An intelligent early warning and fault diagnosis method for large gantry shipbuilding cranes
Through XGBoost prediction model and knowledge graph technology, a small amount of data is used to perform intelligent early warning and fault diagnosis of large cranes, solving the problem of insufficient data and improving the accuracy and efficiency of diagnosis.
Patent Information
- Application Number
- CN202210705724.0
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2022-06-21
- Publication Date
- 2024-11-29
- Estimated Expiration
- 2042-06-21
AI Technical Summary
Large cranes lack real-time and scientific status determination and fault warning in actual operation, resulting in subjectivity and blindness of management and maintenance strategies. The existing intelligent early warning methods rely on a large amount of labeled historical data and cannot provide reliable diagnosis when there is insufficient data.
The XGBoost prediction model is used to combine historical working conditions and sensor data for prediction, and the fault diagnosis is used through knowledge graph technology to achieve intelligent early warning and fault analysis based on a small amount of data.
It realizes reliable intelligent early warning and fault diagnosis of large-scale gantry shipbuilding cranes, improves the accuracy and efficiency of diagnosis, and makes up for the problem of insufficient data.
Smart Images

Figure CN115215215B_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the technical field of crane data analysis and processing, and in particular to an intelligent early warning and fault diagnosis method for a large gantry shipbuilding crane. Background Art
[0002] Shipbuilding is a basic and strategic industry. As an important part of modern manufacturing, its development level reflects the height and quality of a country's modern industrial development. Large cranes, as a major loading and unloading machinery, are an important part of the shipbuilding process and play an important role in the construction of major shipyard projects. Therefore, ensuring the reliable and stable operation of cranes and reducing the downtime of cranes are crucial to the normal operation and work efficiency of shipyards.
[0003] However, in the actual operation of large cranes, there are common problems such as complex working conditions, inadequate equipment information collection, lack of intelligent monitoring and fault warning, etc. During daily use and inspections, operators and managers cannot judge the operating status and health of the crane in real time and scientifically, and can only make rough judgments, resulting in subjective and blind defects in many aspects of crane daily management, inspection and testing, and maintenance strategy formulation.
[0004] Currently, common intelligent early warning or fault prediction methods require a large amount of annotated historical data as support. However, in reality, most cranes have been in use for a long time, and there was no planning for intelligent early warning and fault diagnosis when they were first put into use. Sensors are often installed only when the condition of the crane gradually deteriorates and early warning and diagnosis are needed. Due to the short installation time of the sensors, the historical operation data collected on the crane is limited, and it only contains a small amount of or even no historical fault data of the crane. It cannot provide a large amount of effective historical operation data, and it is impossible to perform reliable and accurate intelligent early warning and fault diagnosis. Summary of the invention
[0005] The purpose of the present invention is to overcome the defects of the above-mentioned prior art and to provide an intelligent early warning and fault diagnosis method for a large gantry shipbuilding crane. Based on a small amount of historical operation data, the intelligent early warning and fault diagnosis of a large gantry shipbuilding crane can be reliably and accurately performed.
[0006] The purpose of the present invention can be achieved by the following technical solution: An intelligent early warning and fault diagnosis method for a large gantry shipbuilding crane comprises the following steps:
[0007] S1. Obtain historical operating data and sensor data of the crane, and perform preprocessing operations on the acquired data to obtain training set data;
[0008] S2. Using the training set data, train the XGBoost prediction model;
[0009] S3, input the actual working condition data of the crane into the XGBoost prediction model, and output the predicted value of the sensor data;
[0010] S4, calculating the time residual sequence between the predicted value of the sensor data and the actual value of the sensor data, and if it exceeds the residual threshold, pushing the warning information;
[0011] S5. Build a knowledge graph based on professional knowledge;
[0012] S6. According to the warning information, a subgraph is extracted from the constructed knowledge graph, and a diagnosis is performed to obtain a diagnosis result;
[0013] S7. Match the information related to the current diagnosis result in the knowledge graph and push it.
[0014] Furthermore, the data acquired in the step S1 includes low-frequency operating signal data and high-frequency sensor data, the operating signal data is specifically PLC control system data, the sensor data includes vibration sensor data and main beam strain sensor data, and the process of preprocessing the acquired data in the step S1 includes feature extraction, missing value filling and normalization processing.
[0015] Furthermore, the specific process of feature extraction is: for high-frequency sensor signal data, Fourier decomposition is used to convert the time domain signal into a frequency domain signal, and then the corresponding amplitude is extracted according to the characteristic frequency of different parts, and the amplitude of the characteristic frequency is used as the signal characteristic value corresponding to different parts, thereby achieving the purpose of aligning high-frequency data to low-frequency data while reducing information loss.
[0016] Furthermore, the missing value filling specifically adopts the KNN (K-Nearest Neighbor) algorithm to perform the missing value filling operation, that is, taking the missing value to be filled as the prediction target, using the non-missing data for fitting, establishing a data prediction model, and then predicting the missing value data;
[0017] The normalization process specifically compresses all data to between (0, 1).
[0018] Furthermore, the training process of the XGBoost prediction model in step S2 includes the following steps:
[0019] S21, initializing sample value;
[0020] S22, define the loss function;
[0021] S23, calculating the derivative of the loss function;
[0022] S24, generating a new tree according to the derivative;
[0023] S25, using the new tree to predict the sample value, and accumulating it with the previous value;
[0024] S26. The decision tree is trained in a loop until the exit value of the loss function is satisfied.
[0025] Furthermore, the specific process of step S3 is: after cleaning the actual PLC control system data of the crane, input it into the XGBoost prediction model, and output the corresponding vibration sensor data prediction value and strain sensor data prediction value.
[0026] Furthermore, the step S4 specifically includes the following steps:
[0027] S41, calculating the residual sequence between the predicted value of the sensor data and the actual value of the sensor data, if it does not exceed the set residual threshold, it is determined that the crane is currently in a normal state, otherwise, executing step S42;
[0028] S42. Calculate the root mean square error of the residual sequence to obtain the root mean square error value of all residuals in each hour. If the fluctuation frequency of the root mean square error value of all residuals is greater than or equal to the set fluctuation threshold, it is determined that the crane is currently faulty and outputs a warning message; otherwise, it is determined that the crane is currently in a normal state.
[0029] Furthermore, the step S5 specifically combines the data sources and characteristics in the field of crane failures, extracts triples by setting rules, builds knowledge into the knowledge graph in the form of triples, and completes the construction of the crane failure knowledge graph.
[0030] Furthermore, the step S6 specifically includes the following steps:
[0031] S61. According to the warning information, all fault record IDs that are simultaneously associated with the warning part and the warning sensor are matched from the knowledge graph, and then the first-degree relationship of these fault IDs is queried as the subgraph related to this warning;
[0032] S62, matching all fault causes in the warning-related subgraph, and calculating the centrality of all fault causes according to the PageRank algorithm, sorting all fault causes according to their centrality, a higher centrality indicates more faults caused by it in the historical records, and taking the fault causes corresponding to the top N centralities as the fault diagnosis results;
[0033] The PageRank calculation method is:
[0034]
[0035] Where: PR(A) is the centrality of A, d is the damping coefficient, which ranges from [0, 1), C(A) is the number of external links of Page A, and n is the number of nodes pointing to A.
[0036] Furthermore, the information pushed in step S7 includes the failure result, failure mode, failure cause, failure phenomenon and impact range.
[0037] Compared with the prior art, the present invention aims to solve the problems of insufficient historical fault data of cranes and unlabeled data, and proposes a solution for predicting sensor data using crane operating data. The solution is based on the XGBoost algorithm and combines historical operating data and sensor data to train an XGBoost prediction model. The predicted value of sensor data can be predicted using actual operating data, and compared with the actual value of sensor data to achieve early warning of the crane. That is, the present invention makes full use of the existing small amount of data and can achieve intelligent early warning of large cranes.
[0038] The present invention is based on knowledge graph technology and applies knowledge graph to intelligent diagnosis of large cranes. By analyzing the warning results of machine learning and combining it with the constructed crane knowledge graph, fault diagnosis of large cranes can be achieved. The present invention replaces the data model with the knowledge model. When fault data is insufficient, a large amount of existing knowledge is used as the basis for diagnosis. The experience and knowledge of industry experts are fully utilized to effectively make up for the deficiencies at the data level, which can effectively improve the accuracy and efficiency of fault diagnosis. BRIEF DESCRIPTION OF THE DRAWINGS
[0039] Figure 1 It is a schematic diagram of the method flow of the present invention;
[0040] Figure 2 It is a schematic diagram of the application process of the embodiment;
[0041] Figure 3a It is a waveform diagram of the predicted value of strain of the main beam in the embodiment;
[0042] Figure 3b It is a waveform diagram of the predicted vibration value of the lifting mechanism in the embodiment;
[0043] Figure 4 Schematic diagram of the knowledge graph structure constructed in the embodiment. DETAILED DESCRIPTION
[0044] The present invention is described in detail below with reference to the accompanying drawings and specific embodiments.
[0045] Example
[0046] like Figure 1 As shown, an intelligent early warning and fault diagnosis method for a large gantry shipbuilding crane comprises the following steps:
[0047] S1. Obtain historical operating data and sensor data of the crane, and perform preprocessing operations on the acquired data to obtain training set data;
[0048] S2. Using the training set data, train the XGBoost prediction model;
[0049] S3, input the actual working condition data of the crane into the XGBoost prediction model, and output the predicted value of the sensor data;
[0050] S4, calculating the time residual sequence between the predicted value of the sensor data and the actual value of the sensor data, and if it exceeds the residual threshold, pushing the warning information;
[0051] S5. Build a knowledge graph based on professional knowledge;
[0052] S6. According to the warning information, a subgraph is extracted from the constructed knowledge graph, and a diagnosis is performed to obtain a diagnosis result;
[0053] S7. Match the information related to the current diagnosis result in the knowledge graph and push it.
[0054] This embodiment applies the above technical solution, such as Figure 2 As shown, it mainly includes:
[0055] Step 1: Obtain the historical operating data and sensor data of the crane, and perform data preprocessing to obtain the training set:
[0056] The crane operating data is the PLC control system data, and the sensor data includes the main beam strain sensor data collected by the strain sensor and the lifting mechanism vibration sensor data collected by the vibration sensor.
[0057] The data preprocessing process includes feature extraction (i.e. data alignment), missing value processing and normalization processing. Specifically, during the feature extraction operation, considering that the collected crane operation data contains high-frequency sensor data and low-frequency operating signal data, for the high-frequency vibration sensor signal, Fourier decomposition is used to convert the time domain signal into the frequency domain, and then the corresponding amplitude is extracted according to the characteristic frequency of different parts, and the amplitude of the characteristic frequency is used as the signal characteristic value corresponding to different parts, so as to achieve the purpose of aligning high-frequency data to low-frequency data while minimizing information loss.
[0058] The vibration signal is converted from the time domain to the frequency domain through Fourier analysis, and then the mechanism model is used to extract the corresponding value of the specific frequency as the characteristic value of the vibration signal within one second, and the characteristic value is used to align with the PLC signal. This not only solves the problem of data misalignment caused by inconsistent data frequency, but also solves the problem of important information loss when aligning data. After converting the vibration signal of the crane into the frequency domain, the key parts of the crane hoisting mechanism are focused on, and the characteristic frequency calculation method of the corresponding parts is obtained based on the relevant professional knowledge of the crane.
[0059] After calculating the corresponding frequencies of different parts, the corresponding amplitudes are found according to the frequencies, and the corresponding amplitudes are used as the monitoring basis values of different parts. This alignment method fully extracts important information from sensor signals by combining professional knowledge and data, making up for the problem of missing information in traditional alignment methods and providing a data basis for subsequent fault prediction based on machine learning.
[0060] After the feature value is extracted, the K nearest neighbor (KNN) algorithm is used to fill the missing values. The basic idea of this algorithm to fill the missing values is to take the value to be filled as the prediction target, use the non-missing data for fitting, establish a data prediction model, and then predict the missing data.
[0061] Finally, normalization is performed to compress the data between (0, 1).
[0062] Step 2: Train the XGBoost prediction model:
[0063] The model is trained based on the processed training set. The training process is as follows: (1) Initialize the sample value; (2) Define the loss function; (3) Calculate the derivative of the loss function; (4) Generate a new tree based on the derivative; (5) Use the new tree to predict the sample value and accumulate it with the previous value; (6) Repeat the training of the decision tree until the exit value of the loss function is met.
[0064] In this embodiment, an acquisition unit and a training unit are constructed to complete the training of the XGBoost prediction model, wherein the acquisition unit is used to acquire the sample set Y:
[0065] Y={(x 1i ,x 2i ,…,x ni ,y 1i ,y 2i ,…,y mi )}(|Y|=o,x ni ∈R p ,y mi ∈R)
[0066] In the formula, x ni is the n-th dimension working condition data value in the ith data, ymi is the characteristic value of the m-th dimension sensor data in the i-th data, o is the number of samples, and p is the number of sample characteristics;
[0067] The training unit is used to construct a gradient boosting tree based on the sample set Y, use the CART regression tree as a weak learner, increase the weak learners in an iterative manner, and finally combine multiple weak learners to form the final strong learner.
[0068] Step 3: Predict sensor data feature values:
[0069] The PLC data of the reaction condition after cleaning is input into the XGBoost prediction model, and the output corresponds to the eigenvalues of the vibration sensor and the eigenvalues of the strain sensor respectively.
[0070] Step 4: Signal evaluation:
[0071] The residual time series method is used to comprehensively estimate the signal index within a period of time. The predicted value and actual value sequence of sensor data within a period of time are used as input for residual calculation. The output is the signal evaluation value within the period of time. The evaluation value is used to replace the instantaneous amplitude as the basis for judgment. The steps are as follows:
[0072] (1) Input the residual sequence of the predicted value and the actual value, and calculate the residual sequence of the two. If the residual does not exceed the threshold, the crane is considered to be in normal condition.
[0073] (2) If the residual exceeds the threshold, the root mean square error (RMSE) of the residual sequence is further calculated:
[0074]
[0075] In the formula, D is the root mean square error, N is the number of samples, z is the predicted value, and z j is the actual value.
[0076] By calculating the RMSE of all residuals within each hour, the operation status of the crane in that short period of time can be comprehensively evaluated. The hourly changes of RMSE are observed. If there are drastic fluctuations, it is predicted that the large gantry shipbuilding crane will fail and maintenance measures need to be taken in time.
[0077] Step 5: Knowledge graph construction:
[0078] Combined with the data sources and characteristics in the field of crane failure, triples are extracted manually or by rules, and knowledge is built into the knowledge graph in the form of triples to complete the construction and application of the crane failure knowledge graph.
[0079] Step 6: Troubleshooting:
[0080] According to the warning information, match all fault record IDs that are simultaneously associated with the warning location and warning sensor, and then query the first-degree relationship of these fault IDs as the subgraph related to this warning;
[0081] Match all the fault causes in the subgraph and calculate the centrality of all fault causes using the PageRank algorithm:
[0082]
[0083] Where PR(A) is the centrality of A, d is a damping coefficient ranging from 0 (inclusive) to 1 (exclusive), C(A) is the number of external links of Page A, and n is the number of nodes pointing to A;
[0084] All fault causes are sorted according to their centrality. The higher the centrality, the more faults caused by it in the historical records. The greater the possibility that the diagnosis result is correct, the fault causes with a certain number of top rankings are finally pushed.
[0085] Step 7: Push fault information:
[0086] According to the diagnosis results, the corresponding information is matched in the knowledge graph and pushed to the user. The pushed information includes fault results, failure modes, failure causes, fault phenomena, and impact range.
[0087] In this embodiment, the existing large-scale gantry shipbuilding crane has a total of 180-dimensional PLC data collected from its operating data, which is used to reflect the working conditions, and the sampling frequency is once per second; 8-dimensional vibration sensor data, the sensors are mainly distributed in the key parts of the lifting mechanism, and are used to monitor the operating status of the lifting mechanism of the large-scale gantry shipbuilding crane, and the sampling frequency is 1000 times per second; 8-dimensional strain sensor data, of which the effective sensor data is 6 dimensions, the sensors are mainly distributed in the key parts of the main beam, and are used to monitor the operating status of the main beam of the large-scale gantry shipbuilding crane, and the sampling frequency is 1000 pieces per second.
[0088] First, the operating condition data and sensor data are subjected to Fourier-based eigenvalue extraction, missing value processing, and data transformation operations to obtain cleaned data, and the cleaned data are split into the training set and test set of the model.
[0089] The working condition data in the training set is used as the input value of the model, and the characteristic value of the sensor is used as the target value to train the prediction model.
[0090] Input the working condition data of the test set into the prediction model and obtain the sensor characteristic values output by the model (respectively Figure 3a and Figure 3bCalculate the time residual sequence of the predicted sensor feature values and the actual sensor feature values in the test set.
[0091] If the residual value exceeds the set threshold, an alarm is issued.
[0092] Collect a large amount of professional knowledge, extract triples from the information using rule-based methods, and build a knowledge graph of large gantry shipbuilding cranes based on these triples (such as Figure 4 shown).
[0093] According to the alarm information, the corresponding subgraph in the knowledge graph of large gantry shipbuilding crane is extracted, and all relevant fault causes are matched.
[0094] Calculate the centrality of the matched fault causes and sort them, push the first specified fault cause in the sorting as the result of this fault diagnosis, and finally match the relevant information in the knowledge graph based on the pushed results.
[0095] In summary, this technical solution is aimed at large gantry shipbuilding cranes. Based on big data, artificial intelligence and knowledge graph technology, it realizes the intelligent early warning and fault diagnosis functions of the lifting mechanism and main beam of large gantry shipbuilding cranes in the absence of historical fault data and unlabeled historical operation data. An intelligent early warning solution for large cranes based on the XGBoost algorithm is proposed. This solution uses the method of predicting sensor data with working condition data to fully utilize the existing data for intelligent early warning of large cranes in response to the current problems of insufficient data volume and unlabeled data. In addition, the knowledge graph is used for intelligent diagnosis of large cranes. By analyzing the warning results of machine learning and combining with the knowledge graph, the intelligent diagnosis function of large cranes is realized. In the case of insufficient data, the experience and knowledge of industry experts can be fully utilized to make up for the deficiencies at the data level.
Claims
1. An intelligent early warning and fault diagnosis method for a large gantry shipbuilding crane, characterized in that: The following steps are involved: S1. Obtain historical operating data and sensor data of the crane, and perform preprocessing operations on the acquired data to obtain training set data; S2. Using the training set data, train the XGBoost prediction model; S3, input the actual working condition data of the crane into the XGBoost prediction model, and output the predicted value of the sensor data; S4, calculating the time residual sequence between the predicted value of the sensor data and the actual value of the sensor data, and if it exceeds the residual threshold, pushing the warning information; S5. Build a knowledge graph; S6. According to the warning information, a subgraph is extracted from the constructed knowledge graph, and a diagnosis is performed to obtain a diagnosis result; S7. Match the information related to the current diagnosis result in the knowledge graph and push it; The data acquired in step S1 includes low-frequency operating signal data and high-frequency sensor data, the operating signal data is specifically PLC control system data, the sensor data includes vibration sensor data and main beam strain sensor data, and the process of preprocessing the acquired data in step S1 includes feature extraction, missing value filling and normalization processing; The step S5 specifically combines the data source and characteristics of the crane failure field, extracts triples by setting rules, constructs the data in the crane failure field into the knowledge graph in the form of triples, and completes the construction of the crane failure knowledge graph; The step S6 specifically comprises the following steps: S61. According to the warning information, all fault record IDs simultaneously associated with the warning part and the warning sensor are matched from the knowledge graph, and then the first-degree relationship of these fault record IDs is queried as the subgraph related to this warning; S62, matching all fault causes in the warning-related subgraph, and calculating the centrality of all fault causes according to the PageRank algorithm, sorting all fault causes according to their centrality, a higher centrality indicates more faults caused by it in the historical records, and taking the fault causes corresponding to the top N centralities as the fault diagnosis results; The PageRank calculation method is: Where: PR(A) is the centrality of A, d is the damping coefficient, which ranges from [0, 1), C(A) is the number of external links of Page A, and n is the number of nodes pointing to A.
2. The intelligent early warning and fault diagnosis method for a large gantry shipbuilding crane according to claim 1 is characterized in that: The specific process of feature extraction is as follows: for high-frequency sensor signal data, Fourier decomposition is used to convert the time domain signal into a frequency domain signal, and then the corresponding amplitude is extracted according to the characteristic frequency of different parts, and the amplitude of the characteristic frequency is used as the signal characteristic value corresponding to different parts, thereby achieving the purpose of aligning high-frequency data to low-frequency data while reducing information loss.
3. The intelligent early warning and fault diagnosis method for a large gantry shipbuilding crane according to claim 1 is characterized in that: The missing value filling specifically adopts the KNN algorithm to perform the missing value filling operation, that is, taking the missing value to be filled as the prediction target, using the non-missing data for fitting, establishing a data prediction model, and then predicting the missing value data; The normalization process specifically compresses all data to between (0, 1).
4. The intelligent early warning and fault diagnosis method for a large gantry shipbuilding crane according to claim 1 is characterized in that: The training process of the XGBoost prediction model in step S2 includes the following steps: S21, initializing sample value; S22, define the loss function; S23, calculating the derivative of the loss function; S24, generating a new tree according to the derivative; S25, using the new tree to predict the sample value, and accumulating it with the previous sample value; S26. The decision tree is trained in a loop until the exit value of the loss function is satisfied.
5. The intelligent early warning and fault diagnosis method for a large gantry shipbuilding crane according to claim 1 is characterized in that: The specific process of step S3 is: after cleaning the actual PLC control system data of the crane, input it into the XGBoost prediction model, and output the corresponding vibration sensor data prediction value and strain sensor data prediction value.
6. The intelligent early warning and fault diagnosis method for a large gantry shipbuilding crane according to claim 5, characterized in that: The step S4 specifically comprises the following steps: S41, calculating the residual sequence between the predicted value of the sensor data and the actual value of the sensor data, if it does not exceed the set residual threshold, it is determined that the crane is currently in a normal state, otherwise, executing step S42; S42. Calculate the root mean square error of the residual sequence to obtain the root mean square error value of all residuals in each hour. If the fluctuation frequency of the root mean square error value of all residuals is greater than or equal to the set fluctuation threshold, it is determined that the crane is currently faulty and outputs a warning message; otherwise, it is determined that the crane is currently in a normal state.
7. The intelligent early warning and fault diagnosis method for a large gantry shipbuilding crane according to claim 6, characterized in that: The information pushed in step S7 includes the failure result, failure mode, failure cause, failure phenomenon and impact range.
Citation Information
Patent Citations
Segmentation of structures for state determination
CA2752370A1
Crawler crane moment limiting system
CN101780926A
