Circulating filtering system fault positioning method based on step-by-step optimization gradient decision tree

Abstract

The invention discloses a circulating filtration system fault positioning method based on a step-by-step optimization gradient decision tree, and the method comprises the following steps: S1, collecting real-time signal data of a circulating filtration system of metal additive manufacturing equipment through a multi-source sensor, and carrying out the normalization index processing and fusion of the collected data; s2, eliminating and extracting features of different signals in the circulating filtration system from the fused data; s3, labels of different signal features are determined; adopting grid search to obtain optimal parameters of a gradient decision tree algorithm so as to construct a fault positioning model; training a fault positioning model; and S4, the implementation activity of the circulating filtration system of the metal additive manufacturing equipment is monitored, and the real-time state of the equipment is determined. According to the method, input data of the multi-source sensor is provided for the fault positioning model, and data redundancy is effectively reduced through data preprocessing and optimization of genetic algorithm hyper-parameters, so that the accuracy and efficiency of the fault diagnosis model are greatly improved.

Description

Technical Field

[0001] The present invention relates to the technical field of equipment fault location diagnosis, and in particular to a method for locating faults of a cyclic filtering system based on a step-by-step optimized gradient decision tree. Background Art

[0002] Metal additive manufacturing (MAM) has attracted much attention in the industrial field because of its ability to manufacture high-performance parts with complex geometries. Selective laser melting (SLM) technology is a process in which a powder-laying device evenly spreads the raw metal powder on a substrate according to a set layer thickness, and melts the metal powder layer by layer according to a predetermined scanning path with the help of a high-energy-density laser beam. Then, the substrate is lowered by a layer thickness, and the processing of the lower layer is repeated. In this way, metal parts are processed layer by layer. However, in actual operation, the rapid changes in pressure and gas content in the equipment bin may lead to defects in part geometry and performance, such as excessive oxygen content causing oxidation of the metal surface, and excessive airflow causing splashing. These problems not only affect the quality and production efficiency of parts, but also may bring safety hazards. Therefore, it is urgent to develop a system for online fault location diagnosis that can detect and solve equipment problems in a timely manner.

[0003] However, metal additive manufacturing equipment is a complex equipment with multiple modules and multiple processes. It is prone to multiple faults and its mathematical model is difficult to establish. In the additive manufacturing process, problems such as sensor errors, noise interference, or the processing methods of the system's internal algorithms will lead to a large amount of redundant data. There is currently no good research method for fault location diagnosis of this type of equipment. Some fault location diagnosis algorithms based on machine learning, such as BP neural networks, support vector machines (SVM), XGBoost, etc., have limitations in their nonlinear fitting capabilities, which makes it difficult to improve the model's diagnostic accuracy for multi-dimensional signals. When processing real-time operating data of metal additive manufacturing equipment, there will be problems in which accuracy and efficiency cannot meet industrial requirements at the same time. Summary of the invention

[0004] In view of the shortcomings of the prior art, it is necessary to provide a fault location diagnosis system for the circulation filtration system of metal additive manufacturing equipment to solve the problems of insufficient accuracy and efficiency of fault location diagnosis in the prior art.

[0005] The present invention provides a method for locating faults in a cyclic filtering system based on a step-by-step optimized gradient decision tree, comprising the following steps:

[0006] S1, data collection and preprocessing: using multi-source sensors (oxygen transmitter, pressure sensor, air flow sensor, anemometer, limit switch), collect real-time signal data of the circulation filtration system of metal additive manufacturing equipment, including oxygen content, pressure, flow, etc., and perform normalized index processing and fusion on the collected data;

[0007] S2, feature extraction: using variance filtering and recursive feature elimination from the fused data to extract the features of different signals in the cyclic filtering system, the signal features include the oxygen content in the processing chamber, the total intake pressure, the processing chamber pressure, etc.;

[0008] S3, model construction: determine the labels based on the extracted different signal features in the loop filtering system; use grid search to obtain the optimal parameters of the gradient decision tree algorithm to build a fault location model; train the fault location model to learn the relationship between the signal features of the loop filtering system and fault location to achieve fault location;

[0009] S4, Fault Monitoring: Use the fault location model to monitor the implementation activities of the circulation filtration system of the metal additive manufacturing equipment to determine the real-time status of the equipment.

[0010] Furthermore, the steps of normalized index processing and fusion in S1 are as follows:

[0011] Feature information x1, x2,...x of different channels k 、x k+1 The new feature information graph is obtained by row feature fusion (element addition), and the formula is as follows:

[0012] U=x1+x2+…+x k +x k+1 (1)

[0013] In the formula, x1, x2…x k and x k+1 represents the input features of different channels, 1, 2, ...k and k+1 represent the 1st channel, 2nd channel, kth channel and k+1 channel respectively, and U represents the value after feature fusion;

[0014] In the second step, the feature information is flattened into one-dimensional features through global average pooling to achieve feature dimensionality reduction, and then the feature information of different channels is weighted by the Softmax function. The formula is as follows:

[0015]

[0016]

[0017] In the formula, and Represents the output value of different channel features after normalization by the Softmax function. and They represent the probability values ​​of the 1st, 2nd, ...k and k+1st channels transitioning to between 0 and 1, It means summing the probability values ​​of all channels.

[0018] The third step is to multiply the output value of the Softmax function and the input feature map to generate a new weighted feature map, and then fuse the new feature map. The formula is:

[0019]

[0020] In the formula, u i represents the weighted feature map generated by the i-th channel, Represents the output of the multi-channel feature fusion module.

[0021] Furthermore, the feature extraction in S2 includes:

[0022] Variance filtering: The variance of different preprocessed data sets is calculated separately, and the two data sets with the smallest variance value σ are eliminated.

[0023]

[0024] Where n is the total number of values; X i is each value, and μ is the mean of this set of values.

[0025] S22, recursive feature elimination: the importance of features is ranked by fault diagnosis through the initial gradient decision tree model. In each round of training, feature parameters with importance less than 2% are eliminated. Iterate until the number of features is half of the number of original collected signals.

[0026] Furthermore, the model construction in S3 includes:

[0027] Define feature labels: Add fault location labels to the feature data T[T1, T2, T3…] retained after preprocessing according to the feature signals of the cyclic filtering system. Each category corresponds to a unique label.

[0028] Model establishment: First build a gradient decision tree, using a fast, distributed, high-performance GBDT-based learner. Calculate the split gain of the tree:

[0029]

[0030] Among them, G L , G R is the sum of the gradients of the left and right subtrees, H L , HR is the sum of the corresponding second-order gradients, and λ is the regularization parameter.

[0031] In the process of iteratively adding decision trees, the gradient g is calculated i :

[0032]

[0033] Among them, Loss is the loss function, which measures the model prediction and the true label y i The gap between. is the prediction for the i-th sample of the model.

[0034] Model optimization: Change the hyperparameters in the gradient decision tree algorithm, including the number of base learners, the maximum depth and number of leaves of the tree, L1 regularization and L2 regularization parameters. The hyperparameter tuning steps are as follows:

[0035] First, determine the number of base learners under a small learning rate. In the GBDT framework, use one-dimensional grid enumeration to change the number of base learners. Perform model training with other parameters set to default. Use the prediction formula to determine the optimal number of base learners:

[0036]

[0037] Among them, y pred is the prediction result of the model, η is the learning rate, f m (x) is the prediction result of the mth tree.

[0038] Then determine the maximum depth and number of leaves of the tree, use a 7*70 matrix grid for model training, and find the maximum depth and number of leaves that make the split gain Gain;

[0039] Then determine the L1 regularization and L2 regularization parameters, use a 10*10 matrix grid for model training, and find the optimal balance point of the regularization objective function to obtain the minimum loss function value. The objective function formula is as follows:

[0040]

[0041] Among them, λ1 is the hyperparameter of L1 regularization, λ2 is the parameter of L2 regularization, and θ is the multidimensional model parameter.

[0042] Finally, the number of base learners is re-determined. After determining the above parameters, the one-dimensional grid is still used to change the number of base learners for model training. The number of base learners with the maximum split gain and the highest accuracy is obtained.

[0043] Further, the fault monitoring in S4 includes:

[0044] The selected real-time data group of characteristic signals is substituted into the fault location model, and judgment is made according to the TP, TN, FN and FP values ​​in the model output confusion matrix. If the TN and TP values ​​increase, it means that the prediction is successful. The set status label is directly read to judge the health status or fault location of the loop filtering system reflected by the real-time data; if the FN and FP values ​​increase, it means that the prediction fails, and the fault location model is rebuilt according to the latest real-time data.

[0045] Among them, according to the confusion matrix and evaluation indicators output by the model, TP and TN are successfully predicted data; FN and FP are wrongly predicted data, which are used to judge the health status or fault location of the loop filtering system reflected by real-time data.

[0046] According to one aspect of the present invention, a storage medium is provided, wherein instructions are stored in the storage medium. When a computer reads the instructions, the computer executes any one of the above-mentioned methods for locating faults in a cyclic filtering system based on a step-by-step optimized gradient decision tree.

[0047] According to another aspect of the present invention, there is provided an electronic device, comprising a processor and the above-mentioned storage medium, wherein the processor executes instructions in the storage medium.

[0048] Compared with the prior art, the present invention has the following beneficial effects:

[0049] 1. The present invention can accurately capture and process signal data from the circulation filtration system, including oxygen content, pressure, etc., through efficient data collection and preprocessing steps. Data feature fusion and preprocessing operations improve the quality of data and the accuracy of subsequent analysis, ensuring that the features extracted from the signal can accurately reflect the accuracy and efficiency of fault location in the circulation filtration system.

[0050] 2. The present invention optimizes the gradient decision tree through the grid search method, and the constructed fault location model can realize the accurate location of equipment faults, which not only improves the accuracy of fault location, but also greatly accelerates the speed of fault diagnosis, and provides strong support for the rapid resolution of metal additive manufacturing equipment circulation filtration system faults. The fault location information greatly reduces the maintenance cost and improves the reliability and safety of metal additive manufacturing equipment. By determining the fault location space information, it can provide accurate guidance for the rapid resolution of faults, which greatly improves the efficiency of metal additive manufacturing equipment maintenance and management. BRIEF DESCRIPTION OF THE DRAWINGS

[0051] Other features, objects and advantages of the present invention will become more apparent from the detailed description of non-limiting embodiments made with reference to the following drawings:

[0052] Figure 1A schematic diagram of a real-time fault location method flow chart according to an embodiment of the present invention;

[0053] Figure 2 This is a schematic diagram of a multi-source sensor experiment in a circulating filtration system of a metal additive manufacturing device according to an embodiment of the present invention;

[0054] Figure 3 This is a schematic diagram of the hyperparameter tuning process of the gradient decision tree in an embodiment of the present invention;

[0055] Figure 4 This is an analysis diagram of the abnormal flow rate of protective window air collected by the circulation filtration system of the metal additive manufacturing equipment in an embodiment of the present invention;

[0056] Figure 5 This is a time domain diagram of the rear side panel air intake flow collected by the circulation filtration system of the metal additive manufacturing equipment in an embodiment of the present invention.

[0057] Figure 6 This is a normal / abnormal analysis diagram of the rear side panel air intake flow collected by the circulation filtration system of the metal additive manufacturing equipment in an embodiment of the present invention. DETAILED DESCRIPTION

[0058] In order to make the purpose, technical solutions and advantages of the present invention more clear, the present invention is described in detail below with reference to the accompanying drawings and specific embodiments.

[0059] like Figure 1 As shown, the real-time fault location method of the circulation filtration system of metal additive manufacturing equipment based on the step-by-step optimization gradient decision tree of the present invention comprises the following steps:

[0060] S1, data collection and preprocessing: using multi-source sensors (oxygen transmitter, pressure sensor, air flow sensor, anemometer, limit switch), collect real-time signal data of the circulation filtration system of metal additive manufacturing equipment, including oxygen content, pressure, flow, etc., and perform normalized index processing and fusion on the collected data;

[0061] S2, feature extraction: using variance filtering and recursive feature elimination from the fused data to extract the features of different signals in the cyclic filtering system, the signal features include the oxygen content in the processing chamber, the total intake pressure, the processing chamber pressure, etc.;

[0062] S3, model construction: determine the labels based on the extracted different signal features in the loop filtering system; use grid search to obtain the optimal parameters of the gradient decision tree algorithm to build a fault location model; train the fault location model to learn the relationship between the signal features of the loop filtering system and fault location to achieve fault location;

[0063] S4, Fault Monitoring: Use the fault location model to monitor the implementation activities of the circulation filtration system of the metal additive manufacturing equipment to determine the real-time status of the equipment.

[0064] Data collection and preprocessing in S1 include:

[0065] like Figure 2 As shown in the figure, signal data is collected by deploying multi-source sensors around the circulating filtration system. In this example, the purge stage of the metal additive manufacturing equipment printing process is selected. The overall purge time is 3067s and the number of pauses is 0. The sampling period is 1s, and PLC data collection is used. A total of 5 types of sensors and 12 parameters are involved. The sampling period is 1s, and PLC data collection is used. The specific sensor configuration is shown in Table 1:

[0066] Table 1 Sensor configuration

[0067]

[0068]

[0069] S12, data preprocessing: normalize and fuse the parameters obtained by multi-source sensors.

[0070] Feature extraction in S2 includes:

[0071] S21, variance filtering: remove the two data sets with the smallest variance values, and remove the data sets of oxygen content 1 in processing bin and oxygen content 2 in processing bin.

[0072] S22, recursive feature elimination: After training, 7 sets of feature data are obtained, namely, the air intake flow of the protective window, the air intake flow of the rear side panel, the wind speed of the return air outlet branch circuit, the wind speed of the return air outlet main circuit, the wind speed of the suction air outlet main circuit, the locked door 1, and the locked door 2.

[0073] Through feature extraction, key signal features can be accurately identified from complex circulation filtration system signal data. These features provide key information for subsequent model construction, real-time monitoring and fault location, greatly improving the accuracy and efficiency of the fault location method of the circulation filtration system of metal additive manufacturing equipment.

[0074] Model building in S3 includes:

[0075] Define feature labels: Add fault location labels according to the characteristic signals of the remaining loop filtering system, as shown in Table 2:

[0076] Table 2 Fault location label

[0077]

[0078]

[0079] Model establishment: GBDT base learner is used. The default number of base learners is 100, the learning rate is 0.1, the L1 regularization term is 1, the L2 regularization term is 1, the sample rate is 1, the maximum depth of the tree is 5, the number of leaves is 50, and the split gain of the tree is Gain = 0.543, the accuracy is 0.98, the recall rate is 0.97, and the F1 is 0.987

[0080] like Figure 3 As shown in the figure, model optimization: changing the hyperparameters in the gradient decision tree algorithm, including the number of base learners, the maximum depth and number of leaves of the tree, L1 regularization and L2 regularization parameters. The hyperparameter tuning steps are as follows:

[0081] First, determine the number of base learners under a small learning rate, use a grid with a parameter range of [50,200] and a step size of 5, and enumerate the number of base learners with the highest evaluation index.

[0082] Then determine the maximum depth and number of leaves of the tree, use the maximum depth parameter range of the tree [3,10], step size 1; the number of leaves parameter range [30,100], step size 1 7*70 matrix grid for model training, and find the maximum depth and number of leaves that make the split gain Gain;

[0083] Then determine the L1 regularization and L2 regularization parameters, the parameter range is [0,1], the step size is 0.1, use a 10*10 matrix grid for model training, and find the optimal balance point of the regularization objective function and the minimum loss function value.

[0084] Finally, the number of base learners is re-determined. After the above parameters are determined, the parameter range [50, 200] and the one-dimensional grid with a step size of 1 are still used to change the number of base learners for model training. The number of base learners with the maximum split gain and the highest accuracy is obtained.

[0085] Through the above process, not only can the fault location of the circulating filtration system of the metal additive manufacturing equipment be achieved with high accuracy, but also the changes in the signal characteristics of the circulating filtration system can be adapted through continuous learning and optimization, thereby maintaining high performance in the dynamically changing metal additive manufacturing process.

[0086] Real-time monitoring and early warning in S4 include:

[0087] Real-time monitoring: achieved through multi-source sensors installed on the circulating filtration system.

[0088] Fault location: The collected real-time data is fed into the fault location model, and the confusion matrix and evaluation indicators are tested to determine whether the equipment is healthy and the fault location.

[0089] Real-time monitoring can quickly identify problems, thereby reducing the downtime of metal additive manufacturing equipment and improving production efficiency. Fault warning can reduce the frequency of failures and maintenance costs, and extend the life of metal additive manufacturing equipment through preventive maintenance.

[0090] The examples described in the present invention are merely descriptions of the preferred implementation modes of the present invention, and are not intended to limit the concept and scope of the present invention. Without departing from the design concept of the present invention, various modifications and improvements made to the technical solutions of the present invention by engineers and technicians in this field should all fall within the protection scope of the present invention.

Claims

1. A method for locating faults in a cyclic filtering system based on a step-by-step optimized gradient decision tree, characterized in that: The steps include: S1, data collection and preprocessing: Use multi-source sensors to collect real-time signal data of the circulation filtration system of metal additive manufacturing equipment, and perform normalized index processing and fusion on the collected data. The multi-source sensors include oxygen transmitters, pressure sensors, air flow sensors, anemometers and limit switches. The real-time signal data include oxygen content, pressure and flow; S2, feature extraction: using variance filtering and recursive feature elimination to extract the features of different signals in the cyclic filtering system from the fused data, wherein the signal features include the oxygen content in the processing chamber, the total intake pressure and the processing chamber pressure; S3, model construction: determine the labels based on the extracted different signal features in the loop filtering system; use grid search to obtain the optimal parameters of the gradient decision tree algorithm to build a fault location model; train the fault location model to learn the relationship between the signal features of the loop filtering system and fault location to achieve fault location; S4, Fault Monitoring: Use the fault location model to monitor the implementation activities of the circulation filtration system of the metal additive manufacturing equipment to determine the real-time status of the equipment.

2. The method according to claim 1, characterized in that In step S1, normalized index processing and fusion include the following steps: The characteristic information x1, x2…x transmitted by multiple source sensors k 、x k+1 The row features are fused to obtain the feature information graph; The feature information graph is flattened into one-dimensional features through global average pooling to achieve feature dimensionality reduction, and the feature information of different channels is weighted through the SoftMax function; Multiply the output value of the SoftMax function and the input feature map to generate a new weighted feature map, and then fuse the new feature map.

3. The method according to claim 1, characterized in that In step S3, a fault location model is constructed The following steps are involved: Define feature labels: Add device status labels to the feature data retained after preprocessing according to the feature signals of the loop filtering system. Each category corresponds to a unique label. Model building: Construct a gradient decision tree and calculate the tree split gain Gain: Among them, G L , G R is the sum of the gradients of the left and right subtrees, H L , H R is the sum of the second-order gradients of the left subtree and the right subtree, and λ is the regularization parameter; In the process of iteratively adding decision trees, the gradient g is calculated i : Among them, Loss represents the loss function, y i represents the true value of the i-th sample, Represents the prediction of the i-th sample of the model; Model optimization: Change the hyperparameters in the gradient decision tree algorithm, where the hyperparameters include the number of base learners, the maximum depth and number of leaves of the tree, L1 regularization and L2 regularization parameters. The hyperparameter tuning steps are as follows: Determine the number of base learners under a small learning rate. In the GBDT framework, use one-dimensional grid enumeration to change the number of base learners. Perform model training with other parameters set to default. Use the prediction formula to determine the optimal number of base learners. The specific formula is: Among them, y pred is the prediction result of the model, η is the learning rate, f m (x) is the prediction result of the mth tree; Determine the maximum depth and number of leaves of the tree, use a 7*70 matrix grid for model training, and find the maximum depth and number of leaves that make the split gain Gain; Determine the L1 regularization and L2 regularization parameters, use a 10*10 matrix grid for model training, and find the optimal balance point of the regularization objective function to obtain the minimum loss function value. The specific formula of the objective function J(θ) is as follows: Among them, λ1 is the hyperparameter of L1 regularization, λ2 is the parameter of L2 regularization, and θ represents the vector θ∈R n , where n is the parameter dimension, |θ i | represents the L1 regularization term, the sum of the absolute values ​​of the model parameters; is the L2 regularization term, the sum of squares of model parameters; Re-determine the number of base learners, use a one-dimensional grid, change the number of base learners for model training, and obtain the number of base learners when the split gain is maximum and the accuracy is highest.

4. The method according to claim 1, characterized in that: In step S4, fault monitoring specifically includes: The selected real-time data group of characteristic signals is substituted into the fault location model, and judgment is made according to the TP, TN, FN and FP values ​​in the model output confusion matrix. If the TN and TP values ​​increase, it means that the prediction is successful. The set status label is directly read to judge the health status or fault location of the loop filtering system reflected by the real-time data; if the FN and FP values ​​increase, it means that the prediction fails, and the fault location model is rebuilt according to the latest real-time data.

5. A storage medium, characterized in that: The storage medium stores instructions, and when a computer reads the instructions, the computer executes the method for locating a fault in a cyclic filtering system based on a step-by-step optimized gradient decision tree as described in any one of claims 1 to 4.

6. An electronic device, characterized in that: The invention comprises a processor and the storage medium as claimed in claim 5, wherein the processor executes instructions in the storage medium.

Images