Model training method, state and residual life prediction method, device and medium

Abstract

The invention provides a model training method, a state and residual life prediction method and device, and a medium. The method comprises the steps of obtaining multi-sensor time sequence data; performing feature extraction on the multi-sensor time sequence data through the first branch to obtain a space-frequency fusion feature; performing bidirectional state capture on the space-frequency fusion feature through a second branch to obtain a bidirectional hidden state sequence; performing historical state association on the bidirectional hidden state sequence through a third branch to obtain an attention weight set; distributing attention weights for all states in the bidirectional hidden state sequence based on the attention weight set; performing feature compression and nonlinear transformation on the updated bidirectional hidden state sequence through a fourth branch, and predicting the affiliated state and residual service life of the sample equipment; and adopting a preset loss function to train the large language model based on the affiliated state and the remaining service life of the sample equipment to obtain an equipment state and remaining service life prediction model. Therefore, equipment fault detection and service life prediction are effectively realized.

Description

Technical Field

[0001] The embodiments of this disclosure relate to the field of artificial intelligence technology, and more specifically, to a model training method, a state and remaining lifetime prediction method, an apparatus, and a medium. Background Technology

[0002] The working efficiency of large-diameter equipment is constrained by multiple factors: In extreme working environments, the combined effects of high temperature and high pressure conditions and high impact loads cause significant recoil loads and mechanical vibrations in the system, leading to an unexpected rate of fatigue damage accumulation in key components, which ultimately results in problems such as deviations in accuracy indicators.

[0003] In related technologies, current equipment support systems generally adopt maintenance strategies based on historical transmission frequency and replacement cycles of vulnerable parts. This model cannot effectively cope with the dynamic damage evolution in complex environments. When faced with high-intensity combat, rapid cross-regional relocation, and other variable operating conditions, traditional methods are difficult to effectively achieve equipment fault detection and lifespan prediction. Summary of the Invention

[0004] The embodiments described herein provide a model training method, a state and remaining lifetime prediction method, an apparatus, and a medium that overcome the aforementioned problems.

[0005] In a first aspect, according to the present disclosure, a training method for a device status and remaining useful life prediction model is provided, which is applied to a large language model, the large language model including: a first branch, a second branch, a third branch and a fourth branch; the first branch is used to extract the spatial features and frequency domain features of the device, and the second branch is used to capture the forward hidden state and backward hidden state of the device during operation through bidirectional temporal modeling; The method includes: Acquire multi-sensor time-series data from the sample device; The first branch is used to extract features from the multi-sensor time-series data of the sample device to obtain spatial-frequency fusion features. The second branch is used to perform bidirectional state capture on the space-frequency fusion features to obtain a bidirectional hidden state sequence; By associating the bidirectional hidden state sequence with historical states through the third branch, an attention weight set for representing different states is obtained. Attention weights are assigned to all states in the bidirectional hidden state sequence based on the attention weight set, so as to update the weights of the bidirectional hidden state sequence. The updated bidirectional hidden state sequence is subjected to feature compression and nonlinear transformation through the fourth branch to predict the state and remaining lifespan of the sample device. The crack propagation physical rules are added as regularization terms to the preset loss function; and the preset loss function is used to train the large language model based on the state and remaining service life of the sample equipment to obtain the equipment state and remaining service life prediction model.

[0006] Secondly, according to the content of this disclosure, a method for predicting equipment status and remaining life is provided, including: Acquire multi-sensor time-series data of the target device; The multi-sensor time-series data of the target device are input into the device status and remaining life prediction model to obtain the status and remaining life of the target device. The equipment status and remaining life prediction model is generated by the method described in the first aspect.

[0007] Thirdly, according to the present disclosure, a training device for a device status and remaining life prediction model is provided, which is applied to a large language model. The large language model includes a first branch, a second branch, a third branch, and a fourth branch. The first branch is used to extract the spatial features and frequency domain features of the device, and the second branch is used to capture the forward hidden state and backward hidden state of the device during operation through bidirectional time series modeling. The device includes: The first acquisition module is used to acquire multi-sensor time-series data of the sample device; The extraction module is used to extract features from the multi-sensor time-series data of the sample device through the first branch to obtain spatial-frequency fusion features; The capture module is used to capture the spatial-frequency fusion features bidirectionally through the second branch to obtain a bidirectional hidden state sequence; The association module is used to associate the bidirectional hidden state sequence with historical states through the third branch to obtain an attention weight set for representing different states. The allocation module is used to allocate attention weights to all states in the bidirectional hidden state sequence based on the attention weight set, so as to update the weights of the bidirectional hidden state sequence. The prediction module is used to perform feature compression and nonlinear transformation on the updated bidirectional hidden state sequence through the fourth branch to predict the state and remaining lifespan of the sample device. The training module is used to incorporate the crack propagation physical rules as regularization terms into a preset loss function; and using the preset loss function, the large language model is trained based on the state and remaining service life of the sample equipment to obtain a prediction model for equipment state and remaining service life.

[0008] Fourthly, according to the present disclosure, a device for predicting equipment status and remaining life is provided, comprising: The second acquisition module is used to acquire multi-sensor time-series data of the target device; The determination module is used to input the multi-sensor time-series data of the target device into the device status and remaining life prediction model to obtain the status and remaining life of the target device. The equipment status and remaining life prediction model is generated by the method described in the first aspect.

[0009] Fifthly, a computer device is provided, including a memory and a processor. The memory stores a computer program, and when the processor executes the computer program, it implements the steps of the device state and remaining lifetime prediction model training method as described in any of the above embodiments, or implements the steps of the device state and remaining lifetime prediction method as described in any of the above embodiments.

[0010] In a sixth aspect, a computer-readable storage medium is provided, on which a computer program is stored, wherein when the computer program is executed by a processor, it implements the steps of the device state and remaining lifetime prediction model training method as described in any of the above embodiments, or implements the steps of the device state and remaining lifetime prediction method as described in any of the above embodiments.

[0011] The device status and remaining life prediction model training method provided in this application embodiment is applied to a large language model, which includes a first branch, a second branch, a third branch, and a fourth branch. The first branch is used to extract the spatial and frequency domain features of the device, and the second branch is used to capture the forward and backward hidden states during the device's operation through bidirectional temporal modeling. The method includes: acquiring multi-sensor time-series data of the sample device; extracting features from the multi-sensor time-series data of the sample device through the first branch to obtain spatial-frequency fusion features; capturing bidirectional states from the spatial-frequency fusion features through the second branch to obtain a bidirectional hidden state sequence; associating the bidirectional hidden state sequence with historical states through the third branch to obtain an attention weight set representing different states; assigning attention weights to all states in the bidirectional hidden state sequence based on the attention weight set to update the weights of the bidirectional hidden state sequence; performing feature compression and nonlinear transformation on the updated bidirectional hidden state sequence through the fourth branch to predict the device's status and remaining life; adding crack propagation physical rules as a regularization term to a preset loss function; and using the preset loss function, training the large language model based on the device's status and remaining life to obtain the device status and remaining life prediction model. Thus, by performing bidirectional state capture on the extracted multi-scale features of the equipment, the forward and backward hidden states during the equipment operation process can be obtained; corresponding attention weights are assigned to different hidden states, and the crack propagation physical rules are added as regularization terms to the preset loss function to ensure that the predicted crack propagation rate conforms to physical laws, improve the model's generalization ability, and thus effectively realize equipment fault detection and service life prediction.

[0012] The above description is merely an overview of the technical solutions of the embodiments of this application. In order to better understand the technical means of the embodiments of this application and to implement them in accordance with the contents of the specification, and to make the above and other objects, features and advantages of the embodiments of this application more obvious and understandable, specific implementation methods of this application are described below. Attached Figure Description

[0013] To more clearly illustrate the technical solutions of the embodiments of this disclosure, the accompanying drawings of the embodiments will be briefly described below. It should be understood that the drawings described below only relate to some embodiments of this disclosure and are not intended to limit this disclosure, wherein: Figure 1 This is a flowchart illustrating a method for training a prediction model of equipment status and remaining life provided in this disclosure.

[0014] Figure 2A This is a schematic diagram of an LSTM structure provided in this disclosure.

[0015] Figure 2B This is a schematic diagram of the structure of a BiLSTM disclosed herein.

[0016] Figure 2C This is a schematic diagram of the structure of an attention mechanism provided in this disclosure.

[0017] Figure 3 This is a flowchart illustrating a method for predicting equipment status and remaining life provided in this disclosure.

[0018] Figure 4 This is a schematic diagram of the structure of a training device for predicting equipment status and remaining life provided in this disclosure.

[0019] Figure 5 This is a schematic diagram of the structure of a device for predicting equipment status and remaining life provided in this disclosure.

[0020] Figure 6 This is a schematic diagram of the structure of a computer device provided in this disclosure.

[0021] It should be noted that the elements in the attached diagram are schematic and not drawn to scale. Detailed Implementation

[0022] To make the objectives, technical solutions, and advantages of the embodiments of this disclosure clearer, the technical solutions of the embodiments of this disclosure will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of this disclosure. All other embodiments obtained by those skilled in the art based on the described embodiments of this disclosure without creative effort are also within the scope of protection of this disclosure.

[0023] Unless otherwise defined, all terms used herein (including technical and scientific terms) shall have the same meaning as commonly understood by one of ordinary skill in the art to which this subject matter pertains. It will be further understood that terms such as those defined in commonly used dictionaries shall be interpreted as having the meaning consistent with their meaning in the context of the specification and in the relevant art, and shall not be interpreted in an idealized or overly formal form unless otherwise explicitly defined herein. As used herein, the statement of “connecting” or “coupling” two or more parts together shall mean that these parts are directly joined together or joined through one or more intermediate components.

[0024] The term "embodiment" as used herein means that a particular feature, structure, or characteristic described in connection with an embodiment may be included in at least one embodiment of this application. The appearance of the phrase "embodiment" in various places throughout the specification does not necessarily refer to the same embodiment, nor is it a separate or alternative embodiment mutually exclusive with other embodiments. It will be explicitly and implicitly understood by those skilled in the art that the embodiments described herein can be combined with other embodiments.

[0025] In this document, the term "and / or" is merely a description of the relationship between related objects, indicating that three relationships can exist. For example, A and / or B can mean: A exists, A and B exist simultaneously, or B exists. Additionally, the character " / " generally indicates that the preceding and following related objects have an "or" relationship. Terms such as "first" and "second" are only used to distinguish one component (or part of a component) from another component (or another part of a component).

[0026] In the description of this application, unless otherwise stated, "multiple" means two or more (including two), and similarly, "multiple groups" means two or more (including two groups).

[0027] To enable those skilled in the art to better understand the present application, the technical solutions in the embodiments of the present application will be clearly and completely described below with reference to the accompanying drawings.

[0028] Figure 1 This is a flowchart illustrating a method for training a device status and remaining lifetime prediction model according to an embodiment of this disclosure. The method is applied to a large language model, which includes a first branch, a second branch, a third branch, and a fourth branch. The first branch extracts the spatial and frequency domain features of the device, and the second branch captures the forward and backward hidden states during device operation through bidirectional temporal modeling. The first branch is a Convolutional Neural Network (CNN) module; the second branch is a Bidirectional Long Short-Term Memory (BiLSTM) module; the third branch is a spatiotemporal attention mechanism module; and the fourth branch is the output layer.

[0029] like Figure 1 As shown, the specific process of training the equipment condition and remaining life prediction model includes: S110. Acquire multi-sensor time-series data of the sample device.

[0030] Among them, the multi-sensor time-series data of the sample device can be used as a pre-constructed seven-dimensional damage feature vector to characterize the degradation state of the sample device's tube. . Indicates the first t Peak temperature of the breech surface per cycle; Indicates the first t Variance of chamber pressure fluctuation after a second fire strike; Indicates the first t The cumulative explosive equivalent of each firing cycle; Indicates the deviation from the baseline trajectory (used to characterize structural integrity). Indicates the residual deviation coefficient of axial stress; Indicates the intensity of radial vibration in the barrel; This indicates the cumulative amount of creep in the tube material.

[0031] In some embodiments, acquiring multi-sensor time-series data of the sample device includes: collecting high-frequency time-series data through multiple types of sensors deployed on the sample device; and performing wavelet transform and feature standardization processing on the high-frequency time-series data to obtain multi-sensor time-series data of the sample device.

[0032] The high-frequency time-series data is obtained in real time from various sensors (vibration sensors, temperature sensors, pressure sensors, strain sensors, etc.) deployed on key parts of equipment (such as artillery barrels and recoil mechanisms). Wavelet transform is performed on the raw signals represented by the high-frequency time-series data to filter out high-frequency battlefield noise (such as dust interference and electromagnetic pulses), while retaining low-frequency trends and transient impact characteristics related to faults. Furthermore, the multi-source sensor data is normalized to eliminate dimensional differences and improve model convergence speed.

[0033] Furthermore, the intelligent design of health management systems for large-diameter equipment needs to fully adapt to the complexity and dynamism of the environment. The AT-CNN-BiLSTM model, by fusing convolutional neural networks, bidirectional long short-term memory networks, and attention mechanisms, can theoretically process multidimensional time-series data of equipment and predict faults. However, the operating conditions of the equipment (such as high noise, data sparsity, and real-time requirements) place higher demands on the model's robustness, computational efficiency, and generalization ability. Based on the characteristics of equipment operating conditions, this paper analyzes the adaptability bottlenecks of the AT-CNN-BiLSTM model, proposes targeted optimization strategies, and designs a virtual-real fusion verification method to ensure the model's applicability in complex environments.

[0034] While performing wavelet transform, it can also model operating conditions and perform spectral analysis on various types of sensor signals (such as vibration, chamber pressure, and temperature signals), facilitating the differentiation between high-frequency sudden impact signals and low-frequency progressive wear characteristics, and establishing battlefield noise interference source models. It provides deep learning models with "prior knowledge" that conforms to physical laws, improving the accuracy and robustness of fault detection.

[0035] For equipment data feature spectrum analysis, the data generated by equipment in complex environments exhibits significant suddenness, mechanical wear characteristics, and sparsity. Taking a certain type of equipment as an example, its key components exhibit multiple failure modes under high temperature, high pressure, and impact loads. Sudden failures, such as mechanical jamming or barrel ablation, are usually caused by transient overloads, resulting in highly abrupt data changes. Through spectral analysis of vibration signals, it was found that the high-frequency band (>1kHz) contains impact response characteristics, reflecting early signs of component ablation; mechanical wear: fatigue damage to components accumulates with the number of shots, manifesting as crack propagation and material property degradation, with the low-frequency band (<100Hz) reflecting this long-term trend; data sparsity: the equipment has a limited firing frequency (a single mission typically <100 shots), resulting in insufficient training data and a high risk of overfitting in the model.

[0036] For the noise interference source model in complex environments, dust, temperature fluctuations, and high-intensity impact loads in complex environments significantly reduce the signal-to-noise ratio (SNR) of sensor data. Taking dust cover as an example, its interference with optical and vibration sensor signals causes the SNR to decrease by about 20%-30%. To model the noise impact, this embodiment assumes that the dust interference follows a Gaussian distribution, and defines the signal-to-noise ratio as shown in the following formula (1).

[0037] (1) In formula (1), Indicates the effective signal power; This represents the noise power. Experiments showed that when the dust cover was 50%, the SNR dropped to 10dB, severely impacting the model's feature extraction capabilities. Furthermore, the coupling interference caused by the impact load complicated the correlation between multi-sensor data, further increasing the difficulty of fault detection. Therefore, a noise-resistant preprocessing module (such as wavelet denoising) needs to be introduced into the model input layer to filter out high-frequency noise and retain the low-frequency trends related to the fault.

[0038] S120. The first branch extracts features from the multi-sensor time-series data of the sample device to obtain spatial-frequency fusion features.

[0039] The CNN module is responsible for extracting multi-scale spatial-temporal features from raw signals from multiple sensors. Specifically, vibration, chamber pressure, and temperature signals, after being modeled for operational characteristics and pre-processed for noise reduction, are fed into the CNN module. The CNN extracts local feature maps reflecting sudden impact (high frequency) and progressive wear (low frequency) through a series of convolution and downsampling operations, and fuses the features from each layer at multiple scales to form a feature sequence for time-series modeling. The CNN module extracts high-frequency impact features (sudden failure) and low-frequency degradation features (progressive wear); extracts and compresses features step by step, removes redundant information, and retains key patterns; uses depthwise separable convolution to reduce computation and improve the real-time performance of edge devices; unfolds the feature maps into temporal vectors, which are used as input to a BiLSTM; the BiLSTM captures time dependencies, and the attention mechanism focuses on key moments, outputting fault diagnosis and remaining useful life (RUL) prediction.

[0040] To adapt to automotive / edge computing platforms, this embodiment prioritizes lightweight convolutional operations (e.g., depthwise separable convolution) and bottleneck structures to reduce the number of parameters and computational overhead. The output of the CNN module can be used as a feature embedding to complete bidirectional degradation pattern modeling and key temporal segment weighting, thereby achieving fault type identification and RUL prediction.

[0041] The basic building blocks of a CNN include an input layer and multiple consecutive convolutional and pooling layers. Convolutional layers perform convolution operations on the input data and the convolutional kernel to extract latent features from the data. The specific operation of a convolutional layer is shown in Equation (2).

[0042] (2) In formula (2), f W is the activation function; W is the weight matrix. b is the bias matrix; * represents the convolution operation.

[0043] The function of the pooling layer is to perform pooling on the output obtained from the convolutional layer, as shown in the following formula (3).

[0044] (3) In formula (3), d i Features extracted by the convolutional layer; m This represents the pooling width.

[0045] In some embodiments, feature extraction is performed on the multi-sensor time-series data of the sample device through the first branch to obtain spatial-frequency fusion features, including: performing multi-scale feature extraction on the multi-sensor time-series data of the sample device through depthwise separable convolution; capturing local transformation patterns in the multi-scale features through the spatial feature extraction unit to obtain spatial features; extracting features of different frequency components in the multi-scale features through the multi-channel convolution kernel in the frequency domain feature extraction unit to obtain frequency domain features; and fusing the spatial features and frequency domain features to obtain spatial-frequency fusion features.

[0046] Among them, depthwise separable convolution is used to extract multi-scale features from the input signal: the spatial feature extraction unit captures local abrupt change patterns (such as the impact waveform of gun barrel erosion and abnormal vibration frequency bands); the frequency domain feature extraction unit extracts features of different frequency components through multi-channel convolution kernels to identify high-frequency impacts and low-frequency wear; and max pooling preserves significant features and reduces computational complexity.

[0047] S130. The spatial-frequency fusion features are captured bidirectionally through the second branch to obtain a bidirectional hidden state sequence.

[0048] Among them, LSTM combines a forget gate and an input gate on the basis of traditional RNN, and its structure is as follows: Figure 2A As shown. The input gate is responsible for filtering and determining which new information at the current time step should be merged into the cell state. This filtering operation is performed based on the following mathematical formulas (4)-(9).

[0049] (4) (5) (6) (7) (8) (9) In formulas (4)-(9), W and b These are the parameters of the LSTM unit; This is the current cell state; It is a new unit Candidate state values; , and These are the forget gate, input gate, and output gate, represented by the Sigmoid function, and their calculations are shown in equations (4), (5), and (6), respectively. The activity of these gating mechanisms depends on the current input. and previous output A strobe value of zero indicates that information transmission is blocked. Specifically, the forget gate... Responsible for deciding whether to inherit from the previous state The data is used to determine the state through formula (8). Updated to status At the same time, formula (7) generates the potential state value. And formula (9) calculates the current output of each LSTM unit. Subsequently, and They are sequentially sent to the next unit, and the cycle repeats in the time series.

[0050] The structure of the BiLSTM module is as follows: Figure 2B As shown, it consists of two LSTM networks, one forward and one backward, which can make bidirectional predictions by utilizing the backward and forward changes of the time series, making the predictions more global and holistic.

[0051] In some embodiments, bidirectional state capture is performed on the space-frequency fusion features through the second branch to obtain a bidirectional hidden state sequence, including: capturing the performance degradation trend of the sample device during historical operation from the space-frequency fusion features using a forward LSTM to obtain a forward hidden state; identifying the hysteresis response and abnormal mutation of the sample device during historical operation from the space-frequency fusion features using a backward LSTM to obtain a backward hidden state; and concatenating the forward hidden state and the backward hidden state to obtain a bidirectional hidden state sequence.

[0052] The forward LSTM models from historical data to the current moment, capturing the performance degradation trend caused by the accumulation of firing counts (such as crack propagation and material fatigue), and recursively calculates the cumulative amount of barrel ablation. Backward LSTM inversion from the past to the present time can identify hysteresis responses and anomalous abrupt changes (such as the delayed effect of bore erosion) and invert crack propagation paths. The forward hidden state \( H_t \) and the backward hidden state \( \overline{H}_t \) are concatenated to form a bidirectional state matrix for comprehensively characterizing the temporal degradation process. That is, a two-way hidden state sequence.

[0053] S140. By associating the bidirectional hidden state sequence with historical states through the third branch, we obtain the attention weight set used to represent different states.

[0054] Attention is a key algorithmic mechanism that mimics human cognitive processes by flexibly assigning different weights to multiple hidden states in a neural network. This allows the model to focus on the most relevant input features, which contribute to the final prediction. By selectively emphasizing certain parts of the input sequence over others, attention enhances the model's ability to capture complex temporal relationships.

[0055] Intelligent screening process such as Figure 2C As shown. In t At time 1, obtain the hidden state of the BiLSTM layer. The correlation between the current hidden state and the historical state is calculated using formula (10). α t,i , α t,i It is a hidden layer state. In time t The attention weight is the weight given to the current output.

[0056] (10) In formula (10), .

[0057] In the AT-CNN-BiLSTM model for health management of large-diameter equipment, the attention mechanism effectively improves the accuracy of remaining service life prediction by dynamically focusing on key time segments in the evolution of fatigue damage in the tube and employing an adaptive weighting strategy based on damage sensitivity. This mechanism generates a weight distribution that conforms to the equipment degradation pattern by quantifying the cumulative impact of each cycle on structural integrity. This allows the model to collaboratively capture the instantaneous fluctuation characteristics of impact loads and the long-term trend evolution of material crack propagation, which is crucial for accurate equipment condition assessment and preventative maintenance decisions in complex environments. The attention mechanism effectively improves the accuracy of remaining service life prediction by dynamically focusing on key time segments in the evolution of fatigue damage in the tube and employing an adaptive weighting strategy based on damage sensitivity.

[0058] S150. Assign attention weights to all states in the bidirectional hidden state sequence based on the attention weight set, so as to update the weights of the bidirectional hidden state sequence.

[0059] Specifically, high weights are assigned to time segments most relevant to the current fault (such as abnormal vibration peaks after a shot) to suppress irrelevant noise. Simultaneously, attention heatmaps can be generated for visualization, aiding in locating damage-sensitive periods and sensor channels.

[0060] S160. The updated bidirectional hidden state sequence is subjected to feature compression and nonlinear transformation through the fourth branch to predict the state and remaining lifespan of the sample device.

[0061] In some embodiments, the updated bidirectional hidden state sequence is subjected to feature compression and nonlinear transformation through a fourth branch to predict the state and remaining lifespan of the sample device. This includes: performing feature compression and nonlinear transformation on the updated bidirectional hidden state sequence through a fully connected layer to obtain nonlinear transformation features; performing binary classification on the nonlinear transformation features to obtain the state of the sample device; and performing regression prediction on the nonlinear transformation features to obtain the remaining lifespan of the sample device.

[0062] The attention-weighted BiLSTM output is input into a fully connected layer for feature compression and nonlinear transformation, used for fault classification and regression prediction. The nonlinear transformation features are used for binary classification to determine if the current state is faulty (e.g., whether a barrel crack exceeds a threshold). Regression prediction is then performed on the nonlinear transformation features, outputting the remaining service life (e.g., the number of remaining shots). This is combined with a physical failure model for confidence calibration. A physical failure model is a mathematical model based on physical mechanisms used to describe the failure process of materials, components, or systems under specific operating conditions (e.g., force, heat, chemical effects). It quantifies the evolution of failure behavior by analyzing physical laws (e.g., mechanics, heat, fatigue, corrosion), and is typically used to predict the remaining service life (RUL) or probability of failure of equipment or structures. The core of a physical failure model is to describe the degradation process of materials or components under specific loads or environmental conditions using mathematical formulas.

[0063] S170. The crack propagation physical rules are added as regularization terms to the preset loss function; and the preset loss function is used to train the large language model based on the state and remaining service life of the sample equipment to obtain the equipment state and remaining service life prediction model.

[0064] To address the issues of sparse artillery firing data and insufficient small sample sizes, this embodiment introduces a deep transfer learning method. This method aligns the feature distributions of simulated and measured data and uses physical failure models (such as the Paris crack propagation formula) as regularization constraints to ensure that the prediction results conform to the laws of mechanical failure. This strategy not only significantly reduces the risk of overfitting but also enhances the model's generalization ability across different operating conditions, enabling fault prediction to maintain high accuracy even with limited training samples.

[0065] To improve the model's ability to generalize to mechanical failures of equipment, a physical failure model is introduced as a constraint. Crack propagation in tubular devices follows the Paris formula, as shown in formula (11) below.

[0066] (11) In formula (11),a The length of the crack; N This represents the number of firing cycles; The stress intensity factor is used to reflect the degree of stress concentration at the crack tip and is usually determined by material properties and load. C and m This is a material constant that reflects the material's sensitivity to fatigue crack propagation; This is the crack propagation rate, which is the length of crack growth per cycle.

[0067] In this embodiment, the Paris formula is transformed into a regularization term of a preset loss function, as shown in the following formula (12).

[0068] (12) In formula (12), This is the mean square error loss; This is a physical constraint term based on the Paris formula, ensuring that the predicted crack propagation rate conforms to physical laws; The weighting coefficient is set to 0.1. Experiments show that after adding physical constraints, the model's MAE for crack propagation prediction decreased from 12% to 8%, significantly improving its generalization ability. Furthermore, using a GAN with physical constraints to generate synthetic fault data compensates for the sparsity of battlefield data, further enhancing the model's generalization ability.

[0069] This embodiment applies the method to a large language model, which includes a first branch, a second branch, a third branch, and a fourth branch. The first branch is used to extract the spatial and frequency domain features of the device, and the second branch is used to capture the forward and backward hidden states during the device's operation through bidirectional temporal modeling. The method includes: acquiring multi-sensor time-series data of the sample device; extracting features from the multi-sensor time-series data of the sample device through the first branch to obtain spatial-frequency fusion features; capturing bidirectional states from the spatial-frequency fusion features through the second branch to obtain a bidirectional hidden state sequence; associating historical states with the bidirectional hidden state sequence through the third branch to obtain an attention weight set representing different states; assigning attention weights to all states in the bidirectional hidden state sequence based on the attention weight set to update the weights of the bidirectional hidden state sequence; performing feature compression and nonlinear transformation on the updated bidirectional hidden state sequence through the fourth branch to predict the state and remaining lifespan of the sample device; incorporating crack propagation physical rules as a regularization term into a preset loss function; and using the preset loss function, training the large language model based on the state and remaining lifespan of the sample device to obtain a device state and remaining lifespan prediction model. Thus, by performing bidirectional state capture on the extracted multi-scale features of the equipment, the forward and backward hidden states during the equipment operation process can be obtained; corresponding attention weights are assigned to different hidden states, and the crack propagation physical rules are added as regularization terms to the preset loss function to ensure that the predicted crack propagation rate conforms to physical laws, improve the model's generalization ability, and thus effectively realize equipment fault detection and service life prediction.

[0070] In addition, a dynamic Dropout strategy can be introduced to deal with the uncertainty of data in complex environments, as shown in the following formula (13).

[0071] (13) By adjusting the network robustness through sensitivity to firing load perturbations and adaptively adjusting the Dropout rate based on the degree of data perturbation, the model's robustness to noise and missing data can be improved, enhancing its adaptability to complex environments. Aligning the feature distributions of simulated and measured data ensures the model's generalization ability across different battlefield environments.

[0072] The efficiency decay coefficient is defined as shown in the following formula (14).

[0073] (14) Damage effectiveness assessment matrix This provides a basis for assessing the sustainable capabilities of intelligent fire control systems.

[0074] In summary, this embodiment organically integrates convolutional neural networks, bidirectional long short-term memory networks, and dynamic spatiotemporal attention mechanisms to achieve joint representation learning of high-dimensional sensor data from devices. The CNN layer optimizes parameter efficiency through depthwise separable convolutions, extracting local abrupt change features of multi-source signals such as vibration and temperature (e.g., local gradient patterns in the ablation impact waveform of a tubular cavity). An improved BiLSTM layer combines physical constraints (Paris crack propagation formula) with bidirectional degradation path modeling to simultaneously capture the cumulative effect of firing counts (forward propagation) and the material fatigue hysteresis response (backward propagation). A spatiotemporal weight allocation mechanism based on damage effectiveness decay coefficients (spatiotemporal joint attention gate) is introduced to dynamically focus on sensor hotspot regions in key cycles (e.g., abnormal frequency bands of vibration in tubular devices). A physical law database of fatigue crack propagation in tubular devices is constructed through finite element simulation, and this database is collaboratively used with a GAN (Generative Adversarial Network) to generate synthetic fault data with physical labels. The Paris crack propagation formula is used as a constraint condition for the GAN generator to ensure that the generated data conforms to the exponential growth law of mechanical failure. Domain adaptation technology is employed to align the feature distributions of simulated and measured data, mitigating cross-scenario generalization errors. By analyzing equipment operating characteristics, including sudden failures, mechanical wear, and data sparsity, bottlenecks in the model's real-time performance, robustness, and small-sample adaptability were identified. To address these issues, optimization strategies including lightweight design, noise reduction enhancement, and physical constraint embedding were proposed, significantly improving the model's real-time inference capabilities on edge devices, fault detection performance in high-noise environments, and the accuracy of modeling mechanical failure patterns. Data augmentation based on finite element simulation and semi-physical testing validated the applicability of the optimized model in complex environments, providing a reliable intelligent solution for equipment health management. The model overcomes the feature coupling modeling bottleneck of traditional neural networks in equipment health management. A network architecture composed of a spatiotemporal attention convolutional-bidirectional long short-term memory network (AT-CNN-BiLSTM) can simultaneously analyze the high-frequency transient impact response and low-frequency gradual damage evolution patterns of equipment through a collaborative design of channel attention weighting and bidirectional temporal correlation feature extraction, enabling fault detection and service life prediction.

[0075] Figure 3 This is a flowchart illustrating a method for predicting equipment status and remaining lifespan according to an embodiment of this disclosure. Figure 3 As shown, the specific process of the equipment condition and remaining life prediction method includes: S310: Acquire multi-sensor time-series data of the target device.

[0076] Among them, the multi-sensor time-series data of the target device can be such as a pre-constructed seven-dimensional damage feature vector for characterizing the degradation state of the target device's tube ( . Indicates the first tPeak temperature of the breech surface per cycle; Indicates the first t Variance of chamber pressure fluctuation after a second fire strike; Indicates the first t The cumulative explosive equivalent of each firing cycle; Indicates the deviation from the baseline trajectory (used to characterize structural integrity). Indicates the residual deviation coefficient of axial stress; Indicates the intensity of radial vibration in the barrel; This indicates the cumulative amount of creep in the tube material.

[0077] Specifically, high-frequency time-series data can be collected by multiple types of sensors deployed on the target device; wavelet transform and feature standardization processing are performed on the high-frequency time-series data to obtain multi-sensor time-series data of the target device.

[0078] S320. Input the multi-sensor time-series data of the target device into the device status and remaining life prediction model to obtain the target device's status and remaining life.

[0079] The equipment status and remaining lifetime prediction model is trained and generated by any of the methods in the above embodiments. The equipment status and remaining lifetime prediction model includes: a first branch, a second branch, a third branch, and a fourth branch; the first branch is used to extract the spatial and frequency domain features of the equipment, and the second branch is used to capture the forward and backward hidden states during equipment operation through bidirectional temporal modeling. The first branch is a convolutional neural network module; the second branch is a bidirectional long short-term memory network module; the third branch is a spatiotemporal attention mechanism module; and the fourth branch is an output layer.

[0080] By inputting multi-sensor time-series data of the target device into the first branch of the device state and remaining lifetime prediction model, spatial-frequency fusion features are obtained. These features are then input into the second branch of the model to obtain a bidirectional hidden state sequence. This sequence is further input into the third branch to obtain attention weight sets representing different states. Attention weights are assigned to all states in the bidirectional hidden state sequence based on these weight sets, updating the sequence's weights. Finally, the updated sequence is input into the fourth branch to predict the target device's state and remaining lifetime. This effectively improves the prediction efficiency of device fault states and remaining lifetime.

[0081] In some embodiments, the method further includes: determining the fault damage level of the target device based on its current status and remaining useful life; and generating a health status assessment and equipment maintenance recommendations for the target device based on the fault damage level.

[0082] Specifically, the remaining service life of the target equipment can be compared with each preset remaining service life threshold range to determine the corresponding fault damage level of the target equipment; or, the corresponding fault damage level of the target equipment can be determined based on a preset correspondence between the current status and the fault damage level. Based on the fault damage level, a corresponding health status assessment template is matched from a preset health status assessment template library to generate a health status assessment for the target equipment. Simultaneously, a corresponding equipment maintenance suggestion template is matched from a preset equipment maintenance suggestion template library, and personalized equipment maintenance suggestions, such as "recommend replacing the gun barrel," are generated based on the actual operating conditions of the target equipment. This provides real-time health status assessment and maintenance suggestions for the fire control system.

[0083] Figure 4 This embodiment provides a training device for a device status and remaining lifetime prediction model, applied to a large language model. The large language model includes a first branch, a second branch, a third branch, and a fourth branch. The first branch is used to extract the spatial and frequency domain features of the device, and the second branch is used to capture the forward and backward hidden states during device operation through bidirectional temporal modeling. The device for training the device status and remaining lifetime prediction model includes: The first acquisition module 410 is used to acquire multi-sensor time-series data of the sample device.

[0084] The extraction module 420 is used to extract features from the multi-sensor time-series data of the sample device through the first branch to obtain spatial-frequency fusion features.

[0085] The capture module 430 is used to capture the spatial-frequency fusion features bidirectionally through the second branch to obtain a bidirectional hidden state sequence.

[0086] The association module 440 is used to associate the bidirectional hidden state sequence with historical states through the third branch to obtain the attention weight set used to represent different states.

[0087] The allocation module 450 is used to assign attention weights to all states in the bidirectional hidden state sequence based on the attention weight set, so as to update the weights of the bidirectional hidden state sequence.

[0088] The prediction module 460 is used to perform feature compression and nonlinear transformation on the updated bidirectional hidden state sequence through the fourth branch to predict the state and remaining lifespan of the sample device.

[0089] Training module 470 is used to add the crack propagation physical rules as regularization terms to the preset loss function; and using the preset loss function, the large language model is trained based on the state and remaining service life of the sample equipment to obtain the equipment state and remaining service life prediction model.

[0090] In this embodiment, optionally, the extraction module 420 is specifically used for: Multi-scale feature extraction is performed on multi-sensor time-series data of sample devices through depthwise separable convolution; spatial features are obtained by capturing local transformation patterns in multi-scale features through spatial feature extraction unit; and frequency domain features are obtained by extracting features of different frequency components in multi-scale features through multi-channel convolution kernel in frequency domain feature extraction unit; spatial and frequency domain features are fused to obtain spatial-frequency fusion features.

[0091] In this embodiment, optionally, the capture module 430 is specifically used for: The forward hidden state is obtained by capturing the performance degradation trend of the sample device during historical operation from the space-frequency fusion features using forward LSTM; the backward hidden state is obtained by identifying the hysteresis response and abnormal mutation of the sample device during historical operation from the space-frequency fusion features using backward LSTM; the forward hidden state and the backward hidden state are spliced ​​together to obtain the bidirectional hidden state sequence.

[0092] In this embodiment, optionally, the prediction module 460 is specifically used for: The updated bidirectional hidden state sequence is subjected to feature compression and nonlinear transformation through a fully connected layer to obtain nonlinear transformation features; binary classification is performed on the nonlinear transformation features to determine the state of the sample device; regression prediction is performed on the nonlinear transformation features to obtain the remaining lifespan of the sample device.

[0093] In this embodiment, optionally, the first acquisition module 410 is specifically used for: High-frequency time-series data are collected by multiple sensors deployed on the sample device; wavelet transform and feature standardization are performed on the high-frequency time-series data to obtain multi-sensor time-series data of the sample device.

[0094] The device for training the equipment status and remaining life prediction model provided in this disclosure can execute the above-described method embodiments. For the specific implementation principle and technical effect, please refer to the above-described method embodiments, which will not be repeated here.

[0095] Figure 5 This embodiment provides a device for predicting equipment status and remaining life. The device includes: The second acquisition module 510 is used to acquire multi-sensor time-series data of the target device.

[0096] The determination module 520 is used to input the multi-sensor time-series data of the target device into the device status and remaining life prediction model to obtain the status and remaining life of the target device.

[0097] The equipment status and remaining life prediction model is generated by any of the methods described in the above embodiments.

[0098] In this embodiment, optionally, a generation module is also included.

[0099] The generation module is used to determine the fault level of the target device based on its current status and remaining service life; and to generate a health status assessment and maintenance recommendations for the target device based on the fault level.

[0100] The device for predicting device status and remaining life provided in this disclosure can execute the above-described method embodiments. For its specific implementation principle and technical effects, please refer to the above-described method embodiments. This disclosure will not repeat them here.

[0101] This application also provides a computer device. Please refer to the following for details. Figure 6 , Figure 6 This is a basic structural block diagram of the computer device in this embodiment.

[0102] The computer device includes a memory 610 and a processor 620 that are interconnected via a system bus. It should be noted that only a computer device with memory 610 and processor 620 is shown in the figure; however, it should be understood that it is not required to implement all the components shown, and more or fewer components may be implemented alternatively. Those skilled in the art will understand that the computer device described herein is a device capable of automatically performing numerical calculations and / or information processing according to pre-set or stored instructions, and its hardware includes, but is not limited to, microprocessors, application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), digital signal processors (DSPs), embedded devices, etc.

[0103] Computer devices can include desktop computers, laptops, handheld computers, and cloud servers. These devices allow for human-computer interaction with users through methods such as keyboards, mice, remote controls, touchpads, or voice-activated devices.

[0104] The memory 610 includes at least one type of readable storage medium, including non-volatile memory or volatile memory, such as flash memory, hard disk, multimedia card, card-type memory (e.g., SD or DX memory), random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM), electrically erasable programmable read-only memory (EEPROM), programmable read-only memory (PROM), magnetic memory, magnetic disk, optical disk, etc. RAM may include static RAM or dynamic RAM. In some embodiments, the memory 610 may be an internal storage unit of a computer device, such as the hard disk or memory of the computer device. In other embodiments, the memory 610 may also be an external storage device of a computer device, such as a plug-in hard disk, smart media card (SMC), secure digital (SD) card, or flash card equipped on the computer device. Of course, the memory 610 may include both internal storage units and external storage devices of the computer device. In this embodiment, the memory 610 is typically used to store the operating system and various application software installed on the computer device, such as the program code of the method described above. In addition, the memory 610 can also be used to temporarily store various types of data that have been output or will be output.

[0105] The processor 620 is typically used to perform the overall operation of a computer device. In this embodiment, the memory 610 is used to store program code or instructions, including computer operation instructions. The processor 620 is used to execute the program code or instructions stored in the memory 610 or to process data, such as program code that runs the methods described above.

[0106] In this article, the bus can be an Industry Standard Architecture (ISA) bus, a Peripheral Component Interconnect (PCI) bus, or an Extended Industry Standard Architecture (EISA) bus, etc. This bus system can be divided into address bus, data bus, control bus, etc. For ease of illustration, only one thick line is used to represent it in the diagram, but this does not mean that there is only one bus or one type of bus.

[0107] Another embodiment of this application also provides a computer-readable medium, which may be a computer-readable signal medium or a computer-readable medium. A processor in a computer reads computer-readable program code stored in the computer-readable medium, enabling the processor to execute the functional actions specified in each step or combination of steps in the above method; and to generate means for implementing the functional actions specified in each block or combination of blocks in the block diagram.

[0108] Computer-readable media include, but are not limited to, electronic, magnetic, optical, electromagnetic, infrared memory or semiconductor systems, devices or apparatuses, or any suitable combination thereof, wherein the memory is used to store program code or instructions, the program code including computer operation instructions, and the processor is used to execute the program code or instructions of the above-described methods stored in the memory.

[0109] The definitions of memory and processor can be found in the description of the foregoing computer device embodiments, and will not be repeated here.

[0110] In the several embodiments provided in this application, it should be understood that the disclosed systems, apparatuses, and methods can be implemented in other ways. For example, the apparatus embodiments described above are merely illustrative; for instance, the division of modules or units is only a logical functional division, and in actual implementation, there may be other division methods. For example, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed. Furthermore, the coupling or direct coupling or communication connection shown or discussed may be through some interfaces; the indirect coupling or communication connection between apparatuses or units may be electrical, mechanical, or other forms.

[0111] In the various embodiments of this application, the functional units or modules can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit. The integrated unit can be implemented in hardware or as a software functional unit.

[0112] If the integrated unit is implemented as a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of this application, in essence, or the part that contributes to the prior art, or all or part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) or processor to execute all or part of the steps of the methods of the various embodiments of this application. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0113] In the claims, any reference signs placed between parentheses should not be construed as limiting the claims. The word "comprising" as described in this application does not exclude the presence of elements or steps not listed in the claims. The word "a" or "an" preceding an element does not exclude the presence of a plurality of such elements. This application can be implemented by means of hardware comprising several different elements and by means of a suitably programmed computer. In the unit claims listing several means, several units of these means may be embodied by the same item of hardware. The use of "first," "second," and "third," etc., does not indicate any order and these words should be interpreted as names. Unless otherwise specified, the steps in the above embodiments should not be construed as limiting the order of execution.

[0114] The above embodiments are only used to illustrate the technical solutions of this application, and are not intended to limit them. Although this application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of this application.

Claims

1. A method for training a prediction model of equipment condition and remaining useful life, characterized in that, It is applied to a large language model, which includes a first branch, a second branch, a third branch, and a fourth branch; the first branch is used to extract the spatial and frequency domain features of the device, and the second branch is used to capture the forward and backward hidden states of the device during operation through bidirectional temporal modeling. The method includes: Acquire multi-sensor time-series data from the sample device; The first branch is used to extract features from the multi-sensor time-series data of the sample device to obtain spatial-frequency fusion features. The second branch is used to perform bidirectional state capture on the space-frequency fusion features to obtain a bidirectional hidden state sequence; By associating the bidirectional hidden state sequence with historical states through the third branch, an attention weight set for representing different states is obtained. Attention weights are assigned to all states in the bidirectional hidden state sequence based on the attention weight set, so as to update the weights of the bidirectional hidden state sequence. The updated bidirectional hidden state sequence is subjected to feature compression and nonlinear transformation through the fourth branch to predict the state and remaining lifespan of the sample device. The crack propagation physical rules are added as regularization terms to the preset loss function; and the preset loss function is used to train the large language model based on the state and remaining service life of the sample equipment to obtain the equipment state and remaining service life prediction model.

2. The method according to claim 1, characterized in that, The step of extracting features from the multi-sensor time-series data of the sample device through the first branch to obtain space-frequency fusion features includes: Multi-scale feature extraction is performed on the multi-sensor time-series data of the sample device using depthwise separable convolution. The spatial features are obtained by capturing local transformation patterns in the multi-scale features through the spatial feature extraction unit; and the frequency domain features are obtained by extracting features of different frequency components in the multi-scale features through the multi-channel convolution kernel in the frequency domain feature extraction unit. The spatial features and the frequency domain features are fused to obtain the spatial-frequency fusion features.

3. The method according to claim 1, characterized in that, The step of performing bidirectional state capture on the space-frequency fusion features through the second branch to obtain a bidirectional hidden state sequence includes: The forward hidden state is obtained by capturing the performance degradation trend of the sample device during historical operation from the space-frequency fusion features using forward LSTM. The backward hidden state is obtained by identifying the hysteresis response and anomalous mutation of the sample device in the historical operation process from the space-frequency fusion features using backward LSTM. By concatenating the forward hidden state and the backward hidden state, the bidirectional hidden state sequence is obtained.

4. The method according to claim 1, characterized in that, The step of performing feature compression and nonlinear transformation on the updated bidirectional hidden state sequence through the fourth branch to predict the state and remaining lifespan of the sample device includes: The updated bidirectional hidden state sequence is subjected to feature compression and nonlinear transformation through a fully connected layer to obtain nonlinear transformation features. The nonlinear transformation features are classified into two categories to determine the state of the sample device. The remaining service life of the sample device is obtained by performing regression prediction on the nonlinear transformation characteristics.

5. The method according to claim 1, characterized in that, The acquisition of multi-sensor time-series data from the sample device includes: High-frequency time-series data are collected using multiple types of sensors deployed on the sample device; Wavelet transform and feature normalization are performed on the high-frequency time series data to obtain the multi-sensor time series data of the sample device.

6. A method for predicting equipment condition and remaining life, characterized in that, include: Acquire multi-sensor time-series data of the target device; The multi-sensor time-series data of the target device are input into the device status and remaining life prediction model to obtain the status and remaining life of the target device. The equipment status and remaining life prediction model is generated by the method described in any one of claims 1-5.

7. The method according to claim 6, characterized in that, Also includes: Based on the status and remaining service life of the target equipment, the fault damage level of the target equipment is determined; Based on the fault damage level of the target equipment, a health status assessment and equipment maintenance recommendations are generated for the target equipment.

8. A training device for a prediction model of equipment condition and remaining useful life, characterized in that, It is applied to a large language model, which includes a first branch, a second branch, a third branch, and a fourth branch; the first branch is used to extract the spatial and frequency domain features of the device, and the second branch is used to capture the forward and backward hidden states of the device during operation through bidirectional temporal modeling. The device includes: The first acquisition module is used to acquire multi-sensor time-series data of the sample device; The extraction module is used to extract features from the multi-sensor time-series data of the sample device through the first branch to obtain spatial-frequency fusion features; The capture module is used to capture the spatial-frequency fusion features bidirectionally through the second branch to obtain a bidirectional hidden state sequence; The association module is used to associate the bidirectional hidden state sequence with historical states through the third branch to obtain an attention weight set for representing different states. The allocation module is used to allocate attention weights to all states in the bidirectional hidden state sequence based on the attention weight set, so as to update the weights of the bidirectional hidden state sequence. The prediction module is used to perform feature compression and nonlinear transformation on the updated bidirectional hidden state sequence through the fourth branch to predict the state and remaining lifespan of the sample device. The training module is used to incorporate the crack propagation physical rules as regularization terms into a preset loss function; and using the preset loss function, the large language model is trained based on the state and remaining service life of the sample equipment to obtain a prediction model for equipment state and remaining service life.

9. A device for predicting equipment status and remaining life, characterized in that, include: The second acquisition module is used to acquire multi-sensor time-series data of the target device; The determination module is used to input the multi-sensor time-series data of the target device into the device status and remaining life prediction model to obtain the status and remaining life of the target device. The equipment status and remaining life prediction model is generated by the method described in any one of claims 1-5.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by the processor, it implements the equipment status and remaining lifetime prediction model training method as described in any one of claims 1 to 5, or implements the equipment status and remaining lifetime prediction method as described in claim 6 or 7.

Images