Equipment fault predictive maintenance method, electronic equipment and storage medium
By applying LSTM models to predict timing and classification of faults in complex equipment such as coating equipment, the problem that the existing technology cannot accurately analyze the cause of faults is solved, and more efficient and accurate fault prediction is achieved, reducing operation and maintenance costs and preventing accidents.
Patent Information
- Application Number
- CN202411654636.8
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2024-11-19
- Publication Date
- 2025-03-28
AI Technical Summary
Existing predictive maintenance methods cannot accurately analyze the causes of failures of complex equipment such as coating equipment, resulting in poor prediction success rate and slowing down product production progress.
The LSTM model is used for timing prediction and fault classification, and the feature extraction is extracted by obtaining real-time data, combining historical detection data for model training, using the timing prediction model to perform multivariate prediction, and fault classification and diagnostic evaluation are performed based on the classification model.
It improves the efficiency and accuracy of equipment failure prediction, can effectively identify the failure mode of complex equipment, reduce operation and maintenance costs, and prevent accidents.
Smart Images
Figure CN119722019A_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the technical field of mechanical health monitoring, and in particular to an equipment failure predictive maintenance method, an electronic device and a storage medium. Background Art
[0002] Predictive Maintenance (PdM) is condition-based maintenance. When the machine is running, it conducts regular (or continuous) condition monitoring and fault diagnosis on its main (or required) parts to determine the equipment's condition and predict the future development trend of the equipment's condition. Based on the equipment's condition development trend and possible failure modes, a predictive maintenance plan is formulated in advance to determine the time, content, method, and necessary technical and material support for the machine's repair. Predictive maintenance integrates equipment condition monitoring, fault diagnosis, fault (condition) prediction, maintenance decision support, and maintenance activities, and is an emerging maintenance method.
[0003] In predictive maintenance, state prediction is to evaluate the current state of components and predict future states based on the operating information of the equipment. Commonly used methods include time series model prediction, grey model prediction and neural network prediction. There are generally three basic approaches to the development of prediction methods: physical models, knowledge systems and statistical models. In practical applications, the three approaches can be combined to form a hybrid fault prediction technology that combines traditional physical models and intelligent analysis methods and can process digital information and symbolic information, which is more effective for achieving predictive maintenance.
[0004] In the actual production process, taking coating equipment as an example, coating equipment is prone to abnormal glue discharge, resulting in spraying sputtering and area changes. The causes include blockage and aging of the glue discharge hose, blockage or deformation of the nozzle, bubbles in the colloid or concentration changes, which in turn lead to abnormal glue discharge. Single hose status collection cannot clearly identify whether there is a fault. Existing maintenance methods usually repair the equipment after detecting an abnormality, which will disrupt normal product production; and conventional predictive maintenance solutions are only applicable to single collection information feedback, and cannot be applied to fault prediction of complex equipment such as coating equipment (equipment with multiple collection information feedback). Summary of the invention
[0005] The present invention provides a method and device for predictive maintenance of equipment failures, which solves the technical problem that the existing predictive maintenance methods cannot accurately analyze the causes of failures of complex equipment such as coating equipment, thereby resulting in a poor prediction success rate and slowing down the progress of product production.
[0006] In order to solve the above technical problems, the present invention provides a method for predictive maintenance of equipment failures, comprising:
[0007] Acquire real-time collected data from the target device and perform feature extraction to obtain real-time feature data;
[0008] Perform multivariate prediction of time series based on the time series prediction model in the LSTM model and the real-time feature data to obtain multi-step prediction data;
[0009] Fault classification is performed based on the classification model in the LSTM model and the multi-step prediction data, and a phased prediction conclusion is output for alarm maintenance;
[0010] The LSTM model is trained based on historical detection data, and diagnostic evaluation is performed in combination with the historical detection data and corresponding historical fault classification results.
[0011] The equipment failure predictive maintenance method proposed in this basic solution uses historical detection data to train the LSTM model, predicts the life of equipment (such as coating equipment), uses a time series prediction model to perform multivariate prediction of time series, outputs multi-step prediction data, and then classifies faults based on the classification model to determine the fault classification results. Finally, the fault classification results and historical detection data are used to diagnose and evaluate the LSTM model. Combined with the model's good learning accuracy and generalization ability, it can directly predict multi-source vibration signals, effectively improve the efficiency and accuracy of life prediction, and reduce operation and maintenance costs and prevent accidents.
[0012] In a further embodiment, the LSTM model is trained based on historical detection data and performs diagnostic evaluation in combination with corresponding historical fault classification results, including:
[0013] Obtain historical detection data of the target device and perform feature extraction to obtain historical feature data;
[0014] Inputting the historical feature data into the time series prediction model to perform multivariate prediction of the time series and output multi-step prediction data;
[0015] Perform fault classification based on the classification model and the multi-step prediction data, and output a phased prediction conclusion as a historical fault classification result;
[0016] The LSTM model is diagnosed and evaluated based on the historical fault classification results and the historical detection data, and the model is applied to perform fault prediction when it is judged to meet the training standards.
[0017] This solution improves the accuracy and effect of the model by autonomously learning and optimizing the model during model training. In addition, it also combines the historical detection data actually collected with the historical fault classification results to achieve diagnostic evaluation of the LSTM model. By objectively reflecting the model's predictive ability on new data, it can discover problems and deficiencies in the model, thereby making targeted improvements and improving the model's prediction accuracy.
[0018] In a further embodiment, the time series prediction model includes a first convolutional layer, a first pooling layer, a second convolutional layer, a second pooling layer, a first Bi-LSTM layer, a second Bi-LSTM layer, a third Bi-LSTM layer, a Dropout layer, an Attention layer, an Add&Norm layer, a first fully connected layer, a second fully connected layer, and a third fully connected layer connected in sequence;
[0019] The Attention layer adopts a multi-head attention mechanism;
[0020] The Add&Norm layer is used to perform residual connection calculation;
[0021] The first Bi-LSTM layer, the second Bi-LSTM layer, and the third Bi-LSTM layer are used to perform multi-step prediction, obtain the predicted state quantity of each node to obtain a fault diagnosis sequence, and output the fault diagnosis sequence as multi-step prediction data.
[0022] The AI time series prediction in this solution adopts the LSTM model, which can effectively solve the problem of long-term dependency between data and allow selective memory and forgetting of information, making it more flexible and efficient in processing complex sequence tasks. It adopts bidirectional LSTM (BiLSTM) cascade to capture context changes and improve the performance of multi-step prediction. At the same time, it introduces a multi-head attention mechanism to further enhance the ability to handle data jumps.
[0023] In a further embodiment, fault classification is performed based on the classification model in the LSTM model and the multi-step prediction data, including:
[0024] Input the multi-step prediction data into a first classification model, obtain a classification result after processing by a normalization layer, and then identify whether there is a first node where a fault occurs; input the multi-step prediction data into a second classification model, obtain a classification result after processing by a normalization layer, and then identify whether there is a second node where a fault occurs;
[0025] When the first node and the second node are identified at the same time, it is determined whether the first node and the second node are both within the prediction time period. If so, a phased prediction conclusion that a faulty node exists within the prediction time period is output as a fault classification result. If not, the next time series fault prediction is entered.
[0026] The first classification model adopts the random forest model; the second classification model adopts the XGBoost model.
[0027] This solution introduces the random forest model and the XGBoost model to simultaneously classify faults. XGBoost uses parallel computing and caching technology to quickly process large-scale data sets and high-dimensional features and provide high-precision prediction results. Random forest can process continuous, discrete, and mixed data without the need for data normalization and is suitable for various types of data sets. The two classification models complement each other and can effectively optimize the prediction success rate.
[0028] In a further embodiment, the LSTM model is diagnosed and evaluated based on the historical fault classification results and the historical detection data, including:
[0029] Acquire a corresponding historical fault node from the historical detection data; perform diagnostic correction on the fault classification result based on the historical fault node to obtain a diagnostic correction value;
[0030] The LSTM model is diagnosed and evaluated based on the diagnostic correction value, and at least one of the prediction accuracy, precision, recall rate and false alarm rate is used as an evaluation indicator. It is judged whether the training standard is met according to the evaluation indicator. If so, it is judged that the model has completed training. If not, the historical detection data of the target device is re-acquired to continue model training.
[0031] This solution performs diagnostic evaluation on the LSTM model based on fault classification results and historical detection data, performs diagnostic correction on the fault classification results through historical fault nodes, and performs diagnostic evaluation on the LSTM model based on the diagnostic correction value. It can determine whether the model can accurately make predictions or classifications, thereby continuously optimizing the model during the iterative development process, thereby ensuring the accuracy and reliability of the model.
[0032] In a further embodiment, the performing diagnostic correction on the fault classification result based on the historical fault node to obtain a diagnostic correction value includes:
[0033] Acquire a predicted fault node from the fault classification result;
[0034] Acquire a fault diagnosis sequence from the multi-step prediction data, and determine the location of the fault node from the fault diagnosis sequence;
[0035] Determine whether the historical fault node is within the preset step length of the fault node. If not, it is determined that the prediction is wrong and the fault node on the fault diagnosis sequence is not processed. If yes, it is determined that the prediction is successful, and the target node corresponding to the historical fault node is determined in the fault diagnosis sequence and the diagnosis correction is performed. The diagnosis correction is as follows:
[0036] If the fault node coincides with the target node, the corresponding fault diagnosis sequence is directly output as the diagnosis correction value; otherwise, the target node is corrected to a fault node, and the original fault node is replaced with a regular node to obtain a corrected fault diagnosis sequence as the diagnosis correction value.
[0037] In the diagnosis and correction process of this scheme, when the fault prediction is confirmed to be successful, if the fault node coincides with the target node, the corresponding fault diagnosis sequence is directly output as the diagnosis correction value; otherwise, the target node is corrected to the fault node, and the original fault node is replaced by a regular node to obtain the corrected fault diagnosis sequence as the diagnosis correction value, which can directly evaluate and help determine whether the model can work effectively in the real world, thereby assisting in the optimization and prediction of the model.
[0038] In a further embodiment, the LSTM model is subjected to a diagnostic evaluation based on the diagnostic correction value, comprising:
[0039] Calculate the predicted life cycle based on the diagnostic correction value, and calculate the actual life cycle based on the historical fault nodes;
[0040] A single life cycle sampling is performed by performing a one-to-one matching between the predicted life cycle and the actual life cycle to obtain a first sampling matching result;
[0041] Uniform sampling is performed by one-to-one matching between the predicted life cycle and the actual life cycle to obtain a second sampling matching result;
[0042] The first sampling matching result and the second sampling matching result are input into a confusion matrix for comprehensive diagnosis evaluation.
[0043] This solution uses average sampling for evaluation, which can measure the absolute accuracy of the algorithm's prediction of the equipment, making the evaluation more representative. It uses single life cycle sampling for evaluation, matching the actual life cycle and the predicted life cycle one by one, making the indicator statistics more consistent with the actual situation. Finally, a comprehensive diagnostic evaluation is performed to improve the accuracy of the evaluation.
[0044] In a further implementation scheme, it also includes: performing an area under the ROC curve test and evaluation on the first classification model and the second classification model, performing a correlation analysis on the false alarm rate and the recall rate, constructing a difference change coefficient in the development process, importing the difference change coefficient into the first classification model and the second classification model, and performing an area under the ROC curve test and evaluation again, so as to accurately optimize the first classification model and the second classification model.
[0045] The ROC curve of this solution can intuitively show the performance of the system at different sensitivity levels by plotting the relationship between the correct detection rate (true positive rate) and the wrong detection rate (false positive rate) of the system under different thresholds, thereby optimizing the system parameters of the first classification model and the second classification model, and being more suitable for actual equipment fault prediction.
[0046] The present invention provides an electronic device, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor implements the above-mentioned equipment failure predictive maintenance method when executing the computer program.
[0047] The present invention also provides a computer-readable storage medium having a computer program stored thereon, and when the computer program is executed by a processor, the above-mentioned method for predictive maintenance of equipment failures is implemented. BRIEF DESCRIPTION OF THE DRAWINGS
[0048] Figure 1 It is a work flow chart of a method for predictive maintenance of equipment failures provided by an embodiment of the present invention;
[0049] Figure 2 is a schematic diagram of multivariate model training provided by an embodiment of the present invention;
[0050] Figure 3 It is a model design diagram of the time series prediction model provided by an embodiment of the present invention;
[0051] Figure 4 is a diagnostic correction schematic diagram of a fault diagnosis sequence provided by an embodiment of the present invention;
[0052] Figure 5 This is a diagram showing the effect of a predictive maintenance algorithm provided by an embodiment of the present invention;
[0053] Figure 6 It is a sampling schematic diagram of single life cycle sampling provided by an embodiment of the present invention. DETAILED DESCRIPTION
[0054] The following specifically illustrates the implementation mode of the present invention in conjunction with the accompanying drawings. The embodiments are provided for illustrative purposes only and cannot be understood as limiting the present invention. The accompanying drawings are provided for reference and illustration only and do not constitute a limitation on the scope of patent protection of the present invention, because many changes can be made to the present invention without departing from the spirit and scope of the present invention.
[0055] Example 1
[0056] An embodiment of the present invention provides a method for predictive maintenance of equipment failures, such as Figure 1 to Figure 6 As shown, in this embodiment, steps S1 to S3 are included:
[0057] S1. Obtain real-time collected data from the target device and perform feature extraction to obtain real-time feature data, including:
[0058] The real-time collected data of the target device is collected in real time, the real-time collected data is classified, and feature extraction is performed to obtain one or more of the maximum value, minimum value, median value, mean value, time length and frequency corresponding to each classification as real-time feature data.
[0059] All real-time feature data are normalized, scaled using MinMaxScaler, and preprocessed using savgol_filter, and finally input into the model for fault prediction.
[0060] S2. Execute multivariate prediction of time series based on the time series prediction model in the LSTM model and the real-time feature data to obtain multi-step prediction data;
[0061] See also Figure 2 For example, the multivariate prediction includes: performing time series prediction (output is 50*8) on the feature data of the above multivariate (input is 500*8).
[0062] In this embodiment, see Figure 3 , the time series prediction model includes a first convolutional layer (Conv1D), a first pooling layer (MAXPooling&Drop), a second convolutional layer (Conv1D), a second pooling layer (MAXPooling&Drop), a first Bi-LSTM layer, a second Bi-LSTM layer, a third Bi-LSTM layer, a Dropout layer, an Attention layer (Multi-Head Attention), an Add&Norm layer, a first fully connected layer (Linear&Relu, Relu activation function), a second fully connected layer (Linear&Relu, Relu activation function), and a third fully connected layer (Linear&Sigmod, Sigmod activation function), which are sequentially connected;
[0063] The Attention layer adopts a multi-head attention mechanism;
[0064] The Add&Norm layer is used to perform residual connection calculation;
[0065] The first Bi-LSTM layer, the second Bi-LSTM layer, and the third Bi-LSTM layer are used to perform multi-step prediction, obtain the predicted state quantity of each node to obtain a fault diagnosis sequence, and output the fault diagnosis sequence as multi-step prediction data.
[0066] In this embodiment, the time series prediction model is a prediction model based on bidirectional LSTM. Bi-LSTM captures the forward and backward dependencies in the sequence by combining two independent LSTM networks, forward and reverse (as shown in the figure below), thereby improving the model's ability to understand sequence data.
[0067] The AI time series prediction in this embodiment adopts the LSTM model, which can effectively solve the long-term dependency problem between data and allow selective memory and forgetting of information, making it more flexible and efficient in processing complex sequence tasks; a bidirectional LSTM (BiLSTM) cascade is used to capture context changes and improve the performance of multi-step prediction; at the same time, a multi-head attention mechanism is introduced to further enhance the ability to handle data jumps.
[0068] S3, performing fault classification based on the classification model in the LSTM model and the multi-step prediction data, and outputting a phased prediction conclusion for alarm maintenance;
[0069] Wherein, performing fault classification based on the classification model in the LSTM model and the multi-step prediction data includes:
[0070] Input the multi-step prediction data into a first classification model, obtain a classification result after processing by a normalization layer, and then identify whether there is a first node where a fault occurs; input the multi-step prediction data into a second classification model, obtain a classification result after processing by a normalization layer, and then identify whether there is a second node where a fault occurs;
[0071] When the first node and the second node are identified at the same time, it is determined whether the first node and the second node are both within the prediction time period. If so, a phased prediction conclusion that a faulty node exists within the prediction time period is output as a fault classification result. If not, the next time series fault prediction is entered.
[0072] The first classification model adopts a random forest model; the second classification model adopts an XGBoost model; specifically, when the random forest model and the XGBoost model simultaneously output that a fault exists within the prediction time period, it is determined that the fault is identified within the prediction time period, otherwise it is considered that the fault is not identified within the prediction time period, and then the model evaluation or prediction conclusion output is entered.
[0073] This embodiment introduces the random forest model and the XGBoost model to perform fault classification simultaneously. XGBoost uses parallel computing and caching technology, which can quickly process large-scale data sets and high-dimensional features and provide high-precision prediction results; random forest can process continuous, discrete and mixed data without data normalization and is applicable to various types of data sets. The two classification models confirm each other and can effectively optimize the prediction success rate.
[0074] In this embodiment, it also includes: performing an area under the ROC curve test and evaluation on the first classification model and the second classification model, performing a correlation analysis on the false alarm rate and the recall rate, constructing a difference change coefficient in the development process, importing the difference change coefficient into the first classification model and the second classification model, and performing an area under the ROC curve test and evaluation again, so as to accurately optimize the first classification model and the second classification model.
[0075] The ROC curve of this embodiment can intuitively display the performance of the system at different sensitivity levels by plotting the relationship between the correct detection rate (true positive rate) and the error detection rate (false positive rate) of the system under different thresholds, thereby optimizing the system parameters of the first classification model and the second classification model, and being more suitable for actual equipment fault prediction.
[0076] Among them, outputting the phased prediction conclusion for alarm maintenance includes: after model prediction, judging "the phased prediction conclusion of the fault node in the prediction time period as the fault classification result" based on the fault diagnosis sequence classification, and timely alarming and pushing relevant information according to the real-time prediction results. Taking the coating equipment as an example, the nozzle failure will occur in the next 24 hours and the air pressure abnormality failure will occur in the next 3 days, and then remind the user to carry out maintenance (such as cleaning the valve body).
[0077] In this embodiment, the LSTM model is trained based on historical detection data, and diagnostic evaluation is performed in combination with the historical detection data and the corresponding historical fault classification results, including steps A1 to A4:
[0078] A1. Obtain historical detection data of the target device and perform feature extraction to obtain historical feature data;
[0079] Taking coating equipment as an example, the specific details are:
[0080] Obtain historical collected data of the target device, analyze the historical collected data to obtain at least one of solder paste thickness data (the coating anomaly "including area anomaly and sputtering anomaly" is highly correlated with solder paste thickness), coating area data, and air pressure data as historical detection data.
[0081] Feature extraction is performed on the historical detection data to obtain historical feature data, where the historical feature data includes one or more of a maximum value, a minimum value, a median value, a mean value, a time length, and a frequency.
[0082] Specifically, air pressure sensors, cameras, current sensors, and voltage sensors are used to collect data, and data such as solder paste thickness, coating area, and air pressure value are analyzed as historical detection data. The data is divided into training sets and validation sets according to a certain ratio, and then all the data are normalized, feature scaled using MinMaxScaler, and preprocessed with savgol_filter filtering, and finally input for model training.
[0083] The Savitzky-Golay filter smoothes the data by fitting a polynomial and retains the high-frequency characteristics of the signal, which can preserve the characteristics of the signal better than other smoothing methods (such as moving average).
[0084] Key feature data related to the failure are extracted in units of product numbers, including maximum value, minimum value, median value, valve opening time length and number of changes (frequency), coating area and solder paste thickness, etc. The correlation between the coating equipment failure and the above data is analyzed, and the data with strong correlation is used as the original input data of the model.
[0085] This embodiment combines historically collected data (such as solder paste thickness, coating area, and air pressure value) with feature extraction, and uses the highly correlated data as the original input data of the model to perform time series prediction and fault category prediction. Based on the obtained abnormal causes, effective guidance can be provided to operation and maintenance personnel to further improve operating efficiency.
[0086] A2. Inputting the historical feature data into the time series prediction model to perform multivariate prediction of the time series and output multi-step prediction data;
[0087] A3. Classify faults based on the classification model and the multi-step prediction data, and output a phased prediction conclusion as a historical fault classification result;
[0088] A4. Perform diagnostic evaluation on the LSTM model based on the historical fault classification results and the historical detection data, and apply the model to perform fault prediction when it is judged to meet the training criteria.
[0089] The LSTM model is diagnosed and evaluated according to the historical fault classification results and the historical detection data, including:
[0090] A41, obtaining a corresponding historical fault node from the historical detection data; and performing diagnostic correction on the fault classification result based on the historical fault node to obtain a diagnostic correction value, including:
[0091] Acquire a predicted fault node from the fault classification result;
[0092] Acquire a fault diagnosis sequence from the multi-step prediction data, and determine the location of the fault node from the fault diagnosis sequence;
[0093] It is determined whether the historical fault node is within the preset step length of the fault node. If not, it is determined that the prediction is wrong and the fault node on the fault diagnosis sequence is not processed; if yes, it is determined that the prediction is successful, and the target node D corresponding to the historical fault node is determined in the fault diagnosis sequence and the diagnosis correction is performed. The diagnosis correction is as follows:
[0094] If the faulty node coincides with the target node D, the corresponding fault diagnosis sequence is directly output as the diagnosis correction value; otherwise, the target node D is corrected to a faulty node, and the original faulty node is replaced with a regular node to obtain a corrected fault diagnosis sequence as the diagnosis correction value.
[0095] See also Figure 4 , set the faulty node in the sequence to 1, and the non-faulty node in the sequence to 0. Obtain the predicted faulty node from the fault classification result as the diagnosis true value sequence; obtain the fault diagnosis sequence from the multi-step prediction data as the diagnosis prediction value sequence.
[0096] At this time, each fault node (1) in the diagnosis prediction value sequence is identified, and within a certain number of steps amend_step before and after it (i.e., within the prediction time period), it is determined whether the diagnosis true value has a fault in the interval; if so, correction is performed, that is, the corresponding position of the predicted position and the true value fault position (i.e., the target node D) is exchanged and corrected to obtain the diagnosis correction value.
[0097] In the diagnosis and correction process of this embodiment, when confirming that the fault prediction is successful, if the fault node coincides with the target node D, the corresponding fault diagnosis sequence is directly output as the diagnosis correction value; otherwise, the target node D is corrected to a fault node, and the original fault node is replaced with a regular node to obtain a corrected fault diagnosis sequence as the diagnosis correction value, which can directly evaluate and help determine whether the model can work effectively in the real world, thereby assisting in the optimization and prediction of the model.
[0098] A42. Perform a diagnostic evaluation on the LSTM model based on the diagnostic correction value, and determine whether it meets the training standards based on the evaluation index. If so, determine that the model has completed training. If not, re-acquire the historical detection data of the target device to continue model training.
[0099] In this embodiment, the evaluation indicators of the model include but are not limited to prediction accuracy, mean absolute error, recall rate, ROC, precision, recall rate and false alarm rate, for example:
[0100]
[0101] Where:
[0102] TP (True Positive): A real fault is correctly detected as a fault;
[0103] TN (True Negative): The real normal is correctly detected as normal;
[0104] FP (False Positive): A true normal state is mistakenly detected as a fault;
[0105] FN (False Negative): A real fault is mistakenly detected as normal.
[0106] In this embodiment, the LSTM model is diagnosed and evaluated based on the diagnosis correction value, including:
[0107] Calculate the predicted life cycle based on the diagnostic correction value, and calculate the actual life cycle based on the historical fault nodes;
[0108] A single life cycle sampling is performed by performing a one-to-one matching between the predicted life cycle and the actual life cycle to obtain a first sampling matching result;
[0109] Uniform sampling is performed by one-to-one matching between the predicted life cycle and the actual life cycle to obtain a second sampling matching result;
[0110] The first sampling matching result and the second sampling matching result are input into a confusion matrix for comprehensive diagnosis evaluation.
[0111] Among them, see Figure 6 The life cycle contains multiple nodes, and the number of remaining uses in each node is the corresponding number. The sampling principle of single life cycle sampling is as follows:
[0112] The predicted life cycle is carried out in sequence and matched with the real life cycle in sequence. When the two have overlapping position intervals, the match is successful;
[0113] After the match is successful, the overlapped position interval is used as the sampling point to calculate the prediction index;
[0114] Then remove this segment from the actual life cycle and continue with the next life cycle until all predicted life cycles are fully counted.
[0115] The sampling principle of uniform sampling is: in the predicted life cycle and the actual life cycle, the remaining usage times are sampled once every T intervals to obtain sampling points.
[0116] This embodiment adopts average sampling for evaluation, which can measure the absolute accuracy of the algorithm's prediction of the device, and the evaluation is more representative; single life cycle sampling is adopted for evaluation, and the actual life cycle and the predicted life cycle are matched one by one, so that the indicator statistics are more consistent with the actual situation, and finally a comprehensive diagnostic evaluation is performed to improve the accuracy of the evaluation.
[0117] This embodiment performs diagnostic evaluation on the LSTM model based on the fault classification results and historical detection data, performs diagnostic correction on the fault classification results through historical fault nodes, and performs diagnostic evaluation on the LSTM model based on the diagnostic correction values. It can be determined whether the model can accurately make predictions or classifications, thereby continuously optimizing the model during the iterative development process, thereby ensuring the accuracy and reliability of the model.
[0118] When training the model, this embodiment improves the accuracy and effect of the model by autonomously learning and optimizing the model. In addition, it also combines the historical detection data actually collected with the historical fault classification results to achieve diagnostic evaluation of the LSTM model, and objectively reflects the model's predictive ability on new data to discover problems and deficiencies in the model, so as to make targeted improvements and improve the model's prediction accuracy.
[0119] See attached Figure 5 In this embodiment, the total number of actual parameter samples is 9622, the number of training samples is 6730, and the number of test samples is 2892. The diagnostic hit rate (TPR) of the diagnostic model in this embodiment for the true value is 93%; the false positive rate (FPR) is 0%, which has a good prediction effect.
[0120] This embodiment only takes coating equipment as an example, but it can also be applied to fault predictive maintenance of other complex equipment, such as vehicle-mounted assembly equipment and performance testing equipment.
[0121] The equipment fault predictive maintenance method proposed in the embodiment of the present invention uses historical detection data to train the LSTM model, predicts the life of equipment (such as complex equipment such as coating equipment), uses a time series prediction model to perform multivariate prediction of time series, outputs multi-step prediction data, and then classifies faults based on the classification model to determine the fault classification results. Finally, the fault classification results and historical detection data are used to diagnose and evaluate the LSTM model. Combined with the model, it has good learning accuracy and generalization ability, can directly predict multi-source vibration signals, effectively improve the efficiency and accuracy of life prediction, and reduce operation and maintenance costs and prevent accidents.
[0122] Example 2
[0123] An embodiment of the present invention provides an electronic device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor implements the equipment failure predictive maintenance method as described in the above embodiment 1 when executing the computer program.
[0124] Example 3
[0125] The present invention also provides a computer-readable storage medium, on which a computer program is stored, and when the computer program is executed by a processor, the method for predictive maintenance of equipment failures as described in the above embodiment 1 is implemented. The storage medium may be a magnetic disk, an optical disk, a read-only memory (ROM) or a random access memory (RAM), etc.
[0126] The above embodiments are preferred implementation modes of the present invention, but the implementation modes of the present invention are not limited to the above embodiments. Any other changes, modifications, substitutions, combinations, and simplifications that do not deviate from the spirit and principles of the present invention should be equivalent replacement methods and are included in the protection scope of the present invention.
Claims
1. A method for predictive maintenance of equipment failure, characterized in that: include, Acquire real-time collected data from the target device and perform feature extraction to obtain real-time feature data; Perform multivariate prediction of time series based on the time series prediction model in the LSTM model and the real-time feature data to obtain multi-step prediction data; Fault classification is performed based on the classification model in the LSTM model and the multi-step prediction data, and a phased prediction conclusion is output for alarm maintenance; The LSTM model is trained based on historical detection data, and diagnostic evaluation is performed in combination with the historical detection data and corresponding historical fault classification results.
2. A method for predictive maintenance of equipment failures according to claim 1, characterized in that: The LSTM model is trained based on historical detection data and performs diagnostic evaluation in combination with corresponding historical fault classification results, including: Obtain historical detection data of the target device and perform feature extraction to obtain historical feature data; Inputting the historical feature data into the time series prediction model to perform multivariate prediction of the time series and output multi-step prediction data; Perform fault classification based on the classification model and the multi-step prediction data, and output a phased prediction conclusion as a historical fault classification result; The LSTM model is diagnosed and evaluated based on the historical fault classification results and the historical detection data, and the model is applied to perform fault prediction when it is judged to meet the training standards.
3. A method for predictive maintenance of equipment failures according to claim 1, characterized in that: The time series prediction model includes a first convolutional layer, a first pooling layer, a second convolutional layer, a second pooling layer, a first Bi-LSTM layer, a second Bi-LSTM layer, a third Bi-LSTM layer, a Dropout layer, an Attention layer, an Add&Norm layer, a first fully connected layer, a second fully connected layer, and a third fully connected layer, which are connected in sequence; The Attention layer adopts a multi-head attention mechanism; The Add&Norm layer is used to perform residual connection calculation; The first Bi-LSTM layer, the second Bi-LSTM layer, and the third Bi-LSTM layer are used to perform multi-step prediction, obtain the predicted state quantity of each node to obtain a fault diagnosis sequence, and output the fault diagnosis sequence as multi-step prediction data.
4. A method for predictive maintenance of equipment failures as claimed in claim 2, characterized in that: Fault classification is performed based on the classification model in the LSTM model and the multi-step prediction data, including: Input the multi-step prediction data into a first classification model, obtain a classification result after processing by a normalization layer, and then identify whether there is a first node where a fault occurs; input the multi-step prediction data into a second classification model, obtain a classification result after processing by a normalization layer, and then identify whether there is a second node where a fault occurs; When the first node and the second node are identified at the same time, it is determined whether the first node and the second node are both within the prediction time period. If so, a phased prediction conclusion that a faulty node exists within the prediction time period is output as a fault classification result. If not, the next time series fault prediction is entered. The first classification model adopts the random forest model; the second classification model adopts the XGBoost model.
5. A method for predictive maintenance of equipment failures as claimed in claim 3, characterized in that: The LSTM model is diagnosed and evaluated according to the historical fault classification results and the historical detection data, including: Acquire a corresponding historical fault node from the historical detection data; perform diagnostic correction on the fault classification result based on the historical fault node to obtain a diagnostic correction value; The LSTM model is diagnosed and evaluated based on the diagnostic correction value, and at least one of the prediction accuracy, precision, recall rate and false alarm rate is used as an evaluation indicator. It is judged whether the training standard is met according to the evaluation indicator. If so, it is judged that the model has completed training. If not, the historical detection data of the target device is re-acquired to continue model training.
6. A method for predictive maintenance of equipment failures as claimed in claim 5, characterized in that: The performing diagnostic correction on the fault classification result based on the historical fault node to obtain a diagnostic correction value includes: Acquire a predicted fault node from the fault classification result; Acquire a fault diagnosis sequence from the multi-step prediction data, and determine the location of the fault node from the fault diagnosis sequence; Determine whether the historical fault node is within the preset step length of the fault node. If not, it is determined that the prediction is wrong and the fault node on the fault diagnosis sequence is not processed. If yes, it is determined that the prediction is successful, and the target node corresponding to the historical fault node is determined in the fault diagnosis sequence and the diagnosis correction is performed. The diagnosis correction is as follows: If the fault node coincides with the target node, the corresponding fault diagnosis sequence is directly output as the diagnosis correction value; otherwise, the target node is corrected to a fault node, and the original fault node is replaced with a regular node to obtain a corrected fault diagnosis sequence as the diagnosis correction value.
7. A method for predictive maintenance of equipment failures as claimed in claim 6, characterized in that: Performing a diagnostic evaluation on the LSTM model based on the diagnostic correction value includes: Calculate the predicted life cycle based on the diagnostic correction value, and calculate the actual life cycle based on the historical fault nodes; A single life cycle sampling is performed by performing a one-to-one matching between the predicted life cycle and the actual life cycle to obtain a first sampling matching result; Uniform sampling is performed by one-to-one matching between the predicted life cycle and the actual life cycle to obtain a second sampling matching result; The first sampling matching result and the second sampling matching result are input into a confusion matrix for comprehensive diagnosis evaluation.
8. A method for predictive maintenance of equipment failures as claimed in claim 4, characterized in that: Also includes: The first classification model and the second classification model are evaluated by the area under the ROC curve test, the false alarm rate and the recall rate are correlated, the difference change coefficient in the development process is constructed, the difference change coefficient is imported into the first classification model and the second classification model, and the area under the ROC curve test is performed again, so as to accurately optimize the first classification model and the second classification model.
9. An electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that: When the processor executes the computer program, the method for predictive maintenance of equipment failures according to any one of claims 1 to 8 is implemented.
10. A storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the method for predictive maintenance of equipment failures as claimed in any one of claims 1 to 8 is implemented.