Search

CN-121981709-A - Equipment fault prediction and maintenance system based on machine learning

CN121981709ACN 121981709 ACN121981709 ACN 121981709ACN-121981709-A

Abstract

The invention relates to the technical field of industrial equipment state monitoring and predictive maintenance, and discloses an equipment fault prediction and maintenance system based on machine learning. The system comprises a data acquisition and feature extraction module, a fault prediction and uncertainty quantification module, a high uncertainty judgment module, an active learning and data enhancement module, a model online updating module and a maintenance decision generation module which are connected in sequence. The method comprises the steps of synchronously outputting fault prediction probability distribution and quantitative indexes representing uncertainty of the fault prediction probability distribution, triggering an active learning mechanism to locate dominant uncertainty characteristics and directives to enhance relevant sensor data acquisition based on comparison of the indexes and dynamic thresholds, updating a model on line by using obtained data, and generating differentiated maintenance decision instructions according to different combination intervals where the prediction probability and the uncertainty indexes are located. According to the invention, data acquisition and model optimization are driven through uncertainty quantification, and the self-adaptive capacity and decision accuracy of the system under complex working conditions are improved.

Inventors

  • HU PEIQI
  • XU ZUNYI

Assignees

  • 山东建筑大学

Dates

Publication Date
20260505
Application Date
20260205

Claims (10)

  1. 1. A machine learning based equipment failure prediction and maintenance system, comprising: the data acquisition and feature extraction module is used for acquiring the running state time sequence data of the target equipment and extracting features to construct a multidimensional feature vector; The fault prediction and uncertainty quantization module is connected with the data acquisition and feature extraction module and comprises a Bayesian neural network-based fault prediction model for receiving the multidimensional feature vector, outputting the prediction probability distribution of a specific fault class of the equipment in a future preset period, and synchronously calculating to generate a quantization index representing the prediction uncertainty; The high uncertainty judging module is connected with the fault prediction and uncertainty quantizing module and is used for comparing the uncertainty quantizing index with a dynamically adjusted decision threshold value and judging that the current prediction is in a high uncertainty state when the quantizing index exceeds the decision threshold value; the active learning and data enhancement module is connected with the high uncertainty judging module and is used for analyzing dominant feature dimensions which lead to the high uncertainty state when judging the high uncertainty state and generating data enhancement acquisition instructions so as to adjust the data acquisition strategy of one or more sensors associated with the dominant feature dimensions; The model online updating module is respectively connected with the active learning and data enhancing module and the fault prediction and uncertainty quantifying module and is used for executing the data enhancing and collecting instruction, obtaining enhanced data and generating a new training sample, and carrying out online updating on the fault prediction model by using the new training sample; The maintenance decision generation module is connected with the fault prediction and uncertainty quantification module and the high uncertainty judgment module and is used for generating and outputting differentiated maintenance decision instructions based on different combination intervals where the prediction probability distribution and the uncertainty quantification index are located.
  2. 2. The machine learning based equipment failure prediction and maintenance system of claim 1, wherein the process of generating the uncertainty quantization index by the failure prediction and uncertainty quantization module calculation comprises the steps of: when the Bayesian neural network model is used for reasoning, a Monte Carlo Dropout method is adopted for reasoning Sub-random forward sampling, where Obtaining Predictive probability vectors for individual fault categories; For each fault class, calculate it Standard deviation of each predicted probability value ; The standard deviation is set As an uncertainty measure for the fault class prediction, and the maximum standard deviation in all fault classes As the uncertainty quantization index of the current prediction.
  3. 3. The machine learning based equipment failure prediction and maintenance system of claim 2, wherein the step of the active learning and data enhancement module analyzing dominant feature dimensions that result in highly uncertain states comprises: s1, calculating the uncertainty quantization index Gradients with respect to feature components in the input multi-dimensional feature vector; s2, making the absolute value of the gradient higher than the preset sensitivity threshold value Is identified as the dominant feature dimension; wherein the data enhancement acquisition instructions are configured to, for an original sensor directly related to the dominant feature dimension, raise its sampling frequency to an original frequency Multiple of at least a duration of Wherein The value of (2) is 1.5 to 3.0, min。
  4. 4. The machine learning based equipment failure prediction and maintenance system of claim 1, wherein the method for updating the dynamically adjusted decision threshold in the high uncertainty determination module comprises the steps of: S1, continuously recording the uncertainty quantitative index of historical prediction and the corresponding subsequent equipment state verification result thereof, wherein the verification result is divided into three types of 'prediction accuracy', 'prediction error' and 'incapability of verification'; S2, counting the proportion of 'prediction error' cases in the prediction judged to be in a high uncertainty state under the given current decision threshold, wherein the proportion is defined as a high uncertainty error rate ; S3, when the high uncertainty error rate is high Continuous and continuous The evaluation period exceeds the upper limit value Raising said decision threshold when said high uncertainty error rate Continuous and continuous With an evaluation period below a lower limit When the decision threshold is lowered, wherein 、 、 Is a preset positive number, and 。
  5. 5. The machine learning based equipment failure prediction and maintenance system of claim 1, wherein the model online update module online updates the failure prediction model using new training samples, comprising: S1, calculating an importance weight matrix of old model parameters for a learned task by adopting an elastic weight merging algorithm before updating the model parameters ; S2, when gradient descent optimization is carried out on the new training sample, the new training sample is subjected to loss function A regularization term is added in the model, and the regularization term is used for punishing the significant change of the old model parameters with high importance weight; The loss function The expression is as follows: , wherein, For the standard prediction error loss, Is the first The parameters to be updated are the same as the parameters to be updated, For its old value it is that, Is that The corresponding importance weight is used to determine the importance of the object, Is a regularized intensity superparameter.
  6. 6. The machine learning based equipment failure prediction and maintenance system of claim 1, wherein the rules in the maintenance decision generation module that generate maintenance decision instructions based on the prediction probability distribution and the uncertainty quantization index comprise: Defining a first interval, predicting probability value Above the action threshold And uncertainty quantization index Below the confidence threshold Correspondingly generating a first type of instruction for executing preventive maintenance; defining a second interval, predicting probability value Above the action threshold But uncertainty quantization index Above a confidence threshold Correspondingly generating a second type instruction for performing deep diagnosis and preparing for maintenance; defining a third interval, predicting probability value Below the action threshold But instability quantization index Is higher than the early warning threshold value Correspondingly generating a third type of instruction for 'strengthening state monitoring and recording'; Wherein, the 、 、 Is a preset threshold parameter and meets 。
  7. 7. The machine learning based equipment failure prediction and maintenance system of claim 6, further comprising a maintenance policy knowledge base connected to the maintenance decision generation module for storing historical execution records of different maintenance decision instructions and their effect evaluations in association; And when the maintenance decision generation module generates a second class or third class instruction, synchronously searching 1 to 3 specific maintenance schemes with optimal historical execution effects under the combination of similar prediction probability and uncertainty from the maintenance strategy knowledge base, and adding the specific maintenance schemes as recommended schemes to the instruction for outputting.
  8. 8. The machine learning based equipment failure prediction and maintenance system of claim 1, further comprising a condition-uncertainty feature map construction module, respectively coupled to the data acquisition and feature extraction module and the active learning and data enhancement module, the method comprising: s1, discretizing the working load, the ambient temperature and the operation mode parameters of equipment into a plurality of working condition units; s2, for each working condition unit, counting the average value of the uncertainty quantization indexes predicted by the fault prediction and uncertainty quantization module under the working condition in the historical data ; S3, in real-time operation, when equipment enters a working condition unit marked as high average uncertainty in the working condition-uncertainty characteristic spectrum, the system pre-triggers the active learning program to adjust a data acquisition strategy in advance, wherein the judging condition of the high average uncertainty is that , Is a preset map threshold value.
  9. 9. The machine learning based equipment failure prediction and maintenance system of claim 1, wherein the failure prediction and uncertainty quantization module uses a loss function for the training process of the bayesian neural network model Consists of two parts: wherein In order to predict the error loss, For the KL-divergence between the model weight distribution and the preset prior distribution, For the super-parameters for balancing the two weights, the super-parameters are gradually increased from 0 to a fixed value in the training process 。
  10. 10. The machine learning based equipment failure prediction and maintenance system of claim 1, further comprising a closed loop optimization module coupled to the maintenance decision generation module, the active learning and data enhancement module, and the model online update module, respectively, wherein the optimization step comprises: S1, associating and recording an execution result of the maintenance decision instruction, a data enhancement acquisition result triggered by the active learning program and a performance change index after model updating to form feedback data; and S2, periodically analyzing the feedback data, evaluating the dynamic decision threshold, the generation logic of the data enhancement acquisition instruction and the effectiveness of the online updating algorithm, and automatically fine-adjusting related algorithm parameters based on an evaluation result.

Description

Equipment fault prediction and maintenance system based on machine learning Technical Field The invention relates to the technical field of industrial equipment state monitoring and predictive maintenance, in particular to an equipment fault prediction and maintenance system based on machine learning. Background In the field of industrial intelligent manufacturing, ensuring continuous and stable operation of key equipment and reducing unplanned shutdown loss are core requirements for improving production benefits and safety level. Predictive maintenance serves as an important evolution direction of traditional preventive maintenance and post-maintenance, and early warning and accurate intervention on potential faults are realized by analyzing equipment operation state data, so that maintenance resource allocation is optimized. Currently, typical plant fault prediction systems are generally built based on static data acquisition strategies and fixed parameter machine learning models. In the model application stage, such systems typically output only a single fault probability value, and lack a measure and evaluation of the reliability of the prediction itself. When the system faces insufficient coverage of training data, obvious change of working conditions of equipment or unknown disturbance of an operating environment, a predicted result may contain a large error, but the system cannot identify and prompt such risks, so that a subsequent maintenance decision may be built on unreliable information. In addition, links such as data acquisition, model learning, maintenance decision generation and the like of the existing system are usually in a linear separation state. The data acquisition strategy is usually fixed and cannot be adjusted in a self-adaptive manner according to the current actual cognitive short board and real-time requirements of the model, the model updating is dependent on a fixed retraining period or accumulation of a large number of new fault samples, a response mechanism is delayed, and the dynamic change of the equipment state or the running mode is difficult to adapt quickly. The architecture lacking the cooperative capability of internal feedback and self-adaption makes the long-term prediction accuracy, decision reliability and overall robustness of the system face continuous challenges when dealing with complex and changeable real industrial scenes with various working conditions. Therefore, the equipment fault prediction and maintenance system based on machine learning is provided for solving the problems of how to enable the equipment fault prediction system to perform self-quantitative evaluation on the reliability of output prediction of the equipment fault prediction system, and on the basis, realizing dynamic coordination and closed-loop interaction of key links such as data acquisition, model optimization and maintenance decision and the like, thereby comprehensively improving the adaptability of the system in the face of unknown working conditions and data scarcity, and the accuracy and the interpretability of final maintenance decision. Disclosure of Invention In order to overcome the above-mentioned drawbacks of the prior art, an embodiment of the present invention provides a machine learning-based equipment failure prediction and maintenance system to solve the above-mentioned problems set forth in the background art. In order to achieve the above purpose, the invention provides a machine learning-based equipment fault prediction and maintenance system, comprising: the data acquisition and feature extraction module is used for acquiring the running state time sequence data of the target equipment and extracting features to construct a multidimensional feature vector; The fault prediction and uncertainty quantization module is connected with the data acquisition and feature extraction module and comprises a Bayesian neural network-based fault prediction model for receiving the multidimensional feature vector, outputting the prediction probability distribution of a specific fault class of the equipment in a future preset period, and synchronously calculating to generate a quantization index representing the prediction uncertainty; The high uncertainty judging module is connected with the fault prediction and uncertainty quantizing module and is used for comparing the uncertainty quantizing index with a dynamically adjusted decision threshold value and judging that the current prediction is in a high uncertainty state when the quantizing index exceeds the decision threshold value; the active learning and data enhancement module is connected with the high uncertainty judging module and is used for analyzing dominant feature dimensions which lead to the high uncertainty state when judging the high uncertainty state and generating data enhancement acquisition instructions so as to adjust the data acquisition strategy of one or more sensors associated with the dominant feature dim