Industrial equipment residual life prediction method, system, equipment and medium

Abstract

The invention provides an industrial equipment residual life prediction method and system, equipment and a medium, and relates to the technical field of artificial intelligence, and the method comprises the steps: carrying out the feature extraction of a preprocessed historical time sequence, so as to obtain the extracted historical features; performing joint training based on the extracted historical features to obtain a trained model; the method comprises the following steps: acquiring a real-time multi-source sensor time sequence signal of industrial equipment in an operation process, and carrying out preprocessing operation on the real-time multi-source sensor time sequence signal to form a preprocessed real-time time sequence; performing feature extraction on the preprocessed real-time time sequence to obtain corresponding real-time features; and inputting the real-time features into a trained model for processing to obtain a residual life prediction result of the industrial equipment. According to the method, the problem of high missing of sensor time sequence data under severe working conditions can be solved, and end-to-end joint learning of data completion and RUL prediction is realized.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of artificial intelligence, in particular to an industrial equipment residual life prediction method, system, device and medium. BACKGROUND

[0002] With the improvement of the intelligent level of core precision industrial production equipment such as numerical control machine tools, the equipment usually integrates a large number of sensors inside to realize real-time state monitoring; the multi-source time series data collected by the sensors has high dimensionality, nonlinear topological relationship, and complex coupling relationship and strong correlation between variables; and in actual industrial data collection, due to sensor hardware failure, communication link interruption or extreme working condition interference, the obtained equipment operation data has different degrees of missing; such missing seriously restricts the traditional deep learning model which depends on complete time series data; how to effectively repair and extract features from incomplete industrial time series data is the key to the multi-task spatio-temporal residual shrinkage network prediction task.

[0003] There are mainly two kinds of existing research methods: 1. Two-stage method, which interpolates or reconstructs the missing data, and then inputs the completed sequence into the prediction model; this method usually completes data completion based on linear interpolation, traditional machine learning or deep learning, and then uses long-term memory network (LSTM), time convolution (TCN) and other models for life prediction; although this paradigm process is clear, the interpolation stage and the prediction task are independent of each other, the interpolation error will be amplified and accumulated to the prediction stage, resulting in a loss of final accuracy.

[0004] 2. End-to-end method, which integrates data processing and RUL prediction into a unified deep learning architecture, directly models the degradation on the missing sequence by introducing missing mask and time decay mechanism; the method alleviates the error transmission problem to some extent, but in the face of the feature selection and insufficient deep feature extraction under the influence of the pseudo-signal caused by data missing, the prediction accuracy is still difficult to be substantially improved.

[0005] Although the existing research methods have achieved certain results in this research field, in the industrial production environment with high missing rate and complex working conditions, the existing models still have deficiencies in stability and feature discrimination ability. SUMMARY

[0006] The technical problem to be solved by the present application is to provide an industrial equipment residual life prediction method which can cope with the high missing problem of sensor time series data under harsh working conditions and realize end-to-end joint learning of data completion and RUL prediction.

[0007] To solve the above technical problems, the technical solutions of the present application are as follows: In a first aspect, an industrial equipment residual life prediction method is provided, and the method comprises the following steps: obtaining historical multi-source sensor time series signals of the industrial equipment during operation, and performing preprocessing operations on the signals to form preprocessed historical time series; extracting features from the preprocessed historical time series to obtain extracted historical features; based on the extracted historical features, performing joint training to obtain a trained model; obtaining real-time multi-source sensor time series signals of the industrial equipment during operation, and performing preprocessing operations on the real-time multi-source sensor time series signals to form preprocessed real-time time series; extracting features from the preprocessed real-time time series to obtain corresponding real-time features; inputting the real-time features into the trained model for processing to obtain a residual life prediction result of the industrial equipment.

[0008] In a second aspect, an industrial equipment residual life prediction system is provided, and the system performs the method described above, and comprises the following steps: an obtaining module configured to obtain historical multi-source sensor time series signals of the industrial equipment during operation, and perform preprocessing operations on the signals to form preprocessed historical time series; an extracting module configured to extract features from the preprocessed historical time series to obtain extracted historical features; a training module configured to perform joint training based on the extracted historical features to obtain a trained model; a prediction module configured to obtain real-time multi-source sensor time series signals of the industrial equipment during operation, and perform preprocessing operations on the real-time multi-source sensor time series signals to form preprocessed real-time time series; extract features from the preprocessed real-time time series to obtain corresponding real-time features; input the real-time features into the trained model for processing to obtain a residual life prediction result of the industrial equipment.

[0009] In a third aspect, a computing device is provided, which comprises a memory and a processor. One or more computer programs are stored in the memory, and the one or more computer programs comprise instructions; when the instructions are executed by the processor, the computing device performs the method of the first aspect.

[0010] In a fourth aspect, a computer readable storage medium is provided, which is used to store a computer program, and the computer program is used to perform the method of the first aspect.

[0011] The above-mentioned scheme of the present application at least has the following beneficial effects: By deeply coupling feature extraction and prediction tasks through joint training, the independent optimization of the completion process and the prediction task in the traditional two-stage method is avoided; for example, in the case of high missing rate of data, the interpolation error of the traditional method will be amplified and passed to the prediction stage, while joint training can dynamically adjust the feature extraction process, so that the model directly learns the feature representation suitable for incomplete data, thereby reducing the risk of error accumulation.

[0012] By seamlessly connecting preprocessing, feature extraction and prediction in a unified feature space, the false signal interference caused by the introduction of missing mask or time decay mechanism is avoided; for example, when dealing with nonlinear data under complex working conditions, joint training can directly optimize the feature extraction layer through gradient back propagation, ensuring that the model focuses on real degradation features rather than false patterns caused by data missing.

[0013] By data cleaning, standardization and time alignment to improve data quality, and by using residual shrinkage network combined with attention mechanism in the feature extraction stage, key degradation features can be automatically selected and noise can be suppressed. For example, in gearbox fault prediction, this method can effectively capture the high-frequency impact features of vibration signals, while filtering the interference caused by load fluctuation or sensor drift.

[0014] By learning the long-term dependence of time series data through the spatiotemporal residual module, and dynamically adjusting the feature weight through the soft threshold function. For example, in bearing degradation prediction, this method can identify the cooperative change trend of multiple source signals such as temperature and vibration, avoiding prediction failure caused by missing of single sensor data; by directly processing incomplete sequences rather than explicit completion, the strong assumption of missing patterns by traditional interpolation methods is avoided. BRIEF DESCRIPTION OF DRAWINGS

[0015] Figure 1 Structure diagram of multi-task spatiotemporal residual shrinkage network; Figure 2 Structure diagram of residual shrinkage construction unit; Figure 3 Structure diagram of multi-task collaborative optimization; Figure 4 Flowchart of industrial equipment remaining life prediction method; Figure 5 Schematic diagram of a computing device. DETAILED DESCRIPTION

[0016] Exemplary embodiments of the present disclosure will be described in greater detail below with reference to the accompanying drawings. Although exemplary embodiments of the present disclosure are shown in the drawings, it should be understood that the present disclosure can be implemented in various forms and should not be limited by the embodiments set forth herein. On the contrary, these embodiments are provided so that the present disclosure can be more thoroughly understood and the scope of the present disclosure can be accurately conveyed to those skilled in the art.

[0017] As Figure 4 shown, the embodiment of the application proposes an industrial equipment residual life prediction method, the method comprises: Step 1, obtaining the historical multi-source sensor time series signal of the industrial equipment in the running process, and performing preprocessing operation on the signal to form the preprocessed historical time series; Step 2, feature extraction is performed on the preprocessed historical time series to obtain the extracted historical features; Step 3, based on the extracted historical features, joint training is performed to obtain a trained model; Step 4, obtaining the real-time multi-source sensor time series signal of the industrial equipment in the running process, and performing preprocessing operation on the real-time multi-source sensor time series signal to form the preprocessed real-time time series; feature extraction is performed on the preprocessed real-time time series to obtain the corresponding real-time features; the real-time features are input into the trained model for processing to obtain the residual life prediction result of the industrial equipment.

[0018] In the embodiment of the application, the historical feature extraction and model training are deeply coupled, rather than the independent step-by-step operation of the traditional two-stage method; in the joint training process, the feature extraction layer parameters are dynamically adjusted through the loss feedback of the prediction task, so that the extracted historical features naturally meet the requirements of residual life prediction, avoiding the error accumulation caused by independent optimization of data completion, feature extraction and prediction, and reducing the prediction inaccuracy risk caused by the deviation of intermediate links; the processing logic of historical data and real-time data is same source and same structure, reducing the system deviation; the historical data and real-time data use the same preprocessing and feature extraction logic, ensuring that the real-time features and the historical features during model training are in the same distribution space, avoiding the feature distribution deviation caused by the difference in processing rules, effectively reducing the random fluctuation of the prediction result, and improving the consistency and reliability of the prediction result. The feature extraction and the prediction task are strongly bound, improving the feature discrimination; the feature extraction is not an isolated operation, but is continuously optimized in the joint training, so that the model can automatically focus on the core features strongly related to equipment degradation, suppress the noise features caused by data loss and working condition fluctuation, and more accurately capture the equipment degradation rules through task-driven feature extraction, thereby strengthening the recognition of the degradation trend; through preprocessing noise reduction and adaptive learning of joint training, the trained model directly learns effective feature representation from the historical time series data containing missing data; when some sensors fail or are interfered to cause data loss, the trained model can automatically use other related sensor data to supplement the information, and still maintains stable prediction performance in the data loss scene.

[0019] In a preferred embodiment of the present application, step 1, the historical multi-source sensor time series signals of the industrial equipment during operation are obtained, and the signals are preprocessed to form preprocessed historical time series, including: Step 11, the historical multi-source sensor time series signals are normalized to obtain normalized historical multi-source sensor time series signals, specifically including: collecting historical time series signals collected by all sensors during the operation of the industrial equipment, these signals contain different types of monitoring variables such as temperature, pressure, and flow, and the values of each variable have different dimensions; then, the statistical characteristics of each sensor time series signal are calculated, specifically the average value and standard deviation of all collected data of the sensor; then, for each data point in the sensor time series signal, the original value of the data point is subtracted from the average value of the sensor signal calculated before, and the difference value is divided by the standard deviation of the sensor signal; through the above operation, the time series signal of each sensor is converted into standardized data with uniform dimensions, and finally the normalized historical multi-source sensor time series signals are obtained, avoiding interference with subsequent modeling due to different value dimensions.

[0020] Step 12, detect and identify missing values and outliers in the normalized historical multi-source sensor time series signals, and use interpolation method to fill the missing values to obtain preprocessed historical time series, specifically including: detecting each point of the normalized historical multi-source sensor time series signals, identifying missing values and outliers in the signals by comparing the numerical difference between each data point and adjacent normal data points, analyzing whether the data point conforms to the normal acquisition range of the sensor, etc.; for the identified outliers, further verify whether they belong to invalid abnormal data according to the information of equipment operation condition, sensor working principle, etc., and treat them as missing values for unified processing; then, use interpolation method to fill all missing values, specifically for the time points with missing values, use the adjacent valid data points before and after the time points as the basis, calculate the intermediate value of the adjacent valid data points to fill the value of the missing position. After filling all missing values, the filled sensor time series signals are obtained.

[0021] First, determine the standard time interval for data acquisition from industrial equipment sensors. This interval must be consistent with the monitoring frequency during normal equipment operation. Then, check the timestamps of each sensor signal after padding. For timestamps that do not meet the standard interval requirements, adjust the time alignment of the signals according to the standard time interval to ensure that the acquisition time of each sensor signal meets the equal interval requirement. Next, integrate all the time-aligned padding signals from all sensors in chronological order to ensure that the data from different sensors at the same time point can correspond one-to-one. Finally, form a time series data containing all sensor information, with uniform time intervals and complete data. This data is the preprocessed historical time series, which can be used for subsequent feature extraction.

[0022] In a preferred embodiment of the present invention, step 2, which involves extracting features from the preprocessed historical time series to obtain extracted historical features, includes: Step 21: Identify key sensor channel features sensitive to industrial equipment degradation based on the preprocessed historical time series; enhance these key sensor channel features to obtain enhanced features, specifically including: Step 211 involves analyzing the preprocessed historical time series to identify key sensor channels and extract the corresponding features for each channel. Specifically, this includes: Obtain the preprocessed historical time series data, specifically the input sequence of the c-th sensor channel. First, calculate the median absolute deviation of the channel sequence. This indicator can robustly reflect the dispersion of data within a channel and effectively avoid interference from outliers; it also calculates the local time gradient of the channel sequence. The short-term dynamic characteristics of the signal are captured by the rate of change of data at adjacent time points. Next, learnable non-negative coefficients are configured for this channel. and ,Will and Multiplying, the average of the local time gradients and Multiply the products, then add them together. Finally, apply the softplus function to the sum, according to the formula... Generate the spatial dynamic threshold of the c-th channel The softplus function ensures If positive, it enables adaptive threshold generation based on robust intra-channel scaling and local temporal gradient.

[0023] Step 212: Calculate the spatial dynamic threshold based on the characteristics of each channel; use the spatial dynamic threshold to perform soft thresholding on the characteristics of each channel to obtain the soft-thresholded characteristics of each channel, specifically including: Obtain the spatial dynamic threshold for each sensor channel and the raw observations of the channel at each time point t. ;against First calculate its absolute value. Then use minus If the difference is greater than 0, then retain the difference and keep it. The original symbol; if the difference is less than or equal to 0, the processing result is set to 0; through this soft thresholding operation, weak pseudo signals generated by missing interpolation or environmental interference in the channel are filtered out, and key features sensitive to equipment degradation are retained, so as to obtain the features of each channel after soft thresholding.

[0024] Step 213 involves fusing the features of each channel after soft thresholding to obtain enhanced features. Specifically, this includes collecting features from all sensor channels after soft thresholding, ensuring complete temporal alignment of features, meaning that the processed features of different channels at the same time t can accurately match. Then, the feature data of each channel at the same time t are integrated to form a multi-channel collaborative feature matrix. This matrix retains key information from individual channels while also fusing the correlation characteristics between different channels, ultimately yielding enhanced features after spatial dimension processing.

[0025] Step 22: Based on the enhanced features, locate key time points in the degradation process of industrial equipment to obtain dynamic thresholds in the time dimension, specifically including: Step 221: Perform global statistical analysis on the enhanced features to obtain global temporal statistical features. In specific applications, this may include: Obtain the enhanced features obtained in step 21, and for each time point t, determine a time window centered at t that covers the period from tw to t+w; then, perform context feature extraction on the enhanced features within this window. The system captures the temporal correlation and changing trends of data within a window to obtain local contextual features at time t. Next, a trainable weight vector is configured for these contextual features. Combine local context features with Perform transpose multiplication calculation ,in, This represents the transpose operation; then it is activated by the sigmoid function. Mapping the calculation results to the 0-1 interval yields the weight parameters for quantile calculation. Finally, the absolute value of the sensor observations within the window is calculated based on this weighting parameter. Quantile statistics Statistical information representing the temporal characteristics of the surrounding area at time t is extracted and used as an important component of global time statistical features.

[0026] Step 222: Based on global time statistical features, generate dynamic thresholds for the time dimension using learnable parameters. Specifically, this includes obtaining the quantile results from the global time statistical features corresponding to each time point t. Configure a scaling factor for the result. Then, the quantile results are compared with... Multiplication, according to the formula Dynamic threshold of time dimension for generation time point t scaling factor The overall threshold size can be adjusted to adapt to different device degradation characteristics and data distribution, ensuring that normal operation signals and critical degradation signals can be effectively distinguished.

[0027] Step 23: Using the enhanced features as input features and combining them with the time-dimensional dynamic threshold for residual shrinkage processing, historical features reflecting the equipment degradation state are obtained as extracted historical features, specifically including: Obtain the raw observation value of channel c at time t in step 21. The spatial dynamic threshold generated in step 211 and the time dynamic threshold generated in step 222 According to the formula right Perform residual shrinkage: First calculate symbols , then calculate and , The difference between the sums If the difference is greater than 0, then the difference is compared with... Multiply to obtain the condensed eigenvalues. If the difference is less than or equal to 0, then... Set to 0; through this operation, pseudo signals are filtered from both spatial and temporal dimensions, redundant interference and invalid information in the time dimension are eliminated, and finally the characteristics that accurately reflect the degradation state of the equipment are obtained.

[0028] In a preferred embodiment of the present invention, step 3, based on the extracted historical features, performs joint training to obtain a trained model, including: Step 31: Perform missing data completion and remaining lifetime prediction on the extracted historical features to obtain preliminary missing data completion results and remaining lifetime prediction results; based on the preliminary missing data completion results, dynamically adjust the loss weights of missing data completion and remaining lifetime prediction to obtain adjusted loss weights, specifically including: After extracting the features from step 2 that reflect the device's degradation state, they are input into the missing data completion task branch and the remaining lifetime prediction task branch, respectively. In the early stages of training, priority is given to the missing data completion task. A multi-task framework is used to complete the missing data to restore the input temporal features, providing a stable feature base for the subsequent remaining lifetime prediction task and reducing prediction bias caused by missing data. During the completion process, the original input data of sensor i at time t is processed... Process according to the following formula: ; Among them, if If the data is already observed, then retain that data directly as the completed result. ;like If the data is missing, the missing data completion function will be called. Combined with the sensor's sequence before time t and the sequence after time t The completed value is calculated. Through the above operations, the optimization of the missing data completion task in the early stage of training is completed, and the optimized result of the missing data completion task, namely the completed temporal features, is obtained. .

[0029] Step 32: Based on the adjusted loss weights, optimize the joint loss for missing data completion and remaining lifetime prediction. When the optimization process reaches convergence, the final trained prediction model is obtained, which specifically includes: Based on the optimization results of the missing data completion task in step 31, a progressive weight adjustment mechanism is introduced to dynamically adjust the loss weights of the missing data completion task and the remaining lifetime prediction task according to the task difficulty. Calculate the reduction in reconstruction loss of the completion task during the t-th training round. The magnitude of RUL loss reduction in the remaining lifetime prediction mission And according to the formula and The weights of the remaining lifetime prediction task are determined respectively. Weights of missing data completion tasks .

[0030] Simultaneously, dynamic weights that change with the training phase are introduced. Perform closed-loop optimization and correction on the weights: In the early stages of training, When the weight is 1, the completion task has a higher weight, so the completion task is optimized first to ensure the stability of the feature representation; as training progresses into the later stages, let The weight of the two tasks is gradually balanced to 0.5, avoiding a single task dominating the training process and allowing the model training focus to gradually shift from data completion to lifespan prediction.

[0031] In the training process after weight adjustment, closed-loop optimization is achieved through regularization constraints. To ensure the consistency between the completed results and the original features and to maintain the smoothness of the time series, temporal consistency constraints and smoothing regularization terms are used, based on the formula... Constructing regularized loss ;in, For the completed eigenvalues, This is an enhancement feature of the ST-RSBU pseudo-signal suppression unit output. The weights for time smoothing regularization are calculated by summing the squared differences of the features at adjacent time points to constrain the magnitude of feature changes at adjacent time points and avoid abrupt changes that are not physically meaningful. The weights for consistency constraints are calculated by summing the squared differences between the completed and enhanced features to ensure consistency between the completed and enhanced features.

[0032] The training process continuously monitors the trends of the loss functions for missing data completion, remaining life prediction, and the total loss function (including regularization loss) throughout the training process. The changes in the total loss function are used to assess the adaptive balance between tasks. When the total loss function stabilizes and no longer decreases significantly (reaching convergence), model training is stopped, resulting in a successfully trained industrial equipment remaining life prediction model.

[0033] In a preferred embodiment of the present invention, step 4 involves acquiring real-time multi-source sensor time-series signals of the industrial equipment during operation, and preprocessing the real-time multi-source sensor time-series signals to form a preprocessed real-time time series; extracting features from the preprocessed real-time time series to obtain corresponding real-time features; and inputting the real-time features into a trained model for processing to obtain the remaining life prediction result of the industrial equipment, including: A real-time data acquisition system for industrial equipment operation is activated. This system maintains synchronous connection with various monitoring sensors deployed on the equipment. These sensors need to cover critical operational aspects of the equipment, specifically including temperature sensors for monitoring the temperature of core components, pressure sensors for collecting pressure in pipelines or cavities, and flow sensors for measuring the flow rate of media. This ensures the acquisition of multiple types of key monitoring variables reflecting the equipment's operating status. The acquisition process must strictly adhere to the standard monitoring frequency set during normal equipment operation to achieve synchronous acquisition of time-series signals from each sensor, avoiding data misalignment due to acquisition time differences. During acquisition, the sensor identifier, acquisition timestamp, and original monitoring value corresponding to each data point must be recorded in real time. Simultaneously, preliminary identification of any missing data (e.g., some sensors failing to transmit data due to temporary signal interruptions) is performed. This raw data, containing multiple types of variables, timestamped, and potentially containing missing values, is used as the initial real-time multi-source sensor time-series signal for subsequent processing.

[0034] The initial real-time multi-source sensor time-series signals acquired are processed according to the standard procedure consistent with historical data preprocessing, specifically in the following three steps: Normalization involves first retrieving the statistical characteristics (i.e., the average and standard deviation of all historical data collected by the sensor) of each sensor channel from the preprocessing stage of the historical multi-source sensor time-series signals of the device. For each raw data point of each sensor channel in the real-time data, the calculation method is "subtract the historical average value of the corresponding sensor from the raw value of the data point, and then divide the difference by the historical standard deviation of the corresponding sensor". Through this operation, real-time sensor data with different dimensions (such as temperature in degrees Celsius, pressure in Pascals, and flow rate in cubic meters per hour) are uniformly converted into dimensionless standardized data, avoiding interference from dimensional differences in subsequent feature extraction and model calculation.

[0035] Missing and outlier handling involves point-by-point detection of the normalized real-time sensor time-series signals. This is achieved by comparing the numerical differences between each data point and adjacent normal data points (e.g., a temperature reading suddenly exceeding the reasonable range at a given moment), considering the sensor's operating principle (e.g., a flow sensor collecting a zero value when the equipment is not running), and the equipment's current operating conditions (e.g., pressure data abnormally low when the equipment is under normal load). This comprehensive approach identifies missing values ​​(e.g., time points where the sensor did not transmit data) and outliers. For identified outliers, further verification is performed using equipment maintenance records (e.g., recent sensor calibration records) and real-time operating parameters (e.g., any instantaneous load fluctuations). Once confirmed as invalid outlier data, it is grouped with missing values ​​for unified processing. Subsequently, interpolation is used to fill in all missing data points: for time points with missing values, the median value of adjacent valid normalized data points is calculated based on the time points before and after that point. This median value is then used to fill in the missing value, ensuring continuous, uninterrupted real-time data for each sensor channel.

[0036] Time alignment and data integration involve retrieving the standard time interval for data acquisition from the equipment's sensors (this interval is consistent with the historical data preprocessing and normal equipment monitoring frequency) and checking the timestamps of the real-time signals from each sensor after missing value imputation. For signals whose timestamps do not meet the standard interval requirements (e.g., a sensor's delayed data transmission causes a certain acquisition time point to exceed the standard interval range), the timestamps of the signals are adjusted according to the standard time interval to ensure that all acquisition time points of the real-time signals from each sensor meet the equal interval requirements. Then, all the time-aligned real-time signals from all sensors are integrated according to the acquisition time sequence, and the data from different sensors at each time point are checked one-to-one to ensure they correspond (e.g., at a certain time t, the real-time data from temperature, pressure, and flow sensors must exist simultaneously with timestamps of t). This ultimately forms a real-time time series containing information on all sensor types, with uniform time intervals and continuous, complete data; this series is the preprocessed real-time time series.

[0037] The preprocessed real-time time series is input into the ST-RSBU unit (spatiotemporal residual shrinkage unit) in the trained multi-task spatiotemporal residual shrinkage network. This unit performs real-time feature extraction, as detailed below: The spatial dimension key channel feature screening and enhancement involves the ST-RSBU unit first performing spatial dimension analysis on the preprocessed real-time time series. For the real-time data of each sensor channel, the dispersion of the data (e.g., the concentration or dispersion of data distribution) and short-term dynamic change trend (e.g., the rate of change of data between adjacent time points) are calculated. Combining these two indicators, key sensor channels sensitive to equipment degradation status are identified (e.g., equipment bearing temperature channels, hydraulic system pressure channels, etc., channels directly related to the aging of core equipment components). Subsequently, soft thresholding is performed on the real-time data of these key channels: based on the real-time data characteristics of each key channel, an appropriate dynamic threshold is set to filter out weak pseudo-signals (e.g., signals with extremely small numerical fluctuations and no actual physical meaning) caused by environmental interference (e.g., workshop electromagnetic interference) or missing value imputation. Only effective signals that truly reflect changes in equipment operating status are retained. Then, the signals of all key channels after soft thresholding are fused to form spatial dimension enhancement features that retain the key information of each channel and integrate the correlation characteristics between channels.

[0038] The ST-RSBU unit then performs time-dimensional analysis on the spatial augmentation features, setting a time window adapted to the equipment's operating cycle (e.g., a window covering multiple acquisition times before and after). For each time point, the ST-RSBU extracts the temporal correlation (e.g., the rising or falling trend of temperature features within a certain time period) and change patterns of the spatial augmentation features within that time window, generating a dynamic time threshold for each time point. This allows the identification of key time points in the real-time data that may reflect the beginning of equipment degradation (e.g., pressure features starting to deviate continuously from the normal range after a certain moment). Next, combining the dynamic thresholds of the spatial and temporal dimensions, residual shrinkage is performed on the spatial augmentation features: by comparing the relationship between each data point and the dual-dimensional thresholds, invalid information in the temporal dimension (e.g., non-degradation signals generated by normal equipment fluctuations) and redundant interference in the spatial dimension (e.g., slight signal changes in non-critical channels) is eliminated, ultimately yielding real-time features that accurately reflect the current degradation state of the equipment.

[0039] The real-time features extracted through the ST-RSBU unit are input into a pre-trained multi-task spatiotemporal residual shrinkage network model. This model, jointly trained with historical data, possesses dual-task processing capabilities for missing data completion and remaining lifetime prediction. Upon receiving the real-time features, the model first verifies the data integrity of the real-time features again through its internal missing data completion branch (ensuring no unprocessed missing or abnormal information). The verified real-time features are then passed to the remaining lifetime prediction branch. Based on the equipment degradation patterns learned during training (such as the correspondence between real-time features and remaining lifetime at different degradation stages), the prediction branch analyzes and calculates the real-time features, ultimately outputting the remaining lifetime prediction result for the current state of the industrial equipment. This result is presented in the form of a specific lifetime duration (such as remaining operating hours).

[0040] In the ST-RSBU unit, to address the heterogeneity issue of different sensor channels, the channel scale is first calculated using the truncated root mean square method, based on the formula... Calculate the truncated root mean square of the c-th channel. ;in, This indicates that the sequence is divided by magnitude, removing the highest and lowest values. The subsequent index set, The truncated root mean square of the c-th channel provides a more robust scaling estimate in the presence of sporadic impulse noise. Based on the truncation scale and distribution skewness information, according to the formula Construct channel thresholds, where, For learnable parameters, This is the sequence skewness coefficient; simultaneously, the local dynamic measure at time t is calculated using the exponential moving average, based on the formula... , and Generate dynamic thresholds in the time dimension. For smoothing coefficients, b are learnable factors. Let represent the input signal value of the sensor in the c-th channel at time t. This indicates that the sensor in the c-th channel is... The input signal value at time; It is the average of the exponentially weighted moving average (EMA).

[0041] Next, multi-level neighborhood information is fused using a gating mechanism, based on the formula... Forming the final spatiotemporal characteristics, ) represents a multi-order neighborhood feature. For the projection matrix, It is a characteristic of the central node itself. yes Neighborhood features; Simultaneously, the residual shrinkage core excitation formula of the ST-RSBU unit is combined. By filtering out spurious signals, the characteristics that ultimately reflect the real-time degradation status of the equipment are obtained. This indicates the input signal at channel c and time step t. The dynamic threshold for residual shrinkage.

[0042] Step 42 involves simultaneously inputting features reflecting the equipment's degradation state into both the missing data completion branch and the remaining lifetime prediction branch of the trained model to obtain lifetime prediction results. Specifically, this includes: The features obtained in step 41 that reflect the device's degradation state (including missing augmented inputs) The missing data completion branch is input into the model; first, a temporal context representation based on a time window is calculated and fused with neighborhood channel information, introducing a spatiotemporal aggregation and masking mechanism: According to the formula Calculate attention weights ,in For observation mask (1 indicates that it has been observed). The time window is half the width. , Let d be the weight matrix and d be the feature dimension. This operation prioritizes the use of observable location information and weakens the contribution of missing points to attention.

[0043] According to the formula ( (As a weight matrix) to obtain the temporal context ; Time sequence context Combined with channel aggregation based on physical / topological adjacency; According to the formula Obtain the completed features ,in For elements of the adjacency matrix, For learnable weight matrix, For the residual autoregressive term, This is a LayerNorm+GELU nonlinear mapping.

[0044] Through the above operations, the complete sensor time series is output after completion. ; The remaining lifetime prediction branch inference will complete the full sequence. The remaining lifetime is input into the model's remaining lifetime prediction branch, which further encodes, aggregates, and regresses the time-series-channel fusion representation: According to the formula The obtained spatiotemporal context is fused at multiple scales after undergoing multi-order adjacency transformation to obtain fused features. .

[0045] According to the formula Through weighted time attention pooling Highlighting the time period most relevant to RUL prediction, we perform temporal pooling on the fused features to obtain pooled features. (T is the time step).

[0046] According to the formula and Heterogeneous uncertainty regression is performed, where For small MLP projection matrix, For bias terms, For remaining lifetime prediction point estimation, To measure the uncertainty in prediction (the softplus function ensures a positive variance), the model ultimately outputs the completed sensor time-series data synchronously. and the corresponding remaining lifetime prediction results (Including uncertainty measurement) This provides complete data support and accurate lifespan references for industrial equipment operation and maintenance decisions.

[0047] This process fully embodies the synergistic effect of the two core components of the present invention framework: the ST-RSBU unit and the multi-task collaborative optimization module. The ST-RSBU simultaneously achieves noise suppression and feature extraction in both time and space dimensions, effectively mitigating redundant information and random interference in high-dimensional industrial time-series data. The multi-task collaborative optimization module, through phased training (the warm-up phase uses the completion task to stabilize the feature distribution and gradually increases the weight of the RUL prediction task), introduces consistency regularization to constrain the differences between completion and prediction features, and time-dimensional smoothing regularization to control non-physical fluctuations. This unifies data completion and RUL prediction under a shared optimization objective, enhances the model's ability to perceive the time-series structure, guides the model to combine completion data with lifetime prediction objectives, and ultimately significantly improves stability and reliability while ensuring real-time prediction.

[0048] Figure 1 The diagram shows the architecture of a multi-task spatiotemporal residual shrinkage network, illustrating the overall process of an industrial equipment remaining life prediction method based on this network. First, multi-source sensor time-series data containing missing values ​​is input and fed into multiple spatiotemporal dual-stream adaptive residual shrinkage units (ST-RSBU) for processing. These units learn dynamic thresholds in both time and space dimensions, filtering out pseudo-signals and extracting key features. Next, the extracted features are aggregated across time and channels using a multi-head attention mechanism, and then input into a temporal convolutional network (TCN) to further capture temporal dependencies. Finally, the features enter a shared feature layer, which then branches into two task branches: One branch is the missing data completion head, which outputs the completed sensor time series data, with a loss of Loss. imp ; Another branch is the remaining lifetime prediction head (Pre-head), which outputs the device's remaining lifetime prediction results, and its loss is Loss. rul .

[0049] Meanwhile, the loss weights of the two tasks are dynamically balanced through a progressive weight adjustment mechanism, enabling joint training and closed-loop optimization of missing data completion and remaining lifetime prediction.

[0050] Figure 2 The figure shows the internal structure of the Spatiotemporal Dual-Stream Adaptive Residual Shrinkage Unit (ST-RSBU). This core structure of the ST-RSBU is used to suppress spurious signals and extract key features in the temporal and spatial dimensions.

[0051] First, the input feature map (dimension C×W×1, where C is the number of channels and W is the time step) undergoes a series of batch normalization (BN), ReLU activation, and convolution (Conv) operations for feature transformation. Then, global average pooling (AbsoluteGAP) is used to compress the time dimension, obtaining channel-level statistical features. Next, these statistical features are processed through fully connected layers (FC), BN, and ReLU activation to generate gating coefficients. a (obtained by Sigmoid activation); the gating coefficients are multiplied element-wise with the original feature map to achieve adaptive weight adjustment for different channel features, that is, the weights of channel features that are not sensitive to device degradation are weakened, while the weights of key channel features that are sensitive to degradation are strengthened.

[0052] Finally, by combining the identity shortcut, the processed features are added to the original input features to complete the residual shrinkage operation, and the final output is the key features after filtering out the pseudo-signals.

[0053] Figure 3 This diagram illustrates the collaborative optimization process of the multi-task learning framework. First, the multi-sensor input is fed into a spatiotemporal dual-stream adaptive residual shrinking unit (ST-RSBU, incorporating both temporal and spatial streams) for feature extraction. The extracted features are then further fused with temporal and channel information via a multi-head attention + temporal convolutional network (TCN) before entering the shared feature layer. Subsequently, the features are divided into two task branches: The missing data completion head outputs the completion result and calculates the completion loss. Loss The remaining lifetime prediction head outputs lifetime prediction results and calculates prediction loss. Loss .

[0054] Next, the loss of the two tasks is summarized by the Loss Calculation module, and the task weights are dynamically adjusted by Progressive Weight Scheduling to achieve phased training (such as prioritizing the optimization of the completion task in the early stage and balancing the two tasks in the later stage). In the Inference Phase, the model infers the operating state of the device and continuously improves the accuracy through the Iterative Refinement process, and finally outputs the Remaining Life Prediction (RUL Output).

[0055] This method can be applied to scenarios such as industrial equipment health management, intelligent manufacturing monitoring, and intelligent operation and maintenance decision-making. It can dynamically capture key dependency structures in multivariate time series data, providing effective technical support for high-precision remaining life prediction of complex industrial equipment. Compared with mainstream technologies, it significantly improves data utilization and prediction accuracy.

[0056] An industrial equipment remaining life prediction system, comprising: The acquisition module is used to acquire historical multi-source sensor time-series signals of industrial equipment during operation, and to preprocess the signals to form a preprocessed historical time series. The extraction module is used to extract features from the preprocessed historical time series to obtain extracted historical features; The training module is used to perform joint training based on the extracted historical features to obtain a trained model. The prediction module is used to acquire real-time multi-source sensor time-series signals of industrial equipment during operation, and to preprocess the real-time multi-source sensor time-series signals to form a preprocessed real-time time series; to extract features from the preprocessed real-time time series to obtain corresponding real-time features; and to input the real-time features into the trained model for processing to obtain the remaining life prediction result of the industrial equipment.

[0057] The prediction system according to embodiments of the present invention can correspond to performing the methods described in the embodiments of the present invention, and the above and other operations and / or functions of each module of the prediction system are respectively for implementing Figure 4 The corresponding process of the method in the illustrated embodiment will not be described in detail here for the sake of brevity.

[0058] This application also provides a computing device. This computing device can utilize a server.

[0059] like Figure 5 As shown in the figure, this is a schematic diagram of a computing device provided in an embodiment of this application. The computing device 700 includes a bus 701, a processor 702, a communication interface 703, and a memory 704. The processor 702, the memory 704, and the communication interface 703 communicate with each other via the bus 701.

[0060] The 701 bus can be a Peripheral Component Interconnect (PCI) bus or an Extended Industry Standard Architecture (EISA) bus, etc. Buses can be categorized as address buses, data buses, control buses, etc. For ease of representation, Figure 5 The bus is represented by a single thick line, but this does not mean that there is only one bus or one type of bus.

[0061] The processor 702 can be any one or more of the following processors: central processing unit (CPU), graphics processing unit (GPU), microprocessor (MP), or digital signal processor (DSP).

[0062] Communication interface 703 is used for external communication. Memory 704 may include volatile memory, such as random access memory (RAM). Memory 704 may also include non-volatile memory, such as read-only memory (ROM), flash memory, hard disk drive (HDD), or solid-state drive (SSD). Memory 704 stores executable code, which processor 702 executes to perform the aforementioned industrial equipment remaining life prediction method.

[0063] Specifically, in implementing the industrial equipment remaining life prediction system shown in the above embodiments, and where each module or unit of the industrial equipment remaining life prediction system described in the above embodiments is implemented by software, the software or program code required to execute the functions of each module / unit in the industrial equipment remaining life prediction system described in the above embodiments can be partially or entirely stored in the memory 704. The processor 702 executes the program code corresponding to each unit stored in the memory 704 to execute the aforementioned industrial equipment remaining life prediction method.

[0064] This application also provides a computer-readable storage medium. The computer-readable storage medium can be any available medium that a computing device can store, or a data storage device such as a data center containing one or more available media. The available medium can be a magnetic medium (e.g., floppy disk, hard disk, magnetic tape), an optical medium (e.g., DVD), or a semiconductor medium (e.g., solid-state drive). The computer-readable storage medium includes instructions that instruct a computing device to execute the aforementioned industrial equipment remaining life prediction method.

[0065] This application also provides a computer program product comprising one or more computer instructions. When the computer instructions are loaded and executed on a computing device, all or part of the processes or functions described in this application are generated.

[0066] The computer instructions may be stored in a computer-readable storage medium or transmitted from one computer-readable storage medium to another. For example, the computer instructions may be transmitted from one website, computer, or data center to another website, computer, or data center via wired (e.g., coaxial cable, fiber optic, digital subscriber line) or wireless (e.g., infrared, wireless, microwave, etc.) means.

[0067] When the computer program product is executed by a computer, the computer performs any of the aforementioned methods for predicting the remaining useful life of industrial equipment. The computer program product can be a software installation package; when any of the aforementioned methods for predicting the remaining useful life of industrial equipment is required, the computer program product can be downloaded and executed on the computer.

[0068] The above description represents the preferred embodiments of the present invention. It should be noted that those skilled in the art can make various improvements and modifications without departing from the principles of the present invention, and these improvements and modifications should also be considered within the scope of protection of the present invention.

Claims

1. A method for predicting the remaining life of industrial equipment, characterized in that, The method includes: The historical multi-source sensor time-series signals of industrial equipment during operation are acquired, and the signals are preprocessed to form a preprocessed historical time series. Feature extraction is performed on the preprocessed historical time series to obtain the extracted historical features; Based on the extracted historical features, joint training is performed to obtain the trained model; The system acquires real-time multi-source sensor time-series signals during the operation of industrial equipment, and performs preprocessing operations on the real-time multi-source sensor time-series signals to form a preprocessed real-time time series; it then extracts features from the preprocessed real-time time series to obtain corresponding real-time features; and finally inputs the real-time features into a trained model for processing to obtain the remaining life prediction result of the industrial equipment.

2. The method for predicting the remaining life of industrial equipment according to claim 1, characterized in that, Acquire historical multi-source sensor time-series signals from industrial equipment during operation, and perform preprocessing operations on the signals to form a preprocessed historical time series, including: The historical multi-source sensor time series signals are normalized to obtain normalized historical multi-source sensor time series signals; Missing and outlier values ​​in the normalized historical multi-source sensor time series signals are detected and identified, and interpolation methods are used to fill in the missing values ​​to obtain the preprocessed historical time series.

3. The method for predicting the remaining life of industrial equipment according to claim 2, characterized in that, Feature extraction is performed on the preprocessed historical time series to obtain extracted historical features, including: Based on the preprocessed historical time series, key sensor channel features sensitive to industrial equipment degradation were identified; these key sensor channel features were then enhanced to obtain enhanced features. Based on the enhanced features, key time points in the degradation process of industrial equipment are located to obtain dynamic thresholds in the time dimension; Using the enhanced features as input features and combining them with the time dimension dynamic threshold for residual shrinkage processing, historical features that reflect the equipment degradation state are obtained as extracted historical features.

4. The method for predicting the remaining life of industrial equipment according to claim 3, characterized in that, Key sensor channel features sensitive to industrial equipment degradation were identified based on preprocessed historical time series. Enhancement processing is performed on key sensor channel features to obtain enhanced features, including: The preprocessed historical time series is analyzed to identify key sensor channels and extract the corresponding features of each channel; Based on the characteristics of each channel, a spatial dynamic threshold is calculated; the spatial dynamic threshold is then used to perform soft thresholding on the characteristics of each channel to obtain the soft thresholded characteristics of each channel. The features of each channel that have undergone soft thresholding are fused together to obtain enhanced features.

5. The method for predicting the remaining life of industrial equipment according to claim 4, characterized in that, Based on enhanced features, key time points in the degradation process of industrial equipment are located to obtain dynamic thresholds in the time dimension, including: Perform global statistical analysis on the enhanced features to obtain global temporal statistical features; Based on global time statistical features, dynamic thresholds for the time dimension are generated through learnable parameters.

6. The method for predicting the remaining life of industrial equipment according to claim 5, characterized in that, Based on the extracted historical features, joint training is performed to obtain a trained model, including: The extracted historical features are subjected to missing data completion and remaining lifetime prediction respectively to obtain preliminary missing data completion results and remaining lifetime prediction results; based on the preliminary missing data completion results, the loss weights of missing data completion and remaining lifetime prediction are dynamically adjusted to obtain the adjusted loss weights. Based on the adjusted loss weights, the joint loss for missing data completion and remaining lifetime prediction is optimized. When the optimization process reaches convergence, the final trained prediction model is obtained.

7. A system for predicting the remaining life of industrial equipment, characterized in that, The system performs the method as described in any one of claims 1 to 6, comprising: The acquisition module is used to acquire historical multi-source sensor time-series signals of industrial equipment during operation, and to preprocess the signals to form a preprocessed historical time series. The extraction module is used to extract features from the preprocessed historical time series to obtain extracted historical features; The training module is used to perform joint training based on the extracted historical features to obtain a trained model. The prediction module is used to acquire real-time multi-source sensor time-series signals of industrial equipment during operation, and to preprocess the real-time multi-source sensor time-series signals to form a preprocessed real-time time series; to extract features from the preprocessed real-time time series to obtain corresponding real-time features; and to input the real-time features into the trained model for processing to obtain the remaining life prediction result of the industrial equipment.

8. A computing device, characterized in that, Including memory and processor; The memory stores one or more computer programs, the one or more computer programs including instructions; when the instructions are executed by the processor, the computing device performs the method as described in any one of claims 1 to 6.

9. A computer-readable storage medium, characterized in that, The computer-readable storage medium is used to store a computer program for performing the method as described in any one of claims 1 to 6.

Images