A hydraulic loading device multi-fault diagnosis method based on a decision tree algorithm

Abstract

A multi-fault diagnosis method for hydraulic loading devices based on decision tree algorithm is proposed. First, the imported data is preprocessed. Second, for multi-channel time-series signals, time-domain statistical features are extracted for each sample and each channel. Then, the extracted time-domain statistical features are standardized. A decision tree algorithm is employed, using the Gini coefficient as a standard to select the optimal splitting feature and threshold at each layer and partitioning the data. The tree grows recursively from the root node, and a discrimination tree for various typical faults of the hydraulic loading device is obtained through training. After training, the tree nodes are traversed, forming a complete discrimination path from the root to each leaf. Finally, the discrimination path and threshold conditions are transformed into a tree diagram to intuitively display the relationship between features and fault categories, improving interpretability. This invention integrates feature construction, decision tree analysis, and rule tree visualization for multi-channel time-series data, improving the interpretability of diagnostic results.

Description

Technical Field

[0001] This invention belongs to the field of hydraulic system fault diagnosis technology, specifically relating to a multi-fault diagnosis method for hydraulic loading devices based on decision tree algorithm. Background Technology

[0002] Hydraulic loading devices are widely used in the aerospace field, primarily to provide stable, controllable, and reliable hydraulic power to apply force and displacement to mechanical systems. A hydraulic loading device is an actuation device, typically composed of a hydraulic pump station, a hydraulic substation, a multi-circuit distributor, and a hydraulic loading actuator. In actual operation, hydraulic loading devices usually contain multiple actuators (servo valves, ball valves, actuators, etc.) and various sensors (pressure, flow, displacement, acceleration, etc.), and involve mechanical, electrical, and hydraulic coupling. These characteristics make hydraulic loading devices highly susceptible to various types and sources of failure under long-term operation and complex working conditions.

[0003] Common faults in hydraulic loading devices include leakage inside the actuator cylinder, leakage in the pipeline, and improper closure of the ball valve. If these faults are not diagnosed in time, they will lead to increased execution errors, decreased loading accuracy, and even equipment failure and damage, posing significant hidden dangers, reducing loading efficiency, and threatening the safety of test personnel and property.

[0004] Traditional fault diagnosis methods (Chao Qun, Gao Haohan, Tao Jianfeng, Wang Yuanhang, Zhou Jian, Liu Chengliang. Adaptive decision-level fusion strategy for the fault diagnosis of axial piston pumps using multiple channels of vibrationsignals. Science China-Technological Sciences, 65(2),2022:470-480. Su Xinping, Wang Taiyong, Wan Shumin. Rapid diagnosis and troubleshooting of hydraulic system faults[J]. Machine Tool & Hydraulics, 2003, (02):240.) often rely on empirical rules or simple threshold judgments. When faced with multi-source signal interference and complex working conditions, it is often difficult to guarantee the accuracy and robustness of the diagnosis. Therefore, it is necessary to introduce machine learning methods with efficient information processing capabilities and interpretability to improve the efficiency and intelligence level of the corresponding hydraulic loading device diagnosis.

[0005] Decision trees, as a typical supervised learning method, are based on the core idea of ​​recursively selecting optimal features and splitting thresholds to continuously divide the sample space, thereby forming a hierarchical tree structure. In classification problems, decision trees can utilize data features as input and generate a tree structure through splitting rules based on the Gini coefficient or information gain. This rule-based discrimination method can not only effectively distinguish different types of fault modes but also intuitively present the diagnostic logic in the form of visual tree diagrams and rule statements. It is suitable for implementing interpretable multi-fault diagnosis in hydraulic loading devices, but no relevant literature has been found. Summary of the Invention

[0006] To overcome the shortcomings of the prior art, the present invention aims to provide a multi-fault diagnosis method for hydraulic loading devices based on decision tree algorithm, which integrates feature construction, decision tree analysis, and rule tree visualization for multi-channel time series data. The diagnostic rule tree generated by the decision tree algorithm can clearly show feature split points and fault categories, thereby improving the interpretability of the diagnostic results.

[0007] To achieve the above objectives, the technical solution adopted by the present invention is as follows:

[0008] A multi-fault diagnosis method for hydraulic loading devices based on decision tree algorithm is proposed. First, the imported data is preprocessed to remove outliers. Second, for multi-channel time-series signals, time-domain statistical features are extracted for each sample and each channel. Then, the extracted time-domain statistical features are standardized to eliminate the influence of different dimensions and magnitudes on the diagnosis. A decision tree algorithm is used, employing the Gini coefficient as a standard to select the optimal splitting feature and threshold at each layer and partitioning the data. The tree grows recursively from the root node, and a discrimination tree for various typical faults of the hydraulic loading device is obtained through training. After training, the tree nodes are traversed, forming a complete discrimination path from the root to each leaf. Finally, the discrimination path and threshold conditions are transformed into a tree diagram to intuitively display the relationship between features and fault categories, improving interpretability.

[0009] A multi-fault diagnosis method for a hydraulic loading device based on a decision tree algorithm includes the following steps:

[0010] Step 1) Data Import and Preprocessing: Import the sample data file and its corresponding tag file. The monitored quantities are as follows: oil tank temperature, pump station outlet pressure, pump station outlet flow rate, pump station return oil pressure, pump station return oil flow rate, high-pressure main pipeline pressure, high-pressure main pipeline flow rate, return oil main pipeline pressure, return oil main pipeline flow rate, hydraulic substation front-end high-pressure pressure, hydraulic substation front-end high-pressure flow rate, hydraulic substation front-end return oil pressure, hydraulic substation front-end return oil flow rate, hydraulic substation rear-end high-pressure pressure, hydraulic substation rear-end return oil pressure, oil distributor rear-end high-pressure pressure, oil distributor rear-end high-pressure flow rate, oil distributor rear-end return oil pressure, oil distributor rear-end return oil flow rate, actuator acceleration, actuator displacement, and actuator load. The number of channels is 22, and the sequence length of each channel is [missing information]. N After importing, perform file alignment and dimension validation: check if the sample size is consistent and if the data dimension is 22× N ;

[0011] Step 2) Target the data in the sample channel The mean, standard deviation, skewness, and kurtosis are calculated as four statistical measures and used as conditional features for splitting decision tree nodes.

[0012] mean The calculation formula is:

[0013]

[0014] Standard deviation The calculation formula is:

[0015]

[0016] Skewness Skewness The calculation formula is:

[0017]

[0018] cliff Kurtosis The calculation formula is:

[0019]

[0020] at standard deviation When the skewness and kurtosis are close to zero, set them to zero.

[0021] Extract the four statistical features mentioned above for each channel signal, and concatenate the statistical features of all channels to form a complete multidimensional feature vector: ;

[0022] Step 3) The Z-score normalization method is used to normalize the feature vectors of different dimensions as follows:

[0023]

[0024] in, It is the standardized feature vector. yes The average value, yes Standard deviation;

[0025] Step 4) Randomly divide all multidimensional feature vector data into training and test sets in an 8:2 ratio, and set the seed of the random number generator to 0. s The node partitioning rule is Gini impurity, and the Gini coefficient is used as the splitting standard to measure the impurity of a node; the maximum depth of the decision tree is set to 5, that is, the tree can split at most 5 levels from the root node to the leaf node.

[0026] Step 5) The starting point of the decision tree is the root node, which contains the entire training dataset. D Recursively split from the root node and calculate its Gini impurity:

[0027]

[0028] in Represents a node The middle belongs to the first The proportion of class samples;

[0029] Step 6) Enumerate all possible thresholds for all features across all channels in the sample. Sort the sample values ​​of this feature in the current node, take the midpoint of adjacent values ​​as a possible split point, and divide the node into... and Then calculate the impurity of the weighted child nodes:

[0030]

[0031] Choose to Minimum features and thresholds As the optimal split for this node;

[0032] Step 7) After the split is complete, the data of the current node is allocated to and Find the left and right child nodes, and then repeat steps 5)-6) on the child nodes.

[0033] Step 8) When the decision tree reaches its maximum depth or all samples within a node belong to the same class... When the stopping condition is met, the further splitting of that node is terminated, and the node is marked as a leaf node; the training process ends when all nodes meet the stopping condition.

[0034] Step 9) Rule Derivation and Interpretability Analysis: After training, traverse the tree nodes, forming a complete discrimination path from the root to each leaf. Each discrimination path consists of several conditions: First, starting from the root node, record the features of the first split and its threshold; when reaching the next level node, add new split conditions; until reaching the leaf node, give the predicted category of the fault, and calculate the proportion of each category in the leaf node as the confidence level; finally, convert the trained decision tree into a tree structure diagram to make the diagnostic process transparent and enhance the interpretability of the results.

[0035] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0036] 1. This invention proposes a multi-fault diagnosis method for multi-channel time-series data that integrates feature construction, standardization, decision tree modeling, and rule tree visualization. The rule tree generated based on the decision tree can clearly display feature split points and fault categories, thereby improving the accuracy of fault diagnosis results for hydraulic loading devices.

[0037] 2. This invention, through integrated feature construction, standardization, and decision tree modeling methods, can not only improve the accuracy of fault diagnosis of hydraulic loading devices, but also enhance the transparency and interpretability of the diagnosis process, providing engineers with clear decision-making basis and helping to achieve intelligent operation and maintenance of hydraulic loading devices. Attached Figure Description

[0038] Figure 1 This is a flowchart of an embodiment of the present invention.

[0039] Figure 2 This is a time-domain waveform diagram of the imported sample data file in an embodiment of the present invention.

[0040] Figure 3 This is a schematic diagram of the rule tree generated in an embodiment of the present invention. Detailed Implementation

[0041] The present invention will be further described in detail below with reference to the embodiments and accompanying drawings.

[0042] Reference Figure 1 A multi-fault diagnosis method for hydraulic loading devices based on decision tree algorithm includes the following steps:

[0043] Step 1) Import the sample data file and its corresponding label file: The monitored quantities and their corresponding labels are as follows: Tank temperature: T 1. Pump station outlet pressure: P 1. Pump station outlet flow rate: Q 1. Pump station return oil port pressure: P _2. Pump station return oil flow rate: Q 2. Pressure of the high-pressure main pipeline: P3. Flow rate of high-pressure main pipeline: Q 3. Main return oil pipeline pressure: P 4. Flow rate of the main return oil pipeline: Q 4. High pressure at the front end of the hydraulic substation: P 5. High-pressure flow rate at the front end of the hydraulic substation: Q 5. Hydraulic substation front-end return oil pressure: P 6. Hydraulic substation front-end return oil flow rate: Q 6. High pressure at the rear end of the hydraulic substation: P 7. Return oil pressure at the rear end of the hydraulic substation: P 8. High pressure at the rear end of the oil distributor: P 9. High-pressure flow rate at the rear end of the oil distributor: Q 7. Oil return pressure at the rear end of the oil distributor: P _10. Oil return flow rate at the rear end of the oil distributor: Q 8. Actuator acceleration: A 1. Displacement of the actuator cylinder: D 1. Load on the actuator cylinder: F _1, number of channels is 22, and the sequence length of each channel is N= 3000; After importing, perform file alignment and dimension verification: check whether the number of samples is consistent and whether the data dimension is 22×3000;

[0044] Step 2) Calculate the time-domain statistical characteristics of each of the 22 sensor channels for each sample, including four statistical measures: mean, standard deviation, skewness, and kurtosis, as conditional features for splitting decision tree nodes;

[0045] mean The calculation formula is:

[0046]

[0047] Standard deviation The calculation formula is:

[0048]

[0049] Skewness Skewness The calculation formula is:

[0050]

[0051] cliff Kurtosis The calculation formula is:

[0052]

[0053] When calculating skewness and kurtosis, if the standard deviation of the channel... Too small ( If ), then skewness and kurtosis are set to 0 to suppress numerical instability;

[0054] Extract the four statistical features mentioned above for each channel signal, and concatenate the statistical features of all channels to form a complete multidimensional feature vector: ;

[0055] Step 3) Due to the differences in dimensions and magnitudes among different sensor channels, the Z-score normalization method is used to normalize the feature vectors of different dimensions as follows:

[0056]

[0057] in, It is the standardized feature vector. yes The average value, yes Standard deviation;

[0058] Step 4) After feature construction and standardization, all multidimensional feature vector data are randomly divided into training and test sets in an 8:2 ratio. The training set is used to learn the feature distributions of different fault modes, and the test set is used to evaluate the model's generalization ability. The node splitting rule is designed as Gini impurity, using the Gini coefficient as the splitting criterion to measure the impurity of a node. To ensure the model can effectively distinguish different faults without overfitting, the maximum depth of the decision tree is set to 5, meaning a maximum of 5 splits from the root node to the leaf node. This ensures the generated rules are concise and easy to interpret, improving the model's interpretability. The seed for the random number generator is set to... s =42, fixed randomness, to ensure reproducibility of results and guarantee the stability of the diagnostic model;

[0059] Step 5) The starting point of the decision tree is the root node, which contains the entire training dataset. D Recursively split from the root node and calculate its Gini impurity:

[0060]

[0061] in Represents a node The middle belongs to the first The proportion of class samples;

[0062] Step 6) Enumerate all possible thresholds for all features across all channels in the sample. Sort the sample values ​​of this feature in the current node, take the midpoint of adjacent values ​​as a possible split point, and divide the node into... and Then calculate the impurity of the weighted child nodes:

[0063]

[0064] Choose to Minimum features and thresholds As the optimal split for this node;

[0065] Step 7) After the split is complete, the data of the current node is allocated to and Find the left and right child nodes, and then repeat steps 5)-6) on the child nodes.

[0066] Step 8) When the decision tree reaches its maximum depth or all samples within a node belong to the same class ( When the stopping condition is met, the further splitting of that node is terminated, and the node is marked as a leaf node; the training process ends when all nodes meet the stopping condition.

[0067] Step 9) Rule Derivation and Interpretability Analysis: After training, traverse the tree nodes to form a complete discrimination path from the root to each leaf. Each discrimination path consists of several conditions: First, start from the root node and record the features and threshold of the first split condition; when reaching the next level node, add a new split condition; until the leaf node is reached, the predicted category of the fault is determined by the majority class of the samples in the leaf node, and the proportion of each class in the leaf node is calculated as the confidence level.

[0068] Reference Figure 2 , Figure 2 The imported raw sample data from 22 sensor channels consists of time-domain waveform sequences, where: T _1 represents the fuel tank temperature, P _1 represents the pump station outlet pressure, Q _1 represents the pump station outlet flow rate, P _2 represents the pressure at the pump station's return oil port. Q _2 represents the return oil flow rate of the pump station. P _3 represents the pressure of the high-pressure main pipeline. Q _3 represents the flow rate of the high-pressure main pipeline. P _4 represents the pressure of the main return oil pipeline. Q _4 represents the flow rate of the main return oil pipeline. P _5 represents the high pressure at the front end of the hydraulic substation. Q _5 represents the high-pressure flow rate at the front end of the hydraulic substation. P _6 represents the return oil pressure at the front end of the hydraulic substation. Q _6 represents the return oil flow rate at the front end of the hydraulic substation. P _7 represents the high-pressure output at the rear end of the hydraulic substation. P _8 represents the return oil pressure at the rear end of the hydraulic substation. P _9 represents the high pressure at the rear end of the oil distributor. Q _7 represents the high-pressure flow rate at the rear end of the oil circuit distributor.P _10 represents the return oil pressure at the rear end of the oil distributor. Q _8 represents the return oil flow rate at the rear end of the oil circuit distributor. A _1 represents the acceleration of the actuator cylinder, D _1 represents the displacement of the actuator cylinder. F _1 represents the load on the actuator cylinder.

[0069] Reference Figure 3 After the model training is completed, this embodiment uses a rule tree visualization method to transform the trained rule tree into a tree structure diagram. The prediction process is as follows: the feature values ​​of the sample to be diagnosed are compared with the thresholds in the rules in turn. For example, if the value of the sample in "mean high-pressure flow rate at the back end of the oil distributor" (Q_7_mean) is less than or equal to -0.437, it enters the sub-node corresponding to this condition. Then, it continues to determine whether "mean high-pressure main pipeline flow rate in the trench" (Q_3_mean) is less than or equal to -0.509. If it is true, it enters the sub-node corresponding to the next layer. And so on, the sample gradually moves down the path until it reaches the leaf node, and the category label corresponding to the leaf node is output to complete the fault diagnosis. Through rule-based expression, not only can the prediction results be given, but also a clear reasoning process can be provided, making the diagnosis process transparent and enhancing the interpretability of the results.

Claims

1. A multi-fault diagnosis method for a hydraulic loading device based on a decision tree algorithm, characterized in that: First, the imported data is preprocessed; second, for multi-channel time series signals, the time-domain statistical features of each sample and each channel are extracted. Then, the extracted time-domain statistical features are standardized; a decision tree algorithm is adopted, using the Gini coefficient as the standard to select the optimal splitting feature and threshold at each layer and divide it, recursively growing from the root node, and obtaining a discrimination tree for various typical faults of the hydraulic loading device through training; After training, the tree nodes are traversed to form a complete discrimination path from the root to each leaf; finally, the discrimination path and threshold conditions are transformed into a tree diagram to intuitively show the relationship between features and fault categories, thereby improving interpretability. The aforementioned method for diagnosing multiple faults in a hydraulic loading device based on a decision tree algorithm includes the following steps: Step 1) Data Import and Preprocessing: Import the sample data file and its corresponding tag file. The monitored quantities are as follows: oil tank temperature, pump station outlet pressure, pump station outlet flow rate, pump station return oil pressure, pump station return oil flow rate, high-pressure main pipeline pressure, high-pressure main pipeline flow rate, return oil main pipeline pressure, return oil main pipeline flow rate, hydraulic substation front-end high-pressure pressure, hydraulic substation front-end high-pressure flow rate, hydraulic substation front-end return oil pressure, hydraulic substation front-end return oil flow rate, hydraulic substation rear-end high-pressure pressure, hydraulic substation rear-end return oil pressure, oil distributor rear-end high-pressure pressure, oil distributor rear-end high-pressure flow rate, oil distributor rear-end return oil pressure, oil distributor rear-end return oil flow rate, actuator acceleration, actuator displacement, and actuator load. The number of channels is 22, and the sequence length of each channel is [missing information]. N After importing, perform file alignment and dimension validation: check if the sample size is consistent and if the data dimension is 22× N ; Step 2) Target the data in the sample channel The mean, standard deviation, skewness, and kurtosis are calculated as four statistical measures and used as conditional features for splitting decision tree nodes. mean The calculation formula is: Standard deviation The formula is: skewness Skewness The formula is: kurtosis Kurtosis The calculation formula is: at standard deviation When the skewness and kurtosis are close to zero, set them to zero. Extract the four statistical features mentioned above for each channel signal, and concatenate the statistical features of all channels to form a complete multidimensional feature vector: ; Step 3) The Z-score normalization method is used to normalize the feature vectors of different dimensions as follows: in, It is the standardized feature vector. yes The average value, yes Standard deviation; Step 4) Randomly divide all multidimensional feature vector data into training and test sets in an 8:2 ratio, and set the seed of the random number generator to 0. s The node partitioning rule is Gini impurity, and the Gini coefficient is used as the splitting standard to measure the impurity of a node; the maximum depth of the decision tree is set to 5, that is, the tree can split at most 5 levels from the root node to the leaf node. Step 5) The starting point of the decision tree is the root node, which contains the entire training dataset. D Recursively split from the root node and calculate its Gini impurity: in Represents a node The middle belongs to the first The proportion of class samples; Step 6) Enumerate all possible thresholds for all features across all channels in the sample. Sort the sample values ​​of this feature in the current node, take the midpoint of adjacent values ​​as a possible split point, and divide the node into... and Then calculate the impurity of the weighted child nodes: Choose to Minimum features and thresholds This is the optimal split for that node; Step 7) After the split is complete, the data of the current node is allocated to and Find the left and right child nodes, and then repeat steps 5)-6) on the child nodes. Step 8) When the decision tree reaches its maximum depth or all samples within a node belong to the same class... When the stopping condition is met, the further splitting of that node is terminated, and the node is marked as a leaf node; the training process ends when all nodes meet the stopping condition. Step 9) Rule Derivation and Interpretability Analysis: After training, traverse the tree nodes to form a complete discrimination path from the root to each leaf. Each discrimination path consists of several conditions: First, starting from the root node, record the features of the first split and its threshold; when reaching the next level node, add new split conditions; until reaching the leaf node, give the predicted category of the fault, and calculate the proportion of each category in the leaf node as the confidence level; finally, convert the trained decision tree into a tree structure diagram to make the diagnostic process transparent and enhance the interpretability of the results.

Images