Equipment intelligent health state evaluation method

Abstract

The invention discloses an equipment intelligent health state evaluation method, which comprises the following steps: S1, collecting data, and preprocessing multi-source heterogeneous data to form a standardized relational data set; s2, performing multi-index gradient assignment, and performing assignment on core indexes through reference value setting and a linear interpolation method; s3, carrying out comprehensive weight distribution on each core index by adopting a mixed weight distribution mechanism; and S4, constructing a CNN-LSTM hybrid prediction model, extracting key time sequence features in the equipment production operation data through a CNN layer, processing time sequence data through an LSTM layer, learning a long-term dependency relationship, predicting a future equipment health state score, and providing a maintenance suggestion. By constructing a multi-dimensional and dynamic equipment health state scoring system, the problems that a traditional evaluation method is high in subjectivity and insufficient in quantification are solved, transformation from passive maintenance to predictive maintenance is achieved, meanwhile, a visual health report can be generated, and a manager can be assisted in rapidly positioning weak links of equipment and making an optimization strategy.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of tobacco equipment management, and in particular to an intelligent health state evaluation method for equipment. BACKGROUND

[0002] The traditional tobacco equipment management method has the following core problems: (1) Single evaluation means: relying on manual experience or single sensor threshold alarm, only focusing on single-point exceptions of equipment operation parameters (such as temperature, vibration, etc. single index), ignoring multi-dimensional parameter coordination effects such as vibration frequency, energy consumption fluctuation, production plan, environmental parameter, process parameter, etc.

[0003] (2) Extensive maintenance strategy: adopting periodic maintenance (PM) or breakdown maintenance (BM), such passive maintenance mode is easy to cause over-maintenance or sudden stop, increasing cost and reducing production efficiency. And the maintenance decision depends on the experience of engineers, lacking data-driven support, and the fault positioning time is long.

[0004] (3) Lack of dynamic scoring system: the equipment health state lacks a unified quantitative standard, only judging the equipment state as "normal / abnormal" binary, which cannot accurately reflect the equipment degradation degree and potential risk. For example, the cutter wear of the forming machine is only judged by manual visual inspection whether to replace, which cannot quantify the specific influence weight of wear degree on production efficiency, and the scoring system does not consider dynamic factors such as equipment aging and process adjustment, resulting in distorted health score of old equipment.

[0005] In the prior art, some enterprises try to introduce machine learning algorithm, but there are the following limitations: Data preprocessing is not sufficient, which makes the model training data deviate from the real scene, the model parameter weight is fixed in the early stage of model training, which is difficult to adapt to the actual situation of complex and variable equipment operation conditions, affecting the model precision and application; the evaluation model relies on a single algorithm, which cannot adapt to the complex nonlinear relationship between equipment health state and multiple parameters, and cannot quantify the influence weight of each index on health score, affecting the maintenance decision efficiency and being difficult to early warn the equipment degradation trend.

[0006] In view of the above problems, the present application aims to propose an intelligent health state evaluation method which fuses multi-dimensional data, index gradient scoring, self-adaptive dynamic weight distribution and CNN-LSTM hybrid prediction model, realizes the intelligent upgrading of equipment whole life cycle management. SUMMARY

[0007] In order to overcome the above problems, the application provides a device intelligent health state evaluation method. The application is a cigarette factory device intelligent health state evaluation method based on multi-dimensional data fusion and deep learning model, which is suitable for real-time state monitoring, health score quantification, fault trend prediction and maintenance decision optimization of key production equipment such as cigarette making machines, packaging machines and forming machines in the tobacco industry, and can promote the transformation of equipment management from 'experience driven' to 'data driven'.

[0008] The technical scheme adopted by the application to solve its technical problems is: The device intelligent health state evaluation method comprises S1 data fusion and preprocessing Data acquisition: integrate multi-dimensional data in MES system, production efficiency early warning system, energy management system and NC system, and acquire device production operation parameters in real time; The data range includes daily output, product quality, energy consumption, material consumption, repair record, repair time proportion, downtime number, locomotive operation process index, rejection amount, internal and external downtime, waiting time and the like, and each item of data is recorded in each system database according to date, shift, brand number, NC, machine type and the like, and the form includes signal, text work order, image, video and the like; The multi-source heterogeneous data is preprocessed, such as data cleaning, alignment and feature extraction, to form a standardized relational data set; Data cleaning: abnormal values are removed by using isolation forest algorithm, and missing values are filled by using linear interpolation method, so as to ensure data continuity; Feature engineering: the standardized data set is constructed by rotation conversion and label classification, and the training set, the validation set and the test set are divided; S2 multi-index gradient scoring Core index: screening the core index in the collected device production operation parameters, including device daily output, product quality, material consumption, locomotive operation process index, repair time proportion and energy consumption; Gradient scoring: Reference value setting: based on historical data and process standard, the optimal interval and warning interval of each core index are determined; Interval division and interpolation: the core index value is mapped to the 0-100 interval by using linear interpolation method, and the out-of-limit value is directly scored as 0 or full score; Dynamic adjustment: according to the aging degree of the device and the process upgrading demand, the core index reference interval is updated regularly; An exponential decay function is constructed with the device service time t as a variable to automatically adjust the index score reference value:

[0009] Among them, The baseline value of a certain core index in the initial ideal state of the equipment, The equipment attenuation coefficient set according to the type of the equipment and experience, The cumulative service life or operating hours; Alternatively, by clustering analysis of historical health scores and core index data, a phased scoring standard is constructed, and the reference value is switched according to the aging state of the equipment; S3 comprehensive weight distribution Objective weight: The information entropy and dispersion of each index data are calculated by entropy weight method, and the index with greater variation degree is given higher weight; Subjective weight: The judgment matrix is constructed by AHP method, and the relative importance between indexes is determined according to expert experience; Comprehensive weight: According to the actual production demand, the weights of entropy weight method and AHP method are weighted and fused, and then the weights of each index are comprehensively adjusted; Specifically, the weight vectors generated by entropy weight method and AHP method are fused by using weighted superposition model: ; Wherein, The subjective weight calculated by AHP method, The objective weight calculated by entropy weight method, It can be set statically according to the actual target concerned by the enterprise or optimized adaptively based on the reward function.

[0010] S4 construct CNN-LSTM hybrid prediction model The prediction model extracts key time sequence features in the equipment production and operation data through CNN, processes time series data through LSTM, learns long-term dependence relationship, and is used to predict future equipment health state score and provide maintenance suggestions; The hybrid prediction model includes CNN layer: used to extract short-term time sequence features in equipment production and operation index data and construct spatial feature representation; LSTM layer: based on time sequence feature sequence to further capture long-term dependence relationship in time dimension and predict health state score in future period; At the same time, attention mechanism is introduced into the model structure, which dynamically focuses on key indicators and time nodes that have greater influence on health state score prediction, and combines historical failure mode association to strengthen the mapping relationship between key features and failures through attention weight; Input data: time series data of historical health score, daily output, product quality, material consumption, locomotive operation process index and other core indexes; Output result: for a future time point, the health score sequence is formed by weighting and summing the predicted values of each index and the dynamic weight, that is, the health score prediction value; the score prediction value is associated with the historical failure mode to provide maintenance optimization suggestions; S5 generate a visual health report The health report content includes: this month's health score overview, health trend prediction curve, weak link analysis, maintenance suggestions and optimization measures.

[0011] The present application brings the following beneficial effects: The present application solves the problem of strong subjectivity and insufficient quantification of traditional evaluation methods by constructing a multi-dimensional and dynamic equipment health score system. Through the mixed weight distribution mechanism (entropy weight method + analytic hierarchy process + dynamic adjustment), the scientificity and applicability of index weighting are improved, the CNN-LSTM hybrid prediction model is used to calculate the equipment health score through multi-dimensional feature fusion, predict the equipment health trend, and use the explainability algorithm to quantify the influence weight of each factor on the score, combine the equipment life cycle state and business experience rules, realize the change from "passive maintenance" to "predictive maintenance", and at the same time generate a visual health report to assist managers to quickly locate the weak link of the equipment and develop optimization strategies. BRIEF DESCRIPTION OF DRAWINGS

[0012] In order to more clearly illustrate the technical solutions in the embodiments of the present application, the following will briefly introduce the drawings needed to be used in the embodiments or prior art description. Obviously, the drawings in the following description are only some embodiments of the present application, and for those skilled in the art, other drawings can be obtained without creative labor.

[0013] Figure 1 It is a flowchart of the evaluation method. Figure 2 It is a schematic diagram of the evaluation system framework. DETAILED DESCRIPTION

[0014] The technical solutions in the embodiments of the present application will be described clearly and completely below with reference to the drawings in the embodiments of the present application. Obviously, the described embodiments are only some embodiments of the present application, not all. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative labor are within the scope of protection of the present application.

[0015] Moreover, the following description is to be understood as not limiting the only embodiments of the application to particular examples described as such, but the intention is to cover all possible embodiments covered by the scope of the application. Numerous specific details of the application are described below in order to provide a thorough understanding. However, it will be apparent to those skilled in the art that the application can be practiced without these specific details. In other instances, well-known methods, procedures, components, and circuits have not been described in detail so as not to unnecessarily obscure aspects of the application.

[0016] The first embodiment of the present application relates to a method for evaluating the intelligent health status of equipment, referring to Figure 1 , comprising S1 data fusion and preprocessing Data collection: by integrating multi-dimensional data in MES system, production efficiency early warning system, energy management system and NC system, real-time acquisition of equipment production operation parameters, data range including daily output, product quality, energy consumption, material consumption, repair record, repair time proportion, downtime number, locomotive running process index, rejection amount, internal and external downtime, waiting time, etc., each data is recorded in the database of each system according to date, shift, brand number, NC, machine type, etc. Marking, including signal (time domain signal), text work order (semantic data), image (image data), video, etc.

[0017] Cleaning, aligning and feature extraction are performed on multi-source heterogeneous data to form a standardized relational data set; Data cleaning: using isolation forest algorithm to remove outliers, using linear interpolation method to fill in missing values, to ensure data continuity; Feature engineering: through rotation conversion, label classification (such as NC coding) to build a standardized data set, and divide the training set, validation set and test set. In this embodiment, the proportion is 6:2:2; S2 multi-index gradient scoring Core index: screening the core index of the collected equipment production operation parameters, including equipment daily output, product quality, material consumption, locomotive running process index, repair time proportion, energy consumption; Gradient scoring: Baseline value setting: based on historical data and process standards, determine the optimal interval of each core index, such as daily output standard value ± 5%, warning interval, such as daily output overproduction 150% or underproduction 30%; Interval division and interpolation: using linear interpolation method to map the core index value to 0-100 interval, and directly assign 0 or full score to the out-of-limit value; Dynamic adjustment: according to the aging degree of equipment and the demand of process upgrading, regularly update the baseline interval of core index; Specifically, an exponential decay function is constructed with the device service time t as the variable, and the benchmark value of the index score (daily output, material consumption, repair time proportion, and locomotive operation process index) is automatically adjusted:

[0018] wherein, is the benchmark value of a certain core index under the initial ideal state of the device, is the device decay coefficient set according to the device type and experience, is the cumulative service life or running hours.

[0019] In addition, dynamic adjustment can also be achieved by clustering analysis of historical health scores and core index data to construct phased scoring standards, and combining device aging state switching to assign reference values.

[0020] S3 Comprehensive Weight Allocation Objective weight: The information entropy and dispersion of each core index data are calculated using the entropy weight method, and the index with greater variation degree is given higher weight; Subjective weight: The judgment matrix is constructed using the analytic hierarchy process, and the relative importance between indexes is determined according to expert experience, such as repair time proportion weight > energy consumption weight; Comprehensive weight: According to the actual production demand, the weights of the entropy weight method and the analytic hierarchy process are weighted and fused to comprehensively adjust the weights of each index; Specifically, the weights generated by the entropy weight method and the analytic hierarchy process are fused by considering the variability of objective indexes and expert experience knowledge. The fusion strategy uses a weighted superposition model: ; wherein, is the subjective weight calculated by the analytic hierarchy process, is the objective weight calculated by the entropy weight method, The actual target of the enterprise can be set statically or optimized adaptively based on the reward function, such as energy saving priority, maximum production capacity, and minimum maintenance cost. For example, taking the device maintenance cost and downtime loss as the reward function, R = −(maintenance cost + 10 × unplanned downtime), the coefficient is dynamically optimized by training to maximize the long-term cumulative reward. At the same time, the influence of each index on the score can be quantified using explainable technology (such as SHAP), and the weight distribution logic can be further refined manually according to the actual production needs.

[0021] S4 Constructing a CNN-LSTM Hybrid Prediction Model The prediction model extracts key time sequence features in the equipment production operation data through a CNN (convolutional neural network), processes time sequence data through an LSTM (long short-term memory network), learns long-term dependencies, and is used to predict future equipment health status scores and provide maintenance recommendations. The hybrid prediction model includes A CNN layer is used to extract short-term time sequence features in the equipment production operation index data, such as daily production abnormal fluctuations, periodic energy consumption increases, etc., and to construct spatial feature representations. The CNN layer adopts a lightweight separated convolution structure optimization, balancing model accuracy and deployment efficiency. Through multiple 1D convolution kernels, the time sequence is operated in a sliding manner, outputting a local feature vector sequence in the form of wherein is the compressed time step, is the number of channels. The output feature sequence is used as input for the LSTM layer for subsequent health score trend modeling.

[0022] An LSTM layer is used to further capture long-term dependencies in the time dimension based on the feature sequence, and to predict health status scores in the future period.

[0023] At the same time, an attention mechanism is introduced into the model structure to avoid equal processing of the importance of multiple features, and to dynamically focus on key indicators and time nodes that have a greater impact on health status score prediction. For example, the weight of the energy consumption parameter during high-load production is increased, and the weight of the vibration parameter of the locomotive running indicator during the aging stage of the equipment is increased, improving the interpretability and accuracy of the model. Further, the historical failure mode correlation can be combined to strengthen the mapping relationship between key features and failures through attention weights.

[0024] Input data: time series data of historical health scores, daily production, product quality, material consumption, and locomotive running process indicators; Output results: for a future time point, the health score sequence is formed by weighting and summing the predicted values of each indicator and the dynamic weights, i.e., the health score prediction value; the score prediction value is associated with business experience, i.e., historical failure mode correlation, and when the score is abnormal (such as continuous decline or falling below the threshold), the maintenance recommendation reasoning module is triggered to locate the problem source based on indicator deviation ordering and to provide maintenance optimization recommendations based on indicator-failure correlation.

[0025] For example, if the "vehicle repair time ratio" is abnormal, the recommendation is to "check repair records and arrange preventive maintenance scheduling".

[0026] In the model training process, the mixed prediction model first inputs the time series data of the normalized historical health score, daily output, product quality, material consumption, locomotive operation process index, repair time proportion, energy consumption, etc. as input into the mixed prediction model composed of a lightweight CNN layer and an LSTM layer. The CNN layer identifies the "short-term pattern" features such as mutations, periodic fluctuations or trend inflection points of each index in the time series, extracts local time sequence features, and then passes the feature vectors to the LSTM layer to learn long-term dependencies, thereby predicting the health score trend in a future time period. In the training process, mean square error (MSE) is used as the loss function, the optimizer is Adam, the initial learning rate is 0.001, the training rounds are set to 100 rounds, the Early Stopping strategy is used to avoid overfitting, and the training set, validation set and test set ratio is 6:2:2.

[0027] Finally, the model can output the health status score prediction value and generate maintenance recommendations combined with the index change trend.

[0028] S5 generates a visual health report The health report content includes: Monthly health score overview: show the health score and index score through a radar chart; Health trend prediction curve, such as a 30-day score trend chart; Weak link analysis, such as excessive material consumption, excessive repair time proportion, etc. Maintenance recommendations and optimization measures, determine the priority and focus of maintenance recommendations in combination with the index weight, equipment life cycle stage, score threshold / drop, etc.

[0029] The second embodiment of the present application relates to a device intelligent health status evaluation system, referring to Figure 2 , comprising A data fusion and preprocessing module is used to integrate multi-dimensional data in MES systems, production efficiency early warning systems, energy management systems and NC systems, to obtain device production operation parameters in real time, and to preprocess multi-source heterogeneous data to form a standardized relational data set. A multi-index gradient scoring module is used to screen core indicators in the collected device production operation parameters, including device daily output, product quality, material consumption, locomotive operation process index, repair time proportion, energy consumption; and to perform gradient scoring on each core indicator. A comprehensive weight allocation module is used to obtain objective weights and subjective weights of each core indicator for device health status evaluation, and to dynamically adjust the index weights using a weighted superposition model according to actual production needs. A mixed prediction model construction module, the mixed prediction model comprises CNN layer: used to extract short-term time series features in equipment production operation index data, output local feature vector sequence as input of LSTM layer; LSTM layer: further capture long-term dependencies in time dimension based on the feature sequence, predict health status score in future period. Introduce attention mechanism in model structure, avoid equal treatment of multiple feature importance.

[0030] Input data: time series data of historical health score, daily output, product quality, material consumption, locomotive operation process index, etc. Output result: health score prediction value; maintenance optimization suggestion.

[0031] Visual health report generation module, used to generate health report according to output of hybrid prediction model; its content can include: this month health score overview, health trend prediction curve, weak link analysis, maintenance suggestion and optimization measures, etc.

[0032] The above examples are only used to illustrate the technical solutions of the present application, but not to limit it; although the present application has been described in detail with reference to the foregoing examples, those skilled in the art should understand that they can still modify the technical solutions recorded in the foregoing examples, or make equivalent replacement for part of the technical features; and these modifications or replacements do not make the essence of the corresponding technical solutions deviate from the spirit and scope of the technical solutions of the embodiments of the present application, and should be included in the protection scope of the present application.

Claims

1. A method for evaluating the intelligent health status of equipment, characterized by: include S1 data fusion and preprocessing Data collection: Integrate multi-dimensional data from the MES system, production efficiency warning system, energy management system, and NC system to obtain equipment production and operation parameters in real time; Preprocess multi-source heterogeneous data to form standardized relational data sets; S2 multi-index gradient scoring Core indicators: Filter the core indicators of the collected equipment production and operation parameters, including equipment daily output, product quality, material consumption, locomotive operation process indicators, repair time ratio, and energy consumption; Gradient scoring: Benchmark value setting: Based on historical data and process standards, determine the optimal range and warning range of each core indicator; Interval division and interpolation: Linear interpolation is used to map the core indicator values ​​to the range of 0-100 points, and values ​​exceeding the limit are directly assigned 0 points or full marks; Dynamic adjustment: regularly update the core indicator benchmark values ​​based on equipment aging and process upgrade requirements; S3 comprehensive weight distribution Objective weight: The entropy weight method is used to calculate the information entropy and dispersion of each indicator data, and indicators with greater variation are given higher weights; Subjective weight: A judgment matrix is ​​constructed using the analytic hierarchy process, and the relative importance of indicators is determined based on expert experience; Comprehensive weight: According to actual production needs, the weights of the entropy weight method and the hierarchical analysis method are weighted and integrated to make comprehensive adjustments to the weights of each indicator; S4 builds a CNN-LSTM hybrid prediction model The hybrid prediction model uses CNN to extract key time series features from equipment production and operation data, processes time series data through the LSTM layer, learns long-term dependencies, and is used to predict future equipment health status scores and provide maintenance recommendations. The hybrid prediction model includes CNN layer: used to extract short-term time series features from equipment production and operation indicator data and construct spatial feature representation; LSTM layer: further captures long-term dependencies in the time dimension based on the time series feature sequence and predicts the health status score in the future cycle; Input data: Time series data of historical health scores and core indicators; Output results: health score prediction value and maintenance optimization suggestions.

2. The device intelligent health status evaluation method according to claim 1, characterized in that: The scope of S1 data collection includes daily output, product quality, energy consumption, material consumption, repair records, repair time ratio, number of downtimes, locomotive operation process indicators, rejection volume, internal and external downtime, waiting time for materials, etc. Each data is recorded in each system database according to the date, shift, brand number, NC, model and other symbols.

3. The device intelligent health status evaluation method according to claim 1, characterized in that: The methods for dynamic adjustment of S2 include: An exponential decay function is constructed with the equipment service time t as the variable to automatically adjust the index score benchmark value: in, It is the benchmark value of a core indicator under the initial ideal state of the equipment. is the equipment attenuation coefficient set according to equipment type and experience, It is the accumulated service years or operating hours.

4. The device intelligent health status evaluation method according to claim 3, characterized in that: Alternatively, by clustering the historical health scores with the core indicator data, a phased scoring standard can be constructed, and the scoring reference values ​​can be switched based on the aging status of the equipment.

5. The device intelligent health status evaluation method according to claim 1, characterized in that: The adjustment method of S3 comprehensive weight is as follows: The weighted superposition model is used to fuse the weight vectors generated by the entropy weight method and the hierarchical analysis method: ； in, is the subjective weight calculated by the AHP method, is the objective weight calculated by the entropy weight method, It can be set statically based on the actual goals of the enterprise or adaptively optimized based on the reward function.

6. The device intelligent health status evaluation method according to claim 1, characterized in that: The hybrid prediction model structure introduces an attention mechanism to dynamically focus on key indicators and time nodes that have a greater impact on health status score prediction. Combined with historical fault mode associations, the mapping relationship between key features and faults is strengthened through attention weights.

7. The device intelligent health status evaluation method according to claim 1, characterized in that: The evaluation method further includes S5 generates visual health reports The health report includes: an overview of the health score for the month, a health trend forecast curve, weak link analysis, maintenance recommendations and optimization measures.

8. The device intelligent health status evaluation method according to claim 1, characterized in that: S1 data preprocessing includes Data cleaning: Use the isolation forest algorithm to remove outliers and linear interpolation to fill missing values ​​to ensure data continuity; Feature Engineering: A standardized data set is constructed through rotation transformation and label classification, and divided into training set, validation set and test set.

Images