CN116560341A - Industrial robot fault diagnosis model and fault diagnosis method - Google Patents
Industrial robot fault diagnosis model and fault diagnosis method Download PDFInfo
- Publication number
- CN116560341A CN116560341A CN202310602754.3A CN202310602754A CN116560341A CN 116560341 A CN116560341 A CN 116560341A CN 202310602754 A CN202310602754 A CN 202310602754A CN 116560341 A CN116560341 A CN 116560341A
- Authority
- CN
- China
- Prior art keywords
- convolution kernel
- feature extraction
- industrial robot
- feature
- fault diagnosis
- Prior art date
- Legal status (The legal status 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 status listed.)
- Pending
Links
- 238000003745 diagnosis Methods 0.000 title claims abstract description 58
- 238000000034 method Methods 0.000 title claims abstract description 25
- 238000000605 extraction Methods 0.000 claims abstract description 81
- 230000006870 function Effects 0.000 claims abstract description 30
- 230000015654 memory Effects 0.000 claims abstract description 29
- 239000013598 vector Substances 0.000 claims abstract description 29
- 238000010586 diagram Methods 0.000 claims abstract description 19
- 238000004519 manufacturing process Methods 0.000 claims description 31
- 238000013528 artificial neural network Methods 0.000 claims description 13
- 238000011176 pooling Methods 0.000 claims description 13
- 230000004913 activation Effects 0.000 claims description 7
- 239000000284 extract Substances 0.000 claims description 5
- 238000007781 pre-processing Methods 0.000 claims description 3
- 238000012549 training Methods 0.000 claims description 3
- 238000003466 welding Methods 0.000 description 35
- 239000010410 layer Substances 0.000 description 10
- 230000002159 abnormal effect Effects 0.000 description 9
- 238000013527 convolutional neural network Methods 0.000 description 6
- 238000012545 processing Methods 0.000 description 6
- 230000005856 abnormality Effects 0.000 description 5
- 230000008901 benefit Effects 0.000 description 5
- 239000003638 chemical reducing agent Substances 0.000 description 5
- 230000004927 fusion Effects 0.000 description 5
- 238000012423 maintenance Methods 0.000 description 5
- 230000008569 process Effects 0.000 description 5
- 230000007812 deficiency Effects 0.000 description 3
- 238000005516 engineering process Methods 0.000 description 3
- 230000000737 periodic effect Effects 0.000 description 3
- 238000012935 Averaging Methods 0.000 description 2
- 238000004458 analytical method Methods 0.000 description 2
- 150000001875 compounds Chemical class 0.000 description 2
- 238000001514 detection method Methods 0.000 description 2
- 230000000694 effects Effects 0.000 description 2
- 238000010801 machine learning Methods 0.000 description 2
- 238000011160 research Methods 0.000 description 2
- 206010063385 Intellectualisation Diseases 0.000 description 1
- 230000009286 beneficial effect Effects 0.000 description 1
- 238000006243 chemical reaction Methods 0.000 description 1
- 238000004891 communication Methods 0.000 description 1
- 239000002131 composite material Substances 0.000 description 1
- 230000006835 compression Effects 0.000 description 1
- 238000007906 compression Methods 0.000 description 1
- 238000013480 data collection Methods 0.000 description 1
- 230000003111 delayed effect Effects 0.000 description 1
- 238000013461 design Methods 0.000 description 1
- 238000011161 development Methods 0.000 description 1
- 238000009826 distribution Methods 0.000 description 1
- 230000036541 health Effects 0.000 description 1
- 239000011159 matrix material Substances 0.000 description 1
- 206010027175 memory impairment Diseases 0.000 description 1
- 238000012986 modification Methods 0.000 description 1
- 230000004048 modification Effects 0.000 description 1
- 238000012544 monitoring process Methods 0.000 description 1
- 238000003058 natural language processing Methods 0.000 description 1
- 238000010606 normalization Methods 0.000 description 1
- 238000002203 pretreatment Methods 0.000 description 1
- 230000035755 proliferation Effects 0.000 description 1
- 230000009467 reduction Effects 0.000 description 1
- 230000006403 short-term memory Effects 0.000 description 1
- 239000002356 single layer Substances 0.000 description 1
- 238000007619 statistical method Methods 0.000 description 1
- 238000006467 substitution reaction Methods 0.000 description 1
- 230000007704 transition Effects 0.000 description 1
Classifications
-
- G—PHYSICS
- G05—CONTROLLING; REGULATING
- G05B—CONTROL OR REGULATING SYSTEMS IN GENERAL; FUNCTIONAL ELEMENTS OF SUCH SYSTEMS; MONITORING OR TESTING ARRANGEMENTS FOR SUCH SYSTEMS OR ELEMENTS
- G05B23/00—Testing or monitoring of control systems or parts thereof
- G05B23/02—Electric testing or monitoring
- G05B23/0205—Electric testing or monitoring by means of a monitoring system capable of detecting and responding to faults
- G05B23/0218—Electric testing or monitoring by means of a monitoring system capable of detecting and responding to faults characterised by the fault detection method dealing with either existing or incipient faults
- G05B23/0243—Electric testing or monitoring by means of a monitoring system capable of detecting and responding to faults characterised by the fault detection method dealing with either existing or incipient faults model based detection method, e.g. first-principles knowledge model
-
- G—PHYSICS
- G05—CONTROLLING; REGULATING
- G05B—CONTROL OR REGULATING SYSTEMS IN GENERAL; FUNCTIONAL ELEMENTS OF SUCH SYSTEMS; MONITORING OR TESTING ARRANGEMENTS FOR SUCH SYSTEMS OR ELEMENTS
- G05B2219/00—Program-control systems
- G05B2219/20—Pc systems
- G05B2219/24—Pc safety
- G05B2219/24065—Real time diagnostics
-
- Y—GENERAL TAGGING OF NEW TECHNOLOGICAL DEVELOPMENTS; GENERAL TAGGING OF CROSS-SECTIONAL TECHNOLOGIES SPANNING OVER SEVERAL SECTIONS OF THE IPC; TECHNICAL SUBJECTS COVERED BY FORMER USPC CROSS-REFERENCE ART COLLECTIONS [XRACs] AND DIGESTS
- Y02—TECHNOLOGIES OR APPLICATIONS FOR MITIGATION OR ADAPTATION AGAINST CLIMATE CHANGE
- Y02P—CLIMATE CHANGE MITIGATION TECHNOLOGIES IN THE PRODUCTION OR PROCESSING OF GOODS
- Y02P90/00—Enabling technologies with a potential contribution to greenhouse gas [GHG] emissions mitigation
- Y02P90/02—Total factory control, e.g. smart factories, flexible manufacturing systems [FMS] or integrated manufacturing systems [IMS]
Abstract
The invention discloses an industrial robot fault diagnosis model, which comprises the following steps: the time sequence feature extraction module is used for extracting time sequence features, splicing and fusing the memory information output by each time window, and taking a time sequence feature diagram obtained by splicing and fusing as an output result of the whole time sequence feature extraction module; the multi-scale convolution module comprises a plurality of feature extraction branches which are arranged in parallel, each feature extraction branch respectively performs feature extraction on the time sequence feature graphs, and fault category feature vectors are obtained after feature graphs extracted by all feature extraction branches are combined; the multi-label classifier comprises a plurality of fault label modules which are arranged in parallel, wherein each fault label module is provided with a classifier in one-to-one correspondence with a part which possibly fails in a corresponding joint, and each classifier is provided with an independent loss function and converts a fault class characteristic vector into a class probability vector. The invention also discloses a fault diagnosis method of the industrial robot.
Description
Technical Field
The invention belongs to the technical field of industrial equipment fault diagnosis, and particularly relates to an industrial robot fault diagnosis model and a fault diagnosis method.
Background
Along with continuous transition and upgrading of automobile manufacturing industry in China to automation and intellectualization, the production scene is increasingly complicated and diversified, and the automobile manufacturing enterprises have high water-rise on performance requirements on all aspects of automatic production equipment, especially in the aspect of safety and reliability. As a core device on a white car body welding production line, a welding robot can disturb the production beat if sudden abnormal fault conditions occur, so that the white car body welding quality is reduced, the production efficiency of enterprises is reduced, and even the health and safety of staff can be threatened. Therefore, the production enterprises have high requirements on the safety reliability and the operation stability of the white body welding robot. The white car body welding robot is a multifunctional, intelligent and special mechanical arm for welding white car body frameworks, has the advantages of high intelligent level, good economic benefit, high safety reliability and the like, and is an essential precision production device in the car production process. However, in complex and diverse production scenes, unpredictable abnormal faults often occur to the welding robot, so that the position accuracy of the welding robot is reduced, the welding quality is not ideal, the whole white car body welding production line is stopped in an unplanned mode, and even the safety of workers is threatened. For enterprises, once the welding robot system has sudden, unexpected abnormal and fault conditions, the production of the whole production line is stopped. If the failed robot is not timely repaired, a huge production accident may develop. In order to prevent unplanned and unexpected abnormality and fault conditions of the welding robot, current automobile production enterprises mostly adopt a traditional mode of regular manual overhaul and maintenance, if the welding robot is found to be abnormal, the abnormal robot needs to be overhauled in all directions, fault components are positioned gradually, and time and labor are consumed. The traditional welding robot abnormality monitoring and fault diagnosis mode wastes enterprise resources very much, and reduces production efficiency. Therefore, the rapid fault diagnosis method research of the welding robot is very necessary, so that engineers are helped to find early abnormal conditions of the robot in time and complete fault diagnosis positioning aiming at the abnormality, and therefore, the targeted component maintenance is performed in time, and effective reference is provided for the planned maintenance work of the production line.
The fault diagnosis research of mechanical equipment is a comprehensive technology of deep crossing in the multidisciplinary field. The fault diagnosis system analyzes the abnormal signals by using technical methods such as signal analysis processing, feature extraction and the like, judges the fault type of the abnormal equipment by combining the fault phenomenon, positions the fault components and finds the fault reasons, and helps engineers to determine corresponding maintenance schemes. The traditional fault diagnosis method comprises characteristic signal analysis, statistical analysis, expert experience system and the like, most of the fault diagnosis methods depend on mature mathematical, physical and experience models, have high requirements on signal data quality, and have certain limitations on applicable scenes and objects. In actual production, deep signals for representing the running state of equipment are often lacking, so that the traditional fault detection method has failure under special scenes, and the variability of the working conditions of the equipment and unavoidable signal noise have great influence on the performance of the traditional method, so that the traditional fault detection method cannot be applied to complex and diverse production scenes. In recent years, with the rapid development of AI technology, large data technology, and high-performance computing clusters, a machine learning and data driven based fault diagnosis method has become a leading-edge subject of fault diagnosis disciplines. And learning massive equipment operation signal data by using machine learning algorithms such as a neural network and the like, extracting the depth characteristics of signals, and realizing intelligent fault diagnosis of the automatic equipment.
Disclosure of Invention
In view of the above, an object of the present invention is to provide an industrial robot fault diagnosis model and a fault diagnosis method capable of realizing rapid diagnosis of an abnormal fault of an industrial robot.
In order to achieve the above purpose, the present invention provides the following technical solutions:
the invention firstly provides an industrial robot fault diagnosis model, which comprises the following steps:
the time sequence feature extraction module is used for extracting time sequence features of current, rotating speed, rotating angle and operation stage signals of each joint of the industrial robot, splicing and fusing the memory information output by each time window, and taking a time sequence feature diagram obtained by splicing and fusing as an output result of the whole time sequence feature extraction module;
the multi-scale convolution module comprises a plurality of feature extraction branches which are arranged in parallel, each feature extraction branch performs feature extraction on the time feature map by adopting different receptive field convolution kernel combinations, and fault class feature vectors obtained after feature maps extracted by all feature extraction branches are combined are used as the output of the multi-scale convolution module;
the multi-label classifier comprises a plurality of fault label modules which are arranged in parallel, wherein the fault label modules are arranged in one-to-one correspondence with joints of the industrial robot, each fault label module comprises a classifier which is arranged in one-to-one correspondence with parts possibly suffering from faults in the corresponding joint, and each classifier is provided with an independent loss function and converts fault category characteristic vectors into category probability vectors.
Further, the timing sequence feature extraction module adopts an LSTM neural network.
Further, five feature extraction branches are arranged in parallel in the multi-scale convolution module, and the feature extraction branches are respectively as follows:
the MSCM_a feature extraction branch comprises a 1 multiplied by 1 convolution kernel to keep the feature information in the time sequence feature diagram as far as possible;
an MSCM_b feature extraction branch for performing a Max-Pooling operation to reduce an input data dimension and extract main feature information;
the MSCM_c feature extraction branch comprises a 1×1 convolution kernel and a 3×3 convolution kernel which are connected in series;
the MSCM_d feature extraction branch comprises a 1×1 convolution kernel and a 5×5 convolution kernel which are connected in series;
the mscm_e feature extraction branch comprises a 1×1 convolution kernel and a 7×7 convolution kernel in series.
Further, in the mscm_c feature extraction branch, the 3×3 convolution kernel is split into one 1×3 convolution kernel and one 3×1 convolution kernel connected in series.
Further, in the mscm_d feature extraction branch, the 5×5 convolution kernel is split into a 3×3 convolution kernel, a 1×3 convolution kernel, and a 3×1 convolution kernel; wherein, the 3×3 convolution kernel is connected in series with the 1×1 convolution kernel in the mscm_d feature extraction branch, and the 1×3 convolution kernel is connected in parallel with the 3×1 convolution kernel and then connected in series with the 3×3 convolution kernel.
Further, in the mscm_e feature extraction branch, the 7×7 convolution kernel is split into two 3×3 convolution kernels, one 1×3 convolution kernel and one 3×1 convolution kernel; wherein, two 3×3 convolution kernels are connected in series, one 3×3 convolution kernel is connected in series with a 1×1 convolution kernel in the mscm_e feature extraction branch, and the 1×3 convolution kernel is connected in parallel with the 3×1 convolution kernel and then connected in series with the other 3×3 convolution kernel.
Further, the multi-scale convolution module further comprises a global average pooling module, wherein the global average pooling module is used for merging feature graphs extracted by all feature extraction branches and obtaining fault category feature vectors.
Further, a sigmoid function is adopted in the classifier as an activation function.
The invention also provides an industrial robot fault diagnosis method, which comprises the following steps:
step one: the method comprises the steps of data acquisition, namely completing acquisition of original data through industrial robot control software and a production line equipment database, preprocessing the acquired original data, and constructing to obtain a data set;
step two: constructing the industrial robot fault diagnosis model;
step three: training an industrial robot fault diagnosis model by utilizing the data set and evaluating the validity of the industrial robot fault diagnosis model;
step four: the current, the rotating speed, the rotating angle and the running stage signals of each joint of the industrial robot are collected in real time through industrial robot control software, so that the fault of the industrial robot is rapidly diagnosed and positioned.
The invention has the beneficial effects that:
the industrial robot fault diagnosis model comprises the steps of firstly, adopting a time sequence feature extraction module to extract time sequence features of each joint current, rotating speed, rotating angle and operation stage signal of a robot, integrating time sequence memory feature quantity of each flexible time window to obtain a two-dimensional time sequence feature map, then further extracting deep space features of operation signals of the robot through a proposed multi-scale convolution module, and finally completing fault diagnosis through a multi-label classifier; the method has the advantages that the multiple feature extraction branches are arranged in the multi-scale convolution module in parallel, each feature extraction branch performs feature extraction on the time feature map by adopting different receptive field convolution kernel combinations so as to obtain multiple features at the same position in the feature map, and the problems that a network only pays attention to a certain specific signal feature and possibly loses certain key signal features due to a specific single convolution kernel, so that differences at similar positions of different types of signal data are difficult to distinguish are effectively avoided; through setting up trouble tag module with industrial robot's joint one-to-one in many tag classifier to set up the classifier in each trouble tag module with the part portion correspondence that probably breaks down in the corresponding joint, each classifier adopts independent loss function and converts trouble category feature vector into category probability vector, does not have to restrict all category probability sum to 1, can make the classifier possess very fine classification granularity with huge single tag classification split into a plurality of two categorised tactics, and effectively promotes the precision.
Drawings
In order to make the objects, technical solutions and advantageous effects of the present invention more clear, the present invention provides the following drawings for description:
FIG. 1 is a flow chart of an embodiment of an industrial robot fault diagnosis method of the present invention;
FIG. 2 is a block diagram of an embodiment of an industrial robot fault diagnosis model of the present invention;
FIG. 3 is a block diagram of a timing feature extraction module;
FIG. 4 is a block diagram of an LSTM neural network;
FIG. 5 is a block diagram of a multi-scale convolution module;
fig. 6 is a block diagram of a multi-label classifier.
Detailed Description
The present invention will be further described with reference to the accompanying drawings and specific examples, which are not intended to limit the invention, so that those skilled in the art may better understand the invention and practice it.
In this embodiment, a welding robot in the welding process of the body-in-white is taken as an example to explain a fault diagnosis method of an industrial robot. Before fault diagnosis, the welding process of the white car body and the structure formed by a production line are required to be analyzed, and the working characteristics, the structure and the performance of a welding robot are analyzed; then, the form and the characteristics of the abnormality and the fault of the welding robot are briefly analyzed, and a general scheme for fault diagnosis of the welding robot is provided. The fault diagnosis is performed based on a fault diagnosis overall scheme, as shown in fig. 1, and the industrial robot fault diagnosis method of the present embodiment includes the following steps:
step one: and (3) data acquisition, namely completing acquisition of original data through industrial robot control software and a production line equipment database, preprocessing the acquired original data, and constructing to obtain a data set.
The operation signal data of the welding robot are collected in real time by the sensors in the joints, are fed back to the controller software and are stored in the production line network cabinet, and the data collection work is completed through the controller software BOS6000 of the welding robot and the equipment database of the welding production line. Because of the problems of difference of models of robots and sensors in a welding production line, fluctuation of signal acquisition frequency and the like, the original data acquired from the production line on site has the phenomena of data value deficiency, abnormality, non-uniform data format and the like, and the original data needs to be subjected to necessary pretreatment, and the pretreatment method comprises grouping, deficiency filling, data conversion and the like.
Step two: an industrial robot fault diagnosis model (ltcnet) was constructed.
As shown in fig. 2, the industrial robot fault diagnosis model of the present embodiment includes a timing feature extraction module, a multi-scale convolution module, and a multi-label classifier.
(1) The time sequence feature extraction module is used for extracting time sequence features of current, rotating speed, rotating angle and operation stage signals of each joint of the industrial robot and outputting memory information h of each time window t And performing splicing and fusion, wherein a time sequence feature diagram obtained by splicing and fusion is used as an output result of the whole time sequence feature extraction module. As shown in fig. 3, in the present embodiment, the timing feature extraction module adopts an LSTM neural network.
The welding robot usually performs the welding operation according to a preset program, so that the robot operation signal data mostly has the time sequence characteristic. In order to mine the time sequence characteristics of each joint signal of the robot, the head of the LtcmNet adopts LSTM to extract the time sequence characteristics of each joint current, rotating speed, rotating angle and operation phase signal of the robot.
A Long Short-Term Memory network (LSTM) is a time-cycled neural network. Due to the unique design structure, LSTM is suitable for processing and predicting very long-spaced and delayed important events in a time series. The structure of the neural network is similar to that of a common circulating neural network, and the memory of longer is realized by introducing 3 gating units, so that the neural network is mainly used for solving the longer short interval dependence. The most predominant unit in LSTM is 3 control gates: an input door, a forget door, and an output door. An input to the gate control network; the forgetting gate is used as the core of the LSTM, decides which knowledge or information needs to be kept in mind and which needs to be forgotten to be removed, and is equivalent to a memory unit in the circulating neural network; the output gate controls the output of the network. The structure of the LSTM neural network is shown in fig. 4.
Structure and principle of LSTM neural network:
(1) forgetting the door: deciding which information to retain and discard from the history "memory", itWill use the output h from the previous time t-1 And input x at the current time t The combination is used as input to output a [0,1] through sigma function]The value between the values acts as a memory weight on the memory c at the previous moment t-1 The degree of memory and forgetfulness is determined. 0 indicates discarding the memory at the previous time and 1 indicates fully retaining the memory at the previous time. Sigma functions often use Sigmoid-like functions with an output interval of 0,1]The memory information selection weight is represented. The mathematical expression of the forgetting gate is:
f t ＝σ(W f ·[h t-1 ,x t ]+b f )
(2) an input door: the updating of the decision memory uses a sigma function whose inputs are the output of the previous moment and the input of the current moment. Like a forget gate, the sigma function returns a value between [0,1] as a memory update weight to determine which new memories need to be remembered and update the memories into the old knowledge base. Therefore, the input gate integrates the memory to be updated screened at the current time and the memory to be reserved at the last time screened by the forgetting gate, and a new memory is obtained. The mathematical expression is as follows:
the result i of equation 4.2 t Andthe element level multiplication is performed to obtain the information to be remembered at the current moment t, and then the information is updated to the old memory by using a formula 4.3 to obtain the latest memory c of the current moment t t 。c t As the latest memory information at the current time and the last time c t-1 As such, it will be passed on to the next moment.
(3) Output door: acting on the current using standard Sigmoid functionsThe latest memory information before obtains probability vector o t And reform the current memorized distribution using the tanh activation function, and output a probability vector o t The multiplication results in an output vector h t . The mathematical expression of the output gate is:
the LSTM network may be formed of a plurality of cells, each cell including the three control gate structures described above, and this unique functional structure allows the LSTM to achieve good results in extracting data timing characteristics.
The time sequence feature extraction module provided by the embodiment takes an LSTM network as a main body, and 20 channel signals X are obtained in total from motor currents, rotating speeds, joint rotating angles and robot operation phases of six joints of the industrial robot, which are acquired by sensors r After normalization processing, the data is used as data input of a time sequence feature extraction module, LSTM neural network is applied to model signal data, time sequence features are extracted, and memory information h output by each time window is obtained t And performing splicing fusion to obtain an output result of the whole time sequence feature extraction module.
(2) The multi-scale convolution module comprises a plurality of feature extraction branches which are arranged in parallel, each feature extraction branch adopts different receptive field convolution kernel combinations to perform feature extraction on the sequential feature images, and fault category feature vectors obtained after feature images extracted by all feature extraction branches are combined are used as the output of the multi-scale convolution module. Specifically, the multi-scale convolution module of this embodiment further includes a global averaging pooling module, where the global averaging pooling module is configured to combine feature graphs extracted by all feature extraction branches and obtain a fault class feature vector. As shown in fig. 5, in this embodiment, five feature extraction branches are arranged in parallel in the multi-scale convolution module, and the feature extraction branches are respectively:
the MSCM_a feature extraction branch comprises a 1 multiplied by 1 convolution kernel to keep the feature information in the time sequence feature diagram as far as possible;
an MSCM_b feature extraction branch for performing a Max-Pooling operation to reduce an input data dimension and extract main feature information;
the MSCM_c feature extraction branch comprises a 1×1 convolution kernel and a 3×3 convolution kernel which are connected in series; in a preferred implementation of this example, the 3×3 convolution kernel is split into one 1×3 convolution kernel and one 3×1 convolution kernel in series;
the MSCM_d feature extraction branch comprises a 1×1 convolution kernel and a 5×5 convolution kernel which are connected in series; in a preferred implementation of the present example, the 5×5 convolution kernel is split into one 3×3 convolution kernel, one 1×3 convolution kernel, and one 3×1 convolution kernel; wherein, the 3×3 convolution kernel is connected in series with the 1×1 convolution kernel in the MSCM_d feature extraction branch, and the 1×3 convolution kernel is connected in parallel with the 3×1 convolution kernel and then connected in series with the 3×3 convolution kernel;
the MSCM_e feature extraction branch comprises a 1×1 convolution kernel and a 7×7 convolution kernel which are connected in series; in a preferred implementation of the present example, the 7×7 convolution kernel is split into two 3×3 convolution kernels, one 1×3 convolution kernel and one 3×1 convolution kernel; wherein, two 3×3 convolution kernels are connected in series, one 3×3 convolution kernel is connected in series with a 1×1 convolution kernel in the mscm_e feature extraction branch, and the 1×3 convolution kernel is connected in parallel with the 3×1 convolution kernel and then connected in series with the other 3×3 convolution kernel.
And the LtcmNet head finishes the task of extracting the signal time sequence characteristics by using the LSTM, and performs splicing and fusion on the memory information acquired by each time window to obtain a time sequence characteristic diagram with complete signal data as the input of a subsequent network model. The memory output strategy of each time step of the LSTM is spliced and fused, which is different from the strategy of only acquiring the last cell output, and the signal shallow layer characteristics can be acquired and more comprehensive information can be reserved, but the useless information in the signal is also collected, so that the signal deep layer characteristics can not be extracted. Therefore, the present embodiment introduces a multi-scale convolution module to remedy the above-mentioned deficiencies.
The CNN slides on the feature vector or matrix through the shared convolution kernel structure to achieve the purpose of extracting the features, part of old features can be discarded in the convolution process, and new features are circled in the convolution process, so that the weight parameter scheduling is reduced, the processing capacity of the network to high-dimensional data is promoted, and the difficulty of feature engineering is greatly reduced. CNN is often applied to scenes with huge feature input characteristics such as voice, video, image and natural language processing, and is also applicable to signal data processing scenes. When the CNN is used for extracting the characteristics of the signals, the traditional mode is to input the signal data into a CNN network after the signal data is unidimensioned to finish the characteristic extraction task, but the signal data of the embodiment is the current, rotating speed and period time sequence data of the operation stages of six joints of the welding robot, and besides the time sequence characteristics are hidden in the time domain of each joint current channel, each rotating speed channel and each operation stage channel, the three channel signals of each joint are also in communication in the space domain. Considering the situation that the current, the rotating speed, the rotating angle and the signal data in the operation stage are possibly covered by strong background noise, so that the relevant characteristic information of faults is weakened, the embodiment adopts a two-dimensional multi-scale characteristic extraction strategy to learn the input signal data in a deeper level, fully extracts the relevant characteristics among the signals of the robot joint, and names the proposed multi-scale convolution module as an MSCM convolution module.
For the periodic signal data of the robot operation in this embodiment, the similarity of the periodic signal data of different fault types is very high, if a specific single convolution kernel is adopted, the network only pays attention to a specific signal feature, and some key signal features may be lost, so that it is difficult to distinguish differences between similar positions of the signal data of different types. In order to improve generalization capability of the model, the MSCM introduces a plurality of size convolution check feature graphs to perform parallel extraction operation, and acquires a plurality of features at the same position in the feature graphs. The MSCM convolution module structure of this embodiment is shown in fig. 5, and is formed by five branches, each branch performs feature extraction on the input feature map, and finally, the feature maps of the five branches are combined and output to obtain a set of depth feature maps. Fig. 3 shows a block diagram of the MSCM module.
In the MSCM module, the first branch adopts a 1X 1 convolution kernel, so that the characteristic information in the characteristic diagram is kept as much as possible. The second branch is a simple Max-Pooling operation, which reduces the dimension of input data and extracts main feature information. The third branch is composed of a 1×1 convolution kernel and a 3×3 convolution kernel, and the 3×3 convolution kernel is disassembled into two asymmetric convolution kernels, the disassembled 1×3 and 3×1 convolution kernels are equivalent to the 3×3 convolution kernel, and the advantage of the convolution kernel disassembly strategy is that network parameters are greatly reduced while the effect is not changed. The fourth branch is composed of a 1×1 convolution kernel and a 5×5 convolution kernel, where a 5×5 convolution kernel with a larger receptive field is used to enhance the spatial location relationship feature extraction capability of the MSCM module. In consideration of the problem of network parameter proliferation caused by increasing receptive fields, a strategy similar to that of a third branch is adopted to split the 5×5 convolution kernel into a 3×3 convolution kernel and a 1×3 and 3×1 asymmetric convolution kernel parallel structure. The fifth branch is composed of a 1×1 convolution kernel and a 7×7 convolution kernel, and the 7×7 convolution kernel is subjected to a corresponding disassembly operation, similar to the fourth branch.
The five branches of the MSCM module respectively adopt different receptive field convolution kernel combinations to extract the characteristics of the characteristic map, and the large convolution kernel is disassembled into a plurality of asymmetric convolution kernel structures. The MSCM module reasonably performs dimension reduction on each branch without destroying the network feature extraction capability, and generates mutually decoupled feature representations through more activated output branches, thereby generating high-order sparse features and accelerating convergence.
(3) The multi-label classifier comprises a plurality of fault label modules which are arranged in parallel, wherein the fault label modules are arranged in one-to-one correspondence with joints of the industrial robot, each fault label module comprises a classifier which is arranged in one-to-one correspondence with parts possibly suffering from faults in the corresponding joint, and each classifier is provided with an independent loss function and converts fault class feature vectors into class probability vectors.
The welding robot has high intelligent level, high production efficiency and high safety, and has obvious economic benefit, and is widely applied to the aspects of automobile welding and assembly. Considering that a welding robot is highly-automatic precision mechanical equipment formed by a plurality of complex components, in the long-time, high-strength and variable-load production process, the situation that multiple components are damaged or single component compound faults are likely to occur, for example, a speed reducer and a motor of a certain joint of the robot are faulty, or a plurality of joints of the robot are faulty at the same time. Although the composite fault may be artificially defined as a single fault signature, for the subject of this embodiment, there may be a speed reducer fault or motor fault in each of the 6 joints of the welder robot, and the fault signatures of each joint are independent of each other, so there are a total of 2-12 categories possible. However, class space of size 2≡12 is too large, and handling such large class space with a simple single-label classifier would greatly increase training pressure and difficulty in model convergence, which is clearly an unworkable option. Aiming at the problems, the embodiment adopts a multi-label classifier as the tail of the LtcmNet, and the function of the multi-label classifier is to receive the characteristic vector output of the MSCM module and complete the multi-label classification task.
Conventional CNN networks often employ Softmax as an activation function in combination with a cross entropy loss function when dealing with multi-classification problems. Where the input of the Softmax function is a k-dimensional row vector and the output is also a k-dimensional row vector, each dimension of the vector being within a (0, 1) interval and summing to 1. The function has the function of converting the class feature vector of the output layer of the convolution network into a corresponding class feature probability vector, enabling the sum of the output values of the classes of the output layer to be 1, enabling the class with the highest return probability value of the final model to serve as the whole model discrimination result, and enabling the mathematical expression of the Softmax function to be:
the essence of softmax operation is to convert the single likelihood of each class prediction result into a probability value in the overall prediction, which is only applicable to single-label multi-fault classification and not to compound fault diagnosis. For this purpose, a plurality of parallel sigmoid activation functions are introduced as the main structure of the multi-label classifier. Sigmoid is a nonlinear activation function whose mathematical expression is:
the sigmoid function image shows an "S" shape, and the sigmoid function can convert the inputted continuous numerical value into a value in the (0, 1) interval, which corresponds to performing numerical compression. The multi-label classifier of the embodiment splits the complex combined fault category of the welding robot into fault labels of motors and reducers of 6 independent joints, respectively and correspondingly sets 6 fault labels, each fault label module comprises a classifier corresponding to parts possibly suffering from faults in the corresponding joints one by one, the parts possibly suffering from faults in the embodiment comprise the reducers and the motors, namely, each fault label module comprises 2 classifiers, and fault diagnosis is carried out on the reducers and the motors of each joint independently. Thus, as shown in FIG. 6, using 12 sigmoid function parallel combinations, instead of a single softmax function classifier, essentially split the original 2-12 sized single-tag class space into 12 single-tag class spaces, each class having its own independent loss function while not necessarily limiting the sum of all class probabilities to 1. The strategy of splitting huge single label classification into a plurality of two classifications enables the classifier to have fine classification granularity and effectively improves accuracy. Fig. 4 is a schematic diagram of a multi-label classifier.
Step three: the industrial robot fault diagnosis model is trained using the data set and the validity of the industrial robot fault diagnosis model is evaluated.
Step four: the current, the rotating speed, the rotating angle and the running stage signals of each joint of the industrial robot are collected in real time through industrial robot control software, so that the fault of the industrial robot is rapidly diagnosed and positioned.
Specifically, in this embodiment, ltcnet is divided into three components, namely a head (timing feature extraction module), a body (multi-scale convolution module) and a tail (multi-label classifier):
(1) Ltcnet head
The LtcmNet head is composed of an LSTM network and is used for completing the time sequence feature extraction of the robot signal data, splicing and fusing the memory information acquired by each time window and outputting a time sequence feature diagram. By adopting a single-layer LSTM network, the number of hidden layers is set to 128 by comparing the experimental results of the next model, and the input dimension is 20. It should be noted that, the welding production line lacks online fault data of the robot, so that the data set of the model contains fault signal data collected by partial engineers when the fault robot performs offline debugging and maintenance. However, the debug access process does not run a fixed debug program, resulting in failure signals that are collected that are not periodic and collection lengths that are staggered. In order to solve the problem of uneven length of the model input signals, the LtcmNet adopts a flexible time window strategy, namely the time window length T in the model head LSTM network can be flexibly adjusted according to the input data length, and the shape of the timing sequence characteristic diagram output by the model head is ensured to be unified to be 1 multiplied by 130 multiplied by 128.
(2) Body of LtcmNet
The ltmcinet body is a multi-scale deep convolution layer formed by stacking a plurality of MSCM multi-scale convolution modules, and is used for completing deep feature extraction of a time sequence feature map, the layer comprises five MSCM modules, namely MSCM_a, MSCM_b, MSCM_c, MSCM_d and MSCM_e, and fault class feature vectors are output through a global average pooling module (Global Average Pooling D, GAP 2D). The table of the ltcnet multiscale convolutional layer structure parameters is shown in table 1.
TABLE 4.1LtcmNet multiscale convolutional layer structure parameters
(3) LtcmNet tail
The model tail is composed of a multi-label classifier and adopts a structure of 12 parallel sigmoid activation functions. The multi-label classifier receives fault class feature vectors (1 multiplied by 12) output by a global average pooling layer (GAP 2D), performs label fusion and fault discrimination, and finally outputs a fault diagnosis discrimination result.
The embodiment builds an LtcmNet model based on LSTM and CNN, and solves the problem of fault diagnosis of the welding robot of the white car body. Firstly, a flexible time window LSTM is provided for extracting a time sequence characteristic diagram of a fault operation signal of a robot, and the width of the flexible time window can be flexibly adjusted according to the length of input signal data, so that an LtcmNet model has the capability of processing non-isometric signal data. Then, an MSCM multi-scale convolution module is provided as a body of the LtcmNet, the MSCM module is formed by five different receptive field parallel convolution branches, correlation characteristics among signals of a robot joint can be fully extracted, information complementation can be realized by the extracted characteristics of each convolution branch, the whole and partial spatial characteristic information of the signals is greatly reserved, and meanwhile, the interference of industrial noise is weakened. And finally, realizing accurate fault diagnosis of the welding robot by the aid of the LtcmNet tail through a multi-label classifier.
The above-described embodiments are merely preferred embodiments for fully explaining the present invention, and the scope of the present invention is not limited thereto. Equivalent substitutions and modifications will occur to those skilled in the art based on the present invention, and are intended to be within the scope of the present invention. The protection scope of the invention is subject to the claims.
Claims (9)
1. An industrial robot fault diagnosis model, which is characterized in that: comprising the following steps:
the time sequence feature extraction module is used for extracting time sequence features of current, rotating speed, rotating angle and operation stage signals of each joint of the industrial robot, splicing and fusing the memory information output by each time window, and taking a time sequence feature diagram obtained by splicing and fusing as an output result of the whole time sequence feature extraction module;
the multi-scale convolution module comprises a plurality of feature extraction branches which are arranged in parallel, each feature extraction branch performs feature extraction on the time feature map by adopting different receptive field convolution kernel combinations, and fault class feature vectors obtained after feature maps extracted by all feature extraction branches are combined are used as the output of the multi-scale convolution module;
the multi-label classifier comprises a plurality of fault label modules which are arranged in parallel, wherein the fault label modules are arranged in one-to-one correspondence with joints of the industrial robot, each fault label module comprises a classifier which is arranged in one-to-one correspondence with parts possibly suffering from faults in the corresponding joint, and each classifier is provided with an independent loss function and converts fault category characteristic vectors into category probability vectors.
2. The industrial robot fault diagnosis model according to claim 1, characterized in that: the time sequence feature extraction module adopts an LSTM neural network.
3. The industrial robot fault diagnosis model according to claim 1, characterized in that: five feature extraction branches are arranged in parallel in the multi-scale convolution module, and are respectively:
the MSCM_a feature extraction branch comprises a 1 multiplied by 1 convolution kernel to keep the feature information in the time sequence feature diagram as far as possible;
an MSCM_b feature extraction branch for performing a Max-Pooling operation to reduce an input data dimension and extract main feature information;
the MSCM_c feature extraction branch comprises a 1×1 convolution kernel and a 3×3 convolution kernel which are connected in series;
the MSCM_d feature extraction branch comprises a 1×1 convolution kernel and a 5×5 convolution kernel which are connected in series;
the mscm_e feature extraction branch comprises a 1×1 convolution kernel and a 7×7 convolution kernel in series.
4. The industrial robot fault diagnosis model according to claim 3, wherein: in the mscm_c feature extraction branch, the 3×3 convolution kernel is split into one 1×3 convolution kernel and one 3×1 convolution kernel connected in series.
5. The industrial robot fault diagnosis model according to claim 3, wherein: in the MSCM_d feature extraction branch, splitting a 5×5 convolution kernel into a 3×3 convolution kernel, a 1×3 convolution kernel and a 3×1 convolution kernel; wherein, the 3×3 convolution kernel is connected in series with the 1×1 convolution kernel in the mscm_d feature extraction branch, and the 1×3 convolution kernel is connected in parallel with the 3×1 convolution kernel and then connected in series with the 3×3 convolution kernel.
6. The industrial robot fault diagnosis model according to claim 3, wherein: in the MSCM_e feature extraction branch, the 7×7 convolution kernel is split into two 3×3 convolution kernels, one 1×3 convolution kernel and one 3×1 convolution kernel; wherein, two 3×3 convolution kernels are connected in series, one 3×3 convolution kernel is connected in series with a 1×1 convolution kernel in the mscm_e feature extraction branch, and the 1×3 convolution kernel is connected in parallel with the 3×1 convolution kernel and then connected in series with the other 3×3 convolution kernel.
7. The industrial robot fault diagnosis model according to claim 1, characterized in that: the multi-scale convolution module further comprises a global average pooling module, wherein the global average pooling module is used for merging feature graphs extracted by all feature extraction branches and obtaining fault category feature vectors.
8. The industrial robot fault diagnosis model according to claim 1, characterized in that: the classifier adopts a sigmoid function as an activation function.
9. An industrial robot fault diagnosis method is characterized in that: the method comprises the following steps:
step one: the method comprises the steps of data acquisition, namely completing acquisition of original data through industrial robot control software and a production line equipment database, preprocessing the acquired original data, and constructing to obtain a data set;
step two: constructing an industrial robot fault diagnosis model according to any one of claims 1-8;
step three: training an industrial robot fault diagnosis model by utilizing the data set and evaluating the validity of the industrial robot fault diagnosis model;
step four: the current, the rotating speed, the rotating angle and the running stage signals of each joint of the industrial robot are collected in real time through industrial robot control software, so that the fault of the industrial robot is rapidly diagnosed and positioned.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202310602754.3A CN116560341A (en) | 2023-05-25 | 2023-05-25 | Industrial robot fault diagnosis model and fault diagnosis method |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202310602754.3A CN116560341A (en) | 2023-05-25 | 2023-05-25 | Industrial robot fault diagnosis model and fault diagnosis method |
Publications (1)
| Publication Number | Publication Date |
|---|---|
| CN116560341A true CN116560341A (en) | 2023-08-08 |
Family
ID=87491401
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN202310602754.3A Pending CN116560341A (en) | 2023-05-25 | 2023-05-25 | Industrial robot fault diagnosis model and fault diagnosis method |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN116560341A (en) |
Cited By (1)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN117232577A (en) * | 2023-09-18 | 2023-12-15 | 杭州奥克光电设备有限公司 | Optical cable distributing box bearing interior monitoring method and system and optical cable distributing box |
-
2023
- 2023-05-25 CN CN202310602754.3A patent/CN116560341A/en active Pending
Cited By (2)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN117232577A (en) * | 2023-09-18 | 2023-12-15 | 杭州奥克光电设备有限公司 | Optical cable distributing box bearing interior monitoring method and system and optical cable distributing box |
| CN117232577B (en) * | 2023-09-18 | 2024-04-05 | 杭州奥克光电设备有限公司 | Optical cable distributing box bearing interior monitoring method and system and optical cable distributing box |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN112149316B (en) | Aero-engine residual life prediction method based on improved CNN model | |
| CN109635677B (en) | Compound fault diagnosis method and device based on multi-label classification convolutional neural network | |
| CN109255505B (en) | Short-term load prediction method of multi-model fusion neural network | |
| Yin et al. | Wasserstein generative adversarial network and convolutional neural network (WG-CNN) for bearing fault diagnosis | |
| CN109297689B (en) | Large-scale hydraulic machinery intelligent diagnosis method introducing weight factors | |
| CN110766143A (en) | Equipment fault intelligent diagnosis method based on artificial neural network | |
| CN109800875A (en) | Chemical industry fault detection method based on particle group optimizing and noise reduction sparse coding machine | |
| CN111460728A (en) | Method and device for predicting residual life of industrial equipment, storage medium and equipment | |
| CN111190349A (en) | Method, system and medium for monitoring state and diagnosing fault of ship engine room equipment | |
| CN116560341A (en) | Industrial robot fault diagnosis model and fault diagnosis method | |
| CN115510950A (en) | Aircraft telemetry data anomaly detection method and system based on time convolution network | |
| CN110851654A (en) | Industrial equipment fault detection and classification method based on tensor data dimension reduction | |
| CN115712873A (en) | Photovoltaic grid-connected operation abnormity detection system and method based on data analysis and infrared image information fusion | |
| Son et al. | Deep learning-based anomaly detection to classify inaccurate data and damaged condition of a cable-stayed bridge | |
| CN115146718A (en) | Depth representation-based wind turbine generator anomaly detection method | |
| Yu-ming et al. | Research on intelligent manufacturing flexible production line system based on digital twin | |
| CN111767324B (en) | Intelligent associated self-adaptive data analysis method and device | |
| CN113343760A (en) | Human behavior recognition method based on multi-scale characteristic neural network | |
| CN113485244A (en) | Numerical control machine tool control system and method based on cutter wear prediction | |
| CN111639609A (en) | Intelligent identification system for metal fracture types based on machine vision and deep learning | |
| CN116842379A (en) | Mechanical bearing residual service life prediction method based on DRSN-CS and BiGRU+MLP models | |
| CN115452376A (en) | Bearing fault diagnosis method based on improved lightweight deep convolution neural network | |
| CN113435055A (en) | Self-adaptive migration prediction method and system in shield cutter head torque field | |
| Chernyshev et al. | Integration of building information modeling and artificial intelligence systems to create a digital twin of the construction site | |
| Tang et al. | Multi-state data-driven gas path analysis method |
Legal Events
| Date | Code | Title | Description |
|---|---|---|---|
| PB01 | Publication | ||
| PB01 | Publication | ||
| SE01 | Entry into force of request for substantive examination | ||
| SE01 | Entry into force of request for substantive examination |