A device life prediction method and device based on multi-modal fusion

Abstract

The application relates to a device life prediction method and device based on multi-modal fusion, which comprises the following steps: acquiring device health time series data; performing quality evaluation based on the time series data and calculating a comprehensive quality score; dynamically adjusting Savitzky-Golay filtering parameters based on the comprehensive quality score; calculating a first residual life prediction value of the device by CUSUM discrimination and Akaike information criterion based on the adjusted Savitzky-Golay filtering parameters; calculating a second residual life prediction value of the device based on the physical mechanism of the device; constructing a Wiener random model based on the multi-sensor of the device; calculating a third residual life prediction value of the device through the Wiener random model; and obtaining a final life prediction value of the device by fusing the first residual life prediction value, the second residual life prediction value and the third residual life prediction value through confidence. The application improves the accuracy of device life prediction and reduces the risk of accidental shutdown by fusing multi-life prediction models.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of equipment detection, and particularly relates to a device life prediction method and device based on multi-modal fusion. BACKGROUND

[0002] With the acceleration of industrial modernization process, equipment health management and predictive maintenance have become the core engine driving the leap in production efficiency and compressing operation and maintenance costs. Equipment remaining useful life (RUL) prediction, as a core technology of predictive maintenance, is of great significance to avoid equipment sudden failure and optimize maintenance strategy. At present, equipment life prediction methods are mainly divided into three categories: physical model-based method, data-driven method and hybrid method. However, the traditional equipment life prediction method has the following technical defects in actual engineering application: 1. Single model limitation is serious; 2. Data quality evaluation system is missing; 3. Degradation phase recognition ability is insufficient; 4. Physical mechanism is seriously insufficient; 5. Real sensor data is not fused.

[0003] Therefore, an intelligent life prediction method that can comprehensively consider data quality, degradation phase, model fusion and physical mechanism is urgently needed to improve prediction accuracy and engineering practicability. SUMMARY

[0004] To realize an intelligent life prediction method that can comprehensively consider data quality, degradation phase, model fusion and physical mechanism, in the first aspect of the application, a device based on multi-modal fusion is provided, comprising: acquiring equipment health degree time series data; performing quality evaluation based on the time series data and calculating a comprehensive quality score; wherein the quality evaluation includes data anomaly check, time check and health degree trend check; dynamically adjusting the Savitzky-Golay filtering parameters based on the comprehensive quality score; calculating the first remaining life prediction value of the equipment by CUSUM and Akaike information criterion based on the adjusted Savitzky-Golay filtering parameters; calculating the second remaining life prediction value of the equipment based on the physical mechanism of the equipment; constructing a Wiener random model based on the multi-sensor of the equipment; calculating the third remaining life prediction value of the equipment through the Wiener random model; and obtaining the final life prediction value of the equipment by fusing the first remaining life prediction value, the second remaining life prediction value and the third remaining life prediction value through confidence.

[0005] In some embodiments of the present application, the first residual life prediction value of the equipment is calculated based on the adjusted Savitzky-Golay filtering parameters, degradation is distinguished by CUSUM, and Akaike information criterion, comprising: filtering the time series data based on the adjusted Savitzky-Golay filtering parameters; extracting a plurality of statistics from the filtered time series data based on an adaptive window; determining a plurality of degradation transition points by variance-gradient joint CUSUM change points based on the plurality of statistics; dividing degradation stages by the plurality of degradation transition points; adaptively selecting a degradation model for each degradation stage based on Akaike information criterion; and calculating the first residual life prediction value of the equipment by the selected degradation model.

[0006] Further, the determination of a plurality of degradation transition points by variance-gradient joint CUSUM change points based on the plurality of statistics comprises: determining change point determination thresholds of variance CUSUM channels and gradient CUSUM channels respectively; marking a plurality of change points based on the change point determination thresholds; mapping all candidate change points to a unified time axis, calculating change point density of each time region and weight of each candidate change point; calculating a comprehensive score of each candidate change point based on the product of the change point density and the corresponding candidate change point weight; and determining the candidate change points with a comprehensive score exceeding a preset threshold as the degradation stage transition points.

[0007] In some embodiments of the present application, the construction of a Wiener random model based on the multi-sensor of the equipment comprises: constructing a first Wiener equation based on the initial health state, drift coefficient and diffusion coefficient of the equipment; modifying the drift coefficient and the diffusion coefficient based on the correlation coefficient of the sensor, the standardized data sequence, the reliability index and the weight coefficient; and constructing a second Wiener equation based on the modified drift coefficient and diffusion coefficient.

[0008] Further, the calculation of the third residual life prediction value of the equipment by the Wiener random model comprises: calculating the third residual life prediction value by the first passage time probability density function based on the second Wiener equation.

[0009] In some embodiments of the present application, the fusion of the first residual life prediction value, the second residual life prediction value and the third residual life prediction value by confidence to obtain the final life prediction value of the equipment comprises: obtaining the confidence, fusion quality and quality score of each residual life prediction value; calculating the comprehensive confidence of the first residual life prediction value, the second residual life prediction value and the third residual life prediction value respectively by the confidence, the quality score and the fusion quality; and calculating the final life prediction value of the equipment by weighting based on each residual life prediction value and the corresponding comprehensive confidence.

[0010] In a second aspect, the application provides a device life prediction device based on multi-modal fusion, comprising: an acquisition module configured to acquire device health time series data; performing quality assessment based on the time series data and calculating a comprehensive quality score; wherein the quality assessment comprises data anomaly checking, time checking and health trend checking; a first calculation module configured to dynamically adjust Savitzky-Golay filter parameters based on the comprehensive quality score; calculate the first residual life prediction value of the device based on the adjusted Savitzky-Golay filter parameters, and Akaike information criterion; a second calculation module configured to calculate the second residual life prediction value of the device based on the physical mechanism of the device; a third calculation module configured to construct a Wiener random model based on the multi-sensor of the device; calculate the third residual life prediction value of the device through the Wiener random model; and a fusion module configured to fuse the first residual life prediction value, the second residual life prediction value and the third residual life prediction value through confidence fusion to obtain the final life prediction value of the device.

[0011] In a third aspect, the application provides an electronic device, comprising: one or more processors; a storage device configured to store one or more programs, when the one or more programs are executed by the one or more processors, the one or more processors implement the device life prediction method based on multi-modal fusion provided in the first aspect of the application.

[0012] In a fourth aspect, the application provides a computer readable medium having a computer program stored thereon, wherein the computer program is executed by a processor to implement the device life prediction method based on multi-modal fusion provided in the first aspect of the application.

[0013] The application has the following beneficial effects:

[0014] 1. Five-dimensional data quality assessment system: a five-dimensional evaluation system including missing values, abnormal values, time regularity, trend rationality and data range checking is established. By comprehensively evaluating the data quality of multiple dimensions, it is ensured that the prediction is based on reliable data, and the limitation of traditional methods considering only single quality index is overcome.

[0015] 2. Variance-gradient joint CUSUM degradation stage identification: an innovative change point detection algorithm is adopted, local statistics are calculated through sliding window, and double discrimination is performed in combination with variance and gradient change. It can automatically identify the transition nodes of the device from the stable period to the linear degradation period and then to the accelerated degradation period, and provide accurate stage information for model selection.

[0016] 3. Multi-model adaptive fusion architecture: A multi-model fusion architecture containing degradation model, integral area and Wiener process is constructed. The optimal model is automatically selected based on AIC criterion, and the advantages of multiple prediction methods are complementary through adaptive weight distribution, which significantly improves the prediction accuracy and stability.

[0017] 4. Uncertainty quantification and confidence interval calculation: Bootstrap resampling technique is used to calculate the confidence interval of prediction results, which quantifies the credibility of prediction. Not only the point prediction result is given, but more importantly, the risk assessment basis is provided to help decision makers make scientific and reasonable maintenance strategies.

[0018] 5. Deep fusion of physical mechanism and data-driven method: The physical mechanism model and data-driven method are innovatively deeply fused. The integral area method is used to calculate the health degree consumption rate, and the Wiener process is used to fuse multi-sensor data. It not only guarantees the physical rationality of prediction, but also gives full play to the advantages of data-driven method. BRIEF DESCRIPTION OF DRAWINGS

[0019] Figure 1 a basic flowchart of the device life prediction method based on multi-modal fusion in some embodiments of the present application;

[0020] Figure 2 a specific flowchart of the device life prediction method based on multi-modal fusion in some embodiments of the present application;

[0021] Figure 3 a structural schematic diagram of the device life prediction device based on multi-modal fusion in some embodiments of the present application;

[0022] Figure 4 a structural schematic diagram of the electronic device in some embodiments of the present application. DETAILED DESCRIPTION

[0023] The principles and characteristics of the present application are described below in conjunction with the drawings, and the examples are only used to explain the present application and are not used to limit the scope of the present application.

[0024] Example 1

[0025] Reference Figure 1 and Figure 2In a first aspect of the present invention, a device based on multimodal fusion is provided, comprising: S100. acquiring time-series health data of the device; performing quality assessment based on the time-series data and calculating a comprehensive quality score; wherein the quality assessment includes data anomaly detection, time detection, and health trend detection; S200. dynamically adjusting Savitzky-Golay filter parameters based on the comprehensive quality score; calculating a first remaining lifetime prediction value of the device based on the adjusted Savitzky-Golay filter parameters and using the CUSUM discrimination criterion for degradation and the Akaike Information Criterion; S300. calculating a second remaining lifetime prediction value of the device based on the physical mechanism of the device; S400. constructing a Wiener stochastic model based on multiple sensors of the device; calculating a third remaining lifetime prediction value of the device using the Wiener stochastic model; S500. fusing the first remaining lifetime prediction value, the second remaining lifetime prediction value, and the third remaining lifetime prediction value by confidence level to obtain a final lifetime prediction value of the device.

[0026] In step S100 of some embodiments of the present invention, device health time-series data is acquired; a quality assessment is performed based on the time-series data, and a comprehensive quality score is calculated; wherein, the quality assessment includes data anomaly check, time check, and health trend check;

[0027] Specifically, to ensure the reliability and accuracy of the healthy lifespan prediction algorithm, this invention first performs a systematic quality assessment on the input health time-series data. Let the health time-series data be H = {h1, h2, ..., h...} n The corresponding timestamp sequence is T={t1,t2,...,t}. n}, where n is the total number of data points. The quality assessment system comprises five core dimensions, and the steps are as follows:

[0028] Data integrity assessment:

[0029] Define the missing value ratio metric:

[0030] , where I(·) is the indicator function. Integrity score The calculation rules are as follows:

[0031] (1),

[0032] Data anomaly assessment:

[0033] Outliers are identified using the interquartile range (IQR) method. Let the valid dataset be H_valid={h i |h i ≠NaN}, calculate:

[0034] (2),

[0035] Outlier determination condition:

[0036] (3),

[0037] Outlier score:

[0038] (4),

[0039] (5),

[0040] Regularity evaluation:

[0041] Calculate the time interval sequence ΔT = {Δt1, Δt2,..., Δtn}, where Δt n-1 = t i - t i+1 - t i . Define the time regularity index:

[0042] (6),

[0043] where σ(·) and μ(·) are the standard deviation and mean functions, respectively.

[0044] Regular score: (7),

[0045] Trend rationality evaluation:

[0046] Fit the health degree change trend using the least squares method:

[0047] (8),

[0048] where = (n+1) / 2, . Trend rationality score:

[0049] (9),

[0050] Numerical validity evaluation: test whether the health degree value is within the standardized interval [0, 1]:

[0051] (10) Numerical validity score:

[0052] (11),

[0053] Comprehensive quality score calculation: calculate the comprehensive quality score using the weighted average method: (12),

[0054] wherein, the weight vector w = [0.25, 0.20, 0.15, 0.15, 0.25] , which embodies the important position of integrity and numerical validity. The final quality score ranges from 0.1 to 1.0, ensuring that the subsequent algorithm has basic data quality assurance. The five-dimensional evaluation system comprehensively characterizes the quality characteristics of health data through quantitative indicators, providing a scientific basis for subsequent data preprocessing, model selection, and prediction result credibility evaluation.

[0055] In step S200 of some embodiments of the application, the first residual life prediction value of the equipment is calculated based on the adjusted Savitzky-Golay filtering parameters, including:

[0056] S201. Filtering the time series data based on the adjusted Savitzky-Golay filtering parameters;

[0057] Specifically, after completing the preliminary data preprocessing and quality evaluation, in order to improve the accuracy and stability of subsequent analysis, an enhanced data smoothing strategy based on dynamic response of data quality is introduced. The core idea of this strategy is to adaptively adjust the parameter configuration of the Savitzky-Golay filter according to the quality score (quality_score) of each time series data, so as to effectively suppress noise interference while preserving key signal features. As a smoothing method based on local polynomial fitting, Savitzky-Golay filtering can better maintain important morphological features such as peaks and inflection points while smoothing data, and is suitable for time series processing in various scenarios such as biological signals, environmental monitoring, and industrial sensing.

[0058] (13),

[0059] wherein, m is the window size, and a is the polynomial order. Specifically, the system first reads the data quality score (quality_score) output in the previous stage, which comprehensively reflects the data integrity, signal-to-noise ratio, sampling consistency, and other multi-dimensional indicators, with a value range of 0 to 1. According to the score result, the system will automatically select the optimal filter parameter combination to achieve the goal of "fine smoothing for high-quality data and robust filtering for low-quality data". The specific parameter configuration is as follows:

[0060] 1) When the data quality score is greater than 0.8, it indicates that the overall noise level of the data is low and the structure is clear. At this time, the subtle changes of the original signal should be preserved as much as possible. Therefore, a small filter window (window size = 5) and a high-order polynomial fitting (polynomial order = 3) are used to achieve high-precision fitting of local trends, improving the smoothness of the data without excessive smoothing.

[0061] 2) When the quality score is greater than 0.6 but not more than 0.8, it indicates that the data has some noise or slight fluctuations, and the smoothing effect needs to be enhanced. In this case, a medium-length filter window (window size = 7) is selected and combined with a second-order polynomial (polynomial order = 2) to ensure a certain fitting ability for curvature changes while improving the ability to suppress mid-frequency noise, thus achieving a good balance between smoothness and fidelity.

[0062] 3) When the quality score is less than or equal to 0.6, it indicates poor data quality and the potential presence of significant noise or anomalous perturbations. To avoid overfitting noise, a larger filtering window (window size = 9) is used, and a first-order linear polynomial (polynomial order = 1) is employed for fitting. While this configuration sacrifices some detail preservation, it effectively enhances the robustness of the filter, suppresses high-frequency noise, and prevents outliers from misleading subsequent modeling. The formula for its smoothed health data is:

[0063] (14),

[0064] The i-th health value after Savitzky-Golay smoothing, Smoothed_health_data[i], is obtained by a locally weighted average of the original health sequence H(·) at position i+j (j from ·). The observations from n to n) are summed using a weighted average. The weights c are... j The Savitzky-Golay filter coefficients are obtained by least-squares fitting of an a-order polynomial.

[0065] It is understandable that by constructing an adaptive sliding window parallel statistical calculation mechanism and combining it with a variance-gradient joint CUSUM change point detection model, accurate four-stage segmentation of device health time series data can be achieved. This method significantly improves the sensitivity and robustness of degradation inflection point identification and effectively supports the segmented modeling and dynamic switching of the prediction model.

[0066] S202. Based on an adaptive window, extract multiple statistics from the filtered time series data;

[0067] Specifically, we first construct a dynamically adjustable sliding window mechanism, where the window size ω is defined as:

[0068] (15),

[0069] Where, = Total length of health time series data. This design ensures that the basic resolution is retained under short sequences, and local features can still be efficiently extracted in long sequences. In the sliding process, for the data segment health_data[start i :end i ] in each window i, the following three key statistics are calculated in parallel:

[0070] Local mean: (used to reflect the central tendency of the current health level of the device)

[0071] (16),

[0072] Local variance: (quantifies the fluctuation intensity of the health state, identifies abnormal disturbances)

[0073] ) 2 (17),

[0074] Local gradient mean: (captures the rate of change of health, identifies the accelerating trend of degradation)

[0075] (18),

[0076] where, represents the gradient.

[0077] The above statistics are generated synchronously through a parallel computing framework, forming three auxiliary time series: local mean sequence {μ i}, local variance sequence { }, and local gradient sequence { }, providing multi-dimensional input for subsequent change point detection.

[0078] S203. Based on the plurality of statistics, a plurality of degradation transition points are determined by a variance-gradient joint CUSUM change point.

[0079] It can be understood that a dual-channel CUSUM type change point detector is constructed, which focuses on fluctuation mutations and degradation rate transitions respectively, and improves the discrimination reliability through fusion strategy.

[0080] Further, the determination of a plurality of degradation transition points based on the plurality of statistics by a variance-gradient joint CUSUM change point includes: determining the change point judgment threshold of the variance CUSUM channel and the gradient CUSUM channel respectively; based on the change point judgment threshold, marking a plurality of change points; mapping all candidate change points to a unified time axis, calculating the change point density of each time region and the weight of each candidate change point; based on the product of the change point density and the corresponding candidate change point weight, calculating the comprehensive score of each candidate change point; the candidate change point whose comprehensive score exceeds the preset threshold is determined as the degradation phase transition point.

[0081] Specifically, the variance CUSUM channel setting reference is the 75th percentile of the local variance sequence as a dynamic threshold. Change point determination condition:

[0082] and and (19),

[0083] The time point meeting the above condition is marked as a variance change point, indicating that the device enters an unstable running state.

[0084] 2) Gradient CUSUM channel setting reference: take the 75th percentile of the local gradient sequence as a reference level. Change point determination condition:

[0085] and (20),

[0086] The time point meeting the condition is marked as a gradient change point, indicating that the degradation process has significantly accelerated.

[0087] 3) Joint discrimination and score fusion mechanism: map the two types of change points to a unified time axis, and count the change point density in a sliding time window and intensity weighted score . (21),

[0088] where + =1, to construct a comprehensive change point score sequence = × , when exceeds the dynamic threshold (k is an adjustment coefficient), it is confirmed as an effective degradation phase transition point.

[0089] S204. Divide the degradation phase through the plurality of degradation transition points; based on the Akaike information criterion, adaptively select a degradation model for each degradation phase;

[0090] Specifically, based on the identified effective change point sequence, this step divides the device full life cycle health degree data into four typical degradation phases:

[0091] 1) Stable running phase (Stage I), characterized by: mean stationary, small variance, and gradient close to zero. Corresponding to the initial service period of the device, the performance is stable.

[0092] 2) Linear degradation phase (Stage II), characterized by: health degree showing an approximately linear slow decline, variance slightly rising, and gradient being a negative constant. Reflecting the normal wear process.

[0093] 3) Accelerated Degradation Stage (Stage III): Characterized by a significantly faster rate of health decline, a substantial increase in variance, and a continuous increase in the absolute value of the gradient. This indicates that the accumulation of internal damage has led to rapid deterioration.

[0094] 4) Stage IV (Critical Failure Stage): Characterized by a sharp decline in health, approaching the failure threshold, peaking variance and gradient, and the system nearing failure. The boundaries of each stage are determined by adjacent variable points. The algorithm automatically outputs the start and end time indices and duration of each stage, forming a structured description of the degradation path.

[0095] A model pool containing three degradation models—linear, exponential, and power functions—was constructed. For each degradation stage, the AIC (Akaike Information Criterion) was used for adaptive model selection to ensure optimal fitting results under different degradation characteristics.

[0096] 1) Linear degradation model: The linear model is suitable for the normal wear and tear stage of equipment, where the health status shows a stable linear downward trend. The formula is as follows: (twenty two),

[0097] Where 'a' represents the degradation slope (usually a negative value) and 'b' represents the initial health level. This model has few parameters and is simple to calculate, making it suitable for describing the uniform wear process of equipment.

[0098] The formula for calculating AIC is: (23), where n is the number of samples, MSE_linear is the mean square error of the linear model, and k_linear=2 is the number of model parameters.

[0099] 2) Exponential degradation model: This model specifically describes the accelerated degradation stage of equipment, where the rate of health decline increases exponentially over time. The formula is as follows:

[0100] (twenty four),

[0101] Where a represents the initial degradation magnitude, b represents the degradation acceleration coefficient, and c represents the asymptotic health lower limit. This model can effectively capture the nonlinear degradation characteristics of equipment after it enters the rapid degradation phase. The AIC calculation formula is: (25)

[0102] in =3 represents the number of parameters in the exponential model.

[0103] 3) Power Function Degradation Model: The power function model is suitable for describing the critical failure stage of equipment, where the health status decays rapidly according to a power law, as shown in the following formula:

[0104] (26),

[0105] where a is the degradation strength coefficient, b is the power exponent (usually less than 1), and c is the failure threshold offset. This model is particularly suitable for describing the sharp degradation behavior of the device when it is close to failure. The AIC calculation formula is:

[0106] (27),

[0107] where = 3 is the number of parameters of the power function model. On the basis of completing the degradation phase division, a candidate model pool containing linear, exponential and power function is constructed for each phase, and the AIC value of each model is calculated, and the minimum AIC is selected as the optimal degradation model of the phase, that is, best_model = argmin(AIC_linear, AIC_exp, AIC_power).

[0108] S205. Calculate the first residual life prediction value of the device through the selected degradation model.

[0109] Based on the selected optimal degradation model, the residual life prediction value RUL_degradation of the degradation model is calculated.

[0110] 1) Linear degradation model:

[0111] (28);

[0112] 2) Exponential degradation model:

[0113] (29);

[0114] 3) Power function degradation model:

[0115] (30),

[0116] where failure_threshold is the failure threshold, current_health is the current health, a is the negative degradation slope, and t_current is the current time point.

[0117] In step S300 of some embodiments of the present application, a second residual life prediction value of the device is calculated based on the physical mechanism of the device;

[0118] Specifically, the device health time series is converted into a quantifiable physical wear index, and the cumulative degradation degree of the device is reflected by the integral area, realizing the effective conversion from data-driven to physical mechanism. Health integral area calculation: (the total area under the health curve is calculated by using the improved trapezoidal integral method, which can accurately capture the time evolution characteristics of the health);

[0119] (31),

[0120] (32),

[0121] where health[i] is the health value at the ith moment, range [0,1]; Δt[i] is the time interval, supporting non-equidistant sampling; max_possible_area is the theoretical maximum area, the integral value when the health is always 1; area_ratio is the area ratio, reflecting the cumulative effect of the overall health level of the device.

[0122] Device type parameterized modeling: According to the physical characteristics and degradation mechanism of different devices, a parameterized life consumption model is established. Through a large amount of historical data statistical analysis, the characteristic parameters of each device type are determined, such as: the compressor is a reciprocating motion device, the mechanical wear is serious, the basic consumption coefficient is high, and the area influence index α=2.5 reflects its sensitivity to health degradation, that is, the compressor device can be set as k_base=0.15, α=2.5, β=1.2; The motor runs relatively smoothly, the basic consumption coefficient is low, and the degradation process is more linear, and the parameters are set as k_base=0.12, α=2.0, β=1.0, which reflects its gradual degradation characteristics. Among them, k_base is the basic consumption coefficient of the device, which reflects the inherent characteristics of the device type; α is the area influence index, which controls the influence degree of health degradation on consumption rate; β is the degradation rate correction coefficient, which adjusts the contribution of degradation trend to consumption rate.

[0123] Life consumption rate calculation:

[0124] ,

[0125] (33),

[0126] where area_ratio is the health integral area ratio, range [0,1], the closer to 1 indicates that the device is healthier; decline_rate is the current degradation rate, which is calculated by the health gradient; current_consumption: current cumulative consumption, range [0,1]; physical is the predicted remaining life, the unit is consistent with the time sequence.

[0127] The method realizes deep fusion of data driving and physical constraints through physical mechanism integral area calculation. The method can not only accurately quantify the degree of cumulative loss of equipment, but also adaptively adjust according to the physical characteristics of different equipment types, and significantly improve the explainability and engineering practicability of the prediction results. As an intuitive representation of physical loss, the integral area provides a clear quantitative basis for maintenance decision-making, effectively making up for the lack of physical constraints in pure data-driven methods.

[0128] In step S400 of some embodiments of the application, the construction of the Wiener random model based on the multi-sensor of the equipment includes:

[0129] S401. Constructing a first Wiener equation based on the initial health state of the equipment, the drift coefficient and the diffusion coefficient;

[0130] Specifically, the degradation process of the equipment is modeled as a Wiener process, the parameters of the Wiener process are estimated through the observed degradation data, and finally the remaining useful life is predicted. This method assumes that the degradation of the equipment is a random process that accumulates damage over time and eventually leads to failure. Compared with traditional deterministic models, the Wiener process model can better capture the randomness and uncertainty in the equipment degradation process. The formula is as follows:

[0131] , (34),

[0132] Where X(t) represents the health state of the equipment at time t, reflecting the current performance level of the equipment; X(0) represents the initial health level of the equipment, which is usually set to 1.0 to represent the perfect state of a new equipment; μ is the drift coefficient, which is used to describe the deterministic degradation trend of the health over time, and its value directly reflects the average degradation rate of the equipment; σ is the diffusion coefficient, which quantifies the intensity of random fluctuations in the degradation process, reflecting the uncertainty and randomness of the performance degradation of the equipment; W(t) is a standard Brownian motion process that satisfies the initial condition W(0)=0, and its increment ΔW(t) follows a normal distribution with mean 0 and variance Δt, i.e. ΔW(t)~N(0,Δt), which introduces a random disturbance term that conforms to the physical reality into the model.

[0133] S402. Correcting the drift coefficient and the diffusion coefficient based on the correlation coefficient of the sensor, the standardized data sequence, the reliability index and the weight coefficient;

[0134] Specifically, considering that the equipment is usually equipped with multiple sensors (such as vibration, temperature, pressure, etc.), this step designs a sensor fusion strategy based on correlation and reliability, which corrects the Wiener process parameters in real time through multi-sensor information, and finally generates a health prediction result considering randomness.

[0135] (35),

[0136] (36),

[0137] (37),

[0138] Where, ρ i Let be the correlation coefficient of the i-th sensor. This is the standardized data sequence of the i-th sensor; Let λ be the reliability index of the i-th sensor. i β i Weighting coefficients related to sensor type; Let Σ be the fusion weight of the i-th sensor, satisfying Σ i=1 m =1, where m is the total number of sensors. Then, sensor drift estimation is used for fusion drift correction. Sensor drift estimation is obtained by calculating the mean of the time derivatives of the data from each sensor, and the original drift parameters are corrected by combining a correlation adjustment factor to improve the model's adaptability.

[0139] Meanwhile, the diffusion parameter is adaptively adjusted to reflect the impact of random fluctuations in sensor data.

[0140] (38),

[0141] (39),

[0142] in, μ fused These are the drift parameters after fusion; For baseline drift estimation; The fusion strength coefficient; For the first The weight of each sensor; For the first Drift estimation for each sensor; For the first Correlation adjustment factor for each sensor; These are the diffusion parameters after fusion; For baseline diffusion (variance) estimation; This is the uncertainty amplification factor; This is an indicator of the overall uncertainty of the sensor.

[0143] S403. Based on the corrected drift coefficient and diffusion coefficient, construct the second Wiener equation.

[0144] Further, the third residual life prediction value of the computing device is calculated by the Wiener stochastic model, including: based on the second Wiener equation, the third residual life prediction value is calculated by the first passage time probability density function.

[0145] Specifically, the health degree prediction result considering randomness is generated based on the corrected Wiener process parameters. Through the calculation of the prediction mean and the prediction distribution, complete prediction information is provided. The residual life prediction is realized by using the first passage time probability density function, and the randomness and uncertainty in the device degradation process are fully considered.

[0146] (40),

[0147] Wherein, F_T(t) is the probability density function of the first passage time T_L, L is the failure threshold, and is used for residual life prediction.

[0148] In step S500 of some embodiments of the present application, the final life prediction value of the device is obtained by fusing the first residual life prediction value, the second residual life prediction value and the third residual life prediction value through the confidence degree, including:

[0149] S501. Obtain the confidence degree, fusion quality and quality score of each residual life prediction value; S502. Calculate the comprehensive confidence degree of the first residual life prediction value, the second residual life prediction value and the third residual life prediction value respectively through the confidence degree, the quality score and the fusion quality;

[0150] The data quality score Q k , the model confidence C k and the sensor fusion quality S k of the four kinds of heterogeneous prediction paths are calculated respectively; then the mapping function f is used to aggregate into the comprehensive confidence W k ; finally, the Softmax method is used for normalization, ensuring that the sum of weights is 1 and the monotonicity is preserved, that is, the higher the confidence, the greater the weight. The comprehensive confidence formula is as follows:

[0151] (41),

[0152] The normalized weight formula is:

[0153] (42),

[0154] Wherein, is a set of prediction methods; is a configurable non-negative monotonic function, such as linear weighting or neural network mapping; is the normalized weight of the kth method.

[0155] After completing the weight normalization, the system obtains four normalized weights that are calibrated by quality-confidence-sensor consistency .

[0156] S503. Based on each remaining life prediction value and the corresponding comprehensive confidence, a final life prediction value of the equipment is calculated.

[0157] Specifically, to avoid any single model producing misleading extreme values in the low-quality section, the present application adopts a "weighted linear fusion" strategy: the remaining life RUL_k and the health degree Health_k of each model output are weighted and summed according to w_k to obtain a globally unique fusion estimate RUL_fusion and Health_fusion.

[0158] The formula is as follows: the remaining life fusion value: (43); the health degree fusion value:

[0159] (44).

[0160] In particular, to quantify the uncertainty of the remaining life prediction without relying on any distribution assumption, the module performs the following standardization-resampling-confidence interval calculation process on the fusion result obtained above. The formula is as follows.

[0161] 1) Residual calculation: (45),

[0162] wherein, is the true remaining life of the i-th sample, in units of h; is the fusion prediction value output by the previous step corresponding to the historical sample, in units of h. After standardization, the obtained standardized residual has a mean of 0 and a variance of 1, and can be directly used for subsequent resampling.

[0163] 2) Bootstrap sample generation:

[0164] Set the resampling number B = 1000. For each loop b = 1, 2, …, B, the Bootstrap prediction value formula is:

[0165] (47),

[0166] wherein, is the fusion point estimate of the current to-be-predicted sample, in units of h; is the historical residual standard deviation, in units of h; is the b-th Bootstrap remaining life prediction, in units of h. After the loop is completed, the Bootstrap sample set The percentile method is used to directly read the confidence boundary without distribution assumption:

[0167] (48),

[0168] wherein represents the alpha percentile of the sample set R, in units of h. The system can schedule routine maintenance according to and prevent high-impact failures according to

[0169] Based on the health fusion value, five health states are classified (healthy, good, warning, failure, and serious failure), and the risk score is calculated by combining the residual life, uncertainty, and health, to divide into three risk levels (low, medium, and high), to generate a comprehensive report containing all prediction results, and to provide specific application guidance and suggestions for equipment maintenance decisions.

[0170] In one specific embodiment of the present application, the following steps are included:

[0171] S1: Obtain equipment health time series data, and perform five-dimensional data quality assessment: missing value check, outlier check, time regularity check, trend rationality check, and data range check, and calculate a comprehensive quality score to provide parameter basis for subsequent processing;

[0172] S2: Dynamically adjust the Savitzky-Golay filter parameters according to the data quality score, and perform smoothing processing on the health time series data to eliminate noise interference;

[0173] S3: Calculate local statistics in parallel within a sliding window, use variance-gradient joint CUSUM to distinguish degradation stages, and use AIC to adaptively select the best from the linear-exponential-power model pool;

[0174] S4: Calculate the health integral area based on the physical mechanism, calculate the life consumption rate combined with the device type parameters, and provide prediction results based on physical constraints;

[0175] S5: Establish a Wiener random process model, estimate the drift and diffusion parameters, fuse multi-sensor data for parameter correction, and provide prediction results considering randomness;

[0176] S6: Integrate linear extrapolation, degradation model, physical mechanism, and Wiener process, and perform weighted fusion prediction based on data quality and model confidence, and output prediction confidence interval in real time Bootstrap residual;

[0177] ​S7: Final prediction result output. The output results include: remaining useful life prediction value and confidence interval, prediction uncertainty quantification index, health state assessment (healthy / early warning / failure) and risk level (low / medium / high).

[0178] It can be understood that by accurately predicting the remaining useful life of the equipment and quantifying the prediction uncertainty, the present application can scientifically evaluate the future operating state and failure risk of the equipment, realize the transition from time-driven maintenance to state-driven maintenance, effectively optimize the maintenance plan and resource allocation, significantly reduce the risk of unexpected downtime, and improve the efficiency and economic benefits of the whole life cycle management of the equipment.

[0179] Embodiment 2

[0180] Reference Figure 3 , the second aspect of the present application provides a device life prediction device 1 based on multi-modal fusion, comprising: an acquisition module 11 for acquiring equipment health time series data; quality assessment based on the time series data, and calculating a comprehensive quality score; wherein the quality assessment includes data anomaly check, time check and health trend check; a first calculation module 12 for dynamically adjusting the Savitzky-Golay filter parameters based on the comprehensive quality score; based on the adjusted Savitzky-Golay filter parameters, the first remaining life prediction value of the equipment is calculated by CUSUM discrimination degradation and Akaike information criterion; a second calculation module 13 for calculating the second remaining life prediction value of the equipment based on the physical mechanism of the equipment; a third calculation module 14 for constructing a Wiener random model based on the multi-sensor of the equipment; the third remaining life prediction value of the equipment is calculated through the Wiener random model; a fusion module 15 for fusing the first remaining life prediction value, the second remaining life prediction value and the third remaining life prediction value by confidence fusion to obtain the final life prediction value of the equipment.

[0181] Further, the first calculation module 12 comprises: a filtering unit for filtering the time series data based on the adjusted Savitzky-Golay filter parameters; an extraction unit for extracting a plurality of statistics from the filtered time series data based on an adaptive window; a division unit for determining a plurality of degradation transition points by variance-gradient joint CUSUM change point based on the plurality of statistics; dividing the degradation stages through the plurality of degradation transition points; a calculation unit for adaptively selecting a degradation model for each degradation stage based on the Akaike information criterion; the first remaining life prediction value of the equipment is calculated through the selected degradation model.

[0182] Embodiment 3

[0183] Reference Figure 4In a third aspect, the present application provides an electronic device comprising: one or more processors; and a memory device storing one or more programs, when executed by the one or more processors, cause the one or more processors to implement the device life prediction method based on multi-modal fusion of the first aspect of the present application.

[0184] The electronic device 500 can include a processing device (e.g., a central processing unit, a graphics processing unit, etc.) 501 that can perform various appropriate actions and processes according to programs stored in a read-only memory (ROM) 502 or loaded into a random access memory (RAM) 503 from a memory device 508. Various programs and data required for the operation of the electronic device 500 are also stored in the RAM 503. The processing device 501, the ROM 502, and the RAM 503 are connected to each other through a bus 504. An input / output (I / O) interface 505 is also connected to the bus 504.

[0185] In general, the following devices can be connected to the I / O interface 505: input devices 506 including, for example, a touch screen, a touch pad, a keyboard, a mouse, a camera, a microphone, an accelerometer, a gyroscope, etc.; output devices 507 including, for example, a liquid crystal display (LCD), a speaker, a vibrator, etc.; a memory device 508 including, for example, a hard disk, etc.; and a communication device 509. The communication device 509 can allow the electronic device 500 to communicate wirelessly or wiredly with other devices to exchange data. Although Figure 4 The electronic device 500 is shown with various devices, but it should be understood that all of the shown devices are not required to implement or have the electronic device. More or less devices can alternatively be implemented. Figure 4 Each block shown in the flowchart of FIG. 6 can represent a device, or a plurality of devices, as necessary.

[0186] In particular, according to embodiments of the present disclosure, the processes described above with reference to the flowcharts can be implemented as a computer software program. For example, embodiments of the present disclosure include a computer program product comprising a computer program carried on a computer readable medium, the computer program containing program code for executing the methods illustrated by the flowcharts. In such embodiments, the computer program can be downloaded and installed from a network by the communication device 509, or installed from the storage device 508, or installed from the ROM 502. When the computer program is executed by the processing device 501, the above-mentioned functions defined in the methods of embodiments of the present disclosure are executed. It should be noted that the computer readable medium described in embodiments of the present disclosure can be a computer readable signal medium or a computer readable storage medium or any combination of the two. The computer readable storage medium may, for example, but is not limited to, an electronic, magnetic, optical, electromagnetic, infrared, or semiconductor system, device or apparatus, or any combination of the above. More specific examples of the computer readable storage medium can include, but are not limited to, an electrical connection having one or more wires, a portable computer diskette, a hard disk, a random access memory (RAM), a read-only memory (ROM), an erasable programmable read-only memory (EPROM or flash memory), an optical fiber, a portable compact disc read-only memory (CD-ROM), an optical storage device, a magnetic storage device, or any suitable combination of the above. In embodiments of the present disclosure, the computer readable storage medium can be any tangible medium that contains or stores a program that can be used by or in connection with an instruction execution system, apparatus or device. In embodiments of the present disclosure, the computer readable signal medium can include a data signal carried in a baseband or as part of a carrier wave, in which the computer readable program code is carried. Such a propagated data signal can take a variety of forms, including but not limited to an electromagnetic signal, an optical signal or any suitable combination of the above. The computer readable signal medium can also be any computer readable medium that is not a computer readable storage medium and that can communicate, propagate or transport a program for use by or in connection with an instruction execution system, apparatus or device. The program code contained on the computer readable medium can be transmitted by any suitable medium, including but not limited to a wire, an optical fiber, an RF (radio frequency) or the like, or any suitable combination of the above.

[0187] The computer readable medium described above can be included in the electronic device described above; or can exist separately from the electronic device and not be assembled into the electronic device. The computer readable medium described above carries one or more computer programs, which, when executed by the electronic device, cause the electronic device to:

[0188] Computer program code for carrying out operations of embodiments of the present disclosure can be written in any combination of one or more programming languages, including an object oriented programming language such as Java, Smalltalk, C++, Python, and conventional procedural programming languages, such as the "C" programming language or similar programming languages. The program code can execute entirely on the user's computer, partly on the user's computer, as a stand-alone software package, partly on the user's computer and partly on a remote computer, or entirely on the remote computer or server. In the latter scenario, the remote computer can be connected to the user's computer through any type of network, including a local area network (LAN) or a wide area network (WAN), or the connection can be made to an external computer (for example, through the Internet using an Internet Service Provider).

[0189] The computer program instructions can also be loaded onto a computer, other programmable data processing apparatus, or other devices to cause a series of operational steps to be performed on the computer, other programmable apparatus or other devices to produce a computer implemented process such that the instructions which execute on the computer or other programmable apparatus provide processes for implementing the functions / acts specified in the flowchart and / or block diagram block or blocks.

[0190] The foregoing is merely illustrative of the principles of the application, and the application should not be limited to such detail. Rather, the spirit and scope of the application are limited solely by the claims.

Claims

1. A method for predicting the service life of a device based on multi-modal fusion, characterized in that, The method comprises the following steps: obtaining equipment health timing data; performing quality assessment based on the timing data and calculating a comprehensive quality score; wherein the quality assessment comprises data anomaly checking, time checking and health trend checking; dynamically adjusting Savitzky-Golay filter parameters based on the comprehensive quality score; calculating a first residual life prediction value of the equipment based on the adjusted Savitzky-Golay filter parameters through CUSUM discrimination and Akaike information criterion; filtering the timing data based on the adjusted Savitzky-Golay filter parameters; extracting a plurality of statistics from the filtered timing data based on an adaptive window; determining a plurality of degradation transition points through variance-gradient joint CUSUM change points based on the plurality of statistics; dividing degradation stages through the plurality of degradation transition points; adaptively selecting a degradation model for each degradation stage based on Akaike information criterion; calculating the first residual life prediction value of the equipment through the selected degradation model; wherein determining a plurality of degradation transition points through variance-gradient joint CUSUM change points based on the plurality of statistics comprises: determining change point determination thresholds of variance CUSUM channels and gradient CUSUM channels respectively; marking a plurality of change points based on the change point determination thresholds; mapping all candidate change points to a unified time axis, calculating change point density of each time region and weight of each candidate change point; calculating a comprehensive score of each candidate change point based on the product of the change point density and the corresponding candidate change point weight; determining the candidate change points with the comprehensive score exceeding a preset threshold as the degradation stage transition points; calculating a second residual life prediction value of the equipment based on the physical mechanism of the equipment; constructing a Wiener random model based on the multi-sensor of the equipment; calculating a third residual life prediction value of the equipment through the Wiener random model; fusing the first residual life prediction value, the second residual life prediction value and the third residual life prediction value through confidence to obtain the final life prediction value of the equipment. 2.The method of claim 1, wherein, The method of constructing a Wiener random model based on the multi-sensor of the equipment comprises: constructing a first Wiener equation based on the initial health state, drift coefficient and diffusion coefficient of the equipment; correcting the drift coefficient and the diffusion coefficient based on the correlation coefficient, standardized data sequence, reliability index and weight coefficient of the sensor; constructing a second Wiener equation based on the corrected drift coefficient and diffusion coefficient. 3.The multi-modal fusion based equipment life prediction method of claim 2, wherein, The method of calculating a third residual life prediction value of the equipment through the Wiener random model comprises: calculating the third residual life prediction value through the first passage time probability density function based on the second Wiener equation. 4.The method of claim 1, wherein, The method of fusing the first residual life prediction value, the second residual life prediction value and the third residual life prediction value through confidence to obtain the final life prediction value of the equipment comprises: obtaining the confidence, fusion quality and quality score of each residual life prediction value; A comprehensive confidence of the first remaining life prediction value, the second remaining life prediction value and the third remaining life prediction value is calculated respectively through the confidence, the quality score and the fusion quality; A final life prediction value of the equipment is calculated by weighting based on each remaining life prediction value and the corresponding comprehensive confidence.

5. A device life prediction apparatus based on multi-modal fusion, characterized by, It comprises: An acquisition module is configured to acquire time series data of equipment health; Quality assessment is performed based on the time series data, and a comprehensive quality score is calculated; wherein the quality assessment comprises data anomaly check, time check and health trend check; A first calculation module is configured to dynamically adjust Savitzky-Golay filtering parameters based on the comprehensive quality score; calculate a first remaining life prediction value of the equipment based on the adjusted Savitzky-Golay filtering parameters through CUSUM degradation discrimination and Akaike information criterion; filter the time series data based on the adjusted Savitzky-Golay filtering parameters; extract a plurality of statistics from the filtered time series data based on an adaptive window; determine a plurality of degradation transition points through variance-gradient joint CUSUM change points based on the plurality of statistics; divide degradation stages through the plurality of degradation transition points; adaptively select a degradation model for each degradation stage based on Akaike information criterion; calculate the first remaining life prediction value of the equipment through the selected degradation model; wherein determining a plurality of degradation transition points through variance-gradient joint CUSUM change points based on the plurality of statistics comprises: determining change point judgment thresholds of variance CUSUM channel and gradient CUSUM channel respectively; marking a plurality of change points based on the change point judgment thresholds; mapping all candidate change points to a unified time axis, calculating change point density of each time region and weight of each candidate change point; calculating a comprehensive score of each candidate change point based on the product of the change point density and the corresponding candidate change point weight; determining the candidate change points whose comprehensive scores exceed a preset threshold as the degradation stage transition points; A second calculation module is configured to calculate a second remaining life prediction value of the equipment based on physical mechanism of the equipment; A third calculation module is configured to construct a Wiener random model based on multiple sensors of the equipment; calculate a third remaining life prediction value of the equipment through the Wiener random model; A fusion module is configured to fuse the first remaining life prediction value, the second remaining life prediction value and the third remaining life prediction value through confidence to obtain a final life prediction value of the equipment.

6. An electronic device, comprising: One or more processors; A storage device is configured to store one or more programs, which make the one or more processors implement a device life prediction method based on multi-modal fusion when the one or more programs are executed by the one or more processors.

7. A computer readable medium having stored thereon a computer program, wherein, The computer program is executed by the processor to implement a device life prediction method based on multi-modal fusion as claimed in any one of claims 1 to 4.

Images