WO2022100046A1 - 一种基于零值比例频谱特征的机械传动故障检测方法 - Google Patents

一种基于零值比例频谱特征的机械传动故障检测方法 Download PDF

Info

Publication number
WO2022100046A1
WO2022100046A1 PCT/CN2021/094141 CN2021094141W WO2022100046A1 WO 2022100046 A1 WO2022100046 A1 WO 2022100046A1 CN 2021094141 W CN2021094141 W CN 2021094141W WO 2022100046 A1 WO2022100046 A1 WO 2022100046A1
Authority
WO
WIPO (PCT)
Prior art keywords
zero
value
data
feature
time series
Prior art date
Application number
PCT/CN2021/094141
Other languages
English (en)
French (fr)
Inventor
詹德川
王魏
李新春
邹联忠
Original Assignee
南京智谷人工智能研究院有限公司
Priority date (The priority date is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the date listed.)
Filing date
Publication date
Application filed by 南京智谷人工智能研究院有限公司 filed Critical 南京智谷人工智能研究院有限公司
Publication of WO2022100046A1 publication Critical patent/WO2022100046A1/zh

Links

Images

Classifications

    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F2218/00—Aspects of pattern recognition specially adapted for signal processing
    • G06F2218/12—Classification; Matching
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F17/00—Digital computing or data processing equipment or methods, specially adapted for specific functions
    • G06F17/10—Complex mathematical operations
    • G06F17/14—Fourier, Walsh or analogous domain transformations, e.g. Laplace, Hilbert, Karhunen-Loeve, transforms
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F17/00—Digital computing or data processing equipment or methods, specially adapted for specific functions
    • G06F17/10—Complex mathematical operations
    • G06F17/18—Complex mathematical operations for evaluating statistical data, e.g. average values, frequency distributions, probability functions, regression analysis
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F18/00—Pattern recognition
    • G06F18/20—Analysing
    • G06F18/24—Classification techniques
    • G06F18/241—Classification techniques relating to the classification model, e.g. parametric or non-parametric approaches
    • G06F18/2411—Classification techniques relating to the classification model, e.g. parametric or non-parametric approaches based on the proximity to a decision surface, e.g. support vector machines
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F18/00—Pattern recognition
    • G06F18/20—Analysing
    • G06F18/24—Classification techniques
    • G06F18/243—Classification techniques relating to the number of classes
    • G06F18/24323—Tree-organised classifiers
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F2218/00—Aspects of pattern recognition specially adapted for signal processing
    • G06F2218/08—Feature extraction
    • G06F2218/10—Feature extraction by analysing the shape of a waveform, e.g. extracting parameters relating to peaks

Definitions

  • the invention relates to a mechanical transmission fault detection method based on a zero-value proportional spectrum feature, and belongs to the field of mechanical transmission fault detection.
  • Machine learning technology requires a large amount of data as training support, especially deep learning technology.
  • the noise tolerance of the model for data and label information is relatively low, and a small change in the data will lead to a huge change in the model prediction.
  • the training data in the field of fan fault detection is relatively small and contains a lot of noise (such as improper installation of sensors); on the other hand, fan fault detection is a time-series signal with high input dimensions, which requires high model capacity. It is easy to lead to overfitting of the model.
  • the previous machine learning technology needs to pass the features designed by experts and assist the shallow model when processing the mechanical transmission timing signal. This method encounters a performance bottleneck, so a set of intelligent solutions is required.
  • the present invention provides a mechanical transmission fault detection method based on zero-value proportional spectral features, which greatly improves the performance and efficiency of the machine learning model through the extracted zero-value proportional features, and uses a hierarchical classifier. While improving the detection rate, the false alarm rate is effectively controlled, and the overall deployment is easy to implement and has strong applicability.
  • a method for detecting mechanical transmission faults based on zero-value proportional spectral characteristics comprising the following steps:
  • Step 1 Data collection: collect the time series characteristics of each monitoring point through sensors;
  • Step 2 Extracting zero-value proportional spectrum features: extracting zero-value proportional spectrum features through spectrum analysis
  • Step 3 Hierarchical classifier training
  • Step 4. Test the new data model.
  • the data collection in step 1 includes the following steps:
  • Step 100 Determine the transmission machinery fault monitoring point
  • Step 101 Deploy signal collection sensors
  • Step 102 collect the timing signal of the transmission machinery speed
  • Step 103 Form the collected data into several groups of data in the form of "sequential signal faulty" and "sequential signal no fault”.
  • the feature extraction of the zero-valued scale spectrum includes the following steps:
  • Step 200 Transform the original transmission mechanical speed time series signal T into frequency domain data F0, and extract the spectrogram F1 in the fixed frequency domain range;
  • Step 201 Traverse different window sizes W1, W2, ..., WN in proportion;
  • Step 202 For a given window size Wi, traverse each window in the spectrogram to take the maximum value, and remove the "burr" phenomenon beside it;
  • Step 203 Get index values corresponding to several peaks
  • Step 204 Calculate the difference between any two peak index values, and sort to obtain the index difference curve C;
  • Step 205 Differentiate the curve C, and the ratio of the statistical difference value of 0 is regarded as the zero-value ratio feature, that is, the Zero Ratio, zr feature;
  • Step 206 Repeat steps 202 to 205 for all window sizes W1, W2,...,WN, and extract the zero-value proportional features zr1, zr2,...,zrN under each window;
  • Step 207 Calculate the mean, variance, maximum, and minimum value of all the above zero-valued scale features, and put all the zero-valued scale features together as the final extracted zero-valued scale feature.
  • step 3 the training of hierarchical classifiers in step 3 includes the following steps:
  • Step 300 Organize the zero-value proportional features extracted from each segment of time series data into a vector V1i;
  • Step 302 Train the classification model M1 according to the training data "V1i, Yi";
  • Step 303 Take the discrete information entropy in the statistical window of each segment of time series data as an additional feature, and incorporate V1i as V2i;
  • Step 304 Train the classification model M2 according to the training data "V2i, Yi".
  • the new data model test in step 4 includes the following steps:
  • Step 400 collect the mechanical transmission time series characteristic data of the component to be predicted
  • Step 401 Extract the zero-value scale spectral feature V1i;
  • Step 402 use the trained model M1 to predict V1i;
  • Step 403 If the predicted value is less than or equal to 0.5, there is no fault in the output;
  • Step 404 If the predicted value is greater than 0.5, extract the discrete information entropy in the statistical window of the time series data, and fuse it with V1i to form V2i;
  • Step 405 use M2 to predict V2i
  • Step 406 if the predicted value is less than or equal to 0.5, there is no fault in the output;
  • Step 407 If the predicted value>0.5, the output is faulty.
  • the extracted zero-value scale spectral features can effectively remove the burr phenomenon, and multiple sliding windows are used to ensure the robustness of the model, and the extracted zero-value scale features can effectively capture fault signals; in addition, a hierarchical classifier is used, a
  • the first-layer classifier can improve the recall rate of the model based on the zero-value proportional feature, that is, to detect the faulty fan as much as possible; then, in order to avoid the false alarm rate being too high, the second-layer classifier is introduced to The samples predicted to be faulty continue to be classified, and the basic features such as information entropy are introduced on the basis of classification, which can greatly reduce the false alarm rate.
  • Embodiment 1 is a mechanical transmission fault detection method based on the zero-value proportional spectrum feature, including the following steps:
  • Step 1 Data collection: collect the time series characteristics of each monitoring point through sensors;
  • Step 2 Extracting zero-value proportional spectrum features: extracting zero-value proportional spectrum features through spectrum analysis
  • Step 3 Hierarchical classifier training
  • Step 4. Test the new data model.
  • the data collection in step 1 includes the following steps:
  • Step 100 Determine the transmission machinery fault monitoring point
  • Step 101 Deploy signal collection sensors
  • Step 102 collect the timing signal of the transmission machinery speed
  • Step 103 Form the collected data into several groups of data in the form of "sequential signal faulty" and "sequential signal no fault”.
  • step 2 the zero-value scale spectral feature extraction in step 2 includes the following steps:
  • Step 200 transform the original transmission mechanical rotational speed timing signal T into frequency domain data F0, and extract the spectrogram F1 in the fixed frequency domain range;
  • Step 201 Traverse different window sizes W1, W2, ..., WN in proportion;
  • Step 202 For a given window size Wi, traverse each window in the spectrogram to take the maximum value, and remove the "burr" phenomenon beside it;
  • Step 203 Get index values corresponding to several peaks
  • Step 204 Calculate the difference between any two peak index values, and sort to obtain the index difference curve C;
  • Step 205 Differentiate the curve C, and the ratio of the statistical difference value of 0 is regarded as the zero-value ratio feature, that is, the Zero Ratio, zr feature;
  • Step 206 Repeat steps 202 to 205 for all window sizes W1, W2,...,WN, and extract the zero-value proportional features zr1, zr2,...,zrN under each window;
  • Step 207 Calculate the mean, variance, maximum, and minimum value of all the above zero-valued scale features, and put all the zero-valued scale features together as the final extracted zero-valued scale feature.
  • the hierarchical classifier training in step 3 includes the following steps:
  • Step 300 Organize the zero-value proportional features extracted from each segment of time series data into a vector V1i;
  • Step 302 Train the classification model M1 according to the training data "V1i, Yi";
  • Step 303 Take the discrete information entropy in the statistical window of each segment of time series data as an additional feature, and incorporate V1i as V2i;
  • Step 304 Train the classification model M2 according to the training data "V2i, Yi".
  • the new data model test in step 4 includes the following steps:
  • Step 400 collect the mechanical transmission time series characteristic data of the component to be predicted
  • Step 401 Extract the zero-value scale spectral feature V1i;
  • Step 402 use the trained model M1 to predict V1i;
  • Step 403 If the predicted value is less than or equal to 0.5, there is no fault in the output;
  • Step 404 If the predicted value is greater than 0.5, extract the discrete information entropy in the statistical window of the time series data, and fuse it with V1i to form V2i;
  • Step 405 use M2 to predict V2i
  • Step 406 if the predicted value is less than or equal to 0.5, there is no fault in the output;
  • Step 407 If the predicted value>0.5, the output is faulty.
  • the data collection steps are as follows: select the bearing as the main measurement point to deploy the rotational speed sensor, and measure the low-frequency vibration in three directions: horizontal, vertical and axial (step 100, step 101), and the acquisition is faulty
  • the rotational speed sequence signal corresponding to the non-faulty fan step 102
  • each segment of the sequence signal is collected for 4s-30s, and the data is organized into the form of "sequence signal, whether there is a fault", that is, "sequence signal is faulty" and "" Timing Signal No Fault” form (step 103).
  • the steps of extracting the spectral features of the zero-value scale are as follows: transform the original transmission mechanical speed time series signal T into frequency domain data F0, and extract a spectrogram F1 in a fixed frequency domain range (step 200). Traverse different window sizes W1, W2, .
  • step 203 take the index values corresponding to several peaks (step 203), calculate the difference between any two peak index values, and sort to obtain the index difference curve C (step 204), perform a difference on the curve C, and the statistical difference is
  • the ratio of 0 is regarded as the zero-valued ratio feature, that is, the Zero Ratio, zr feature (step 205); for all window sizes W1, W2, ..., WN, repeat steps 202-205 to extract the zero-valued ratio feature zr1 under each window ,zr2,...,zrN (step 206); calculate the mean, variance, maximum, and minimum value of all the above zero-valued scale features, and put all the zero-valued scale features together as the final extracted zero-valued scale feature (step 207).
  • the classification model M2 is trained according to the training data "V2i, Yi" (step 304); in this embodiment, the classification model M1 and the classification model M2 adopt the support vector machine model, and in practical application, the classification model M1 , the classification model M2 can also use the random forest model.
  • the new data model testing steps are as follows: collect the mechanical transmission time series data of the component to be predicted (step 400), extract the zero-value proportional spectrum feature V1i (step 401), and use the trained model M1 to predict V1i (step 402), if If the predicted value is less than or equal to 0.5, there is no fault in the output (step 403), and if the predicted value is greater than 0.5, the discrete information entropy feature in the statistical window of time series data is extracted, and it is fused with V1i into V2i (step 404), and M2 is used to predict V2i ( Step 405), if the predicted value is less than or equal to 0.5, the output has no fault (step 406), and if the predicted value is greater than 0.5, the output has a fault (step 407).

Landscapes

  • Engineering & Computer Science (AREA)
  • Physics & Mathematics (AREA)
  • Theoretical Computer Science (AREA)
  • General Physics & Mathematics (AREA)
  • Data Mining & Analysis (AREA)
  • Mathematical Physics (AREA)
  • General Engineering & Computer Science (AREA)
  • Pure & Applied Mathematics (AREA)
  • Mathematical Optimization (AREA)
  • Mathematical Analysis (AREA)
  • Computational Mathematics (AREA)
  • Evolutionary Biology (AREA)
  • Artificial Intelligence (AREA)
  • Bioinformatics & Computational Biology (AREA)
  • Bioinformatics & Cheminformatics (AREA)
  • Life Sciences & Earth Sciences (AREA)
  • Computer Vision & Pattern Recognition (AREA)
  • Algebra (AREA)
  • Evolutionary Computation (AREA)
  • Databases & Information Systems (AREA)
  • Software Systems (AREA)
  • Operations Research (AREA)
  • Probability & Statistics with Applications (AREA)
  • Signal Processing (AREA)
  • Testing Of Devices, Machine Parts, Or Other Structures Thereof (AREA)

Abstract

一种基于零值比例频谱特征的机械传动故障检测方法，包括数据采集步骤、零值比例频谱特征提取步骤、层级分类器训练步骤和新数据模型测试步骤；首先通过在齿轮、轴承等监控点的传感器收集部件运转的时序信号，然后通过频谱分析方法对信号进行变换处理，并提取零值比例频谱特征；最后通过提取的特征建立层级分类器模型训练和测试。该方法一方面可以通过提取的零值比例特征大幅度提升机器学习模型的性能和效率，另一方面通过层级分类器提升检出率的同时控制误报率，易实现部署且适用性强。

Description

一种基于零值比例频谱特征的机械传动故障检测方法 技术领域
本发明涉及一种基于零值比例频谱特征的机械传动故障检测方法，属于机械传动故障检测领域。
背景技术
由于风力发电具有清洁可再生的优点，受到了来自世界范围内的广泛关注。但是，风机常常工作在恶劣的环境中，并且工作强度极其高。一方面，从风叶、塔架到发电箱，甚至是轴承、齿轮的损坏都会给生产过程带来巨大的安全威胁和经济损失；另一方面，通过故障诊断专家进行人工判断一台风机是否发生故障具有很高的时间和人力成本。这些都是风机故障检测和维修的难点，因此出现了使用机器学习技术对风机故障进行检测和分析的方式。
机器学习技术需要大量的数据作为训练支持，尤其是深度学习技术。此外，模型对于数据和标记信息的噪声容忍度比较低，当数据发生微小的变化即会导致模型预测发生巨大的变化。但是，一方面，风机故障检测领域的训练数据比较少且带有很多噪声(比如传感器安装位置不当)；另一方面，风机故障检测是一个时序信号，输入维度高，对模型的容量要求高，容易导致模型的过拟合。以往的机器学习技术在处理机械传动时序信号时需要通过专家设计的特征并辅助以浅层模型，这种方法遇到了性能瓶颈，因此需要一套智能化的解决方案。
发明内容
为解决现有技术的不足，本发明提供一种基于零值比例频谱特征的机械传动故障检测方法，通过提取的零值比例特征大幅度提升了机器学习模型的性能和效率，并且通过层级分类器提升检出率的同时有效的控制了误报率，整体易实现部署且适用性强。
本发明所采用的技术方案为：
一种基于零值比例频谱特征的机械传动故障检测方法，包括如下步骤：
步骤一、数据收集：通过传感器收集各个监控点的时序特征；
步骤二、零值比例频谱特征提取：通过频谱分析提取零值比例频谱特征；
步骤三、层级分类器训练；
步骤四、新数据模型测试。
优选的是，步骤一中数据收集包括如下步骤：
步骤100：确定传动机械故障监控点；
步骤101：部署信号收集传感器；
步骤102：采集传动机械转速时序信号；
步骤103：将采集的数据组成“时序信号有故障”和“时序信号无故障”形式的若干组数据。
进一步的优选，步骤二中零值比例频谱特征提取包括如下步骤：
步骤200：将原生传动机械转速时序信号T变换为频域数据F0，提取固定频域范围内的频谱图F1；
步骤201：按照比例遍历不同的窗口大小W1，W2，…，WN；
步骤202：对于给定窗口大小Wi，依次遍历频谱图中的每一个窗口取最大值，去除旁边的“毛刺”现象；
步骤203：取若干个高峰对应的索引值；
步骤204：计算高峰索引值任意两个之间的差值，并排序得到索引差值曲线C；
步骤205：对曲线C进行差分，统计差分值为0的比例当做零值比例特征，即Zero Ratio,zr特征；
步骤206：对所有窗口大小W1,W2,…,WN，重复步骤202-步骤205，提取每个窗口下的零值比例特征zr1,zr2,…,zrN；
步骤207：计算以上所有零值比例特征的均值、方差、最大、最小值，和所有零值比例特征放在一起当做最终提取的零值比例特征。
进一步的优选，步骤三中层级分类器训练包括如下步骤：
步骤300：将每段时序数据提取的零值比例特征组织成一个向量V1i；
步骤301：该时序数据对应的部件没有故障，设置Yi＝0；反之，该时序数据对应的部件有故障，设置Yi＝1；
步骤302：根据训练数据“V1i，Yi”训练分类模型M1；
步骤303：对每段时序数据统计窗口内离散信息熵作为额外特征，融入V1i作为V2i；
步骤304：根据训练数据“V2i,Yi”训练分类模型M2。
进一步的优选，步骤四中新数据模型测试包括如下步骤：
步骤400：收集待预测部件的机械传动时序特征数据；
步骤401：提取零值比例频谱特征V1i；
步骤402：利用训练好的模型M1对V1i进行预测；
步骤403：如果预测值≤0.5，输出无故障；
步骤404：如果预测值大于＞0.5，提取时序数据统计窗口内离散信息熵，与V1i融合为V2i；
步骤405：使用M2对V2i进行预测；
步骤406：如果预测值≤0.5，输出无故障；
步骤407：如果预测值＞0.5，输出有故障。
本发明的有益效果在于：
提取的零值比例频谱特征可以有效地去除毛刺现象，采用了多个滑动窗口保证模型的鲁棒性，并且提取的零值比例特征可以有效地捕获故障信号；此外，使用了层级分类器，一方面，第一层分类器基于零值比例特征可以提高模型召回率，即尽可能检测出有故障的风机；然后为了避免误报率太高，引入了第二层分类器对第一层分类器预测为有故障的样本继续分类，分类的依据引入了信息熵等基本特征，可以大幅度降低误报率。
具体实施方式
下面结合实施例对本发明做具体的介绍。
实施例1：本实施例是一种基于零值比例频谱特征的机械传动故障检测方法，包括如下步骤：
步骤一、数据收集：通过传感器收集各个监控点的时序特征；
步骤二、零值比例频谱特征提取：通过频谱分析提取零值比例频谱特征；
步骤三、层级分类器训练；
步骤四、新数据模型测试。
首先，步骤一中数据收集包括如下步骤：
步骤100：确定传动机械故障监控点；
步骤101：部署信号收集传感器；
步骤102：采集传动机械转速时序信号；
步骤103：将采集的数据组成“时序信号有故障”和“时序信号无故障”形式的若干组数据。
其次，步骤二中零值比例频谱特征提取包括如下步骤：
步骤200：将原生传动机械转速时序信号T变换为频域数据F0，提取固定频域范围内的 频谱图F1；
步骤201：按照比例遍历不同的窗口大小W1，W2，…，WN；
步骤202：对于给定窗口大小Wi，依次遍历频谱图中的每一个窗口取最大值，去除旁边的“毛刺”现象；
步骤203：取若干个高峰对应的索引值；
步骤204：计算高峰索引值任意两个之间的差值，并排序得到索引差值曲线C；
步骤205：对曲线C进行差分，统计差分值为0的比例当做零值比例特征，即Zero Ratio,zr特征；
步骤206：对所有窗口大小W1,W2,…,WN，重复步骤202-步骤205，提取每个窗口下的零值比例特征zr1,zr2,…,zrN；
步骤207：计算以上所有零值比例特征的均值、方差、最大、最小值，和所有零值比例特征放在一起当做最终提取的零值比例特征。
然后，步骤三中层级分类器训练包括如下步骤：
步骤300：将每段时序数据提取的零值比例特征组织成一个向量V1i；
步骤301：该时序数据对应的部件没有故障，设置Yi＝0；反之，该时序数据对应的部件有故障，设置Yi＝1；
步骤302：根据训练数据“V1i，Yi”训练分类模型M1；
步骤303：对每段时序数据统计窗口内离散信息熵作为额外特征，融入V1i作为V2i；
步骤304：根据训练数据“V2i,Yi”训练分类模型M2。
再然后，步骤四中新数据模型测试包括如下步骤：
步骤400：收集待预测部件的机械传动时序特征数据；
步骤401：提取零值比例频谱特征V1i；
步骤402：利用训练好的模型M1对V1i进行预测；
步骤403：如果预测值≤0.5，输出无故障；
步骤404：如果预测值大于＞0.5，提取时序数据统计窗口内离散信息熵，与V1i融合为V2i；
步骤405：使用M2对V2i进行预测；
步骤406：如果预测值≤0.5，输出无故障；
步骤407：如果预测值＞0.5，输出有故障。
在实际应用时，数据收集步骤依次为：将轴承处选为主要测量点部署转速传感器，对于低频振动，在水平、垂直和轴向三个方向进行测量(步骤100，步骤101)，采集有故障和无故障的风机对应的转速时序信号(步骤102)，每段时序信号采集4s-30s，并将数据组织成“时序信号，是否有故障”的形式保存，即“时序信号有故障”和“时序信号无故障”形式(步骤103)。
零值比例频谱特征提取步骤依次为：将原生传动机械转速时序信号T变换为频域数据F0，提取固定频域范围内的频谱图F1(步骤200)。按照比例遍历不同的窗口大小W1，W2，…，WN(步骤201)，对于给定窗口大小Wi，依次遍历频谱图中的每一个窗口取最大值，去除旁边的“毛刺”现象(步骤202)，取若干个高峰对应的索引值(步骤203)，计算高峰索引值任意两个之间的差值，并排序得到索引差值曲线C(步骤204)，对曲线C进行差分，统计差分值为0的比例当做零值比例特征，即Zero Ratio,zr特征(步骤205)；对所有窗口大小W1，W2，…，WN，重复步骤202-步骤205，提取每个窗口下的零值比例特征zr1,zr2，…,zrN(步骤206)；计算以上所有零值比例特征的均值、方差、最大、最小值，和所有零值比例特征放在一起当做最终提取的零值比例特征(步骤207)。
层级分类器训练步骤依次为：将每段时序数据提取的零值比例特征组织成一个向量V1i(步骤300)，如果该时序数据对应的部件没有故障，设置Yi＝0，反之Yi＝1(步骤301)，根据训练数据“V1i，Yi”训练分类模型M1，可以采用支持向量机、随机森林等模型(步骤302)，对每段时序数据统计窗口内离散信息熵等特征作为额外特征，融入V1i作为V2i(步骤303)，根据训练数据“V2i,Yi”训练分类模型M2(步骤304)；本实施例中，分类模型M1、分类模型M2采用支持向量机模型，在实际应用时，分类模型M1、分类模型M2也可以采用随机森林模型。
新数据模型测试步骤依次为：收集待预测部件的机械传动时序数据(步骤400)，提取零值比例频谱特征V1i(步骤401)，利用训练好的模型M1对V1i进行预测(步骤402)，如果预测值小于或等于0.5，则输出无故障(步骤403)，预测值大于0.5，则提取时序数据统计窗口内离散信息熵特征，与V1i融合为V2i(步骤404)，使用M2对V2i进行预测(步骤405)，如果预测值小于或等于0.5，输出无故障(步骤406)，预测值大于0.5则输出有故障(步骤407)。
以上所述仅是本发明专利的优选实施方式，应当指出，对于本技术领域的普通技术人员 来说，在不脱离本发明专利原理的前提下，还可以做出若干改进和润饰，这些改进和润饰也应视为本发明专利的保护范围。

Claims (5)

  1. 一种基于零值比例频谱特征的机械传动故障检测方法，其特征在于：包括如下步骤：
    步骤一、数据收集：通过传感器收集各个监控点的时序特征；
    步骤二、零值比例频谱特征提取：通过频谱分析提取零值比例频谱特征；
    步骤三、层级分类器训练；
    步骤四、新数据模型测试。
  2. 根据权利要求1所述的一种基于零值比例频谱特征的机械传动故障检测方法，其特征在于，所述步骤一中数据收集包括如下步骤：
    步骤100：确定传动机械故障监控点；
    步骤101：部署信号收集传感器；
    步骤102：采集传动机械转速时序信号；
    步骤103：将采集的数据组成“时序信号 有故障”和“时序信号 无故障”形式的若干组数据。
  3. 根据权利要求2所述的一种基于零值比例频谱特征的机械传动故障检测方法，其特征在于，所述步骤二中零值比例频谱特征提取包括如下步骤：
    步骤200：将原生传动机械转速时序信号T变换为频域数据F0，提取固定频域范围内的频谱图F1；
    步骤201：按照比例遍历不同的窗口大小W1，W2，…，WN；
    步骤202：对于给定窗口大小Wi，依次遍历频谱图中的每一个窗口取最大值，去除旁边的“毛刺”现象；
    步骤203：取若干个高峰对应的索引值；
    步骤204：计算高峰索引值任意两个之间的差值，并排序得到索引差值曲线C；
    步骤205：对曲线C进行差分，统计差分值为0的比例当做零值比例特征，即Zero Ratio,zr特征；
    步骤206：对所有窗口大小W1,W2,…,WN，重复步骤202-步骤205，提取每个窗口下的零值比例特征zr1,zr2,…,zrN；
    步骤207：计算以上所有零值比例特征的均值、方差、最大、最小值，和所有零值比例特征放在一起当做最终提取的零值比例特征。
  4. 根据权利要求3所述的一种基于零值比例频谱特征的机械传动故障检测方法，其特征 在于，所述步骤三中层级分类器训练包括如下步骤：
    步骤300：将每段时序数据提取的零值比例特征组织成一个向量V1i；
    步骤301：该时序数据对应的部件没有故障，设置Yi＝0；反之，该时序数据对应的部件有故障，设置Yi＝1；
    步骤302：根据训练数据“V1i，Yi”训练分类模型M1；
    步骤303：对每段时序数据统计窗口内离散信息熵作为额外特征，融入V1i作为V2i；
    步骤304：根据训练数据“V2i,Yi”训练分类模型M2。
  5. 根据权利要求4所述的一种基于零值比例频谱特征的机械传动故障检测方法，其特征在于，所述步骤四中新数据模型测试包括如下步骤：
    步骤400：收集待预测部件的机械传动时序特征数据；
    步骤401：提取零值比例频谱特征V1i；
    步骤402：利用训练好的模型M1对V1i进行预测；
    步骤403：如果预测值≤0.5，输出无故障；
    步骤404：如果预测值大于＞0.5，提取时序数据统计窗口内离散信息熵，与V1i融合为V2i；
    步骤405：使用M2对V2i进行预测；
    步骤406：如果预测值≤0.5，输出无故障；
    步骤407：如果预测值＞0.5，输出有故障。
PCT/CN2021/094141 2020-11-10 2021-05-17 一种基于零值比例频谱特征的机械传动故障检测方法 WO2022100046A1 (zh)

Applications Claiming Priority (2)

Application Number Priority Date Filing Date Title
CN202011244962.3 2020-11-10
CN202011244962.3A CN112347944B (zh) 2020-11-10 2020-11-10 一种基于零值比例频谱特征的机械传动故障检测方法

Publications (1)

Publication Number Publication Date
WO2022100046A1 true WO2022100046A1 (zh) 2022-05-19

Family

ID=74362581

Family Applications (1)

Application Number Title Priority Date Filing Date
PCT/CN2021/094141 WO2022100046A1 (zh) 2020-11-10 2021-05-17 一种基于零值比例频谱特征的机械传动故障检测方法

Country Status (2)

Country Link
CN (1) CN112347944B (zh)
WO (1) WO2022100046A1 (zh)

Families Citing this family (2)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN112347944B (zh) * 2020-11-10 2023-10-13 南京大学 一种基于零值比例频谱特征的机械传动故障检测方法
CN113255783B (zh) * 2021-05-31 2021-09-14 湖南安存科技有限公司 基于无监督学习的传感器故障检测方法和装置

Citations (5)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN108181107A (zh) * 2018-01-12 2018-06-19 东北电力大学 计及多分类目标的风电机组轴承机械故障诊断方法
CN110132598A (zh) * 2019-05-13 2019-08-16 中国矿业大学 旋转设备滚动轴承故障噪声诊断算法
US20200200648A1 (en) * 2018-02-12 2020-06-25 Dalian University Of Technology Method for Fault Diagnosis of an Aero-engine Rolling Bearing Based on Random Forest of Power Spectrum Entropy
CN111680665A (zh) * 2020-06-28 2020-09-18 湖南大学 基于数据驱动的采用电流信号的电机机械故障诊断方法
CN112347944A (zh) * 2020-11-10 2021-02-09 南京智谷人工智能研究院有限公司 一种基于零值比例频谱特征的机械传动故障检测方法

Patent Citations (5)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN108181107A (zh) * 2018-01-12 2018-06-19 东北电力大学 计及多分类目标的风电机组轴承机械故障诊断方法
US20200200648A1 (en) * 2018-02-12 2020-06-25 Dalian University Of Technology Method for Fault Diagnosis of an Aero-engine Rolling Bearing Based on Random Forest of Power Spectrum Entropy
CN110132598A (zh) * 2019-05-13 2019-08-16 中国矿业大学 旋转设备滚动轴承故障噪声诊断算法
CN111680665A (zh) * 2020-06-28 2020-09-18 湖南大学 基于数据驱动的采用电流信号的电机机械故障诊断方法
CN112347944A (zh) * 2020-11-10 2021-02-09 南京智谷人工智能研究院有限公司 一种基于零值比例频谱特征的机械传动故障检测方法

Also Published As

Publication number Publication date
CN112347944B (zh) 2023-10-13
CN112347944A (zh) 2021-02-09

Similar Documents

Publication Publication Date Title
CN102798529B (zh) 大型风力机组轴承故障诊断方法及系统
WO2022100046A1 (zh) 一种基于零值比例频谱特征的机械传动故障检测方法
CN102944416B (zh) 基于多传感器信号融合技术的风电机组叶片故障诊断方法
CN101782475B (zh) 基于风力发电机组振动的叶片故障诊断方法
CN102937522B (zh) 一种齿轮箱复合故障诊断方法及系统
CN103940611B (zh) 一种风电机组变工况下滚动轴承自适应异常检测方法
CN106441896A (zh) 滚动轴承故障模式识别及状态监测的特征向量提取方法
CN110595778B (zh) 一种基于mmf与igra风电机组轴承故障诊断方法
CN111209934A (zh) 风机故障预报警方法及系统
CN106055734A (zh) 一种风电齿轮箱故障诊断方法
CN106777611A (zh) 复杂周转轮系微弱故障识别与性能退化监测系统及方法
Cao et al. Intelligent fault diagnosis of wind turbine gearbox based on Long short-term memory networks
CN111442927B (zh) 一种基于多尺度数学形态学的滚动轴承故障诊断方法
Li et al. Mobile robot motor bearing fault detection and classification on discrete wavelet transform and LSTM network
CN113255458A (zh) 基于多视图关联特征学习的轴承故障诊断方法
CN104792520A (zh) 风力发电机组齿轮箱故障诊断方法
CN112177865B (zh) 一种用于解决风机故障检测中标记噪音和标记不足的方法
CN105987809A (zh) 基于随机共振的离心式压缩机半开式叶轮裂纹检测方法
CN103149047A (zh) 一种基于非线性混合模型的冷却塔声学诊断方法
CN117093938A (zh) 一种基于深度学习的风机轴承的故障检测方法和系统
Wang et al. An improved time-varying morphological filtering and its application to bearing fault diagnosis
CN112378605B (zh) 一种基于emd分解自学习的风电机组叶片故障识别方法
CN114781466A (zh) 基于旋转机械振动信号谐波基频的故障诊断方法及系统
CN112364762B (zh) 一种基于阶梯误差频谱特征的机械传动故障检测方法
Zheng et al. Wavelet packet decomposition and neural network based fault diagnosis for elevator excessive vibration

Legal Events

Date Code Title Description
121 Ep: the epo has been informed by wipo that ep was designated in this application

Ref document number: 21890578

Country of ref document: EP

Kind code of ref document: A1

NENP Non-entry into the national phase

Ref country code: DE

122 Ep: pct application non-entry in european phase

Ref document number: 21890578

Country of ref document: EP

Kind code of ref document: A1