Equipment health degree assessment method for staged model deployment

Abstract

The invention discloses an equipment health degree assessment method for staged model deployment, and belongs to the technical field of specific calculation models, and the method comprises the steps: sequentially dividing a model deployment process into an initial stage, a middle stage and a later stage; for the initial stage, obtaining an equipment health degree score according to the collected sample data through an expert evaluation method, and setting an evaluation rule; in the middle stage, according to the collected sample data, an equipment health degree score is obtained through an evaluation rule, and the sample data is divided for artificial intelligence algorithm training; and for the later stage, obtaining an equipment health degree scoring model according to the artificial intelligence algorithm obtained by training. The staged model deployment strategy covers all stages of the equipment from no data to sufficient data, so that the continuity and availability of health degree assessment are ensured. High-quality sample data is provided for training of an artificial intelligence algorithm through an expert evaluation method, the model training quality and efficiency are improved, and the problem of high data dependence in the prior art is solved.

Description

Technical Field

[0001] The present invention belongs to the technical field of computer systems based on specific computing models, and particularly relates to a method for evaluating equipment health degree with phased model deployment. Background Art

[0002] With the rapid development of intelligent manufacturing, equipment health degree evaluation has become a key technology for ensuring production safety and reducing operation and maintenance costs. In modern industrial production, the efficient operation of equipment is directly related to product quality, production efficiency, and energy consumption. However, during the long-term operation of equipment, it will inevitably experience performance degradation or even sudden failures due to factors such as wear, aging, and environmental interference. Traditional equipment maintenance methods (such as regular inspections or repairs after failures) can no longer meet the requirements of modern industry for equipment reliability and economy.

[0003] Therefore, calculating equipment health degree scores by means of artificial intelligence models and big data analysis technology has become the mainstream method in the field of equipment health degree analysis. For example, Chinese Patent CN118296511A provides a method, device, and medium for evaluating equipment health degree based on artificial intelligence, and Chinese Patent CN118428926A provides a method, device, and monitoring system for evaluating equipment health degree. Using supervised learning or deep learning models (such as SVM, LSTM, YOLOv5, etc.) to model time series data or image data and directly output health degree scores.

[0004] However, this method has an important prerequisite, that is, it needs to rely on a large amount of data accumulation as support. When actually carrying out equipment health degree analysis and corresponding deployment, it will inevitably go through a gradual data accumulation stage, that is, starting from no available data at first, gradually transitioning to having a certain amount of data, and finally accumulating sufficient data resources.

[0005] Existing methods generally assume that equipment operation data has been fully accumulated and cannot adapt to the gradual nature of data from scratch in actual deployment. For example, when the equipment is first put into use, the insufficient data volume makes it difficult for the AI model to take effect.

[0006] Therefore, there is an urgent need for a method for evaluating equipment health degree for health degree management throughout the entire life cycle of equipment in industrial scenarios from no data to sufficient data and then to large-scale deployment. Summary of the Invention

[0007] The present invention provides a method for evaluating equipment health degree with phased model deployment, which solves the problems of strong data dependence and lack of phased strategies in the prior art through a phased strategy, realizes the continuity of equipment health degree evaluation, and is applicable to health degree management throughout the entire life cycle of equipment in industrial scenarios from no data to sufficient data and then to large-scale deployment.

[0008] The technical solution adopted by the present invention is as follows: A method for evaluating equipment health degree with phased model deployment, comprising: The model deployment process is sequentially divided into an initial stage, a middle stage, and a later stage, wherein the initial stage and the middle stage are divided by the number of samples collected, and the middle stage and the later stage are divided by the total number of algorithms trained; For the initial stage, through the expert evaluation method, according to the sample data collected, the equipment health degree score is obtained, and the evaluation rules are set; For the middle stage, according to the sample data collected, through the evaluation rules, the equipment health degree score is obtained, and the sample data is divided for artificial intelligence algorithm training; For the later stage, according to the artificial intelligence algorithm obtained by training, an equipment health degree scoring model is obtained to evaluate the equipment health degree according to the sample data collected in real time.

[0009] The method for evaluating equipment health degree with phased model deployment disclosed in the present invention also has the following additional technical features: The initial stage and the middle stage are divided by the number of samples collected, specifically: The sample data includes visual recognition type and / or data monitoring type; For the sample data of the visual recognition type, by measuring the number of object types, the number of single-type scenarios, and the number of single-scenario samples, the first threshold of the sample quantity is obtained; For the sample data of the data monitoring type, by the number of sample categories, the number of single-category scenarios, and the signal sequence time step, the first threshold of the sample quantity is obtained; When the sample quantity is less than the first threshold of the sample quantity, it is in the initial stage; when the sample quantity is greater than or equal to the first threshold of the sample quantity, it is in the middle stage.

[0010] The middle stage and the later stage are divided by the total number of algorithms trained, specifically: For the sample data of the visual recognition type, by measuring the number of object types, the number of single-type scenarios, and the number of single-scenario algorithms, the threshold for dividing the total number of algorithms is obtained; For the sample data of the data monitoring type, by the number of sample categories, the number of single-category scenarios, and the number of single-scenario algorithms, the threshold for dividing the total number of algorithms is obtained; When the number of algorithms is less than the threshold for dividing the total number of algorithms, it is in the middle stage; when the number of algorithms is greater than or equal to the threshold for dividing the total number of algorithms, it is in the later stage.

[0011] Through the expert evaluation method, based on the collected sample data, obtain the equipment health score, and set the evaluation rules, specifically: Through the expert evaluation method, based on the collected sample data, obtain the equipment health score; Through the analysis of the patterns in the sample data, obtain the scoring indicators, and divide the indicator levels according to the influence degree of the scoring indicators on the equipment health score, so as to correspondingly set the indicator weights.

[0012] The scoring indicators are specifically: The sample data includes visual recognition type and / or data monitoring type; For the sample data of the visual recognition type, the score of the scoring indicator is determined according to the size of the target image; For the sample data of the data monitoring type, the score of the scoring indicator is obtained based on data fluctuations or probability distribution differences.

[0013] The artificial intelligence algorithm training is specifically: Based on each scoring indicator and the equipment health score in the sample data, with the indicator weight as the initial value, conduct the training of the artificial intelligence algorithm; According to different sample data categories, select different artificial intelligence algorithms, Among them, for the sample data of the visual recognition type, select the algorithm in the field of machine vision, and for the sample data of the data monitoring type, select the classification algorithm and / or regression algorithm.

[0014] The division of the sample data is specifically: According to the sample data, divide it into a training set, a validation set, and a test set according to a fixed ratio; The sample data in the training set, the validation set, and the test set are isolated from each other, and all contain the sample data of different sample data types and different scenarios.

[0015] The process of model deployment is successively divided, and it also includes: The process of model deployment being successively divided also includes a promotion stage, The later stage and the promotion stage are divided according to the number of samples used for training the equipment health score model.

[0016] The later stage and the promotion stage are divided according to the number of samples used for training the equipment health score model, specifically: Through the number of samples used for artificial intelligence algorithm training in the middle stage, the difference degree between transferable fields, and the transfer efficiency factor, obtain the second threshold of the number of samples; When the number of samples for training the device health score model is less than the second threshold of the number of samples, it is in the later stage; when the number of samples is greater than or equal to the second threshold of the number of samples, it is in the promotion stage.

[0017] The present invention also discloses a device health assessment system for phased model deployment, including: A data acquisition module for real-time acquisition of sample data; A data preprocessing module for preprocessing the sample data real-time acquired by the data acquisition module; A data transmission module for transmitting the sample data preprocessed by the data preprocessing module to a data storage module; A data storage module for storing and managing the sample data; An artificial intelligence algorithm platform for performing phased deployment of the device health score model according to the sample data in the data storage module; A health score module for performing device health assessment through the device health score model according to the sample data real-time acquired.

[0018] Due to the adoption of the above technical solutions, the beneficial effects achieved by the present invention are: 1. In the present invention, the model deployment process is sequentially divided into an initial stage, a middle stage, and a later stage. For the initial stage, through the expert evaluation method, according to the acquired sample data, the device health score is obtained and the evaluation rules are set.

[0019] The strategy in the initial stage is applicable to the scenario of data scarcity. When the device is just put into use, or in the stage where data collection is not carried out in time after the device is put into use, the amount of data is small at this time. Not only can the device health assessment in the data scarcity scenario be carried out through the expert evaluation method to avoid the interruption of the assessment due to data loss, but also because the amount of data is small, it will not cause the problem of too long time-consuming for expert evaluation, improving the applicability of the present invention. In addition, the sample data after expert evaluation provides high-quality sample data for the training of the model algorithm, and can avoid the problems of slow model training and difficulty in taking effect due to data scarcity.

[0020] For the intermediate stage, based on the collected sample data, through the evaluation rules, the equipment health score is obtained, and the sample data is divided for artificial intelligence algorithm training. At this time, the sample data has been accumulated and gradually transitions to artificial intelligence algorithm training, which can also reduce the time-consuming of expert evaluation. At this time, through the evaluation rules obtained by expert evaluation, the continuously collected sample data is evaluated to provide more sample data for the training of artificial intelligence algorithms. Moreover, the evaluation rules obtained by the expert evaluation method can also guide the starting point of artificial intelligence algorithm training, which can not only reduce the time consumption of artificial intelligence algorithm training, but also improve the scientificity of artificial intelligence algorithm training.

[0021] For the later stage, based on the trained artificial intelligence algorithm, an equipment health score model is obtained to evaluate the equipment health according to the sample data collected in real time. At this time, multiple artificial intelligence algorithms have been trained for multiple types of data, and a health score model is constructed based on the trained artificial intelligence algorithms to achieve real-time automated evaluation. In a scenario with sufficient data, through automated evaluation, not only the real-time performance of equipment health evaluation is improved, but also the dependence on manual work can be reduced.

[0022] Generally speaking, through the phased model deployment strategy, the present invention covers all stages of the equipment from having no data to having sufficient data, ensuring the continuity and usability of health evaluation. Moreover, the transition from manual to automatic is realized. The expert evaluation method provides high-quality sample data for the training of artificial intelligence algorithms, improves the quality and efficiency of model training, and solves the problems of strong data dependence and lack of phased strategies in the prior art. BRIEF DESCRIPTION OF THE DRAWINGS

[0023] The drawings described herein are used to provide a further understanding of the present invention and constitute a part of the present invention. The illustrative embodiments of the present invention and their descriptions are used to explain the present invention and do not constitute an improper limitation of the present invention. In the drawings: Figure 1 It is a schematic flowchart of the equipment health evaluation method with phased model deployment according to an embodiment of the present invention; Figure 2 It is an architecture diagram of the equipment health evaluation system with phased model deployment according to an embodiment of the present invention. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0024] In order to more clearly illustrate the overall concept of the present invention, the following will be described in detail by way of examples in combination with the drawings of the specification.

[0025] In the following description, many specific details are set forth in order to provide a thorough understanding of the present invention. However, the present invention may also be implemented in other ways different from those described herein. Therefore, the scope of protection of the present invention is not limited by the specific embodiments disclosed below.

[0026] As Figure 1 shown, a method for evaluating the health of a device with phased model deployment includes: S100: Divide the model deployment process into an initial stage, a middle stage, and a late stage in sequence, where the initial stage and the middle stage are divided by the number of samples collected, and the middle stage and the late stage are divided by the total number of algorithms trained.

[0027] The main purpose of this step is to ensure that the device health evaluation method can adapt to the progressive accumulation process of data from scratch through a phased strategy, so as to achieve the health management of the entire life cycle of the device. Specifically: Adapt to different stages of data accumulation. When a new device is put into use, the amount of data is small, and it is impossible to directly use a complex AI model for evaluation. Through phased division, a preliminary evaluation can still be carried out in the case of insufficient data, and gradually transition to a more accurate AI model as the amount of data increases.

[0028] Improve the flexibility and operability of the system. Different evaluation methods are used in different stages, which can flexibly meet the requirements in various scenarios and avoid evaluation interruption or inaccuracy caused by insufficient data.

[0029] The initial stage and the middle stage are divided by the number of samples collected. For the insufficient data in the initial stage, a first threshold of the number of samples (such as 500 - 2000 images or data points) is set according to the device type and application scenario. When the number of samples collected is less than this threshold, it is in the initial stage.

[0030] In the initial stage, due to insufficient data, the expert evaluation method (such as manual rule reasoning) is used to score the health. For example, for a numerically controlled machine tool, in the initial stage, the expert can evaluate the appearance quality (such as cracks, scratches) and simple electrical parameters (such as voltage fluctuation range) to generate a health score.

[0031] In the intermediate stage, data accumulates gradually. When the number of collected samples reaches or exceeds the first threshold, it enters the intermediate stage. At this time, the number of samples is sufficient to support a certain degree of artificial intelligence algorithm training. In the intermediate stage, using the collected sample data and the initially set evaluation rules, the dataset is further divided (e.g., 70% for training, 20% for validation, and 10% for testing), and the artificial intelligence algorithm starts to be trained. For example, in the intermediate stage, the existing sample data can be used to train the YOLOv5 model for visual recognition (such as detecting appearance defects), and at the same time, the LSTM model is trained to process time-series data (such as vibration signals).

[0032] The intermediate stage and the later stage are divided by the total number of trained algorithms. A threshold for the total number of algorithms is set according to the device type and application scenario (e.g., 4 algorithms are trained). When the total number of trained algorithms reaches or exceeds this threshold, it enters the later stage. In the later stage, a health score model is constructed based on the trained artificial intelligence algorithms to achieve real-time evaluation. For example, in the later stage, the fully trained YOLOv5 and LSTM models can be deployed to monitor the appearance quality and operating status of the device in real time and output a comprehensive health score.

[0033] It can be understood that after the new device is put into use, it takes time to collect data, and a sufficient number of data is required to support complex analysis. Through the phased strategy, it is ensured that preliminary evaluation can still be carried out when the data is insufficient, and the evaluation method is gradually optimized as the data accumulates. Through the sample quantity limit, in the initial stage, the expert evaluation method avoids unnecessary complex calculations when the data is insufficient and saves computing resources. In the intermediate stage, AI algorithms are gradually introduced, avoiding premature investment of a large amount of resources in model training. Through the algorithm quantity limit, it is ensured that the finally deployed AI model has high accuracy and reliability and is suitable for equipment health management in actual industrial scenarios.

[0034] This step, through stage division, ensures that the equipment health evaluation method can adapt to the gradual accumulation process of data from scratch and realizes the health management of the entire life cycle of the equipment.

[0035] S200: For the initial stage, through the expert evaluation method, according to the collected sample data, the equipment health score is obtained, and the evaluation rules are set.

[0036] The main purpose of this step is to construct a preliminary health score framework using expert experience in the initial stage when the data accumulation is insufficient, provide basic rules and weight allocation for subsequent model training, and at the same time ensure that a scientific and operable health evaluation result can still be provided when the data is scarce.

[0037] By inviting experts in the equipment field (such as maintenance engineers and technicians from equipment manufacturers) to score the collected sample data (such as images and electrical parameters), and through rule analysis, key indicators are identified. For example, the dimensional accuracy deviation and appearance cracks of a numerically controlled machine tool may be listed as high-priority indicators.

[0038] According to expert experience, clarify the scoring criteria for each indicator. For example, for dimensional accuracy deviation: 100 points are given when the deviation is within ±0.05 mm, and 50 points are given when the deviation exceeds ±0.1 mm. For voltage fluctuation: 100 points are given when the fluctuation range is within ±1%, and 60 points are given when the fluctuation exceeds ±2%.

[0039] Through multiple rounds of discussion, score the indicators and perform normalization processing. For example: The weight of mechanical properties (such as accuracy retention) is 0.4, the weight of electrical properties is 0.3, and the weight of the maintenance status is 0.1.

[0040] According to the scoring logic of each indicator, and summing the scores of each indicator weighted by their weights, a comprehensive health score is generated to obtain a preliminary evaluation rule. For example, for a certain equipment, the mechanical properties score 80 points (weight 0.4), and the electrical properties score 75 points (weight 0.3), and the comprehensive score is 80×0.4 + 75×0.3 = 32 + 22.5 = 54.5 points.

[0041] In this step, when the data is insufficient, scores are quickly generated through expert experience to avoid evaluation blanks. For example, when a certain stamping machine is just put into use, only 500 images are needed to generate a preliminary score through manual inspection and rule reasoning to ensure the continuity of evaluation. Traditional methods (such as AI models) require a large amount of data support, while the expert evaluation method can provide a preliminary reference when the data is scarce.

[0042] Through weight allocation and rule standardization, initial parameters are provided for the AI model training in the intermediate stage, which can reduce the blindness of AI model training and accelerate model iteration and optimization. For example, the algorithm development in the intermediate stage can be initialized based on the initial weights, saving training time. It can also avoid investing a large amount of resources in complex AI training when the data is insufficient.

[0043] Generally speaking, in this step, a scoring framework is constructed through the expert evaluation method when the data is insufficient, ensuring the continuity of health assessment and providing standardized rules for the subsequent AI model.

[0044] S300: For the intermediate stage, according to the collected sample data, through the evaluation rule, obtain the equipment health score, and divide the sample data for artificial intelligence algorithm training.

[0045] The main purpose of this step is to score the sample data through evaluation rules and reasonably divide the data set when the data accumulation reaches a certain amount, providing a high-quality data basis for the training of artificial intelligence algorithms, and thus gradually transitioning to the automated evaluation in the later stage.

[0046] In the mid-term stage, the scoring rules in the initial stage (such as grading and deducting points for appearance defects) are continued to score and label the collected sample data, so as to increase the number of samples for algorithm training and improve the training effectiveness of samples.

[0047] Divide the sample data according to a fixed ratio (such as 70% training set, 20% validation set, 10% test set). For example, if 100,000 image data are collected in the mid-term stage, then there are 70,000 in the training set, 20,000 in the validation set, and 10,000 in the test set.

[0048] Select an algorithm according to the sample data. For visual recognition, select a machine vision algorithm (such as YOLOv5). For example, for equipment appearance defect detection, select the YOLOv5m version to balance accuracy and speed. For data monitoring, select a classification algorithm (such as SVM) or a regression algorithm (such as LSTM). For example, for vibration signal analysis, select the LSTM model to capture time series features.

[0049] Train the corresponding algorithm with the sample data in the training set, adjust the algorithm hyperparameters and monitor the generalization ability of the algorithm during the training process with the sample data in the validation set, and use the sample data in the test set to finally evaluate the performance of the algorithm.

[0050] The number of samples reaches the first threshold (such as 500,000 images or 500,000 sets of time series data), but the total number of algorithms required in the later stage is not yet met (such as 4 algorithms have not been developed). At this time, it is in the mid-term stage. Start algorithm training when the data volume is sufficient to avoid the model being invalid due to insufficient data in the initial stage. For example, at least 100,000 sets of vibration data are required in the mid-term stage of a numerically controlled machine tool to support the training of the LSTM algorithm.

[0051] This step divides the mid-term stage through the number of samples, realizing a smooth transition from manual scoring to AI training.

[0052] S400: For the later stage, according to the trained artificial intelligence algorithm, obtain a device health score model to evaluate the device health according to the sample data collected in real time.

[0053] The main purpose of this step is to deploy the trained AI model on the premise of sufficient data and completed algorithm training, realizing real-time and high-precision device health evaluation.

[0054] Sample data (images, time series signals) are collected in real time through devices such as sensors (vibration, temperature), cameras, etc. Convolutional feature extraction is performed on the image data, and frequency domain analysis or time series modeling is performed on the time series data. The features are input into a trained AI model to output a device health score (such as 0-100 points). For example, the LSTM model analyzes the time series features of vibration signals to judge the wear degree of the main shaft and score it.

[0055] Processing real-time data through the AI model significantly improves the evaluation accuracy, and the real-time performance meets the requirements of industrial scenarios. For example, the LSTM model only needs 50 ms to process vibration signals within 1 second, meeting the real-time monitoring of high-frequency devices. Avoid missed detection of faults caused by manual detection delays. For example, the main shaft fault of a certain stamping machine can be predicted 24 hours in advance through real-time vibration analysis, reducing downtime losses.

[0056] When the total number of algorithms reaches the threshold and enters the later stage, ensure that the AI model is fully trained, avoid overfitting or underfitting caused by insufficient data, and can also perform targeted processing and prediction on various types of sample data through multiple algorithms. For example, the LSTM model requires at least 500,000 sets of vibration data to capture complex time series features.

[0057] This step realizes the automation and high precision of device health assessment through model deployment and real-time evaluation in the later stage.

[0058] As a preferred embodiment of the present invention, the initial stage and the intermediate stage are divided by the number of samples collected, specifically: The sample data includes visual recognition types and / or data monitoring types; For the sample data of the visual recognition type, the first threshold of the number of samples is obtained by measuring the number of object types, the number of single-type scenarios, and the number of samples in a single scenario; For the sample data of the data monitoring type, the first threshold of the number of samples is obtained by the number of sample categories, the number of single-category scenarios, and the signal sequence time step; When the number of samples is less than the first threshold of the number of samples, it is in the initial stage; when the number of samples is greater than or equal to the first threshold of the number of samples, it is in the intermediate stage.

[0059] The main purpose of this embodiment is to scientifically quantify the accumulation degree of sample data, clearly define the boundary between the initial stage and the intermediate stage, ensure the use of the manual evaluation method (initial stage) when the data is insufficient, and start the artificial intelligence algorithm training (intermediate stage) when the data is sufficient, so as to achieve the dynamic adaptability and resource optimization of device health assessment.

[0060] According to the differences in sample data, it can be divided into two major categories of indicators: "visual recognition type" and "data monitoring type". The "visual recognition type" indicators specifically include the health detection of "appearance quality" and "precision indicators". The "appearance quality" inspection examines whether the appearance of mechanical equipment is beautiful and clean, and whether there are obvious scratches, deformations, color differences, knocks, paint peeling and other defects. For example, if there are obvious scratches on the outer surface of a CNC machine tool, it will affect its aesthetics.

[0061] The "precision indicators" category specifically includes "dimensional accuracy" and "form accuracy". "Dimensional accuracy" refers to the deviation range between the dimensions of parts processed by machining mechanical equipment, such as the parts processed by a lathe and the designed required dimensions. For example, if the required diameter of a shaft part to be processed is 50mm, the actual processed dimension error within ±0.05mm is considered qualified in terms of precision. "Form accuracy" refers to indicators such as flatness and cylindricity. For example, for the plane processed by a grinding machine, its flatness error needs to meet the process requirements of subsequent assembly, etc.

[0062] The "data monitoring type" indicators specifically include the health detection of "mechanical properties" and "electrical properties". "Mechanical properties" evaluate the operating ability of mechanical equipment in the working state, including motion performance, vibration performance, temperature performance, sound performance, oil analysis, etc. "Electrical properties" include parameters such as voltage, current, power, resistance, inductance, energy consumption, etc., as well as electrical safety performance indicators such as whether the grounding of the equipment is good and whether the electrical insulation performance meets the standards.

[0063] The initial stage and the middle stage are divided according to the sample quantity. For visual recognition type samples, the first sample quantity threshold = the number of measurement object types × the number of single-type scenarios × the number of single-scenario samples.

[0064] For example, for the "appearance quality" type of indicators, the test object is defects, and it is required that the collected photos cover at least 5 - 10 typical defect types (such as cracks, scratches, dents, etc.), and each defect needs to come from 3 - 5 different acquisition scenarios (different lighting, angles, equipment) to avoid overfitting of the model.

[0065] In specific industrial practices, for small and medium-sized enterprises, in a single scenario, each defect requires 500 - 2000 images (the total sample data volume is about 10,000 - 50,000 images). For large industrial scenarios, that is, across multiple production lines, in a single scenario, each defect requires 5000 - 20000 images, and the total sample data volume needs to be 100,000 - 1,000,000 images, covering different equipment, materials, and defect severity levels.

[0066] For the "precision index" category, measurement object types such as circular apertures, rectangular components, irregular surfaces, etc., require that the collected image data cover at least 5-10 typical measurement objects, and each object needs to come from 3-5 different acquisition scenarios (different measurement distances, viewing angles, device resolutions) to ensure that the model can adapt to diverse measurement environments.

[0067] In specific industrial practices, for small and medium-sized enterprises, in a single scenario, each measurement object requires 800-3000 images (total data volume of about 20,000-80,000 images). For large industrial scenarios, that is, across multiple production lines, each measurement object requires 5000-60000 images, with a total of 150,000-1.5 million images, covering different device models, material reflection characteristics, and measurement accuracy requirements.

[0068] For the "data monitoring" index, the first threshold of the sample quantity = the number of sample categories × the number of single-category scenarios × the time step of the signal sequence. At this time, sample categories such as "strength qualified / unqualified" and "wear level" of mechanical properties. The time step of the signal sequence, for example, when the sampling frequency of motor vibration data is 10 kHz, 1 second of data corresponds to 10,000 time points.

[0069] For example, in the motor vibration signal detection scenario (2 categories of qualified / unqualified, single scenario). Each sample contains 10,000 time points, and the total data volume = 2×1×10000 = 10000 time-series samples.

[0070] When the actually collected sample quantity is less than the calculated first threshold, the device is in the initial stage. The expert evaluation method is used to generate a health score, and evaluation rules are set (such as manual rule inference).

[0071] When the actually collected sample quantity is greater than or equal to the first threshold, the device enters the mid-stage. The sample data is used to divide the data set (training set, validation set, test set), and the artificial intelligence algorithm training is started (such as YOLOv5, LSTM).

[0072] This implementation method quantifies and calculates the threshold to clarify the critical point of stage conversion, avoiding resource waste caused by subjective judgment. For example, if no threshold is set, AI training may be mistakenly started when the data is insufficient, resulting in overfitting or underfitting of the model, thus leading to poor model performance.

[0073] Visual recognition categories need to cover a variety of object types and scenarios (such as the normal / defective states of spindles, cutters, and housings), and the sample quantity needs to meet the diversity requirements; data monitoring categories need to capture time-series characteristics (such as the frequency change of vibration signals), and the sample volume needs to be sufficient to reflect the dynamic pattern.

[0074] When the number of samples in the initial stage is small, the expert evaluation method is used to save computing resources; when the number of samples increases in the middle stage, training is started in a timely manner to avoid the increase in storage costs caused by data redundancy.

[0075] In this embodiment, by scientifically quantifying the accumulation degree of sample data, the boundary between the initial stage and the middle stage is clearly defined, ensuring that the expert evaluation method is adopted when the data is insufficient and the AI algorithm training is started when the data is sufficient.

[0076] As a preferred embodiment of the present invention, through the expert evaluation method, according to the collected sample data, the equipment health score is obtained, and the evaluation rules are set, specifically: Through the expert evaluation method, according to the collected sample data, the equipment health score is obtained; By analyzing the laws of the sample data, scoring indicators are obtained, and according to the influence degree of the scoring indicators on the equipment health score, the indicator levels are divided to correspondingly set the indicator weights.

[0077] The main purpose of this embodiment is to construct a scientific and quantifiable health score rule through the expert evaluation method in the initial stage when the data accumulation is insufficient, providing a standardized framework for subsequent AI algorithm training.

[0078] Deeply analyze various aspects of the equipment's characteristics and operating parameters, and combine factors such as the equipment's structure, function, and common failure modes in the past to determine a series of indicators that can reflect the equipment's health.

[0079] Furthermore, the determined numerous indicators are classified according to different categories for easy subsequent consideration and management. For example, the first-level indicators are divided into mechanical performance, electrical performance, operating stability, and maintenance status. Taking mechanical performance as an example, the second-level indicators under it are divided into accuracy retention, component wear, and load capacity. Continuing with accuracy retention as an example, the third-level indicators under it are divided into dimensional accuracy deviation and shape accuracy deviation.

[0080] Determine the weights of each indicator. Different health score indicators have different importance to the overall health of the equipment, so corresponding weights need to be assigned to each indicator to reflect its relative importance in the comprehensive evaluation. The weight assignment can be determined in various ways, such as: importance evaluation method, analytic hierarchy process, and data-driven method.

[0081] Among them, in the importance evaluation method, the importance of each indicator is evaluated and scored, and then the opinions of each expert are synthesized to determine the weight. For example, for a numerically controlled machine tool, it is considered that the index of machining part dimensional accuracy deviation is crucial to the overall performance of the equipment, and a weight of 0.3 is given; while the equipment appearance quality has a relatively smaller impact on the core operation, a weight of 0.05 is given, etc.

[0082] The analytic hierarchy process constructs the equipment health evaluation system into a hierarchical structure model, including the goal layer (equipment health), the criterion layer (different categories of indicators, such as mechanical performance indicators, electrical performance indicators, etc.), and the scheme layer (specific scoring indicators). By constructing a pairwise comparison judgment matrix, the relative weights of each indicator with respect to the goal layer are calculated.

[0083] The data-driven method collects a large amount of historical equipment operation data and corresponding information such as fault records and maintenance conditions, and uses data analysis methods (such as correlation analysis, principal component analysis, etc.) to analyze the degree of association between each indicator and the actual health status of the equipment (such as whether a fault occurs and the severity of the fault), and determines the weight according to the strength of the association. For example, through data analysis, it is found that the current fluctuation of the spindle motor has a strong correlation with the occurrence of equipment faults, so a higher weight is assigned to this indicator.

[0084] In addition, after the weights are set, the weights of each level are normalized, and finally it is necessary to ensure that the sum of the weights of all indicators at each level is 1. For example, after the normalization of the first-level indicator weights, the mechanical performance weight is 0.4, the electrical performance weight is 0.3, the operation stability weight is 0.2, and the maintenance status weight is 0.1. Taking mechanical performance as an example, after the normalization of the second-level indicator weights, the precision retention weight is 0.4, the component wear weight is 0.3, and the load capacity weight is 0.3. Taking precision retention as an example, after the normalization of the third-level indicator weights, the dimensional accuracy deviation weight is 0.6, and the shape accuracy deviation weight is 0.4.

[0085] Based on the scores and weights of each indicator, a comprehensive health evaluation rule is constructed.

[0086]

[0087] In a specific embodiment, among the first-level indicators, the mechanical performance score is 80 points, the electrical performance score is 75 points, the operation stability score is 85 points, and the maintenance status score is 90 points. The contribution of mechanical performance to the overall health is 80×0.4 = 32 points, the contribution of electrical performance is 75×0.3 = 22.5 points, the contribution of operation stability is 85×0.2 = 17 points, and the contribution of maintenance status is 90×0.1 = 9 points. The comprehensive health evaluation score based on the first-level indicators is 32 + 22.5 + 17 + 9 = 80.5 points.

[0088] The scoring indicators are specifically: The sample data includes visual recognition type and / or data monitoring type; For the sample data of the visual recognition type, the scoring of the scoring indicators is determined according to the size of the target image; For the sample data of the data monitoring type, the scoring of the scoring index is obtained based on data fluctuations or probability distribution differences.

[0089] The main purpose of this step is to accurately quantify the key features of equipment health through a differential scoring method, and improve the scientificity and practicality of the evaluation. For the visual recognition type of sample data: the scoring index is determined based on the size of the target image (such as defect length, area); for the data monitoring type of sample data: the scoring index is determined based on data fluctuations (such as variance, standard deviation) or probability distribution differences (such as KL divergence, JS divergence).

[0090] For visual recognition (such as image defects), physical dimensions (such as scratch length) need to be concerned, while for data monitoring (such as vibration signals), statistical characteristics (such as distribution differences) need to be concerned. Through targeted scoring rules, confusion of different data types is avoided (such as misjudging the importance of data fluctuations by using size scoring).

[0091] Specifically, for defect size grading, the scoring interval is divided according to the physical size of the defect (such as length, area). For example, scratch length: <1mm → 90 points (minor defect), 1 - 5mm → 70 points (medium defect), 5mm → 50 points (severe defect).

[0092] Area ratio: For component wear or corrosion, calculate the proportion of the defect area to the whole. For example, <5% → 95 points, 5% - 20% → 80 points, 20% → 60 points.

[0093] For data fluctuation analysis, calculate the statistics of time - series data (such as standard deviation, variance). For example, voltage fluctuation: standard deviation <0.5V → 100 points (stable), 0.5 - 1.0V → 80 points (slight fluctuation), 1.0V → 60 points (severe fluctuation).

[0094] For probability distribution differences, measure the difference between the current data and the normal state through KL divergence or JS divergence. For example, vibration signal analysis: KL divergence <0.1 → 100 points (normal), 0.1 - 0.5 → 70 points (potential anomaly), 0.5 → 50 points (fault).

[0095] In this embodiment, the visual recognition type relies on physical features (such as size), while the data monitoring type relies on statistical features (such as fluctuations). The visual recognition type quantifies the defect severity through size (such as scratch length 3mm → 70 points), which is intuitive and easy to interpret; the data monitoring type captures dynamic anomalies through statistics or distribution differences (such as KL divergence 0.3 → 70 points), avoiding misjudgment due to a single threshold. Through differential scoring rules, this embodiment realizes the accurate evaluation of visual recognition type and data monitoring type samples.

[0096] Generally speaking, through weight assignment and rule standardization, this embodiment provides initial parameters for the AI algorithm training in the intermediate stage. For example, the YOLOv5 algorithm in the intermediate stage can directly reuse the index weights set in the initial stage (such as the appearance quality weight of 0.05), which can reduce the blindness of AI algorithm training, accelerate the iterative optimization of the model, and save training time.

[0097] As a preferred embodiment of the present invention, the sample data is divided as follows: According to the sample data, a training set, a validation set, and a test set are obtained by dividing according to a fixed ratio; The sample data in the training set, the validation set, and the test set are isolated from each other, and all contain the sample data of different sample data types and different scenarios.

[0098] The main purpose of this embodiment is to provide a representative and generalizable data basis for subsequent model training and evaluation by reasonably dividing the sample data set.

[0099] The collected data samples are divided into a training set, a validation set, and a test set. Usually, 70% of the data is selected as the training set for model training, 20% of the data is used as the validation set for adjusting the model hyperparameters and monitoring the generalization ability of the model during training, and the remaining 10% of the data is used as the test set for finally evaluating the performance of the model.

[0100] In addition, the larger the data volume and the more comprehensive the covered device states (including normal operating states, pre-fault states of different degrees, and states during various fault occurrences, etc.), the more conducive it is to training an accurate and reliable health monitoring model. When dividing the training set, the validation set, and the test set, it is necessary to ensure that the distribution of each category in different data sets is relatively balanced to avoid data bias affecting the model effect.

[0101] Specifically, each set contains visual recognition classes (such as device appearance defect images) and data monitoring classes (such as sensor data of voltage, temperature, vibration, etc.). Ensure that each set contains sample data from different working conditions (such as normal operation, early stage of fault, high temperature and high humidity, etc.). Stratified sampling technology can be used to keep the internal category distribution of each subset consistent.

[0102] By evenly distributing the sample data under different working conditions in the training set, the validation set, and the test set, the model can learn more comprehensive feature expressions. For example, having seen high-temperature samples during training, the model will perform more stably when facing similar scenarios during deployment. If the model is only trained with normal-temperature samples, it may lead to misjudgment or missed detection in high-temperature environments, affecting the actual application effect.

[0103] The training set and the test set are completely isolated to prevent the model from "memorizing" the test samples, thus obtaining a more realistic performance evaluation. For example, the accuracy on the test set can reflect the performance of the model on unknown data. The validation set is used to adjust hyperparameters (such as the learning rate and the number of network layers), and the test set is used for the final performance evaluation, forming a closed-loop feedback mechanism. For example, if it is found through the validation set that the recognition rate of the current model decreases under low light conditions, relevant samples can be added in the next round of training.

[0104] Generally speaking, in this embodiment, by scientifically dividing the sample data, it is ensured that the training set, the validation set, and the test set do not overlap with each other, and at the same time cover different data types and operating scenarios, which is an important prerequisite for building a high-quality AI model. It not only improves the generalization ability and evaluation accuracy of the model, but also provides a solid foundation for the algorithm training in the intermediate stage.

[0105] As a preferred embodiment under this embodiment, the artificial intelligence algorithm training is specifically as follows: Based on each scoring index and the device health score in the sample data, using the index weight as the initial value, the artificial intelligence algorithm is trained; According to different sample data categories, different artificial intelligence algorithms are selected, Among them, for sample data of visual recognition type, algorithms in the field of machine vision are selected, and for sample data of data monitoring type, classification algorithms and / or regression algorithms are selected.

[0106] The main purpose of this embodiment is to realize the automated scoring prediction of the device health through reasonable selection and training of artificial intelligence algorithms, improve the evaluation efficiency and accuracy, and provide a model basis for subsequent deployment.

[0107] For sample data of visual recognition type (such as images), the characteristics are high-dimensional and unstructured, and rely on spatial feature extraction. It is recommended to use algorithms such as convolutional neural network (CNN), YOLO series (for defect detection), ResNet (for classification), etc. For example, use YOLOv5 to detect the crack position on the device shell and quantify its size to provide input for scoring.

[0108] For sample data of data monitoring type (such as time series data such as voltage, temperature, vibration signals, etc.), the characteristics are structured and time series, and focus on trend changes or statistical distributions. Therefore, the recommended algorithms include, for classification tasks: XGBoost, LightGBM, SVM, random forest; for regression tasks: LSTM, GRU, linear regression, ridge regression, etc. For example, use LSTM to model the motor current fluctuation and predict whether it is in an abnormal state.

[0109] Use the pre - divided training set data (including visual images and sensor data) in the early stage; each sample is attached with a health score label generated by the expert evaluation method and the weights of each index. Take the index weights obtained by the expert evaluation method as part of the initialization of the model parameters. For example, set the initial values of the weighted loss function or attention mechanism in the multi - modal fusion model; avoid the model learning from scratch, accelerate the convergence speed, and can significantly shorten the algorithm training cycle.

[0110] For visual algorithms (such as YOLOv5), take images as input, output defect sizes or classification results, and perform end - to - end training in combination with the score labels; for monitoring algorithms (such as LSTM), take sensor time - series data as input, output health scores or failure probabilities. Visual recognition - type and data monitoring - type samples are respectively modeled using suitable algorithms, which can more accurately capture the key features in their respective data.

[0111] At the same time, use the validation set to adjust hyperparameters (such as learning rate, batch size); use the test set to evaluate the performance of the final model (such as mean square error, accuracy, F1 value, etc.).

[0112] This embodiment realizes the automatic evaluation of equipment health by scientifically selecting and training artificial intelligence algorithms.

[0113] As a preferred embodiment of the present invention, the middle stage and the later stage are divided by the total number of trained algorithms. Specifically: For the sample data of the visual recognition type, obtain the total number of algorithm division thresholds by measuring the number of object types, the number of single - type scenarios, and the number of algorithms in a single scenario; For the sample data of the data monitoring type, obtain the total number of algorithm division thresholds by measuring the number of sample categories, the number of single - category scenarios, and the number of algorithms in a single scenario; When the number of algorithms is less than the total number of algorithm division thresholds, it is in the middle stage; when the number of algorithms is greater than or equal to the total number of algorithm division thresholds, it is in the later stage.

[0114] The core purpose of this embodiment is to stage - divide the system development stage through the objective index of the number of algorithms, so as to provide a decision - making basis for subsequent resource investment, model optimization direction, and deployment strategy.

[0115] For the visual recognition type, the total number of algorithm division thresholds = the number of object types × the number of single - type scenarios × the number of algorithms in a single scenario. Among them, the number of object types: such as gearboxes, motors, cutting tools, etc.; the number of single - type scenarios: such as normal operation, high temperature, high humidity, dust, etc.; the number of algorithms in a single scenario: multiple algorithms may be required in each scenario (such as detection, positioning, classification, scoring).

[0116] For data monitoring categories, the threshold for dividing the total number of algorithms = the number of sample categories × the number of single-category scenarios × the number of algorithms for a single scenario. Among them, the number of sample categories: such as voltage anomaly, vibration exceeding the standard, temperature deviation, etc.; the number of single-category scenarios: such as mild anomaly, moderate anomaly, severe anomaly; the number of algorithms for a single scenario: such as regression prediction, classification discrimination, trend warning, etc.

[0117] When the number of algorithms that have been trained and deployed is less than this threshold, the system is in the intermediate stage; when it is equal to or exceeds this threshold, the system enters the later stage. By setting a clear "intermediate / later" boundary, the team can focus on allocating development resources, avoid waste of resources caused by blind development, and ensure that each newly added algorithm has practical application value.

[0118] When the number of algorithms reaches a certain scale, the system can introduce a model building collaboration mechanism (such as ensemble learning) to improve the overall scoring stability. A single algorithm is difficult to handle complex working conditions, while model fusion can improve accuracy without significantly increasing computing power.

[0119] In a specific embodiment, for a compressor health assessment system in a chemical plant, for visual recognition tasks, the number of measurement object types: 3 types (motor, bearing, seal); the number of single-type scenarios: 5 types (normal temperature, high temperature, high pressure, low oil pressure, vibration anomaly); the number of algorithms for a single scenario: 2 (defect detection + dimension measurement); the threshold for dividing the total number of visual algorithms = 3×5×2 = 30.

[0120] For data monitoring tasks, the number of sample categories: 4 types (temperature, pressure, current, vibration); the number of single-category scenarios: 3 types (normal, warning, alarm); the number of algorithms for a single scenario: 2 (classification + regression); the threshold for dividing the total number of monitoring algorithms = 4×3×2 = 24.

[0121] If a total of 25 algorithms are currently deployed (18 visual + 7 monitoring) < total threshold 54 → the system is still in the intermediate stage; if further developed later and the cumulative number of deployed algorithms reaches 54 → enter the later stage.

[0122] This embodiment realizes a scientific division of the system development stage by quantifying the relationship between the number of algorithms and the task complexity. In the intermediate stage, it focuses on the identification of core failure modes, the algorithm deployment is scattered, and independent model calls are used; in the later stage, a unified model platform is introduced to support multi-algorithm collaborative reasoning and functions, forming a complete health assessment system.

[0123] As a preferred embodiment of the present invention, the model deployment process is divided successively, and further includes: The successive division of the model deployment process further includes a promotion stage. The later stage and the promotion stage are divided according to the number of samples trained by the equipment health score model.

[0124] The main purpose of this embodiment is to clarify the application scope and applicability of the model in different development stages by considering the number of model training samples, so as to ensure that the model can operate stably and efficiently in the industrial field and gradually expand its application scope.

[0125] It can be understood that in a large industrial system, there are often many groups of devices with similar hardware structures, key components, similar operating environment factors, and similar working modes. For example, stamping machines are often produced according to unified design specifications and manufacturing processes. They have similar mechanical structures, such as the same stamping die installation method, the matching structure of the slider and the guide rail, and the key component structures of the hydraulic or mechanical transmission systems are basically the same. These similar hardware components mean that in the operation process of the equipment, the basic physical factors affecting its health status are the same. For example, the wear of the die, the stress condition of the transmission components, and the lubrication requirements all follow similar principles. They usually also face some common failure modes. For example, long-term stamping operations are likely to cause wear and fatigue cracks in the die, the hydraulic system may have leaks and unstable pressures, and the mechanical transmission components may become loose due to frequent stress.

[0126] So that the data model trained based on the data collected from one device and used to identify health-related features can perform effective feature matching and analysis on the data of other similar devices, and further support the promotion and use of the health score system in the same industrial system.

[0127] In the later stage, the model has been verified within a certain range (such as specific types of devices or working conditions), and has been trained and optimized based on a relatively large sample data set. At this stage, the model is mainly used for internal testing or small-scale pilot applications; it is deeply optimized for specific devices or working conditions to ensure that the performance of the model in this field reaches the best.

[0128] The model has been fully verified and is applicable not only to specific devices or working conditions, but also can be extended to other similar devices or working conditions for use.

[0129] The later stage and the promotion stage are divided according to the number of samples trained by the equipment health score model. Specifically: The second threshold of the number of samples is obtained through the number of samples used for artificial intelligence algorithm training in the middle stage, the difference degree between transferable domains, and the transfer efficiency factor; When the number of samples for training the device health score model is less than the second threshold of the number of samples, it is in the later stage; when the number of samples is greater than or equal to the second threshold of the number of samples, it is in the promotion stage.

[0130] In this embodiment, collecting a sufficient number of qualified target domain data is used as the sign of the end of the later stage and entering the promotion stage.

[0131] The second threshold N of the number of samples is specifically:

[0132] Wherein, is the number of samples used for artificial intelligence algorithm training in this field, is the difference degree between this field and the transferable field; is the transfer efficiency factor.

[0133] The value of S depends on the transfer strategy of transfer learning (such as frozen layer transfer, fine-tuning transfer, domain adaptation, etc.). Among them, only the classification head (such as the fully connected layer) is modified in the frozen feature extraction layer, and there is no need to adjust the pre-trained backbone network; for fine-tuning transfer, the last 1-3 layers of the backbone network are unfrozen and jointly trained with the classification head. For domain adaptation, the entire pre-trained model is unfrozen and retrained on the target data. For frozen layer transfer, S = 5, and for fine-tuning transfer, S = 1.

[0134] In a specific embodiment, if the number of samples used for artificial intelligence algorithm training in this field is 100,000, the domain difference degree is 0.3, and frozen layer transfer is used, then the newly added data volume of the target domain needs to be approximately N=(100000×0.3) / 5 = 6000.

[0135] This embodiment uses the number of samples as the division criterion, and clarifies the specific path of the model from development to actual application and then to full promotion. At the same time, according to the difference degree and transfer strategy, the number of samples that need to be newly prepared is obtained, which improves the model transfer efficiency and transfer feasibility, and ensures the smooth completion of tasks in each stage. It clarifies the specific path of the model from development to actual application and then to full promotion.

[0136] Such as Figure 2 shown, the present invention provides again a device health assessment system for phased model deployment, including: A data acquisition module for real-time acquisition of sample data; A data preprocessing module for preprocessing the sample data real-time acquired by the data acquisition module; A data transmission module for transmitting the sample data preprocessed by the data preprocessing module to the data storage module; A data storage module for storing and managing the sample data; An artificial intelligence algorithm platform for phased deployment of a device health score model based on sample data in the data storage module; A health score module for evaluating the device health through a device health score model based on the sample data collected in real time.

[0137] Therefore, any effects of the device health assessment method capable of phased model deployment can be achieved, which will not be elaborated here.

[0138] In the present invention, those not described can be implemented by adopting or referring to existing technologies.

[0139] Each embodiment in this specification is described in a progressive manner. For the same or similar parts among the embodiments, reference can be made to each other. Each embodiment focuses on the differences from other embodiments.

[0140] The above are only embodiments of the present invention and are not used to limit the present invention. For those skilled in the art, the present invention can have various changes and modifications. Any modifications, equivalent replacements, improvements, etc. made within the spirit and principle of the present invention shall be included within the scope of the claims of the present invention.

Claims

1. A method for evaluating the device health status with phased model deployment, characterized in that, Including: The model deployment process is sequentially divided into an initial stage, a middle stage, and a later stage. The initial stage and the middle stage are divided by the number of samples obtained through collection, and the middle stage and the later stage are divided by the total number of algorithms trained; For the initial stage, through the expert evaluation method, based on the sample data obtained through collection, the equipment health score is obtained, and the evaluation rules are set; For the middle stage, based on the sample data obtained through collection, through the evaluation rules, the equipment health score is obtained, and the sample data is divided for artificial intelligence algorithm training; For the later stage, based on the artificial intelligence algorithm obtained through training, an equipment health score model is obtained to evaluate the equipment health according to the sample data obtained in real-time collection.

2. The method for evaluating the device health degree by staged model deployment according to claim 1, characterized in that, The initial stage and the middle stage are divided by the number of samples obtained through collection, specifically: The sample data includes visual recognition type and / or data monitoring type; For the sample data of the visual recognition type, by measuring the number of object types, the number of single-type scenarios, and the number of samples in a single scenario, the first threshold of the sample quantity is obtained; For the sample data of the data monitoring type, by the number of sample categories, the number of single-category scenarios, and the signal sequence time step, the first threshold of the sample quantity is obtained; When the number of samples is less than the first threshold of the sample quantity, it is in the initial stage; when the number of samples is greater than or equal to the first threshold of the sample quantity, it is in the middle stage.

3. The method for evaluating the device health degree by staged model deployment according to claim 2, wherein, The middle stage and the later stage are divided by the total number of algorithms trained, specifically: For the sample data of the visual recognition type, by measuring the number of object types, the number of single-type scenarios, and the number of algorithms in a single scenario, the threshold for dividing the total number of algorithms is obtained; For the sample data of the data monitoring type, by the number of sample categories, the number of single-category scenarios, and the number of algorithms in a single scenario, the threshold for dividing the total number of algorithms is obtained; When the number of algorithms is less than the threshold for dividing the total number of algorithms, it is in the middle stage; when the number of algorithms is greater than or equal to the threshold for dividing the total number of algorithms, it is in the later stage.

4. The device health assessment method for phased model deployment according to claim 1, characterized in that Through the expert evaluation method, based on the sample data obtained through collection, the equipment health score is obtained, and the evaluation rules are set, specifically: Through the expert evaluation method, based on the sample data obtained through collection, the equipment health score is obtained; Through the analysis of the rules of the sample data, the scoring indicators are obtained, and according to the influence degree of the scoring indicators on the equipment health score, the indicator levels are divided to correspondingly set the indicator weights.

5. The method for evaluating the device health degree by staged model deployment according to claim 4, characterized in that, The scoring indicators are specifically: The sample data includes visual recognition type and / or data monitoring type; For the sample data of the visual recognition type, the scoring of the scoring indicator is determined according to the size of the target image; For the sample data of the data monitoring type, the scoring of the scoring indicator is obtained according to the data fluctuation or the probability distribution difference.

6. The method for evaluating the device health degree by phased model deployment according to claim 5, wherein The artificial intelligence algorithm training is specifically: Based on each scoring indicator and the equipment health score in the sample data, with the indicator weight as the initial value, the artificial intelligence algorithm is trained; According to different sample data categories, different artificial intelligence algorithms are selected. Among them, for the sample data of visual recognition, algorithms in the field of machine vision are selected, and for the sample data of data monitoring, classification algorithms and / or regression algorithms are selected.

7. The method for evaluating the device health degree by staged model deployment according to claim 5, characterized in that, Divide the sample data, specifically: According to the sample data, divide the training set, validation set, and test set according to a fixed ratio; The sample data in the training set, the validation set, and the test set are isolated from each other and all contain the sample data of different sample data types and different scenarios.

8. The device health assessment method for phased model deployment according to claim 1, characterized in that, Dividing the model deployment process sequentially also includes: The sequential division of the model deployment process also includes the promotion stage. The later stage and the promotion stage are divided by the number of samples used for training the equipment health score model.

9. The method for evaluating the device health degree by phased model deployment according to claim 8, wherein The later stage and the promotion stage are divided by the number of samples used for training the equipment health score model, specifically: Obtain the second threshold of the number of samples through the number of samples used for training the artificial intelligence algorithm in the mid-term stage, the difference degree between transferable fields, and the transfer efficiency factor; When the number of samples used for training the equipment health score model is less than the second threshold of the number of samples, it is in the later stage; when the number of samples is greater than or equal to the second threshold of the number of samples, it is in the promotion stage.

10. A device health assessment system for phased model deployment, characterized in that It includes: A data acquisition module for real-time acquisition of sample data; A data preprocessing module for preprocessing the sample data real-time acquired by the data acquisition module; A data transmission module for transmitting the sample data preprocessed by the data preprocessing module to the data storage module; A data storage module for storing and managing the sample data; An artificial intelligence algorithm platform for performing phased deployment of the equipment health score model according to the sample data in the data storage module; A health score module for evaluating the equipment health through the equipment health score model according to the sample data real-time acquired.

Images