Big data driven predictive maintenance method and system for industrial equipment
Through the predictive maintenance method of industrial equipment driven by big data, using yolov5 and particle swarm-optimized BP neural network, a lightweight defect detection and fault prediction model is built, which solves the problem of insufficient prediction accuracy and maintenance delay in complex failure modes of aluminum smelting furnace equipment. , realizes the precise maintenance and efficient operation of aluminum smelting furnace.
Patent Information
- Application Number
- CN202411095410.9
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2024-08-12
- Publication Date
- 2024-11-15
- Estimated Expiration
- 2044-08-12
AI Technical Summary
During aluminum smelting, the operating conditions of the smelting furnace are extreme, the equipment is aging and fault mode are complex, the prediction accuracy of the existing prediction model is limited, and the signs of failure cannot be detected in advance, and the maintenance system lacks efficient real-time monitoring and intelligent analysis capabilities, resulting in delays in troubleshooting.
Using a predictive maintenance method of industrial equipment driven by big data, integrating the attribute parameters, operating performance data and equipment defect images of aluminum smelting furnaces through edge computing, and using yolov5 to build a lightweight defect detection model and particle swarm optimization BP neural network to build fault prediction Model, evaluate equipment defect type and failure probability, set maintenance score thresholds, and decide whether to perform downtime maintenance.
It realizes accurate predictive maintenance of aluminum smelting furnaces, improves the reliability and efficiency of equipment operation, shortens fault location and processing time, and improves maintenance efficiency and production efficiency.
Smart Images

Figure CN118628487B_ABST
Abstract
Description
Technical Field
[0001] The present invention belongs to the technical field of equipment maintenance, and in particular relates to a big data-driven industrial equipment predictive maintenance method and system. Background Art
[0002] As the pace of manufacturing industry's transformation towards intelligence accelerates, aluminum smelting, as a key link in the production of basic materials, is crucial to ensuring the stability of the supply chain and product quality through its efficient and stable operation. However, aluminum smelting is an energy-intensive, technology-intensive industrial process involving many complex and interdependent mechanical equipment, including smelting furnaces, reduction furnaces, refining equipment, casting equipment, and various types of transportation and control systems.
[0003] For example, a Chinese patent with authorization announcement number CN110705133B discloses a predictive maintenance method and predictive maintenance equipment, the method comprising: receiving first data sent by a data transmission device; classifying the first data, wherein the types of the first data include normal data, abnormal data and fault data; and establishing a prediction model using a variance algorithm and the first data.
[0004] For example, a Chinese patent with authorization announcement number CN109460567B discloses a maintenance method and system for multi-component equipment, which includes the following steps: (1) collecting status monitoring data of each component during equipment operation and initial status monitoring data after each maintenance, and establishing equipment operation and maintenance records; (2) establishing a random degradation model and a random maintenance quality model for each component of the equipment based on the operation and maintenance records, and establishing a maintenance optimization model for the equipment based on all random degradation models, random maintenance quality models and selected optimization indicators; (3) solving the optimal abnormal point threshold, optimal failure threshold and optimal monitoring interval of the equipment based on the maintenance optimization model, and then regularly monitoring the equipment according to the optimal monitoring interval, and maintaining the equipment according to the optimal abnormal point threshold and optimal failure threshold.
[0005] The above existing technologies have the following problems: 1) During the aluminum smelting process, the operating conditions of the smelting furnace are extreme (high temperature, corrosive environment, etc.), and the equipment aging and failure modes are complex and changeable. The existing prediction model may not be able to accurately reflect this complexity, and the prediction accuracy is limited, making it difficult to detect signs of failure in advance; 2) In the face of sudden failures or performance degradation, the existing maintenance system may lack efficient real-time monitoring and intelligent analysis capabilities, and cannot quickly locate problems, assess impacts, and formulate response strategies, thereby delaying the best time for fault handling. In order to solve the above problems, the present invention provides a big data-driven predictive maintenance method and system for industrial equipment. Summary of the invention
[0006] In view of the shortcomings of the prior art, the present invention proposes a big data-driven predictive maintenance method and system for industrial equipment. The method collects the attribute parameters, operating performance data and equipment defect images of an aluminum smelting furnace, uses edge computing for preprocessing and integration, and forms an equipment maintenance data set. The defect detection model is constructed by yolov5 and the fault prediction model is constructed by a BP neural network optimized by a particle swarm. The equipment defect type and failure probability are evaluated respectively. The system sets a maintenance score threshold, calculates the maintenance score by combining the single defect score and the failure probability, and decides whether to perform shutdown maintenance accordingly. The method realizes accurate predictive maintenance of aluminum smelting furnaces and improves the reliability and efficiency of equipment operation.
[0007] To achieve the above object, the present invention provides the following technical solutions:
[0008] Big data-driven predictive maintenance methods for industrial equipment include:
[0009] Step S1, obtaining aluminum smelting furnace attribute parameters, operating performance data and equipment defect image data, preprocessing the collected data using edge computing methods, and integrating the different preprocessed data to obtain equipment maintenance data sets;
[0010] Step S2: construct a lightweight defect detection and evaluation model through yolov5, input the preprocessed image data into the defect detection and evaluation model for training, and obtain a single defect score at the corresponding position;
[0011] Step S3, constructing a fault prediction model based on the BP neural network optimized by particle swarm optimization, and inputting the preprocessed attribute parameters and operating performance data into the fault prediction model for training, so as to obtain the probability of a single defect causing a smelting furnace failure;
[0012] Step S4: setting a single maintenance score threshold and a comprehensive maintenance score threshold, and calculating a single maintenance score and a comprehensive maintenance score according to the obtained single defect score of the corresponding position and the probability of the corresponding single defect causing a failure;
[0013] Step S5, compare the obtained single maintenance score with the single maintenance score threshold, if at least one single maintenance score is greater than or equal to the single maintenance score threshold, the aluminum smelting furnace is shut down for maintenance, if the single maintenance scores are all less than the single maintenance score threshold, compare the comprehensive maintenance score with the comprehensive maintenance score threshold, if the comprehensive maintenance score is greater than or equal to the comprehensive maintenance score threshold, the aluminum smelting furnace is shut down for maintenance, if the comprehensive maintenance score is less than the comprehensive maintenance score threshold, keep the current working state of the aluminum smelting furnace, and monitor the operating state of the aluminum smelting furnace in real time.
[0014] Specifically, the specific steps of obtaining the single defect score at the corresponding position in step S2 include:
[0015] S201, filtering and enhancing the preprocessed equipment defect image data, and performing multi-level annotation on the enhanced equipment defect image to obtain the annotated equipment defect image data; the multi-level annotation includes primary annotation, secondary annotation, tertiary annotation and quaternary annotation, the primary annotation is the defect anchor frame annotation; the secondary annotation is the defect position annotation under the corresponding anchor frame, the tertiary annotation is the defect type annotation under the corresponding position, and the quaternary annotation is the defect score annotation under the corresponding defect type;
[0016] S202, inputting the annotated device defect image data into the slice layer in the lightweight defect detection and evaluation model, segmenting the input device defect image to obtain a device defect image input sequence X of 640*640*3 in size;
[0017] S203, inputting the acquired defect image sequence X into the lightweight backbone layer in the lightweight defect detection and evaluation model, performing preliminary feature extraction, and obtaining the initial defect features of the equipment;
[0018] S204, inputting the acquired initial defect features of the equipment into the defect similarity clustering layer, clustering the defect features belonging to the same defect type, and obtaining N clustered defect feature spaces;
[0019] S205, inputting the obtained N cluster defect feature spaces into the bottleneck layer in the lightweight defect detection and evaluation model to perform feature attention fusion, and obtaining the equipment defect fusion cluster feature;
[0020] S206: Input the acquired equipment defect fusion clustering features into the prediction layer in the lightweight defect detection and evaluation model to obtain the predicted defect type, defect anchor frame position and single defect score of the corresponding anchor frame position.
[0021] Specifically, the specific steps of obtaining the single defect score at the corresponding position in S2 also include:
[0022] S207, constructing a comprehensive loss function for defect detection, and performing training according to the process of S201-206 to obtain a trained lightweight defect detection evaluation model;
[0023] S208. Deploy the acquired trained lightweight defect detection and evaluation model to the smelting furnace visual sensor, and perform defect scoring and positioning based on the equipment defect images collected in real time by the visual sensor.
[0024] Specifically, the loss function L in S207 is:
[0025] ,
[0026] Where C represents the total number of equipment defect types, represents the one-hot encoded vector of the cth true defect type label, represents the probability of predicting the defect to be the cth defect type, and represents a hyperparameter; represents the intersection-over-union ratio of the nth predicted defect anchor box and the corresponding true label anchor box, represents the Mahalanobis distance between the center point of the n-th predicted defect anchor box and the center point of the corresponding true label anchor box, represents the balance weight coefficient, represents the penalty factor for underestimation of the defect score corresponding to the nth defect, represents the penalty factor for overestimation of defect score corresponding to the nth defect, represents the nth defect prediction evaluation score, represents the true evaluation score of the nth defect; Represents the weighted coefficient of the comprehensive loss function.
[0027] Specifically, in step S3, the probability of a single defect causing a smelting furnace failure is obtained, and the specific steps include:
[0028] S301, input the aluminum smelting furnace attribute parameters and operating performance data in the equipment maintenance data set into the principal component analysis algorithm, calculate the attribute parameters and operating performance variables with a cumulative contribution rate greater than 80%, and use them as effective prediction data;
[0029] S302, input the acquired valid prediction data into the BP neural network in the fault prediction model, and initialize a weight and bias vector space V consisting of I particles for the BP neural network through the particle swarm algorithm, and configure an update speed and position for each particle, where: , , represents the i-th particle, represents the jth weight coefficient in the ith particle, represents the bias coefficient in the i-th particle, and m represents the total number of weight coefficients in the i-th particle;
[0030] S303: Constructing a fitness function based on the cross entropy loss function of the fault prediction model and the operating performance data , the specific formula includes:
[0031] ,
[0032] in, represents the probability of the nth defect predicted at time t leading to the failure of the smelting furnace, represents the true probability that the nth defect causes the smelting furnace to fail, It represents the amount of aluminum products produced by the smelting furnace at time t, represents the output at time t The energy consumption corresponding to aluminum products, N represents the total number of defects detected by the smelting furnace in real time, represents a hyperparameter, Represents the weight coefficient of the fitness function;
[0033] S304, setting a fitness threshold and a maximum iteration cycle, iterating S302-S303, when the fitness function value is lower than the fitness threshold or reaches the maximum iteration cycle, stopping the iteration, and outputting the weight and bias vector corresponding to the optimal particle;
[0034] S305 , feeding back the acquired weight and bias vector to the BP neural network, and predicting the probability of a single defect in the next operation cycle of the smelting furnace causing a smelting furnace failure.
[0035] Specifically, the comprehensive maintenance score in S4 includes the following formula:
[0036] ,
[0037] Among them, F represents the comprehensive maintenance score, represents the evaluation score corresponding to the nth defect predicted at time t, It represents the probability that the nth defect predicted at time t will lead to the failure of the smelting furnace.
[0038] Big data-driven industrial equipment predictive maintenance system, including: data processing module, defect detection module, fault prediction module and maintenance identification module;
[0039] The fault prediction module is used to construct a fault prediction model through a BP neural network optimized by a particle swarm according to the attribute parameters and the operating performance data, and predict the probability of a single defect causing a smelting furnace failure;
[0040] The data processing module includes a data collection unit and a preprocessing unit; the data collection unit is used to obtain attribute parameters, operating performance data and equipment defect image data from sensors of the aluminum smelting furnace;
[0041] The preprocessing unit is used to use edge computing methods to perform cleaning, denoising, and normalization preprocessing operations on the collected data, and integrate data in different formats into a device maintenance data set in a unified format.
[0042] Specifically, the defect detection module includes a model building unit and a defect detection unit;
[0043] The model building unit is used to build and train a lightweight defect detection and evaluation model based on the yolov5 algorithm;
[0044] The defect detection unit is used to predict the single defect score, defect type and defect position of the corresponding position based on the preprocessed smelting furnace image data and the trained defect detection evaluation model.
[0045] Specifically, the maintenance discrimination module includes a scoring calculation unit and a decision response unit;
[0046] The score calculation unit is used to set a maintenance score threshold, and calculate a single maintenance score and a comprehensive maintenance score according to the output results of the defect detection module and the fault prediction module; the decision response unit is used to compare the threshold set by the score calculation unit and the single maintenance score with the comprehensive maintenance score, and decide whether to perform a shutdown maintenance operation according to the comparison result.
[0047] A computer-readable storage medium stores computer instructions, which, when executed, execute a big data-driven predictive maintenance method for industrial equipment.
[0048] Compared with the prior art, the present invention has the following beneficial effects:
[0049] In view of the deficiencies of the prior art, the present invention can efficiently analyze the equipment defect images in real time, quickly locate the problem, and evaluate its impact on the equipment performance through a lightweight defect detection and evaluation model based on yolov5, which greatly shortens the time for fault location and processing and improves maintenance efficiency. Secondly, the fault prediction model constructed by the BP neural network based on particle swarm optimization in this method can predict the probability of a single defect causing a smelting furnace failure based on the property parameters and operating performance data of the equipment, and calculate the maintenance score based on the predicted defect score and the predicted failure probability, and use the maintenance score to determine whether to perform maintenance operations. This intelligent prediction method can more accurately reflect the future operation and fault status of the equipment, can better grasp the time point of equipment maintenance, and improve production efficiency and equipment utilization. BRIEF DESCRIPTION OF THE DRAWINGS
[0050] Figure 1 This is a flow chart of a method for predictive maintenance of industrial equipment driven by big data according to Embodiment 1 of the present invention;
[0051] Figure 2 This is a logical structure diagram of a big data driven industrial equipment predictive maintenance method according to Embodiment 1 of the present invention;
[0052] Figure 3 A structural diagram of a lightweight backbone layer in a lightweight defect detection and evaluation model in Example 1 of the present invention;
[0053] Figure 4 This is a module structure diagram of a big data-driven industrial equipment predictive maintenance system according to Example 2 of the present invention. DETAILED DESCRIPTION
[0054] Example 1
[0055] See also Figure 1 and Figure 2 The present invention provides an embodiment: a big data driven industrial equipment predictive maintenance method, the specific steps include:
[0056] Step S1, obtaining aluminum smelting furnace attribute parameters, operating performance data and equipment defect image data, preprocessing the collected data using edge computing methods, and integrating the different preprocessed data to obtain equipment maintenance data sets;
[0057] Step S2: construct a lightweight defect detection and evaluation model through yolov5, input the preprocessed image data into the defect detection and evaluation model for training, and obtain a single defect score at the corresponding position;
[0058] Step S3, constructing a fault prediction model based on the BP neural network optimized by particle swarm optimization, and inputting the preprocessed attribute parameters and operating performance data into the fault prediction model for training, so as to obtain the probability of a single defect causing a smelting furnace failure;
[0059] Step S4, setting a single maintenance score threshold and a comprehensive maintenance score threshold, and calculating a single maintenance score and a comprehensive maintenance score according to the obtained single defect score of the corresponding position and the probability of the corresponding single defect causing a fault; in this embodiment, the single maintenance score threshold and the comprehensive maintenance score threshold are set by those skilled in the art based on the failure and maintenance record data of the aluminum smelting furnace;
[0060] Step S5, compare the obtained single maintenance score with the single maintenance score threshold, if at least one single maintenance score is greater than or equal to the single maintenance score threshold, the aluminum smelting furnace is shut down for maintenance, if the single maintenance scores are all less than the single maintenance score threshold, compare the comprehensive maintenance score with the comprehensive maintenance score threshold, if the comprehensive maintenance score is greater than or equal to the comprehensive maintenance score threshold, the aluminum smelting furnace is shut down for maintenance, if the comprehensive maintenance score is less than the comprehensive maintenance score threshold, keep the current working state of the aluminum smelting furnace, and monitor the operating state of the aluminum smelting furnace in real time.
[0061] Furthermore, the specific steps of obtaining the single defect score at the corresponding position in step S2 include:
[0062] S201, filtering and enhancing the pre-processed equipment defect image data, and performing multi-level annotation on the enhanced equipment defect image to obtain the annotated equipment defect image data; the multi-level annotation includes primary annotation, secondary annotation, tertiary annotation and quaternary annotation, the primary annotation is the defect anchor frame annotation; the secondary annotation is the defect position annotation under the corresponding anchor frame, the tertiary annotation is the defect type annotation under the corresponding position, and the quaternary annotation is the defect score annotation under the corresponding defect type; the equipment defect types include furnace lining erosion, cracks or damage to smelting furnace components, structural defects of smelting furnace, leakage of smelting furnace, damage to electrical components of smelting furnace, etc.;
[0063] S202, input the labeled device defect image data into the slicing layer in the lightweight defect detection and evaluation model, segment the input device defect image, and obtain a device defect image input sequence X of 640*640*3 in size; further, the slicing layer of this embodiment is implemented by a configured convolutional network, such as MobileNetV2, ShuffleNet, etc.;
[0064] S203, input the acquired defect image sequence X into the lightweight backbone layer in the lightweight defect detection and evaluation model, perform preliminary feature extraction, and obtain the initial defect features of the equipment; further, refer to Figure 3 The lightweight backbone layer of this embodiment uses EfficientViT to replace the backbone layer network of the original yolov5, specifically including a standard convolution (Conv), a dilated convolution (DSConv), 4 depthwise separable convolutions (MBConv), specifically depthwise separable convolution 1, depthwise separable convolution 2, depthwise separable convolution 3 and depthwise separable convolution 4, two EfficientViT components, specifically EfficientViT component 1 and EfficientViT component 2, EfficientViT component 1 and EfficientViT component 2 are the same, and both include a lightweight MSA module and an MBConv module. The lightweight MSA mainly extracts the context information of the input image, while the MBConv module mainly extracts local information; MSA balances performance and efficiency, uses ReLU-based attention to achieve a global receptive field, and has better performance on mobile devices. Replacing the backbone network of YOLOv5 with the EfficientViT network can reduce the number of parameters and the amount of calculation of the model, so that the model has good performance when deployed on edge devices with limited hardware resources.
[0065] S204, input the acquired initial defect features of the equipment into the defect similarity clustering layer, cluster the defect features belonging to the same defect type, and obtain N clustered defect feature spaces; further, the clustering algorithm used in this embodiment is k-means clustering, and the number of clusters is the number of defect types, and the specific number is determined by the defects detected in actual work;
[0066] S205, inputting the obtained N clustered defect feature spaces into the bottleneck layer in the lightweight defect detection and evaluation model for feature attention fusion, and obtaining the equipment defect fusion clustering features; in this embodiment, the bottleneck layer adopts the network structure in the YOLOv5 original network to fuse the input features on different paths;
[0067] S206, input the acquired equipment defect fusion clustering features into the prediction layer in the lightweight defect detection and evaluation model to obtain the predicted defect type, defect anchor box position and the single defect score of the corresponding anchor box position; the prediction layer includes 4 prediction heads, namely prediction head 1, prediction head 2, prediction head 3 and prediction head 4, wherein prediction head 1 is used to output the defect anchor box, prediction head 2 is used to output the defect type corresponding to the defect anchor box, prediction head 3 is used to output the defect position corresponding to the defect type, and prediction head 4 is used to output the defect score corresponding to the defect position;
[0068] S207, constructing a comprehensive loss function for defect detection, and training according to the process of S201-206 to obtain a trained lightweight defect detection evaluation model, the loss function L is specifically:
[0069] ,
[0070] Where C represents the total number of equipment defect types, represents the one-hot encoded vector of the cth true defect type label, represents the probability of predicting the defect to be the cth defect type, and Represents hyperparameters, which are used to control the weight and difficulty of difficult-to-classify samples. represents the intersection-over-union ratio of the nth predicted defect anchor box and the corresponding true label anchor box, represents the Mahalanobis distance between the center point of the n-th predicted defect anchor box and the center point of the corresponding true label anchor box, represents the balance weighting coefficient, which is determined by technicians in this field through a large number of experiments. It indicates the penalty factor for underestimation of the defect score corresponding to the nth defect. Underestimation means that the predicted evaluation score is lower than the actual evaluation score. It represents the penalty factor for overestimation of the defect score corresponding to the nth defect. Overestimation means that the predicted evaluation score is higher than the actual evaluation score. represents the nth defect prediction evaluation score, represents the true evaluation score of the nth defect; represents the weighted coefficient of the comprehensive loss function, The corresponding values are determined by those skilled in the art through a large number of experiments;
[0071] S208. Deploy the acquired trained lightweight defect detection and evaluation model to the smelting furnace visual sensor, and perform defect scoring and positioning based on the equipment defect images collected in real time by the visual sensor.
[0072] Through the design of multi-level annotation and lightweight backbone layer, the model can accurately capture the defect characteristics of equipment, and use clustering algorithm to cluster similar defect characteristics to improve the effectiveness of feature representation. At the same time, the introduced bottleneck layer and prediction layer structure further improve the detection accuracy and scoring accuracy of the model. Finally, by constructing a comprehensive loss function for training, the model can achieve efficient and accurate defect detection and scoring under limited hardware resources, effectively ensuring the safe operation of the smelting furnace equipment.
[0073] Furthermore, in step S3, the probability of a single defect causing a smelting furnace failure is obtained, and the specific steps include:
[0074] S301. Input the aluminum smelting furnace attribute parameters and operation performance data in the equipment maintenance data set into the principal component analysis algorithm, calculate the attribute parameters and operation performance variables with a cumulative contribution rate greater than 80%, and use them as effective prediction data; in this embodiment, the aluminum smelting furnace attribute parameters include furnace body specifications, energy consumption, aluminum loss index, smelting temperature range and furnace body material; the operation performance data include smelting time, production efficiency, real-time energy consumption data, failure rate of each part of the smelting furnace components and smelting furnace maintenance records;
[0075] S302, input the acquired valid prediction data into the BP neural network in the fault prediction model, and initialize a weight and bias vector space V consisting of I particles for the BP neural network through the particle swarm algorithm, and configure an update speed and position for each particle, where: , , represents the i-th particle, represents the jth weight coefficient in the ith particle, represents the bias coefficient in the i-th particle, and m represents the total number of weight coefficients in the i-th particle; further, in this embodiment, the neural network node of the BP algorithm is calculated by the following formula: , where p represents the number of hidden layer nodes, Indicates that the number of input layer nodes is 2, Indicates that the number of output layer nodes is 2, The offset is set to 2 here; therefore, in this embodiment ;
[0076] S303: Constructing a fitness function based on the cross entropy loss function of the fault prediction model and the operating performance data , the specific formula includes:
[0077] ,
[0078] in, represents the probability of the nth defect predicted at time t leading to the failure of the smelting furnace, It indicates the true probability that the nth defect will cause the smelting furnace to fail. The true probability is obtained through historical data. When a failure occurs, the probability , when no failure occurs, the probability , It represents the amount of aluminum products produced by the smelting furnace at time t, represents the output at time t The energy consumption corresponding to aluminum products, N represents the total number of defects detected by the smelting furnace in real time, Represents a hyperparameter used to control the intensity of the energy efficiency impact. When it is large, a small change in energy efficiency will cause a large change in the fitness function. and Represents the weight coefficient of the fitness function; It represents the output of aluminum products under unit energy consumption. The larger the value, the higher the production efficiency. When the value is lower than the production level threshold, it means that the performance of the smelting furnace is seriously aged and needs to be shut down for maintenance. The production level threshold is set by the minimum value of the output of aluminum products per unit energy consumption in the aluminum smelting industry. Furthermore, the cross entropy loss function L of the fault prediction model in this embodiment is: ;
[0079] S304, set the fitness threshold and maximum iteration cycle, perform iterations on S302-S303, when the fitness function value is lower than the fitness threshold or reaches the maximum iteration cycle, stop the iteration, and output the weight and bias vector corresponding to the optimal particle ;The maximum iteration cycle of the particle swarm algorithm is set to 200 times;
[0080] S305 , feeding back the acquired weight and bias vector to the BP neural network, and predicting the probability of a single defect in the next operation cycle of the smelting furnace causing a smelting furnace failure.
[0081] Furthermore, the comprehensive maintenance score in S4 includes the following formula:
[0082] ,
[0083] Among them, F represents the comprehensive maintenance score, represents the evaluation score corresponding to the nth defect predicted at time t, It represents the probability of the nth defect causing the failure of the smelting furnace predicted at time t. Since many defects on the smelting furnace are located in close proximity, it is difficult to predict the probability of each defect causing the failure of the smelting furnace in actual implementation. Therefore, in this embodiment, the average value of the probability of failure of each equipment component in the smelting furnace is used instead of the probability of each defect on the component causing the failure of the smelting furnace to perform the actual calculation process, thereby reducing the difficulty in the actual calculation.
[0084] Example 2
[0085] See also Figure 4 , another embodiment provided by the present invention: a big data driven industrial equipment predictive maintenance system, comprising: a data processing module, a defect detection module, a fault prediction module and a maintenance discrimination module;
[0086] A data processing module is used to collect multi-dimensional data of the aluminum smelting furnace and pre-process the collected data; the data processing module includes a data collection unit and a pre-processing unit; the data collection unit is used to obtain attribute parameters, operating performance data and equipment defect image data from sensors and log records of the aluminum smelting furnace; the pre-processing unit is used to use edge computing methods to perform pre-processing operations such as cleaning, denoising and normalization on the collected data, and integrate data in different formats into a unified format of equipment maintenance data set;
[0087] The defect detection module is used to perform defect detection through a lightweight defect detection and evaluation model based on the smelting furnace image data; the defect detection module includes a model building unit and a defect detection unit; the model building unit is used to build and train a lightweight defect detection and evaluation model based on the yolov5 algorithm; the defect detection unit is used to predict the single defect score, defect type and defect position of the corresponding position based on the preprocessed smelting furnace image data and the trained defect detection and evaluation model;
[0088] The fault prediction module is used to build a fault prediction model based on attribute parameters and operating performance data through a BP neural network optimized by a particle swarm, and predict the probability of a single defect causing a smelting furnace failure;
[0089] The maintenance discrimination module is used to calculate the maintenance score according to the defect score and the predicted failure probability, and use the calculated maintenance score to make a maintenance decision and trigger the corresponding maintenance operation; the maintenance discrimination module includes a score calculation unit and a decision response unit; the score calculation unit is used to set the maintenance score threshold, and calculate the single maintenance score and the comprehensive maintenance score according to the output results of the defect detection module and the fault prediction module; the decision response unit is used to compare the threshold set by the score calculation unit and the single maintenance score with the comprehensive maintenance score, and decide whether to perform the shutdown maintenance operation according to the comparison result.
[0090] Example 3
[0091] A computer-readable storage medium stores computer instructions, which, when executed, execute a big data-driven predictive maintenance method for industrial equipment.
[0092] The embodiments of the present invention are described above in conjunction with the accompanying drawings, but the present invention is not limited to the above-mentioned specific implementation modes, which are merely illustrative rather than restrictive. Under the guidance of the present invention, ordinary technicians in the field may also change, modify, replace and modify the above-mentioned embodiments without departing from the scope of protection of the purpose of the present invention and the claims, and all of these are within the protection of the present invention.
Claims
1. A big data driven predictive maintenance method for industrial equipment, characterized in that: include: Step S1, obtaining aluminum smelting furnace attribute parameters, operating performance data and equipment defect image data, preprocessing the collected data using edge computing methods, and integrating the different preprocessed data to obtain equipment maintenance data sets; Step S2: construct a lightweight defect detection and evaluation model through yolov5, input the preprocessed image data into the defect detection and evaluation model for training, and obtain a single defect score at the corresponding position; Step S3, constructing a fault prediction model based on the BP neural network optimized by particle swarm optimization, and inputting the preprocessed attribute parameters and operating performance data into the fault prediction model for training, so as to obtain the probability of a single defect causing a smelting furnace failure; Step S4: setting a single maintenance score threshold and a comprehensive maintenance score threshold, and calculating a single maintenance score and a comprehensive maintenance score according to the obtained single defect score of the corresponding position and the probability of the corresponding single defect causing a failure; Step S5, compare the obtained single maintenance score with the single maintenance score threshold, if at least one single maintenance score is greater than or equal to the single maintenance score threshold, the aluminum smelting furnace is shut down for maintenance, if the single maintenance scores are all less than the single maintenance score threshold, compare the comprehensive maintenance score with the comprehensive maintenance score threshold, if the comprehensive maintenance score is greater than or equal to the comprehensive maintenance score threshold, the aluminum smelting furnace is shut down for maintenance, if the comprehensive maintenance score is less than the comprehensive maintenance score threshold, keep the current working state of the aluminum smelting furnace, and monitor the operating state of the aluminum smelting furnace in real time.
2. The big data driven predictive maintenance method for industrial equipment according to claim 1, characterized in that: The specific steps of obtaining the single defect score at the corresponding position in step S2 include: S201, filtering and enhancing the preprocessed equipment defect image data, and performing multi-level annotation on the enhanced equipment defect image to obtain the annotated equipment defect image data; the multi-level annotation includes primary annotation, secondary annotation, tertiary annotation and quaternary annotation, the primary annotation is the defect anchor frame annotation; the secondary annotation is the defect position annotation under the corresponding anchor frame, the tertiary annotation is the defect type annotation under the corresponding position, and the quaternary annotation is the defect score annotation under the corresponding defect type; S202, inputting the annotated device defect image data into the slice layer in the lightweight defect detection and evaluation model, segmenting the input device defect image to obtain a device defect image input sequence X of 640*640*3 in size; S203, inputting the acquired defect image sequence X into the lightweight backbone layer in the lightweight defect detection and evaluation model, performing preliminary feature extraction, and obtaining the initial defect features of the equipment; S204, inputting the acquired initial defect features of the equipment into the defect similarity clustering layer, clustering the defect features belonging to the same defect type, and obtaining N clustered defect feature spaces; S205, inputting the obtained N cluster defect feature spaces into the bottleneck layer in the lightweight defect detection and evaluation model to perform feature attention fusion, and obtaining the equipment defect fusion cluster feature; S206: Input the acquired equipment defect fusion clustering features into the prediction layer in the lightweight defect detection and evaluation model to obtain the predicted defect type, defect anchor frame position and single defect score of the corresponding anchor frame position.
3. The big data driven predictive maintenance method for industrial equipment according to claim 2, characterized in that: The specific steps of obtaining the single defect score at the corresponding position in step S2 also include: S207, constructing a comprehensive loss function for defect detection, and performing training according to the process of S201-206 to obtain a trained lightweight defect detection evaluation model; S208. Deploy the acquired trained lightweight defect detection and evaluation model to the smelting furnace visual sensor, and perform defect scoring and positioning based on the equipment defect images collected in real time by the visual sensor.
4. The big data driven predictive maintenance method for industrial equipment according to claim 3, characterized in that: The loss function L in S207 is specifically: , Where C represents the total number of equipment defect types, represents the one-hot encoded vector of the cth true defect type label, represents the probability of predicting the defect to be the cth defect type, and represents a hyperparameter; represents the intersection-over-union ratio of the nth predicted defect anchor box and the corresponding true label anchor box, represents the Mahalanobis distance between the center point of the n-th predicted defect anchor box and the center point of the corresponding true label anchor box, represents the balance weight coefficient, represents the penalty factor for underestimation of the defect score corresponding to the nth defect, represents the penalty factor for overestimation of defect score corresponding to the nth defect, represents the nth defect prediction evaluation score, represents the true evaluation score of the nth defect; Represents the weighted coefficient of the comprehensive loss function.
5. The big data driven predictive maintenance method for industrial equipment according to claim 4, characterized in that: The probability of a single defect causing a smelting furnace failure is obtained in step S3, and the specific steps include: S301, input the aluminum smelting furnace attribute parameters and operating performance data in the equipment maintenance data set into the principal component analysis algorithm, calculate the attribute parameters and operating performance variables with a cumulative contribution rate greater than 80%, and use them as effective prediction data; S302, input the acquired valid prediction data into the BP neural network in the fault prediction model, and initialize a weight and bias vector space V consisting of I particles for the BP neural network through the particle swarm algorithm, and configure an update speed and position for each particle, where: , , represents the i-th particle, represents the jth weight coefficient in the ith particle, represents the bias coefficient in the i-th particle, and m represents the total number of weight coefficients in the i-th particle; S303: Constructing a fitness function based on the cross entropy loss function of the fault prediction model and the operating performance data , the specific formula includes: , in, represents the probability of the nth defect predicted at time t leading to the failure of the smelting furnace, represents the true probability that the nth defect causes the smelting furnace to fail, It represents the amount of aluminum products produced by the smelting furnace at time t, represents the output at time t The energy consumption corresponding to aluminum products, N represents the total number of defects detected by the smelting furnace in real time, represents a hyperparameter, Represents the weight coefficient of the fitness function; S304, setting a fitness threshold and a maximum iteration cycle, iterating S302-S303, when the fitness function value is lower than the fitness threshold or reaches the maximum iteration cycle, stopping the iteration, and outputting the weight and bias vector corresponding to the optimal particle; S305 , feeding back the acquired weight and bias vector to the BP neural network, and predicting the probability of a single defect in the next operation cycle of the smelting furnace causing a smelting furnace failure.
6. The big data driven predictive maintenance method for industrial equipment according to claim 5, characterized in that: The specific formula for the comprehensive maintenance score in step S4 includes: , Among them, F represents the comprehensive maintenance score, represents the evaluation score corresponding to the nth defect predicted at time t, It represents the probability that the nth defect predicted at time t will lead to the failure of the smelting furnace.
7. A big data driven industrial equipment predictive maintenance system, which is implemented based on the big data driven industrial equipment predictive maintenance method according to any one of claims 1 to 6, characterized in that: include: Data processing module, defect detection module, fault prediction module and maintenance identification module; The fault prediction module is used to construct a fault prediction model through a BP neural network optimized by a particle swarm according to the attribute parameters and the operating performance data, and predict the probability of a single defect causing a smelting furnace failure; The data processing module includes a data collection unit and a preprocessing unit; the data collection unit is used to obtain attribute parameters, operating performance data and equipment defect image data from sensors of the aluminum smelting furnace; The preprocessing unit is used to use edge computing methods to perform cleaning, denoising, and normalization preprocessing operations on the collected data, and integrate data in different formats into a device maintenance data set in a unified format.
8. The big data driven industrial equipment predictive maintenance system according to claim 7, characterized in that: The defect detection module includes a model building unit and a defect detection unit; The model building unit is used to build and train a lightweight defect detection and evaluation model based on the yolov5 algorithm; The defect detection unit is used to predict the single defect score, defect type and defect position of the corresponding position based on the preprocessed smelting furnace image data and the trained defect detection evaluation model.
9. The big data driven industrial equipment predictive maintenance system according to claim 8, characterized in that: The maintenance discrimination module includes a scoring calculation unit and a decision response unit; The score calculation unit is used to set a maintenance score threshold, and calculate a single maintenance score and a comprehensive maintenance score according to the output results of the defect detection module and the fault prediction module; the decision response unit is used to compare the threshold set by the score calculation unit and the single maintenance score with the comprehensive maintenance score, and decide whether to perform a shutdown maintenance operation according to the comparison result.
10. A computer-readable storage medium, characterized in that: Computer instructions are stored thereon, and when the computer instructions are executed, the big data driven industrial equipment predictive maintenance method described in any one of claims 1-6 is executed.
