial evaluation model is feasible to obtain the validation set.

[0029] After dividing the modeling dataset into a training set, the time-series features in the training set can be input into the LSTM temporal encoding layer. The LSTM temporal encoding layer encodes the input time-series features to output a hidden state sequence representing the temporal dynamics. Specifically, after dividing the modeling dataset into a training set, the time-series feature data of the training set can be input into the LSTM temporal encoding layer of a pre-defined neural network model for encoding, obtaining a hidden state sequence representing the temporal dynamics. The pre-defined neural network model can be a hybrid architecture model of Long Short-Term Memory Network (LSTM) and Global Attention Mechanism (Transformer). This retains the advantage of LSTM in capturing gradual long-term dependencies while also taking advantage of Transformer in accurately capturing long-distance mutations. Compared to a single LSTM model or Transformer model, the evaluation performance obtained by evaluating the remaining lifetime of components using the hybrid architecture model of this application will be significantly improved.

[0030] The time series features corresponding to the training set are multimodal input tensors X∈R. （T×D） Where T is the time step and D is the fused feature dimension. That is, this application can combine the features of time series indicators, spectral indicators and operating parameters to form a multimodal input tensor, which is then input into the LSTM temporal coding layer of a preset neural network model for encoding, thereby obtaining a hidden state sequence that represents the temporal dynamic characteristics.

[0031] Understandably, the degradation rate and behavior patterns of equipment components differ under different operating conditions. Failure characteristics are strongly correlated with operating condition characteristics, such as parameters like load, speed, vibration, and current. Therefore, the latent state sequence characterizing time-series dynamic properties can include one or more parameter characteristics such as vibration parameters and current under operating conditions.

[0032] In other words, the LSTM temporal coding layer of the preset neural network model in this application is an LSTM layer, which has the advantage of being good at capturing progressive long-term dependencies. The LSTM layer can be a single layer or a double layer, without any restriction. A single-layer LSTM is suitable for data feature analysis scenarios with small or medium data scale, relatively simple and straightforward temporal patterns, such as those mainly relying on recent history, and relatively basic tasks, such as simple anomaly detection and trend prediction. A double-layer LSTM is suitable for data feature analysis scenarios with a large amount of high-quality training data, very complex temporal patterns, and high abstraction requirements, such as natural language understanding, speech recognition, and health assessment of complex mechanical systems using multi-source heterogeneous sensor fusion, extremely long sequences, and crucial long-term dependencies, such as predicting the remaining lifespan of a device throughout its entire life cycle, and where computational resources are sufficient and insensitivity to inference speed is not a major concern.

[0033] A single-layer or two-layer LSTM processes the multimodal input tensor X to obtain the final hidden state h. T ∈R d h T The long-range temporal dependencies have been condensed and used as the "prior anchor" for subsequent attention, i.e., h. T A hidden state sequence that characterizes the temporal dynamics.

[0034] The query mapping layer maps the hidden state sequence to a single query vector, meaning that h can be... T Perform linear mapping W q We obtain the unique query term (Query) for Transformer self-attention. This unique query term forms a single query vector Q, i.e., W. q :h T →Q R dq dq represents the feature dimension corresponding to the fused query terms.

[0035] After obtaining a single query vector and a sequence of hidden states, the keys and values ​​obtained from the linear transformation of the single query vector and the sequence of hidden states can be input into the Transformer attention mechanism layer. The Transformer attention mechanism layer performs weighted fusion through a self-attention mechanism to generate a fused feature vector. That is, the key and value can be obtained from the single query vector Q and the sequence of hidden states h. T It is obtained through a linear transformation. Specifically, we can first process the hidden state sequence h. T via linear mapping W k W v We obtain, keeping K, V∈R (T×dk) T is the time step, and dk is the feature dimension corresponding to the fused key and value. Understandably, since the number of queries is reduced from T to 1, the computational complexity of attention is reduced from O(T²·d) to O(T·d).

[0036] The Transformer attention mechanism layer of the pre-defined neural network model may include a single-head or multi-head attention layer, and the single-head or multi-head attention layer may include formula (1): Attention(Q,K,V)=softmax(Q ) V..............Formula (1) The output of the above formula (1) is then processed by the feedforward neural network and residual connections of the Transformer attention mechanism layer to obtain the global context vector g, g∈R.d In other words, g is the fused feature vector. The role of the feedforward neural network is to independently transform the features at each location, increase non-linear expressive power, and fuse relational information with local features to generate a richer representation. The role of residual connections includes preserving the "relational information" extracted by attention, preventing information loss in multi-layer transmission, and achieving a convenient path for "identity mapping".

[0037] Next, the fused feature vector g is input into the regression output layer. The regression output layer maps the fused feature vector g into a scalar form of the component's remaining useful life prediction value, thus predicting the component's remaining useful life (RUL). pred RUL pred ∈R 1 This forms the initial evaluation model.

[0038] After obtaining the initial evaluation model, the time-series features from the validation set can be input into the initial evaluation model to obtain the prediction results of the validation set. Based on the prediction results of the validation set and the actual remaining service life of the component, the operating parameters of the initial evaluation model are optimized using a preset loss function to construct the final evaluation model. That is, in the process of training the evaluation model, this application can use the total loss L of the validation set as the evaluation index of the model performance. By adjusting the operating parameters of the initial evaluation model, the evaluation model with the lowest total loss is determined as the optimal evaluation model according to the preset loss function, thereby ensuring the generalization ability and prediction accuracy of the evaluation model on the prediction data.

[0039] In detail, when constructing the evaluation model, a cosine annealing learning rate scheduling strategy can be adopted, setting T... max The training cycle is 50 epochs, and early stopping on the validation set is used to dynamically adjust the training process by monitoring the total loss L on the validation set, ensuring that the evaluation model achieves optimal performance within a limited training period.

[0040] In summary, the model training method proposed in this application adopts a cascaded attention coupling scheme, and the evaluation model is constructed in a two-step manner of first time series and then global domain. This can not only reduce the amount of computation, but also highlight the weak early features, thereby improving the evaluation efficiency and accuracy of the true remaining service life of the component.

[0041] Thus, the evaluation model trained by the model training method of this application can intelligently quantify and predict the remaining service life of equipment components based on equipment operation data, replacing manual experience judgment and achieving an objective, fair, and accurate assessment of the remaining service life. This avoids premature replacement of equipment components before they fail, prevents performance redundancy waste, and reduces the production cost of equipment components. Simultaneously, the model of this application integrates LSTM and Transformer in a cascaded structure. LSTM is used to extract long-range temporal dependencies and generate global prior anchors. A single query vector significantly reduces computational complexity while retaining sensitivity to key degradation features. This application employs a cascaded attention coupling scheme, constructing the evaluation model through a two-step approach: first temporal, then global. This not only compresses computation but also highlights early, weak features, further improving evaluation efficiency and accuracy.

[0042] Please see Figure 3 In some implementations, step 02 includes: 021: Based on the operating parameters of components during equipment operation, multi-dimensional feature data is obtained through processing; 022: Based on the component's historical failure events and historical failure timestamps, sample data of the component's remaining useful life is constructed; 023: Associate multidimensional feature data with sample data to obtain the modeling dataset.

[0043] Please combine Figure 2 Steps 021, 022, and 023 can be implemented by the dataset construction module 20. The dataset construction module 20 is used to process multidimensional feature data based on the operating parameters of components during equipment operation; to construct sample data of the remaining service life of components based on historical failure events and historical failure timestamps; and to associate the multidimensional feature data with the sample data to obtain the modeling dataset.

[0044] Specifically, the operating parameters of components collected during equipment operation in this application cover multi-dimensional features such as vibration signals (acceleration, velocity, displacement), temperature (e.g., ambient temperature, component surface temperature), rotational speed, torque, current, power consumption, and load, comprehensively covering key indicators of equipment operation and facilitating equipment status assessment. Therefore, based on the operating parameters of components during equipment operation, multi-dimensional feature data including vibration signals, temperature, rotational speed, torque, current, power consumption, and load can be obtained.

[0045] Based on the historical failure events and timestamps of components, sample data of the remaining service life of the components is constructed. That is, this application employs a combination of high-frequency sampling (e.g., frequency ≥ 1kHz) and low-frequency sampling (e.g., 1Hz) to meet the requirements of different signal characteristics. The collected multidimensional feature data is output in the form of raw time-series data streams, with the component's historical failure timestamps and historical failure events appended, and stored in local cache and cloud databases for subsequent processing and analysis.

[0046] After obtaining the multidimensional feature data of the component and the sample data of the component's remaining service life, the multidimensional feature data and the sample data are associated to obtain the modeling dataset. This ensures that the feature data corresponds one-to-one with the target remaining service life value, providing high-quality training data for subsequent supervised learning modeling.

[0047] Thus, the model training method of this application can aggregate the multidimensional feature data obtained by processing the operating parameters of the components during the operation of the device and the sample dataset containing RUL labels to form a modeling dataset, providing high-quality training data for subsequent training.

[0048] Please see Figure 4 In some implementations, step 021 includes: 0211: Preprocess the operating parameters of the component to obtain the processed operating parameters, wherein the preprocessing includes at least one of resampling and alignment, outlier handling, missing value handling, standardization and / or Fourier denoising; 0213: Extract feature data from multiple dimensions of the processed running data to form multidimensional feature data; among which, the multidimensional feature data includes the first dimension data corresponding to the vibration signal, the second dimension data corresponding to the current signal, and the third dimension data corresponding to the temperature signal.

[0049] Please combine Figure 2 Steps 0211 and 0213 can be implemented by the dataset construction module 20. The dataset construction module 20 is used to preprocess the operating parameters of the component to obtain processed operating data. The preprocessing includes at least one of resampling and alignment, outlier handling, missing value handling, standardization, and / or Fourier denoising. Multiple-dimensional feature data are extracted from the processed operating data to form multi-dimensional feature data. The multi-dimensional feature data includes first-dimensional data corresponding to the vibration signal, second-dimensional data corresponding to the current signal, and third-dimensional data corresponding to the temperature signal.

[0050] Specifically, after collecting the operating parameters of a component at a certain moment, the operating parameters are preprocessed to obtain processed operating parameters. The preprocessing includes at least one of resampling and alignment, outlier handling, missing value handling, standardization, and / or Fourier denoising. This means that the operating parameters can be preprocessed using one or more of these methods to obtain more reliable operating parameters.

[0051] The resampling and alignment process involves first interpolating operating parameters from different sampling frequencies (e.g., 10 kHz, 1 kHz, or 1 Hz) to the same time point, such as a time stamp of 1 second. Then, the operating parameters from different sampling frequencies are combined into a single complete feature. This unifies sensor data from different sources, frequencies, and time bases to the same time base and sampling frequency, thereby constructing a synchronized and consistent multivariate time series for subsequent fusion analysis and modeling.

[0052] The outlier handling process is as follows: for each sequence's operating parameters, outliers are identified using three standard deviations as the cutoff, and the identified outliers are replaced with the median. This prevents the model from learning incorrect patterns and improves the model's robustness.

[0053] The missing value imputation process is as follows: For time-series data, linear interpolation can be used for imputation. Specifically, if n consecutive values ​​are missing from time t, then n+1 equal division points are taken from the two values ​​before and after the missing value and used to imput the missing values. Imputing missing values ​​with reasonable values ​​ensures that the algorithm can run normally and guarantees data integrity.

[0054] The standardization process involves standardizing the data for each sequence's parameters using maximum and minimum value standardization methods. Specifically, the standardized sequence value is X = (X... max X)∕(X X min The purpose of standardizing data is to give features from different sensors and at different scales an equal voice in the model.

[0055] The Fourier noise reduction process is as follows: Figure 5As shown, a Fast Fourier Transform (FFT) is performed on the original sequence parameters to obtain their frequency domain representation. Then, a threshold is set to distinguish and remove high-frequency noise components in the frequency domain. Finally, an Inverse Fourier Transform is performed on the remaining components in the frequency domain to obtain the denoised time-domain sequence. This method effectively suppresses random noise and irrelevant interference while preserving key fault characteristics.

[0056] After obtaining the processed operating parameters, multi-dimensional feature data can be extracted from these parameters to form multi-dimensional feature data. This multi-dimensional feature data includes first-dimensional data corresponding to vibration signals, second-dimensional data corresponding to current signals, and third-dimensional data corresponding to temperature signals. For example, for vibration signals, first-dimensional data such as RMS, peak value, kurtosis, and spectral energy distribution can be extracted. For current signals, second-dimensional data such as RMS value and harmonic distortion rate (THD) can be extracted. For temperature signals, feature data such as mean and temperature rise rate can be extracted. For rotational speed and load, third-dimensional data such as mean and fluctuation coefficient can be extracted. By analyzing these multi-dimensional feature data, this application can comprehensively reflect the degradation trend of the equipment, facilitating accurate assessment of the remaining service life of equipment components through a trained evaluation model.

[0057] In one example, a 20-dimensional multi-dimensional feature vector can be constructed, covering degradation trends in mechanical, electrical, and thermodynamic aspects, ensuring that the model captures equipment degradation characteristics from multiple angles and improves prediction accuracy.

[0058] Multidimensional feature data can be output as a feature matrix, where each row corresponds to a time window and each column corresponds to a feature dimension, providing structured input for subsequent model training and facilitating model learning and prediction.

[0059] Thus, the model training method of this application can first preprocess the operating parameters of the components during the operation of the equipment, and then extract multi-dimensional feature data from the processed operating parameters to obtain multi-dimensional feature data, which can more comprehensively reflect the degradation trend of the equipment, so as to achieve accurate assessment of the remaining service life of the equipment components through the trained evaluation model.

[0060] Please see Figure 6 In some implementations, step 022 includes: 0221: Based on the historical failure events and timestamps of the component, a timeline is constructed for the component from the start of operation to the failure of the component. Within the time interval defined by the timeline, the timeline is divided according to a preset fixed time interval to obtain multiple observation moments. For each observation moment, the time difference between each observation moment and the corresponding failure timestamp is used as the remaining service life of the component at the observation moment. In some implementations, step 023 includes 0231: Align the multidimensional feature data with the remaining service life of the component according to the timestamp of the observation time; 0232: Based on the aligned multidimensional feature data and the corresponding remaining lifetime, a modeling dataset is constructed.

[0061] Please combine Figure 2 Steps 0221, 0231, and 0232 can be implemented by the dataset construction module 20. The dataset construction module 20 is used to construct a timeline for the component from the start of its operation to its failure, based on the component's historical failure events and timestamps. Within the time interval defined by the timeline, the timeline is divided according to a preset fixed time interval to obtain multiple observation moments. For each observation moment, the time difference between each observation moment and the corresponding failure timestamp is used as the component's remaining service life at that observation moment. The multidimensional feature data and the component's remaining service life are aligned one-to-one according to the timestamps of the observation moments. Based on the aligned multidimensional feature data and the corresponding remaining service life, a modeling dataset is constructed.

[0062] Specifically, the fixed time interval can be 7 minutes, 8 minutes, 9 minutes, 10 minutes, 11 minutes, 12 minutes, 13 minutes, 14 minutes, 15 minutes, or 16 minutes, without restriction. For example, if the fixed time interval is 10 minutes, the multiple observation times can be 0, 10 minutes, 20 minutes, 30 minutes, 40 minutes, 50 minutes, 60 minutes, 70 minutes, 80 minutes, 90 minutes, or 100 minutes, and so on. The time interval defined by the time axis can be the time interval from the start of component operation to the time of component failure.

[0063] In other words, this application can construct a timeline for a component from its first operation to failure based on historical failure events and historical failure timestamps, and calculate the historical remaining useful life (RUL) to failure using the current time as a reference, providing a clear target value for supervised learning. For example, historical failure events can be sampled and generated at fixed time intervals, such as sampling and generating historical failure events every 10 minutes, generating RUL tags, with the tag unit being hours or the number of cycles.

[0064] Aligning multidimensional feature data with the remaining useful life (RUL) of components according to the timestamp of the observation time means aligning the multidimensional feature data with the RUL labels of the components according to the timestamp of the observation time to construct "feature-RUL" paired samples. Understandably, the resulting modeling dataset containing RUL labels can be used for subsequent supervised learning modeling to ensure that the evaluation model can learn the degradation patterns of the equipment.

[0065] Thus, the model training method of this application adopts an alignment method that aligns multidimensional feature data with the remaining service life of the component, which can ensure that the features correspond one-to-one with the target remaining service life value, and can provide high-quality training data for subsequent supervised learning modeling.

[0066] Please see Figure 7 In some implementations, prior to step 09, the model training method further includes: 081: The mean squared error loss function term, the physical impossibility penalty term, and the dynamic weight decay term of the preset loss function are calculated based on the sample data; 082: The mean squared error loss function term, the physical impossibility penalty term, and the dynamic weight decay term are summed to obtain the preset loss function.

[0067] Please see Figure 8 The model training device 100 also includes a loss function construction module 81. Steps 081 and 082 can be implemented by the loss function construction module 81. The loss function construction module 81 is used to calculate the mean squared error loss function term, the physical impossibility penalty term, and the dynamic weight decay term of the preset loss function based on the sample data; and to accumulate the mean squared error loss function term, the physical impossibility penalty term, and the dynamic weight decay term to obtain the preset loss function.

[0068] Understandably, the remaining useful life (RUL) of an industrial component cannot physically exceed its theoretical maximum useful life (RUL). max However, when the initial evaluation model built from the pre-defined neural network model is run under extreme conditions or in small sample scenarios, it may output RUL. pred >RUL max Even negative values ​​can lead to distortions and safety hazards when using the initial assessment model to predict remaining useful life.

[0069] In view of this, the model training method of this application introduces two additional mechanisms, namely a physical impossibility penalty term and a dynamic weight decay term, in addition to the original mean squared error (MSE) loss function term, so as to construct a composite loss function that takes into account both fitting accuracy and physical consistency.

[0070] Specifically, the preset loss function Ltotal It can be calculated using the following formula (2): L total =L MSE +λ(epoch) L phy ............Formula (2) Among them, L MSE = In the formula, This is a predicted value for the remaining useful life. The remaining useful life is the historical actual value, N is the sample size, and L is the number of samples. MSE This is the mean squared error function. In other words, the overall fitting accuracy can be guaranteed using the traditional mean squared error function.

[0071] L phy =β· In the formula, This is a predicted value for the remaining useful life. The maximum value of the remaining useful life, when RUL pred The closer to or beyond RUL max When the time is right, the punishment is amplified exponentially; β controls the intensity of the punishment, and τ controls the sensitivity. RUL max The lifespan of the component is determined by the design life or fatigue test, and can be adjusted according to the operating conditions. Preferably, β∈[0.8, 1.5], τ∈[5, 15].

[0072] λ(epoch)=λ0 e ( 0.02epoch) In the early stages of training, λ is relatively large, forcing the network to adhere to physical boundaries. As the number of epochs increases, λ decays exponentially, and the network gradually focuses on data fitting, avoiding underfitting due to over-constraint. λ(epoch) is a function that returns a scaling factor based on the current epoch number, used to adjust the learning rate. λ0 can be 0.5, or its value can be fine-tuned using early stopping on the validation set.

[0073] Thus, the model training method of this application introduces two additional mechanisms, namely a physical impossibility penalty term and a dynamic weight decay term, in addition to the original mean squared error loss function term. The resulting preset loss function can balance fitting accuracy and physical consistency.

[0074] Please see Figure 9 In some implementations, the model training method further includes: 091: The test set is obtained by partitioning the modeling dataset; 092: Evaluate the feasibility of the model based on the test set, and output the feasibility results of the model. 093: Improve the evaluation model based on the feasibility results to obtain the improved evaluation model.

[0075] Please combine Figure 2 Steps 091, 092, and 093 can be implemented by the optimization module 90. That is, the optimization module 90 is used to obtain a test set based on the modeling dataset; to test the feasibility of the evaluation model based on the test set and output the feasibility results of the evaluation model; and to improve the evaluation model based on the feasibility results to obtain the improved evaluation model.

[0076] Specifically, in addition to dividing the modeling dataset into training and validation sets, a test set can also be created. Understandably, this application can divide the modeling dataset into three parts, serving as the training set, validation set, and test set, respectively. For example, a 70% training set, 15% validation set, and 15% test set ratio can be used to ensure sufficient data for learning, optimization, and evaluation at each stage of the model. Taking a total of 100,000 samples as an example, the final dataset would be divided into a 70,000-sample training set, a 15,000-sample validation set, and a 15,000-sample test set, ensuring that each subset has sufficient data and does not overlap.

[0077] In detail, this application can employ a random shuffle strategy to avoid temporal order bias, ensure that the data of the same component appears in only a subset, prevent information leakage, and ensure the fairness of model evaluation.

[0078] In other words, the test set of this application is used to test the feasibility of the evaluation model constructed from the training set and the validation set. If the feasibility of the evaluation model is satisfactory, the feasibility evaluation framework of the evaluation model can be evaluated from three dimensions: technology, deployment and operation and maintenance, and business.

[0079] If the feasibility result indicates that the evaluation model is not feasible, specifically for example, if the evaluation model is deployed in the cloud but the infrastructure for the evaluation model is not set up properly, then the infrastructure for the evaluation model can be improved based on the feasibility result to obtain a better evaluation model.

[0080] Thus, the model training method of this application can divide a test set from the modeling dataset, test the feasibility of the evaluation model based on the test set to improve the evaluation model, and ensure stable use when evaluating the remaining service life of equipment components.

[0081] Please see Figure 10 In some implementations, the model training method further includes: 094: Obtain test data to determine the remaining service life of the component; 095: Process the test data according to the preset data stream window length and preset step size to obtain the model test data; 096: Input the model test data into the evaluation model to obtain the test results; 097: Perform drift detection on the test results. If the drift value of the test results is greater than the preset drift threshold, trigger retraining and evaluation of the model.

[0082] Please see Figure 11 The model training device 100 also includes a retraining module 91. Steps 094, 095, 096, and 097 can be implemented by the retraining module 91. That is, the retraining module 91 is used to acquire the test data of the remaining service life of the component; process the test data according to the preset data stream window length and preset step size to obtain model test data; input the model test data into the evaluation model to obtain test results; perform drift detection on the test results, and if the drift value of the test results is greater than the preset drift threshold, trigger retraining of the evaluation model.

[0083] Specifically, the test data for obtaining the remaining service life of a component can be obtained by acquiring the operating parameter data of the component of the industrial equipment in real time through industrial sensors at a certain current moment.

[0084] The preset data stream window length can be 120s for the Flink window and the preset step size is 10s. That is, this application can construct a sliding time window with a Flink window length of 120s and a step size of 10s to obtain model test data.

[0085] Then, the model test data is input into the evaluation model, and the test results are output, which can be the test results with the component failure time calculation and the component remaining life label RUL.

[0086] The test results are then pushed to the operation and maintenance layer of the model training device 100. This layer calculates the relative entropy (Kullback-Leibler Divergence, KL) divergence every 24 hours to monitor whether data drift has occurred. For example, a preset drift threshold of 0.3 can be used. If the drift value obtained from monitoring data drift using KL divergence is greater than 0.3, retraining is triggered. KL divergence is an indicator that measures the difference between two probability distributions PP and QQ.

[0087] In other words, the model training method of this application can trigger retraining of the evaluation model when data drift occurs in the test results of the evaluation model, thus timely mitigating the data drift phenomenon and training a better-performing evaluation model without data drift. Data drift refers to the phenomenon where, throughout the entire lifecycle of a machine learning model, the statistical characteristics of the data encountered by the model in the production environment (i.e., online data) change compared to the statistical characteristics of the data used during model training (i.e., training data), leading to a decline in model performance.

[0088] Thus, the model training method of this application can detect in real time whether data drift occurs during the evaluation model testing process, thereby triggering retraining of the evaluation model and improving the performance of the evaluation model in a timely manner.

[0089] In other embodiments of this application, an offline task can be set to periodically and automatically trigger retraining of the evaluation model to ensure its performance. For example, the offline task could check at 2 PM every day whether the time interval since the last successful training is greater than or equal to 30 days. If so, a full retraining would be automatically triggered; otherwise, retraining of the evaluation model would not be necessary.

[0090] In other embodiments of this application, retraining can also be triggered when performance degradation of the evaluation model is detected. Performance degradation can be manifested as a 7-day moving average of the online mean absolute error (MAE) increasing by more than 20% from the baseline, or an early warning hit rate of less than 85%, which triggers retraining of the evaluation model.

[0091] In other embodiments of this application, retraining can also be manually triggered. For example, when a maintenance engineer discovers false or missed data, they can manually input the RUL tag with one click to fine-tune the data increment for the day, and then manually trigger a full retraining of the evaluation model the next day.

[0092] In other words, the model training method of this application can trigger the retraining of the evaluation model when one or more of the above four retraining triggering conditions are met, so as to ensure the performance of the evaluation model.

[0093] Please see Figure 12 This application also provides a repair method, which uses an evaluation model trained using the model training method described in any of the above embodiments. The repair method includes: S11: Obtain the operating data of the component to be evaluated in the equipment; S12: Based on the operating data and evaluation model of the component to be evaluated, assess the remaining useful life of the component to be evaluated; S13: Draw the remaining service life curve based on the remaining service life of the component to be evaluated; S14: Analyze and process the remaining service life of the component to be evaluated to obtain early warning information; S15: Based on the remaining service life curve and / or early warning information, control the component to be evaluated to be shut down for maintenance or to be replaced; and / or, display the remaining service life curve and / or early warning information of the component to be evaluated in real time.

[0094] Please see Figure 13 This application also provides a maintenance device 200. The maintenance device 200 includes an acquisition module 210, an evaluation module 220, a curve plotting module 230, an analysis module 240, and a control and display module 250. Step S11 can be implemented by the acquisition module 210, step S12 can be implemented by the evaluation module 220, step S13 can be implemented by the evaluation module 220, step S14 can be implemented by the analysis module 240, and step S15 can be implemented by the control and display module 250. That is, the acquisition module 210 is used to acquire the operating data of the component to be evaluated in the equipment; the evaluation module 220 is used to evaluate the remaining service life of the component to be evaluated based on the operating data and evaluation model; the curve plotting module 230 is used to plot the remaining service life curve based on the remaining service life of the component to be evaluated; the analysis module 240 analyzes and processes the remaining service life of the component to be evaluated to obtain early warning information; the control and display module 250 is used to control the component to be evaluated to perform shutdown maintenance or replace the component to be evaluated based on the remaining service life curve and / or early warning information; and / or, to display the remaining service life curve and / or early warning information of the component to be evaluated in real time.

[0095] Specifically, the evaluation model of this application can be trained using the model training method described above. There can be one or more components to be evaluated; that is, the maintenance method of this application can evaluate the remaining useful life of only a single component, or it can evaluate the remaining useful life of multiple components simultaneously.

[0096] This application utilizes industrial-grade sensors to acquire operational data of the components under evaluation during equipment operation. Specifically, by deploying industrial-grade sensors on the components under evaluation and connecting these sensors to the sensor detection interface provided by the equipment manufacturer, real-time data acquisition of equipment operating status can be achieved. This ensures the real-time nature and accuracy of the acquired operational parameters, providing a reliable data foundation for subsequent remaining service life assessment. The acquired operational parameters of the components under evaluation during equipment operation encompass multi-dimensional characteristics such as vibration signals (acceleration, velocity, displacement), temperature (e.g., ambient temperature, component surface temperature), rotational speed, torque, current, power consumption, and load.

[0097] After obtaining the operating data of the component to be evaluated in the device, the remaining useful life of the component is assessed based on the operating data and the assessment model. In other words, this application can input the operating data of the component to be evaluated into the assessment model to obtain the remaining useful life of the component.

[0098] This application can input the evaluation prediction results of the remaining service life of the component to be evaluated from the evaluation model into a preset curve drawing tool for curve drawing processing to obtain the remaining service life curve, and analyze and process the remaining service life of the component to be evaluated to obtain early warning information.

[0099] It should be noted that, based on the remaining service life curve and / or warning information, the component to be evaluated is controlled to be shut down for maintenance or replacement; and / or, the remaining service life curve and / or warning information of the component to be evaluated is displayed in real time, so that maintenance personnel can shut down the component to be evaluated for maintenance or replacement according to the remaining service life curve and / or warning information, including the following situations: (1) The maintenance method of this application can automatically control the component to be evaluated to be shut down for maintenance and automatically control the replacement of the component to be evaluated based solely on the remaining service life curve and warning information; (2) The maintenance method of this application can directly display the remaining service life curve and warning information on the display interface, so that maintenance personnel can determine whether manual shutdown maintenance and manual replacement of the component to be evaluated are required based on the remaining service life curve and warning information; (3) The maintenance method of this application can display the remaining service life curve and warning information on the display interface while automatically controlling the component to be evaluated to be shut down for maintenance and automatically controlling the replacement of the component to be evaluated based on the remaining service life curve and warning information.

[0100] Among them, the following three situations apply to the shutdown maintenance or replacement of the component to be evaluated based on the remaining service life curve and / or early warning information: (1) the shutdown maintenance or replacement of the component to be evaluated based on the remaining service life curve; (2) the shutdown maintenance or replacement of the component to be evaluated based on early warning information; and (3) the shutdown maintenance or replacement of the component to be evaluated based on the remaining service life curve and early warning information.

[0101] For example, if the maintenance module 240 can know from the remaining service life curve or early warning information that the remaining service life of a certain component to be evaluated is close to 0 or has already reached 0, it can control the component to be evaluated to automatically stop for maintenance, or control the automatic replacement of a new component, thereby achieving more intelligent control, eliminating the need for shutdown and maintenance of each unit, avoiding premature replacement of the component to be evaluated, and making full use of the available life of the component to be evaluated.

[0102] In other words, the maintenance method of this application can control the component to be evaluated to automatically stop for maintenance or automatically replace new components based on the remaining service life and warning information displayed on the interface, thereby achieving more intelligent control, eliminating the need for shutdown and maintenance of each unit, avoiding premature replacement of the component to be evaluated, and making full use of the available life of the component to be evaluated.

[0103] In addition, after the remaining service life curve is plotted, it can be output to the display layer of the model training device 100 to display the remaining service life curve in real time. This allows users to more intuitively understand the changing trend of the remaining service life of the component to be evaluated in the current device, and enables users to manually stop the evaluation of the component for maintenance or replacement based on the remaining service life curve.

[0104] Furthermore, when analyzing the remaining service life of the component to be evaluated to obtain early warning information, for example, if the analysis shows that the remaining service life of a certain component of the current equipment is close to or has already reached 0, an early warning message can be output. The early warning message could be, for example, "The remaining service life of component 1 is close to 0, please replace it in time." At this time, the maintenance method of this application can also output early warning information to the display layer, displaying the early warning information of the component in real time, so that users can manually stop the machine for maintenance or replacement of the component to be evaluated based on the remaining service life curve.

[0105] In other words, in the maintenance method of this application, maintenance personnel can perform shutdown operations only on the equipment to be maintained based on the remaining service life and warning information displayed on the interface, eliminating the need for shutdown and maintenance of each unit, avoiding premature replacement of the parts to be evaluated, and making full use of the available life of the parts to be evaluated.

[0106] Thus, the maintenance method of this application uses an evaluation model to quantify the remaining service life of the component to be evaluated. Based on the remaining service life curve and early warning information obtained from the obtained remaining service life, a component replacement plan can be formulated in advance. According to the differentiated downtime maintenance strategy, only the industrial manufacturing equipment to be maintained is shut down, thereby shortening the downtime and improving the overall utilization rate of the equipment.

[0107] This application also provides a computer device. The computer device includes a memory and a processor. The memory stores a computer program, and when the computer program is executed by the processor, it implements the model training method and the maintenance method described in any of the above embodiments.

[0108] Thus, the evaluation model trained by the computer equipment application model training method of this application can intelligently quantify and effectively predict the remaining service life of equipment components based on equipment operation data, replacing manual experience judgment and achieving an objective, fair, and accurate assessment of the remaining service life. This avoids premature replacement of equipment components before they fail, prevents performance redundancy waste, and reduces the production cost of equipment components. Furthermore, the model of this application integrates LSTM and Transformer in a cascaded structure. LSTM is used to extract long-range temporal dependencies and generate global prior anchors. A single query vector significantly reduces computational complexity while retaining sensitivity to key degradation features. This application employs a cascaded attention coupling scheme, constructing the evaluation model through a two-step approach: first temporal, then global. This not only compresses computation but also highlights early, weak features, further improving evaluation efficiency and accuracy.

[0109] This application also provides a computer-readable storage medium. The computer-readable storage medium stores a computer program that, when executed by one or more processors, implements the model training method and the maintenance method described in any of the above embodiments.

[0110] Thus, the evaluation model trained by the computer-readable storage medium application model training method of this application can intelligently quantify and effectively predict the remaining service life of equipment components based on equipment operation data, replacing manual experience judgment and achieving an objective, fair, and accurate assessment of the remaining service life. This avoids premature replacement of equipment components before they fail, prevents performance redundancy waste of equipment components, and reduces the production cost of equipment components. Simultaneously, the model of this application integrates LSTM and Transformer, adopting a cascaded structure. LSTM is used to extract long-range temporal dependencies and generate global prior anchor points. Then, a single query vector significantly reduces computational complexity while retaining sensitivity to key degradation features. This application employs a cascaded attention coupling scheme, constructing the evaluation model through a two-step approach: first temporal, then global. This not only compresses computation but also highlights early weak features, further improving evaluation efficiency and accuracy.

[0111] The above embodiments merely illustrate several implementation methods of this application, and while the descriptions are relatively specific and detailed, they should not be construed as limiting the scope of this patent application. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of this application, and these all fall within the protection scope of this application. Therefore, the protection scope of this patent application should be determined by the appended claims.

Claims

1. A model training method for training an evaluation model based on a base model to assess the remaining service life of equipment components, characterized in that, The basic model includes: an LSTM temporal encoding layer, a query mapping layer, a Transformer attention mechanism layer, and a regression output layer; the model training method includes: Obtain the operating parameters of the component during its operation, the historical failure events of the component, and the historical failure timestamps of the component; Based on the operating parameters, the historical failure events, and the historical failure timestamps, a modeling dataset containing time series features and the actual remaining service life of the components corresponding to the time series features is constructed. The modeling dataset is divided into a training set and a validation set; The time series features in the training set are input into the LSTM temporal coding layer, and the LSTM temporal coding layer encodes the input time series features to output a hidden state sequence that represents the temporal dynamic characteristics. The query mapping layer maps the hidden state sequence into a single query vector; The single query vector and the keys and values ​​obtained by linear transformation of the hidden state sequence are input into the Transformer attention mechanism layer. The Transformer attention mechanism layer performs weighted fusion through a self-attention mechanism to generate a fused feature vector. The fused feature vector is input into the regression output layer, which maps the fused feature vector into a scalar form of the remaining useful life prediction value of the component, so as to predict the remaining useful life of the component and form an initial evaluation model. The time series features in the validation set are input into the initial evaluation model to obtain the prediction results of the validation set; Based on the prediction results of the validation set and the actual remaining service life of the component, the operating parameters of the initial evaluation model are optimized using a preset loss function to construct the final evaluation model.

2. The model training method according to claim 1, characterized in that, Based on the operating parameters, the historical failure events, and the historical failure timestamps, a modeling dataset is constructed that includes time-series features and the actual remaining service life of the components corresponding to those time-series features. Based on the operating parameters of the components during the operation of the equipment, multidimensional feature data is obtained; Based on the historical failure events and historical failure timestamps of the component, sample data of the remaining service life of the component is obtained; The modeling dataset is obtained by associating the multidimensional feature data with the sample data.

3. The model training method according to claim 2, characterized in that, The multidimensional feature data obtained by processing the operating parameters of the components during the operation of the device includes: The operating parameters of the component are preprocessed to obtain the processed operating parameters, wherein the preprocessing includes at least one of resampling and alignment, outlier handling, missing value handling, standardization, and Fourier denoising. The multidimensional feature data is formed by extracting feature data from the processed operating parameters in multiple dimensions; wherein, the multidimensional feature data includes first-dimensional data corresponding to vibration signals, second-dimensional data corresponding to current signals, and third-dimensional data corresponding to temperature signals.

4. The model training method according to claim 3, characterized in that, The sample data for obtaining the remaining service life of the component based on its historical failure events and timestamps includes: Based on the historical failure events and historical failure timestamps of the component, a timeline is constructed for the component from the start of operation to the failure of the component. Within the time interval defined by the timeline, the timeline is divided according to a preset fixed time interval to obtain multiple observation moments. For each observation moment, the time difference between each observation moment and the corresponding failure timestamp is taken as the remaining service life of the component at the observation moment. The step of associating the multidimensional feature data with the sample data to obtain the modeling dataset includes: The multidimensional feature data and the remaining service life of the component are aligned one-to-one according to the timestamp of the observation time; The modeling dataset is constructed based on the aligned multidimensional feature data and the corresponding remaining lifetime.

5. The model training method according to claim 2, characterized in that, Before constructing the final evaluation model by optimizing the parameters of the initial evaluation model using a preset loss function based on the prediction results of the validation set and the actual remaining service life of the component, the method further includes: Based on the sample data, the mean squared error loss function term, the physical impossibility penalty term, and the dynamic weight decay term of the preset loss function are calculated. The preset loss function is obtained by summing the mean square error loss function term, the physical impossibility penalty term, and the dynamic weight decay term.

6. The model training method according to claim 1, characterized in that, The model training method also includes: The test set is obtained by dividing the modeling dataset. Test the feasibility of the evaluation model based on the test set, and output the feasibility results of the evaluation model; The evaluation model is improved based on the feasibility results to obtain the improved evaluation model.

7. The model training method according to claim 1, characterized in that, The model training method also includes: Obtain the test data for the remaining service life of the component; The test data is processed according to the preset data stream window length and preset step size to obtain model test data; The test data of the model is input into the evaluation model to obtain the test results; The test results are subjected to drift detection. If the drift value of the test results is greater than a preset drift threshold, the evaluation model is retrained.

8. A maintenance method, comprising an evaluation model trained using the model training method according to any one of claims 1 to 7, characterized in that, The repair method includes: Obtain operational data of the components to be evaluated in the equipment; Based on the operating data of the component to be evaluated and the evaluation model, the remaining service life of the component to be evaluated is evaluated; The remaining service life curve is plotted based on the remaining service life of the component to be evaluated; The remaining service life of the component to be evaluated is analyzed and processed to obtain early warning information; Based on the remaining service life curve and / or the early warning information, control the component to be evaluated to be shut down for maintenance or to be replaced; and / or, display the remaining service life curve and / or the early warning information of the component to be evaluated in real time.

9. A computer device, characterized in that, The computer device includes a memory and a processor. The memory stores a computer program, which, when executed by the processor, implements the model training method according to any one of claims 1-7 and the maintenance method according to claim 8.

10. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a computer program that, when executed by one or more processors, implements the model training method of any one of claims 1-7 and the maintenance method of claim 8.

Images