CN-120805064-B - Digital twinning-combined multi-mode equipment dimension checking knowledge intelligent recommendation system

CN120805064BCN 120805064 BCN120805064 BCN 120805064BCN-120805064-B

Abstract

The invention discloses an intelligent recommendation system for multi-mode equipment maintenance and inspection knowledge combined with digital twinning, and belongs to the technical field of equipment maintenance and inspection; the method is used for solving the technical problem that the static recommendation strategy and the global edge cloud collaborative optimization are difficult to consider in the existing scheme, through synchronous acquisition of multimode data driven by digital twin, a long-short-term memory network model fused with an attention mechanism is subjected to strong correlation feature screening and dynamic weighting based on mutual information entropy through an attenuation rate deviation weighting loss function, a knowledge graph with physical entity correlation precision and causal reasoning capacity is built by utilizing a time-varying feature matrix to capture the evolution rule of the state of equipment along with time, full-link interpretability from data, features, entities, causal and decision is realized, and the contradiction between real-time processing requirements of high-frequency data and global knowledge graph dependence in industrial equipment maintenance can be solved by the three-layer framework of edge-end high-frequency response-cloud global optimization-federal learning parameters.

Inventors

XIA YUAN
GONG JINPENG
SUN QUNLI
DU LIN
ZHANG YUCHAO
ZHANG XINXIN
LI SHUWEI

Assignees

江阴怡源智信运维技术股份有限公司

Dates

Publication Date: 20260505
Application Date: 20250902

Claims (10)

1. An intelligent recommendation system for multi-modal equipment dimension inspection knowledge combined with digital twinning is characterized by comprising: The multi-modal feature processing module predicts the residual life of the key component through a long-short-term memory network based on the equipment operation data of the digital twin body in real time and extracts vibration spectrum, temperature gradient and image texture features which are strongly associated with life attenuation, and constructs a time-varying feature matrix; wherein, constructing an LSTM network integrating an attention mechanism, inputting the LSTM network into a preprocessed multi-mode time sequence characteristic, and outputting the LSTM network into a residual life predicted value ; The loss function is designed as: Wherein MSE () is a mean square error function; Weighting coefficients for MSE loss terms; the weight coefficient is the attenuation rate deviation term; To predict the attenuation rate With true decay rate Is used for the deviation of (a), Representing the rate of life decay; the service life is really remained for the equipment; And calculating mutual information entropy of each modal feature and RUL F is a single characteristic variable in the multi-mode time sequence characteristic set, and screening is carried out Strong correlation characteristics of not less than 0.8; The causal relation enhancement processing module is used for fusing predictive features and historical maintenance data to construct a double-layer knowledge graph containing a fault causal link, wherein the bottom layer is a physical entity association layer, the upper layer is a causal reasoning layer, and causal strength among entities is quantified through a causal Bayesian network; The mixed recommendation strategy dynamic analysis module is used for dynamically adjusting the weight duty ratio of causal reasoning and collaborative filtering based on the content by adopting reinforcement learning based on the current RUL value and the knowledge map causal path of the equipment, analyzing the RUL value and dynamically triggering a causal reasoning priority mode, and carrying out active supervision evaluation and management control on the implementation effect of the dynamic triggering causal reasoning priority mode based on the calculated recommendation adoption degree; The recommendation acceptance degree is obtained by calculating the ratio of the total number of effective recommendation acceptance to the total number of ineffective recommendation acceptance, wherein the effective recommendation acceptance refers to the selection of the recommendation items in the candidate recommendation list; Data analysis is carried out on the recommended adoption degree, when the implementation effect of the dynamic triggering causal reasoning priority mode is determined, if the recommended adoption degree is larger than or equal to a recommended adoption threshold value, the implementation effect of the dynamic triggering causal reasoning priority mode is judged to be normal, and the application of the existing priority mode is maintained; if the recommended adoption degree is smaller than the recommended adoption threshold, judging that the implementation effect of the dynamic triggering causal reasoning priority mode is abnormal, and optimizing and adjusting the application of the existing priority mode; The edge cloud collaborative optimization module is used for processing the high-frequency sensor data in real time and generating preliminary recommendation, and the cloud performs causal path mining of the knowledge graph and recommendation model parameter optimization based on the global dimension detection data, so that parameter synchronization is realized through federal learning.
2. The intelligent recommendation system for multi-modal device dimension inspection knowledge in combination with digital twinning according to claim 1, wherein time sequence sensing data, visual data and text data are collected based on real-time data mapping of digital twinning and physical devices; preprocessing and feature enhancement are carried out on the acquired multi-mode data to obtain a multi-mode time sequence feature set; The method comprises the steps of constructing a long-short-time memory network integrating an attention mechanism, inputting the long-short-time memory network into a preprocessed multi-mode time sequence feature, outputting the multi-mode time sequence feature into a residual life predicted value, and generating a full life cycle fault data training model through digital twin simulation.
3. The intelligent recommendation system for multi-modal device dimension inspection knowledge in combination with digital twinning according to claim 2, wherein the time window is used for constructing the time-varying feature matrix The weighted feature vectors are arranged according to time sequence as units to construct a time-varying feature matrix Wherein, the method comprises the steps of, N is the number of matrix columns; is the initial point in time.
4. The intelligent recommendation system for multi-modal equipment dimension inspection knowledge combining digital twinning according to claim 3, wherein the physical entities and the association relations are identified based on the fusion data, and an entity-relation network is constructed through entity extraction, association relation definition and entity association matrix construction.
5. The intelligent recommendation system for multi-modal device dimension inspection knowledge in combination with digital twinning according to claim 4, wherein causal reasoning layers are constructed based on causal bayesian networks to mine causal links of faults and quantify causal strength among entities.
6. The intelligent recommendation system for multi-modal device dimension inspection knowledge in combination with digital twinning according to claim 5, wherein the average causal effect ACE is calculated by a conditional probability table of a causal bayesian network, and the formula is: Wherein, the method comprises the steps of, For intervening operations, representing mandatory cause variables Occurring and taking it to a value of 1; representing the forced cause variable Occurring and taking it to a value of 0; To intervene in Time result variable Probability of occurrence; To intervene in Time result variable Probability of occurrence; Is causal intensity.
7. The intelligent recommendation system for multi-modal device dimension inspection knowledge combining digital twinning according to claim 6, wherein when the double-layer map is fused, the upper-layer causal variable is associated with the bottom-layer physical entity, and each batch of dimension inspection data or prediction features are added, the causal strength is updated by the following formula: Wherein, the method comprises the steps of, Is a historical weight coefficient; The causal strength calculated before the data is newly added; for the causal strength calculated based on the newly added data.
8. The intelligent recommendation system for multi-modal device dimension inspection knowledge in combination with digital twinning as claimed in claim 7, wherein a reward function is defined Quantifying the effect of the recommendation strategy under the current weight combination, and guiding the agent to learn the optimal weight adjustment strategy: ; Wherein, the The fault prediction accuracy of the recommended maintenance strategy is calculated; Normalizing the value for maintenance cost; To indicate the function when When 1 is taken, otherwise 0 is taken, For the current predicted remaining life-time, A threshold value for triggering a causal inference priority mode; is causal path intensity; are weight coefficients.
9. The intelligent recommendation system for multi-modal device dimension inspection knowledge in combination with digital twinning according to claim 8, wherein the time-varying feature matrix is extracted based on compressed data Calling a lightweight CB model to generate a recommendation: Wherein, the method comprises the steps of, The weights are fixed at the edge ends of the causal reasoning model and the collaborative filtering model respectively; Causal reasoning similarity and collaborative filtering scores of the kth recommendation item are respectively obtained.
10. The intelligent recommendation system for multi-mode equipment dimension inspection knowledge combining digital twinning according to claim 9, wherein when cloud recommendation model parameter optimization and federation learning synchronization are carried out, the cloud optimizes recommendation model parameters based on a global knowledge-graph, and sends the optimized parameters to edge nodes through federation learning, and cloud-edge parameter synchronization is achieved through FedProx algorithm.

Description

Digital twinning-combined multi-mode equipment dimension checking knowledge intelligent recommendation system Technical Field The invention relates to the technical field of equipment maintenance and inspection, in particular to an intelligent recommendation system for multi-mode equipment maintenance and inspection knowledge combining digital twinning. Background Along with the improvement of the intelligent degree of industrial equipment, predictive maintenance (PHM) becomes a core technology for guaranteeing the safe operation of the equipment. When the technical scheme is implemented, the defects that the real-time performance and the global optimization are difficult to be achieved due to poor fault tracing interpretability and edge cloud data processing and splitting caused by insufficient multi-mode feature fusion and residual life RUL prediction precision and lack of causal reasoning capability of a knowledge graph exist. Aiming at the defects, the invention provides an intelligent recommendation system for multi-mode equipment dimension checking knowledge combining digital twinning, which is used for solving the defects existing in the prior art. Disclosure of Invention The invention aims to provide a digital twinning-combined multi-mode equipment dimension checking knowledge intelligent recommendation system which is used for solving the technical problem that the stationarity of a recommendation strategy and the cooperative global optimization of an edge cloud are difficult to consider in the existing scheme. The aim of the invention can be achieved by the following technical scheme: an intelligent recommendation system for multi-modal device dimension inspection knowledge in combination with digital twinning, comprising: The multi-modal feature processing module predicts the residual life of the key component through a long-short-term memory network based on the equipment operation data of the digital twin body in real time and extracts vibration spectrum, temperature gradient and image texture features which are strongly associated with life attenuation, and constructs a time-varying feature matrix; wherein, constructing an LSTM network integrating an attention mechanism, inputting the LSTM network into a preprocessed multi-mode time sequence characteristic, and outputting the LSTM network into a residual life predicted value The loss function is designed as: Wherein MSE () is a mean square error function, ω 1 is a weight coefficient of MSE loss term, ω 2 is a weight coefficient of attenuation rate deviation term, Δλ=λ pred-λtrue is a deviation of predicted attenuation rate λ pred and real attenuation rate λ true, RUL true is the true remaining lifetime of the device; The causal relation enhancement processing module is used for fusing predictive features and historical maintenance data to construct a double-layer knowledge graph containing a fault causal link, wherein the bottom layer is a physical entity association layer, the upper layer is a causal reasoning layer, and causal strength among entities is quantified through a causal Bayesian network; The mixed recommendation strategy dynamic analysis module is used for dynamically adjusting the weight duty ratio of causal reasoning and collaborative filtering based on the content by adopting reinforcement learning based on the current RUL value and the knowledge map causal path of the equipment, analyzing the RUL value and dynamically triggering a causal reasoning priority mode, and carrying out active supervision evaluation and management control on the implementation effect of the dynamic triggering causal reasoning priority mode based on the calculated recommendation adoption degree; The edge cloud collaborative optimization module is used for processing the high-frequency sensor data in real time and generating preliminary recommendation, and the cloud performs causal path mining of the knowledge graph and recommendation model parameter optimization based on the global dimension detection data, so that parameter synchronization is realized through federal learning. Preferably, the time sequence sensing data, the visual data and the text data are collected based on real-time data mapping of the digital twin body and the physical equipment; preprocessing and feature enhancement are carried out on the acquired multi-mode data to obtain a multi-mode time sequence feature set; The method comprises the steps of constructing a long-short-time memory network integrating an attention mechanism, inputting the long-short-time memory network into a preprocessed multi-mode time sequence feature, outputting the multi-mode time sequence feature into a residual life predicted value, and generating a full life cycle fault data training model through digital twin simulation. Preferably, when constructing the time-varying feature matrix, the weighted feature vectors are arranged in time sequence with the time window Δt as a unit, and the time-varying feature